Full Code of isi-nlp/Zoph_RNN for AI

master 1e3e7da688cf cached

119 files

125.4 MB

3.0M tokens

528 symbols

1 requests

Copy disabled (too large) Download .txt

Showing preview only (11,984K chars total). Download the full file to get everything.

Repository: isi-nlp/Zoph_RNN
Branch: master
Commit: 1e3e7da688cf
Files: 119
Total size: 125.4 MB

Directory structure:
gitextract_7rhl_yau/

├── .gitattributes
├── .gitignore
├── README.md
├── README_XING.md
├── executable/
│ ├── ZOPH_RNN_1
│ ├── ZOPH_RNN_2
│ ├── ZOPH_RNN_3
│ ├── ZOPH_RNN_4
│ └── ZOPH_RNN_XING
├── sample_data/
│ ├── dev_english.txt.tok.lc
│ ├── dev_french.txt.tok.lc
│ ├── ptb.train.txt
│ ├── ptb.valid.txt
│ ├── test_english.tok.lc
│ ├── test_french.tok.lc
│ ├── train_english.txt.tok.lc.10k
│ └── train_french.txt.tok.lc.10k
├── scripts/
│ ├── att_unk_rep.py
│ ├── berk_aligner/
│ │ ├── berkeleyaligner.jar
│ │ ├── run_aligner.sh
│ │ └── unk_replace.conf
│ ├── bleu_format.py
│ ├── bleu_format_valid.py
│ ├── compile.sh
│ ├── compile.xing.sh
│ ├── create_vocab_mapping_file.py
│ ├── create_vocab_mapping_file_preinit.py
│ ├── fsa/
│ │ ├── convert.py
│ │ ├── demo.sh
│ │ ├── enc1.txt
│ │ ├── enc2.txt
│ │ ├── fsa.txt
│ │ ├── generate_fsa.py
│ │ ├── source.singleline.txt
│ │ ├── source.test.txt
│ │ ├── source.train.txt
│ │ ├── source.valid.txt
│ │ ├── target.test.txt
│ │ ├── target.train.txt
│ │ └── target.valid.txt
│ ├── generate_train_decode.py
│ ├── load_lstm.py
│ ├── load_model.py
│ ├── multi-bleu.perl
│ ├── pretrain.py
│ ├── read_hpc_output.py
│ ├── translate/
│ │ ├── f2e_decode.sh
│ │ └── f2e_train.sh
│ └── unk_format.py
└── src/
├── BZ_CUDA_UTIL.h
├── Eigen_Util.h
├── Full_Input_To_Hidden_Layer.h
├── Hidden_To_Hidden_Layer.h
├── Hidden_To_Hidden_Layer.hpp
├── Input_To_Hidden_Layer.h
├── Input_To_Hidden_Layer.hpp
├── LSH_WTA.h
├── LSTM.h
├── LSTM.hpp
├── LSTM_HH.h
├── LSTM_HH.hpp
├── NCE.h
├── NCE.hpp
├── NCE_node.h
├── add_model_info.h
├── attention_combiner.h
├── attention_combiner.hpp
├── attention_combiner_node.h
├── attention_combiner_node.hpp
├── attention_layer.h
├── attention_layer.hpp
├── attention_node.h
├── attention_node.hpp
├── base_layer.h
├── base_layer.hpp
├── base_loss.h
├── bi_encoder.h
├── bi_encoder.hpp
├── charCNN_node.h
├── char_file_helper.h
├── conv_char.h
├── conv_char.hpp
├── custom_kernels.h
├── decoder.h
├── decoder_model_wrapper.h
├── decoder_model_wrapper.hpp
├── encoder_multi_source.h
├── encoder_multi_source.hpp
├── ensemble_factory.h
├── ensemble_factory.hpp
├── fileHelper.h
├── fileHelper_source.h
├── file_helper_char_decoder.h
├── file_helper_decoder.h
├── format.cc
├── format.h
├── fsa.h
├── fsa.hpp
├── global_params.h
├── gpu_info_struct.h
├── highway_network.h
├── highway_network.hpp
├── highway_node.h
├── input_file_prep.h
├── input_file_prep.hpp
├── logger.h
├── main.cu
├── memory_util.h
├── model.h
├── model.hpp
├── multinomial.h
├── prev_states.h
├── softmax.h
├── softmax.hpp
├── softmax_node.h
├── transfer_layer.h
├── tree_LSTM.h
├── tree_LSTM.hpp
└── trunc_softmax.h

================================================
FILE CONTENTS
================================================

================================================
FILE: .gitattributes
================================================
executable/ZOPH_RNN_XING filter=lfs diff=lfs merge=lfs -text

================================================
FILE: .gitignore
================================================
.DS_Store
*~
[#].[#]
.[#]*
*[#]

*pyc
.history

================================================
FILE: README.md
================================================
# Zoph\_RNN: A C++/CUDA toolkit for training sequence and sequence-to-sequence models across multiple GPUs

This is [Barret Zoph's](http://barretzoph.github.io/) code for Zoph\_RNN
Send any questions or comments to barretzoph@gmail.com

This toolkit can successfully replicate the results from the following papers (the multi-gpu parallelism, which is explained in the tutorial, is similar to 6)

1. [Multi-Source Neural Translation](http://www.isi.edu/natural-language/mt/multi-source-neural.pdf)
2. [Simple, Fast Noise Contrastive Estimation for Large RNN Vocabularies](http://www.isi.edu/natural-language/mt/simple-fast-noise.pdf)
3. [Transfer Learning for Low-Resource Neural Machine Translation](http://arxiv.org/pdf/1604.02201v1.pdf)
4. [Effective Approaches to Attention-based Neural Machine Translation](http://stanford.edu/~lmthang/data/papers/emnlp15_attn.pdf)
5. [Addressing the Rare Word Problem in Neural Machine Translation](http://stanford.edu/~lmthang/data/papers/acl15_nmt.pdf)
6. [Sequence to Sequence Learning with Neural Networks](http://arxiv.org/pdf/1409.3215.pdf)
7. [Recurrent Neural Network Regularization](http://arxiv.org/pdf/1409.2329.pdf)

# Instructions for compilation/using the code
The code for Zoph\_RNN is provided in the `src/` directory. Additionally, a precompiled binary (named `ZOPH_RNN`) is provided that will work on 64 bit linux for cuda 7.5, so it is not necessary to compile the code.

If you just want to use the executable, then run the following command `cat executable/ZOPH_RNN_1 executable/ZOPH_RNN_2 executable/ZOPH_RNN_3 executable/ZOPH_RNN_4 > ZOPH_RNN`. Then `ZOPH_RNN` will be the executable that you can use. To run the executable you need to be sure your path variable includes the location to CUDA. This is a sample command of putting cuda into your PATH variable `export PATH=/usr/cuda/7.5/bin:$PATH`

If you want to compile the Zoph\_RNN code run `bash scripts/compile.sh`, which will compile the code given you set a few environmental variables. The variables that need to be set are below:

1. `PATH_TO_CUDA_INCLUDE` (example value: `/usr/cuda/7.5/include/` )
2. `PATH_TO_BOOST_INCLUDE` (example value: `/usr/boost/1.55.0/include/` )
3. `PATH_TO_CUDA_LIB_64` (example value: `/usr/cuda/7.5/lib64/` )
4. `PATH_TO_BOOST_LIB` (example value: `/usr/boost/1.55.0/lib/` )
5. `PATH_TO_CUDNN_V4_64` (example value: `/usr/cudnn_v4/lib64/` )
6. `PATH_TO_EIGEN` (example value: `/usr/eigen/` )
7. `PATH_TO_CUDNN_INCLUDE` (example value: `/usr/cudnn_v4/include/` )

### The acceptable versions for the libraries above

Note that cuda version greater than 7.0 is required to run the code, while the rest are required to compile the code

* cuda version greater than 7.0
* gcc version greater than 4.8.1, but not greater than 4.9
* CuDNN version = 4
* Boost version = 1.51.0 or 1.55.0
* Any version of Eigen

# Tutorial
For this tutorial `ZOPH_RNN` represents the executable to run the code. Also all the scripts in the `scripts` folder require python 3 to run.

This command will bring up the program's help menu showing all the flags that the program can be run with:

```
./ZOPH_RNN -h
```

There are two different kinds of models this code can train

1. Sequence models (Ex: Language Modeling)
2. Sequence-to-Sequence models (Ex: Machine Translation)

The commands for these two different architectures are almost the same, all that needs to change is adding a `-s` flag if you want to use the sequence model. The sequence-to-sequence model is used by default.

In the `sample_data` directory there is sample data provided that shows the proper formatting for files.

### Training a seq-to-seq model:
Lets step through an example that trains a basic sequence-to-sequence model. The following code will train a sequence-to-sequence model with the source training data `/path/to/source_train_data.txt` and the target training data `/path/to/target_train_data.txt`. These are placeholder names that will be replaced with your data files when you are training your own model. The resulting model will be saved to `model.nn`, but this can be named whatever the user wants. Training data always needs to consist of one training example per line, with tokens separated by spaces.

```
./ZOPH_RNN -t /path/to/source_train_data.txt /path/to/target_train_data.txt model.nn
```

By default the source sentences will always be fed in the reversed direction as in [Sequence to Sequence Learning with Neural Networks](http://arxiv.org/pdf/1409.3215.pdf). If you want to feed in the source sentences in the forward direction then simply preprocess your source data, so that it is in the reversed direction.

There are many flags that can be used to train more specific architectures. Lets say we want to train a model with 3 layers (default is 1), 500 hiddenstates (default is 100), and a minibatch of size 64 (default is 8). The following command does this:

```
./ZOPH_RNN -t /path/to/source_train_data.txt /path/to/target_train_data.txt model.nn -N 3 -H 500 -m 64
```

Lets also make the model have 20,000 source vocabulary and 10,000 target vocabulary (by default the code makes the source vocabulary equal to the number of unique tokens in the source training data, and the target vocab does the same). Also lets apply dropout with a keep probability of 0.8 to the model, where dropout is applied as specified in [Recurrent Neural Network Regularization](http://arxiv.org/pdf/1409.2329.pdf).

```
./ZOPH_RNN -t /path/to/source_train_data.txt /path/to/target_train_data.txt model.nn -N 3 -H 500 -m 64 --source-vocab-size 20000 --target-vocab-size 10000 -d 0.8
```

Additionally, lets change the learning rate to 0.5 (default is 0.7), add the local-p attention model with feed input as in [Effective Approaches to Attention-based Neural Machine Translation](http://stanford.edu/~lmthang/data/papers/emnlp15_attn.pdf).

To monitor the training we also want to be able to monitor the performance of the model during training on some held out set of data (developement/validation). Lets do this in the code and also add the option that if perplexity (better is lower) on the held out set of data increased since it was previously checked, then we multiply the current learning rate by 0.5.

```
./ZOPH_RNN -t /path/to/source_train_data.txt /path/to/target_train_data.txt model.nn -N 3 -H 500 -m 64 --source-vocab-size 20000 --target-vocab-size 10000 -d 0.8 -l 0.5 --attention-model true --feed-input true -a /path/to/source_dev_data.txt /path/to/target_dev_data.txt -A 0.5
```

During training the code needs to produce temporary files. By default these will be put in the directory where the code is launched from, but we can change this to whatever we want. Additionally, we can make all of the output that is typically printed to standard out (the screen) also be printed to a file.

Typically during training only one model will be output at the end of training. To make the code output the best model during training according to the perplexity on your heldout data specificed by the `-a` flag we can add the `-B` flag.

Or if you want to save all models every half epoch we can do that with the `--save-all-models` flag.

By default the code will throw away any sentences in training and in the held out data longer than some fixed length which is 100 by default. We can change this to whatever we want, but be careful as it will greatly increase memory usage. Lets change it to 500.

By default the code uses an MLE objective function, which can be very computationally expensive if the target vocabulary is big. To alleviate this issue we can train with NCE instead of MLE by using the `--NCE` flag. This is the same NCE as in [Simple, Fast Noise Contrastive Estimation for Large RNN Vocabularies](http://www.isi.edu/natural-language/mt/simple-fast-noise.pdf). A good number of noise samples is usually around 100. Note that the `--NCE` flag only has to be specified during training and not during force-decode or decode.

One feature of this code is that is supports model parallelism across multiple gpus. To see the number of available gpu's on your node you can type `nvidia-smi`. The `-M` flag allows our model to put each layer on a gpu of our choosing along with the softmax. -M 0 1 2 3 means put layer 1 on GPU 0, layer 2 on GPU 1, layer 3 on GPU 2 and the softmax on GPU 3. By default the code does -M 0 0 0 0, putting everything on the default GPU 0. We can also change up the specification depending how many gpus we have on the node, so we could do `-M 0 0 1 1` if we only have 2 gpus on our node.

### Supplying your own vocabulary mapping file
The `--source-vocab-size N` and the `--target-vocab-size N` flags create a vocabulary mapping file that will replace all words not in the top N most frequent words with <unks>'s. The code will create an integer mapping that is stored in the top of the model file. If you want to supply your own mapping file you can do this using the `--vocab-mapping-file /path/to/my_mapping.nn`. The `my_mapping.nn` can be a previously trained model, in that case it will use the exact same vocabulary mapping as that model. This is useful because if you want to ensemble models using the `--decode` flag, then the models must have exactly the same target vocabulary mapping file for it to work. In the `scripts/` directory there is a python script called `create_vocab_mapping_file.py`. We can use this to create a mapping file, which then gets fed into the training using the following command:

```
python scripts/create_vocab_mapping_file.py /path/to/source_training_data.txt /path/to/target_training_data.txt 5 my_mapping.nn
```

This will create a mapping file named `my_mapping.nn`, which we can then use for training a model using the following command.

```
./ZOPH_RNN -t /path/to/source_training_data.txt /path/to/target_training_data.txt model.nn --vocab-mapping-file my_mapping.nn
```

Instead of using the `create_vocab_mapping_file.py` script, we can also use an existing model as the input for the `--vocab-mapping-file` flag

```
./ZOPH_RNN -t /path/to/source_train_data.txt /path/to/target_train_data.txt model.nn --vocab-mapping-file my_old_model.nn
```

This `--vocab-mapping-file` flag only needs to be specified during training. This can also be used for sequence models in the same way.

### Force-Decoding a seq-to-seq model
Once the model finished training we can use the model file (model.nn, best.nn or any of the models output from `--save-all-best` in the previous training example) for getting the perplexity for a set of source/target pairs or do beam decoding to get the best target outputs given some source sentences. Lets do the former first. We will specify the source and target data we want to get the perplexity for along with the per line log probabilities of each sentece. The output file we specify (`/path/to/output/perp_output.txt`) will contain the per line log probabilities and the total perplexity will be output to standard out. Additionally, we can use the `--logfile` flag as before if we also want standard out to be put to a file too and the `-L` flag to change what the longest sentence the code will accept.

```
./ZOPH_RNN -f /path/to/source_perp_data.txt /path/to/target_perp_data.txt model.nn /path/to/output/perp_output.txt --logfile /path/to/log/logfile.txt -L 500
```

If we trained the model using NCE then we can use the `--NCE-score` flag, which will make the model get the per line log probabilities using an unnormalized softmax. This greatly speeds up force-decode as now a normalization over the softmax does not have to be done, but now it does not represent a distribution that sums to 1. The reason we can do this is because the NCE training objective makes the normalization constant close to 1, so we can get a reasonably good approximation.

### Kbest Decoding for a seq-to-seq model
Lets have the model output the most likely target translation given the source using beam decoding. This can be done with the `--decode` (`-k`) flag. The `model.nn` file will be the trained neural network, `kbest.txt` is where we want the output to be put to and and `source_data.txt` is the file containing the source sentences that we want to be decoded. Once again short sentences are thrown out, so we can change that using the `-L` flag.

```
./ZOPH_RNN -k 1 model.nn kbest.txt --decode-main-data-files /path/to/source_data.txt -L 500
```

By default the model uses beam decoding with a beam size of 12. We can change this using the `-b` flag.

```
./ZOPH_RNN -k 1 model.nn kbest.txt --decode-main-data-files /path/to/source_data.txt -L 500 -b 25
```

We can also output the log probabilities of each sentence being decoded and have it saved in kbest.txt using the `--print-score` flag

```
./ZOPH_RNN -k 1 model.nn kbest.txt --decode-main-data-files /path/to/source_data.txt -L 500 -b 25 --print-score true
```

Another default during decoding is that the output sentences can only be within the length range [0.5\*(length of source sentence),1.5\*(length of source sentence)]. This can be changed with this --dec-ratio flag.

```
./ZOPH_RNN -k 1 model.nn kbest.txt --decode-main-data-files /path/to/source_data.txt -L 500 -b 25 --print-score true --dec-ratio 0.2 1.8
```

### Ensemble decoding for a seq-to-seq model
In the above example we only decoded a single model. In this code you have the option of ensembling multiple outputs using the `--decode` flag. All of the models you want to ensemble must have the same target vocabulary mappings, so you must use the `--vocab-mapping-file` flag as specified above. We can ensemble together 8 models below, but any number of models can be specified by the user.

```
./ZOPH_RNN -k 1 model1.nn model2.nn model3.nn model4.nn model5.nn model6.nn model7.nn model8.nn kbest.txt --decode-main-data-files /path/to/source_data1.txt /path/to/source_data2.txt /path/to/source_data3.txt /path/to/source_data4.txt /path/to/source_data5.txt /path/to/source_data6.txt /path/to/source_data7.txt /path/to/source_data8.txt -L 500 -b 25 --print-score true --dec-ratio 0.2 1.8
```

Note that now we pass in 8 different model files and 8 different source data files. The reason for the 8 different source files is that the source vocabularies could be different for all 8 models, so different types of data can be passed in. If you want the same data passed in for all 8 model, then simply copy `/path/to/source_data.txt` 8 times as the input to `--decode-main-data-files`.

### Training a seq model
Training a sequence model is much like training a sequence-to-sequence model. Now we must employ the `-s` flag to denote that we want to train a sequence model. Lets train a model with slighly different parameters from the sequence-to-sequence model above. This model will have a hiddenstate size of 1000, minibatch size of 32, 2 layers, dropout rate of 0.3 and a target vocabulary size of 15K. Also note that now we only need to pass in one data file for training and for dev since it is only a sequence model and not a sequence-to-sequence model.

```
./ZOPH_RNN -s -t /path/to/training_data.txt model.nn -H 1000 -m 32 -l 0.2 -N 2 -M 0 1 2 -d 0.7 --target-vocab-size 15000 -a /path/to/dev_data.txt -A 0.5 --tmp-dir-location /path/to/tmp/ --logfile /path/to/log/logfile.txt --save-all-models true -L 500
```

### Force-Decoding a seq model
To force decode the model it is almost the same as force-decoding a sequence-to-sequence model. In the seq model you can also use the `-m` flag to speedup the batching process, but it will no longer output the per line log probability if `-m` is not set to 1.

```
./ZOPH_RNN -s -f /path/to/dev_data.txt model.nn /path/to/output/perp_output.txt -L 500 --logfile /path/to/log/logfile.txt
```

### Decoding a seq model
This is not a feature in the code.

### Training Multi-Source model
Lets train a multi-source model like in [Multi-Source Neural Translation](http://www.isi.edu/natural-language/mt/multi-source-neural.pdf). In this model we will have two source encoders and one target encoder. This means we need to have 3-way parallel data. The two source training files in this example are: `source1_train_data.txt` and `source2_train_data`.txt. The target training file is: `target_train_data.txt`. All 3 of these files must have the same number of lines. Notice that we must now add the `--multi-source` flag which specifies the second source training file. Additionally, we must specify it a second neural network file name that will be created just like `model.nn`.

```
./ZOPH_RNN -t /path/to/source1_train_data.txt /path/to/target_train_data.txt model.nn --multi-source /path/to/source2_train_data.txt src.nn
```

By default the model combines the two source encoders using the "Basic" method as specified in [Multi-Source Neural Translation](http://www.isi.edu/natural-language/mt/multi-source-neural.pdf). To use the "Child-Sum" method we can add the following flag `--lstm-combine 1`.

```
./ZOPH_RNN -t /path/to/source1_train_data.txt /path/to/target_train_data.txt model.nn --multi-source /path/to/source2_train_data.txt src.nn --lstm-combine 1
```

Additionally, we can use the multi-source attention model from the above paper by adding the three following flags `--attention-model 1 --feed-input 1 --multi-attention 1`. All three flags must be specified to use the multi-source attention model.

Now lets have the model use a dev set for learning rate monitoring like before.

```
./ZOPH_RNN -t /path/to/source1_train_data.txt /path/to/target_train_data.txt model.nn --multi-source /path/to/source2_train_data.txt src.nn --lstm-combine 1 --attention-model 1 --feed-input 1 --multi-attention 1 -a /path/to/source1_dev_data.txt /path/to/target_dev_data.txt /path/to/source2_dev_data.txt -A 0.5
```

### Force-Decoding a Multi-Source model
To force-decode a multi-source model the `--multi-source` flag must be specified when using the -f flag.

```
./ZOPH_RNN -f /path/to/source1_perp_data.txt /path/to/target_perp_data.txt model.nn /path/to/output/perp_output.txt --logfile /path/to/log/logfile.txt -L 500 --multi-source /path/to/source2_perp_data.txt src.nn
```

### Kbest Decoding a Multi-Source model
To decode a multi-source model two additional flags needs to be specified.

```
./ZOPH_RNN -k 1 model.nn kbest.txt --decode-main-data-files /path/to/source1_data.txt --decode-multi-source-data-files /path/to/source2_data.txt --decode-multi-source-vocab-mappings src.nn
```

### Training a Preinit Model
Lets train a model using tranfer learning as specified in [Transfer Learning for Low-Resource Neural Machine Translation](http://arxiv.org/pdf/1604.02201v1.pdf). First we need to have parent data (source and target) and child data (source and target) where the parent and child models must have the same target language. In the paper the shared target language was English.

Also note that this can only be done with seq-to-seq models and not seq models or multi-source models.

First we must make a mapping file that was shown in the "Supplying your own vocabulary mapping file" section. We can use the script `create_vocab_mapping_file_preinit.py` in the `scripts/` folder. Run the following command to create a vocabulary mapping file where all words that appear less than 5 times will be replaced by <unk> (the 5 can be changed to whatever the user wants):

```
python scripts/create_vocab_mapping_file_preinit.py /path/to/source_child_data.txt /path/to/target_child_data.txt 5 my_mapping.nn /path/to/source_parent_data.txt
```

Now we have created a mapping file `mapping.nn`, which can now be used for training. Now lets train the parent model

```
./ZOPH_RNN -t /path/to/source_parent_data.txt /path/to/target_parent_data.txt parent_model.nn --vocab-mapping-file my_mapping.nn
```

Once the parent model finished training, we can now train the child model using the following script in the `scripts/` folder.

```
python scripts/pretrain.py --parent parent_model.nn --trainsource /path/to/source_child_data.txt --traintarget /path/to/target_child_data.txt --devsource /path/to/source_child_dev_data.txt --devtarget /path/to/target_child_dev_data.txt --rnnbinary ZOPH_RNN --child child.nn
```

Once the above arguements are supplied other normal parameter flags can be added just like in the `ZOPH_RNN` executable.

### Unk Replacement in seq-to-seq model
To do Unk replacement as specified in [Addressing the Rare Word Problem in Neural Machine Translation](http://stanford.edu/~lmthang/data/papers/acl15_nmt.pdf) there is a python script provided in the `scripts/` directory. The following commands need to be run if you want to do unk replacement. This can only be done with attention seq-to-seq models.

1. When decoding (-k or --decode flags) add in the following flag `--UNK-decode /path/to/unks.txt`.

The `unks.txt` file will be generated during decoding, so save it somewhere that it can be accessed later.

2. Run the Berkeley aligner to in order to generate a t-table. The Berkeley aligner is available at: https://code.google.com/archive/p/berkeleyaligner/. A sample parameter file is provided in the scripts/berk_aligner directory. The run_aligner.sh script will run the berkeley aligner, what needs to be changed on the user's end is the unk_replace.conf file. The `execDir` field must be changed to the directory that you want the Berkeley aligner to output all of its files to. The `trainSources` field must give a path to the source training data. The `testSources` field must be changed to the path of the dev data for both source and target. The suffixes must also be changed accordinly depending on what the files were named.

A sample of what the data/train/ and the data/test/ directories should contain are below (if the foreign and english suffixes are u and e respectively):

`ls data/train`
results in
`train.e train.u`

`ls data/test`
results in
`test.e test.u`

Once the Berkeley aligner finishes running you need to take the ttable (there are two given by the berkeley aligner, one of P(source | target) and the other P(target | source)) that corresponds to P(target | source). If you make the target language english then the name of the ttable is: `stage2.2.params.txt` else `stage2.1.params.txt`.

Now we can run the following command to decode a seq-to-seq model using unk replacement. `/path/to/unks.txt` is where additional information will be stored during decoding when using unk replacement.

```
./ZOPH_RNN -k 1 model.nn kbest.txt --decode-main-data-files /path/to/source_data.txt -L 500 -b 25 --print-score true --dec-ration 0.2 1.8 --UNK-decode /path/to/unks.txt
```

Next we will run the `scripts/unk_format.py` script to convert the output of the ZOPH_RNN code into correct format for the `scripts/att_unk_rep.py` script.

```
python scripts/unk_format.py kbest.txt kbest.txt.formatted
```

Next we will run the final `scripts/att_unk_rep.py` script.

```
python scripts/att_unk_rep.py /path/to/source_data.txt kbest.txt.formatted stage2.2.params.txt kbest.txt.formatted.unkrep
```

Now the `kbest.txt.formatted.unkrep` will contain the decoded sentences with the rare words replaced. The format is 1 output per line.

### Models from papers:
Here are sample commands that can be run to create models in the papers above:

For the paper [Multi-Source Neural Translation](http://www.isi.edu/natural-language/mt/multi-source-neural.pdf). Here is the command to train a multi-source attention model with german and french as the input and english as the output. If you dont want attention remove `--attention-model 1 --feed-input 1 --multi-attention 1` and if you want to use the basic method of combination instead of the child-sum method then change `--lstm-combine 1` to `--lstm-combine 0`.

```
./ZOPH_RNN -t german_train_data.txt train_english_data.txt model.nn -n 15 -B best.nn -m 128 -H 1000 -l 0.7 -w 5 -a german_dev_data.txt english_dev_data.txt french_dev_data.txt -A 1 -v 50000 -V 50000 --clip-cell 50 1000 -N 4 -M 0 1 1 2 3 --multi-source french_train_data.txt src.nn -d 0.8 -L 65 --logfile log.txt --screen-print-rate 15 --fixed-halve-lr-full 11 -P -0.08 0.08 --lstm-combine 1 --attention-model 1 --feed-input 1 --multi-attention 1
```

For the paper [Simple, Fast Noise Contrastive Estimation for Large RNN Vocabularies](http://www.isi.edu/natural-language/mt/simple-fast-noise.pdf) the command below will train the billion word language model.

```
./ZOPH_RNN --logfile log.txt -a english_dev_data.txt -s -t english_train_data.txt model.nn -B best.nn --NCE 100 --screen-print-rate 300 -N 4 -M 0 1 2 3 3 -l 0.7 -P -0.08 0.08 -A 0.5 -d 0.8 -n 20 -c 5 -H 2048 --vocab-mapping-file my_mapping.nn -L 205
```

For the paper [Transfer Learning for Low-Resource Neural Machine Translation](http://arxiv.org/pdf/1604.02201v1.pdf) the following command wil train the parent model and the child model (with the child language being Uzbek).

```
python scripts/create_vocab_mapping_file_preinit.py uzbek_train_data.txt english_child_train_data.txt 5 my_mapping.nn french_train_data.txt

./ZOPH_RNN -t french_train_data.txt english_parent_train_data.txt -H 750 -N 2 -d 0.8 -m 128 -l 0.5 -P -0.08 0.08 -w 5 --attention-model 1 --feed-input 1 --screen-print-rate 30 --logfile log.txt -B best.nn -n 10 -L 100 -A 0.5 -a french_dev_data.txt english_parent_dev_data.txt --vocab-mapping-file my_mapping.nn
```

Once the parent model finishes training then run:

```
python scripts/pretrain.py --parent best.nn --trainsource uzbek_train_data.txt --traintarget english_child_train_data.txt --devsource uzbek_dev_data.txt --devtarget english_child_dev_data.txt --rnnbinary ZOPH_RNN --child child.nn -d 0.5 -l 0.5 -A 0.9 -P -0.05 0.05 -w 5 -L 100 -m 128 -n 100 --attention_model True --feed_input True
```

For the paper [Effective Approaches to Attention-based Neural Machine Translation](http://stanford.edu/~lmthang/data/papers/emnlp15_attn.pdf) the command below will train the "Base + reverse + dropout + local-p attention (general) + feed input" from Table 1.

```
./ZOPH_RNN --logfile log.txt -a english_dev_data.txt german_dev_data.txt -t english_train_data.txt german_train_data.txt model.nn -B best.nn --screen-print-rate 300 -N 4 -M 0 1 2 2 3 -L 50 -l 1 -P -0.1 0.1 --fixed-halve-lr-full 9 -A 1 -d 0.8 -n 12 -w 5 --attention-model 1 --feed-input 1 --attention-width 10 -v 50000 -V 50000
```

For the paper [Sequence to Sequence Learning with Neural Networks](http://arxiv.org/pdf/1409.3215.pdf) the following command will train the "Single reversed LSTM"

```
./ZOPH_RNN -t source_train_data.txt target_train_data.txt model.nn -H 1000 -N 4 -v 160000 -V 80000 -P -0.08 0.08 -l 0.7 -n 8 --fixed-halve-lr 6 -m 128 -w 5 -L 100
```

# Changes from previous version
- The flag (`--HPC-output`) has been renamed to (`--logfile`)
- The flag (`--source-vocab`) has been renamed to (`--source-vocab-size`)
- The flag (`--target-vocab`) has been renamed to (`--target-vocab-size`)
- The flag (`--random-seed`) now takes in an integer to use as the fixed random seed, or by default now seeds with the time
- The flag (`--save-all-best`) has been renamed to (`--save-all-models`)
- The flag (`--feed_input`) has been renamed to (`--feed-input`)
- The default minibatch size was changed from 128 to 8
- The default hiddenstate size was changed from 1000 to 100
- Added attention, multi-source, NCE, unk-replacement, transfer learning

# License
MIT

================================================
FILE: README_XING.md
================================================
# Decoding with Finate State Acceptor (FSA), Locality Sensitive Hashing (LSH) and Word Alignemnt (WA)

This is the description for additional features built on top of Zoph\_RNN by Xing Shi.
Please contact [Xing Shi](http://xingshi.me)(shixing19910105@gmail.com) for any questions

All the following papers are based on this code:

1. [Why Neural Translations are the Right Length](http://xingshi.me/data/pdf/EMNLP2016short.pdf)
2. [Does String-Based Neural MT Learn Source Syntax?](http://xingshi.me/data/pdf/EMNLP2016long.pdf)
3. [Generating Topical Poetry](http://xingshi.me/data/pdf/EMNLP2016poem.pdf) (for FSA decoding)
4. [Hafez: an Interactive Poetry Generation System](http://xingshi.me/data/pdf/ACL2017demo.pdf) (for FSA decoding)
5. [Speeding up Neural Machine Translation Decoding by Shrinking Run-time Vocabulary](http://xingshi.me/data/pdf/ACL2017short.pdf) (for LSH and WA decoding)

# Instructions for compilation/using the code
The source code is provided in `src/` directory. You can compile the code into a standalone executable by:

```
bash scripts/compile.xing.sh
```

The executable `ZOPH_RNN_XING` will appear in the root folder. A pre-compiled executable `ZOPH_RNN_XING` is in folder `executable\`.

You should set the 7 environment variables and have the required libraries mentioned in `README.md` before compilation.
Two additional requirements:

* CUDA version greater than 8.0
* gcc version greater than 4.9.3

# Decoding with FSA

This section will describe how you can constrain the RNN decoding with a FSA so that the every output will be accepted by the FSA.

The FSA file format should follow the format defined by [Carmel](http://www.isi.edu/licensed-sw/carmel/).

All the script and sampled data file related with this section are in folder `scripts/fsa/`.

## Example data preparation

We will focus on a simple task : Translate numbers into letters, and use a fsa file to force the output to have following words: `["lstm","is","great","slow","luckily","we","make","it","fast","enough","and","with","fsa"]`

To generate fsa file and coresponding train and dev set:

```
cd scripts/fsa/
python generate_fsa.py
```

It will generate the following:

* source.train.txt : 6 number sentences.
* target.train.txt : 6 letter sentences.
* source.valid.txt : 2 number sentences.
* target.valid.txt : 2 letter sentences.
* fsa.txt

Here, we define `EXEC=../../../executable/ZOPH_RNN_XING`.

### [Train] Train the translation model

```
$EXEC -t source.train.txt target.train.txt model.nn -L 15 -H 40 -N 1 -M 0 1 -B best.nn -a source.valid.txt target.valid.txt -d 0.5 -P -0.05 0.05 -w 5 -m 20 -n 20 -l 1 -A 0.8
```

### [Decode] decode the top 10 for the source.valid.txt

```
$EXEC -k 10 best.nn kbest.txt --print-score 1 -b 20 --decode-main-data-files source.valid.txt
```

## Batch Mode
`Batch Mode` means we will translate all the sentences in the `source.valid.txt` file with the same FSA file `fsa.txt`

### [Decode + FSA] Decode the top 10 with fsa

```
$EXEC -k 10 best.nn kbest_fsa.txt --print-score 1 -b 5 --fsa fsa.txt --decode-main-data-files source.valid.txt
```

### [Decode + FSA + Beam Info]
To see the beam cells during decoding, use flag `--print-beam 1`
```
$EXEC -k 10 best.nn kbest_fsa.txt --print-score 1 -b 5 --fsa fsa.txt --print-beam 1 --decode-main-data-files source.valid.txt
```

### [Decode + Fsa + Beam Info + encourage-list + repeat-penalty + adjacent-repeat-penalty + alliteration + wordlen ]

Beside FSA, we also provide several other weights to control the output:

1. Encourage Lists and Encouange Weights: Each encourage list is a file contains either a list of words (like `enc1.txt`) or a list of words with weights (like `enc2.txt`). You feed the encourange lists by flag `--encourage-list enc1.txt enc2.txt` For each encourage list, you should also assign a weight by flag `--encourage-weight 1.0,-1.0`.
2. Repeat penalty: to prevent producing repeated words during decoding. Use flag `--repeat-penalty -1.0`.
3. Adjacent repeat penalty: to prevent producing consectuive repeated wrods. Use flag `--adjacent-repeat-penalty -1.0`.
4. Alliteration: Use flag `--alliteration-weight 1.0`.
5. Word length weight: Use flag `--wordlen-weight 1.0`.

Please refer the paper [Hafez: an Interactive Poetry Generation System](http://xingshi.me/data/pdf/ACL2017demo.pdf) for detailed description of these style controls.

Put all the flags together, we got:

To reproduce the results in [Hafez: an Interactive Poetry Generation System](http://xingshi.me/data/pdf/ACL2017demo.pdf), please also checkout the [Rhyme Generation code](https://github.com/Marjan-GH/Topical_poetry) and [Web Interface code](https://github.com/shixing/poem).

## Interactive Mode

Usually, you want to decode different sentence with different FSA. With `Batch Mode`, you'll have to create a `soruce.txt` and reload the whole RNN model from disk for each sentence, which can takes 1-2 minutes.

Thus, we provide the `Interactive Mode`, where you can provide different `source.txt` and `fsa.txt` without reloading the RNN model.

To enable `Interactive Mode`, use flag `--interactive 1`

```
$EXEC -k 10 best.nn kbest_fsa.txt --print-score 1 -b 5 --fsa fsa.txt --print-beam 1 --decode-main-data-files source.valid.txt --repeat-penalty -1.0 --adjacent-repeat-penalty -1.0 --interactive 1
```
Once loaded the RNN model, it will print the following instructions:
```
Please input k:<k> source_file:<source_file> fsa_file:<fsa_file> repetition:<repetition_weight> alliteration:<alliteration_weight> wordlen:<wordlen_weight> encourage_list_files:<file1>,<file2> encourage_weights:<weight1>,<weight2>
```
Follow the instruction and input the following in STDIN:
```
k:1 soruce_file:source.single.txt fsa_file:fsa repetition:-1.0 alliteration:1.0 wordlen:1.0 encourage_list_files:enc1.txt,enc2.txt encourage_weights:1.0,-1.0
```
It will decode the print the results in STDOUT. Then you can type another commend into STDIN to decode another sentence. NOTE: `source.single.txt` should contains only one sentence.

## Interactive Line Mode

This mode is specially for the interactive poem generation task where human and machine compose a poem line by line alternatively. Use flag `--interactive-line 1` to enable this mode.

```
$EXEC -k 10 best.nn kbest_fsa.txt --print-score 1 -b 5 --fsa fsa.txt --print-beam 1 --decode-main-data-files source.valid.txt --interactive-line 1 --interactive 1
```
You can choose one of the following three commend to type in STDIN:

1. `source <source_file>` : process the source-side forward propagation.
2. `words word1 word2 word3` feed the target-side RNN with words sequence `word1 owrd2 word3`. This is supposed to be the line that human composed.
3. `fsaline <fsa_file> encourage_list_files:enc1.txt,enc2.txt encourage_weights:1.0,-1.0 repetition:0.0 alliteration:0.0 wordlen:0.0` Let the RNN to continue decode with FSA.

Both step 2 and 3 will start from the previous hidden states and cell states of target-side RNN.

You can also ensemble two models `best.nn.1` and `best.nn.2` by:

```
$EXEC -k 10 best.nn.1 best.nn.2 kbest_fsa.txt --print-score 1 -b 5 --fsa fsa.txt --print-beam 1 --decode-main-data-files source.valid.txt source.valid.txt --interactive-line 1 --interactive 1
```

and addtionally, you can use `words_ensemble` option to provide two different human inputs for the two models:

4. `words_ensemble word11 word12 word13 ___sep___ word21 word22 word23 ___sep___` feed the target-side RNN with words sequence `word11 owrd12 word13` for `best.nn.1` and `word21 word22 word23` for `best.nn.2` These are supposed to be the lines human composed.

# Decoding with Word Alignment

Suppose we are translating from French to English, we could use the word alignment information to speed up the decoding. Please find details in 5. [Speeding up Neural Machine Translation Decoding by Shrinking Run-time Vocabulary](http://xingshi.me/data/pdf/ACL2017short.pdf).

The commend to decode with word alignment information is:

```
$EXEC --decode-main-data-files $SRC_TST -b 12 -L 200 -k 1 $fe_nn_attention $OUTPUT --target-vocab-shrink 2 --f2e-file $ALIGNMENT_FILE --target-vocab-cap $NCAP
```
where `$ALIGNMETN_FILE` is the file that contains word alignment information with the following format:
```
<Source Word Id> <Target Candidate Word Id 1> <Target Candidate Word Id 1> ... <Target Candidate Word Id 10>
```
Each line starts with `<Source Word Id>` and follows with `$NCAP` candidate target word ids. The source word id and target word id should be consistant with the word ids used by model `$fe_nn_attention`.

# Decoding with Locality Sensitive Hashing

To decode with Winer-Take-All LSH:

```
$EXEC --decode-main-data-files $SRC_TST -L 100 -k 1 $fe_nn $OUTPUT --lsh-type 1 --WTA-K 16 --WTA-units-per-band 3 --WTA-W 500 --WTA-threshold 5 --target-vocab-policy 3 -b 12 > $LOG
```

================================================
FILE: executable/ZOPH_RNN_1
================================================
[File too large to display: 41.8 MB]

================================================
FILE: executable/ZOPH_RNN_2
================================================
[File too large to display: 22.5 MB]

================================================
FILE: executable/ZOPH_RNN_3
================================================
[File too large to display: 38.4 MB]

================================================
FILE: executable/ZOPH_RNN_4
================================================
[File too large to display: 11.4 MB]

================================================
FILE: executable/ZOPH_RNN_XING
================================================
version https://git-lfs.github.com/spec/v1
oid sha256:9331caaa8c6bf6ac7eb07a2d59e0706ad70f610ead632953de49015229cb0b57
size 269

================================================
FILE: sample_data/dev_english.txt.tok.lc
================================================
parliament does not support amendment freeing tymoshenko
today , the ukraine parliament dismissed , within the code of criminal procedure amendment , the motion to revoke an article based on which the opposition leader , yulia tymoshenko , was sentenced .
the amendment that would lead to freeing the imprisoned former prime minister was revoked during second reading of the proposal for mitigation of sentences for economic offences .
in october , tymoshenko was sentenced to seven years in prison for entering into what was reported to be a disadvantageous gas deal with russia .
the verdict is not yet final ; the court will hear tymoshenko &apos;s appeal in december .
tymoshenko claims the verdict is a political revenge of the regime ; in the west , the trial has also evoked suspicion of being biased .
the proposal to remove article 365 from the code of criminal procedure , upon which the former prime minister was sentenced , was supported by 147 members of parliament .
its ratification would require 226 votes .
libya &apos;s victory
the story of libya &apos;s liberation , or rebellion , already has its defeated .
muammar kaddafi is buried at an unknown place in the desert . without him , the war is over .
it is time to define the winners .
as a rule , islamists win in the country ; the question is whether they are the moderate or the radical ones .
the transitional cabinet declared itself a follower of the customary sharia law , of which we have already heard .
libya will become a crime free country , as the punishment for stealing is having one &apos;s hand amputated .
women can forget about emancipation ; potential religious renegades will be executed ; etc .
instead of a dictator , a society consisting of competing tribes will be united by koran .
libya will be in an order we cannot imagine and surely would not ask for .
however , our lifestyle is neither unique nor the best one and would not , most probably , be suitable for , for example , the people of libya .
in fact , it is a wonder that the islamic fighters accepted help from the nonbelievers .
their only excuse is that from their point of view , nato raids were not controlled by an american general , but allah , the greatness of whom they celebrated with each strike .
when winners are sought for in libya , the west is not on the list .
we participated in a shooting , served islam , and our politicians got rid of a dictator , their political ally , without any guarantee of profit .
given the recent bad experience from afghanistan and iraq , we decided against military occupation of libya .
the locals thus do not have to worry about expelling of the &quot; crusaders . &quot;
without the occupation , will the oil companies get libya &apos;s black gold for next to nothing , though ?
it is quite likely that they will not . the west will then finally be able to boast about its selfless protection of human rights .
unless the followers of the sharia law put a stop to that .
new era of crisis commences
france and the rest of europe are now in the investors &quot; spotlight .
a statistical storm blew over europe on tuesday .
individual countries released their 3q gdp data .
between quarters , the czech economy stagnated ; germany and france are growing .
yet the debit market is breaking records today .
investors are no longer looking at only italy ; they now focus on spain , france , austria , and others .
eurozone statistical offices in germany , czech republic , and some other countries , released their preliminary estimates of economic development in the third quarter .
germany &apos;s gdp has grown by the expected 0.5 % ; and france &apos;s by 0.4 % .
both countries also adjusted their second quarter data , where germany did better than expected ( a 0.3 % increase instead of the original 0.1 % ) . france reported a slight decrease , by 0.1 % ( the original information suggested stagnation ) .
in this year &apos;s third quarter , the czech economy showed an interim increase of 1.5 % .
between the quarters though , the gdp stagnated . according to the czech statistical office , this confirmed the trend of a gradual slowing of economic growth .
the results are worse than estimates by analysts .
key events :
in this year &apos;s third quarter , czech economy showed an interim increase of 1.5 % ; however , it stagnated between the quarters .
based on preliminary and seasonal data , in the third quarter , the german economy grew by 0.5 % .
between july and september , the economy of both the eurozone and the european union grew , compared to the previous quarter , by 0.2 % , which corresponds with estimates by analysts .
the greek economy showed an interim decrease of 5.2 % .
the earnings on its 10-year bonds are 28.45 % .
earnings of both italian and spanish bonds are growing towards critical limits .
even austria , netherlands , france , and belgium , etc. are beginning to experience problems .
despite the fact that the situation in the region is getting worse , earnings of czech 1 - year bonds stay under 4 % .
voices are coming from germany , suggesting that ecb be the last resort creditor .
czechs qualified for euro .
in montenegro they won 1 : 0 and celebrate a 200 million windfall .
football representation successfully managed the toughest task of the season !
thanks to both the great čech and beautiful goal of petr jiráček , they won 1 : 0 .
in the second half , czech football was lucky to maintain its luxury lead from the home return match , and will again feature at the european championship .
ihned.cz watched the progress match in a detailed report .
nearly no excitement in the first half , and a great portion of luck in the other .
czech representation team , partially due to the home team &apos;s failure in the close ; the great čech ; and heroic jiráček ; won 1 : 0 in montenegro , and celebrates moving on to the european championship .
perfect czech team defence only allowed the montenegro team to threaten in the second half .
damjanovic and vucinic became the match &apos;s least lucky couple after losing two great opportunities .
the wonderful petr čech is one of the reasons the czech representation team will not miss the next year &apos;s european championship . and czech football will thus gain nearly 200 million czech crowns .
from the opening minutes , the game was very pugnacious .
both teams found it difficult to join in the combination ; the game remained around the mid-field line most of the time .
that was also why there was no pressure from the home team in the first half . long kicks behind the defence were neatly averted by the czech defence .
iranian students are prepared to die for their country &apos;s nuclear program
they created a live shield .
several hundred university students created a live chain around the nuclear power plant in iran .
the gesture is to express their support of their country &apos;s nuclear program against a possible israel attack .
at tuesday noon , students &apos; prayers in front of the building were followed by their chanting of death to both the usa and israel .
in isfahan , iran , several hundred university students created a live chain around isfahan uranium conversion facility .
this way , they express their readiness to protect their country &apos;s nuclear program with their own lives .
in its recent report , the international atomic energy agency ( iaea ) stated that everything leads to the conclusion that iran has not stopped striving for a nuclear weapon .
possible attack on iran nuclear facilities has been discussed in israel .
at tuesday noon , students &apos; prayers in front of the entrance into isfahan uranium conversion facility were followed by their chanting of death to both the usa and israel .
in case of an attack , aimed at the highest possible paralysation of the iranian nuclear program , the targets would be , based on israel press , the facilities in both isfahan and natanz , where the uranium is being enriched .
an obese child &apos;s diet : no breakfast and sausage for dinner
over a third of children between 9 and 13 are overweight ; while 9 % of schoolchildren are overweight , 5 % are obese .
as a recent research shows , obese children usually come from families with unhealthy lifestyle .
again , we can see the importance of the environment and models surrounding a human being for his or her healthy development .
for example , sausages occur up to half more frequently on the tables of obese children compared to the families of children with standard weight .
over 19 % of obese children had fried chips for dinner the day before the research connected to the educational event called obesity is no accident .
moreover , the families of obese children do not practise sports . however , this concerns a majority of families of current schoolchildren - only 15 % of the children said their parents find time for sports activities at least once a week .
the research was performed among 900 children .
the research took place at the end of last year &apos;s school year ; it included almost 900 elementary school children from throughout bohemia .
it was the second year of the obesity is no accident event supported by both všeobecná zdravotní pojišťovna insurance company and unilever .
so far , the project has been joined by over 12 thousand school children from 4th to 8th grades .
proper diet and sufficient sports activities represent key defining factors causing dramatically growing number of children who are either overweight or obese .
&quot; for example , only half of those inquired said they have five meals a day , in the breakfast-snack-lunch-snack-dinner pattern , one of the basic preconditions of correct diet , &quot; one of the results of the research says .
no lunch and sausage in the evening
experts warn that even relative details are crucial for healthy development ; such as a regular time of family breakfast .
in the czech republic , only 23 % of schoolchildren &apos;s families have breakfast together .
in a majority of families , everyone has breakfast separately .
the worst discovery was the fact that in one tenth of households people do not have breakfast at all .
yet breakfast is the core of proper daily diet .
a survey has shown that fries are consumed more often in the families of obese children .
nearly everyone - regardless of their weight - drinks sweet soft drinks every day ( 70 % ) .
though over 40 % of children do sports regularly , the number drops down to less than a half - 15 % - when it comes to their parents .
compose your own menu
another part of the survey allowed selected children ( 204 boys and 172 girls from 13 schools ) to compose their own menu based on selection offered and their taste and opinions .
&quot; diet preferences of czech children between 12 and 15 are far from ideal as regards healthy diet . &quot;
their favourite meals include large number of flour , meat , and sweet dishes .
&quot; especially alarming is the high popularity of sweet soft drinks , &quot; the final report of the survey concludes .
as the main dish for lunch , pasta won very tightly over poultry .
however , children showed very little interest in legumes , fish , and vegetable lunches .
tendency towards health
the survey also showed that children are willing to adjust their eating habits provided they are given correct information .
&quot; many participants from the survey control group showed a significant shift of preferences towards healthy food when completing a questionnaire after the end of the obesity is no accident educational lecture , &quot; project results show .
this was quite significant , for example , in case of sweet soft drinks , where more children would begin to prefer plain water .
in addition , for their main course , children ticked poultry and fish more often , limiting sweet meals and dumplings .
a stylist must also be a good psychologist , professionals claim
in the opening article on stylist training , i was complaining a bit about the amount of information in the introductory lessons .
i had no idea what was coming .
lectures by three experts from the field ; dressing tips for difficult figures ; and a seminar on how it works in practise .
and more is yet to come .
the closer we get to the final exam , the more information and practical tasks await .
the training takes place every saturday and sunday , but it is really packed with information .
this time , together with alenka , the director , ( there is not a single man in the course ) we discussed how to dress a woman whose body is not in perfect symmetry .
a majority of women are not perfectly symmetrical .
one has a short neck , the other has a boyish figure without proper curves , yet another is overweight .
a stylist must know how to optically soften wide shoulders , make the neck look longer , or add the curves necessary .
we were given tips on how to accentuate the figure of women such as jennifer lopez , i.e. with bigger breasts and hips and narrow waist , and not to dress them too provocatively .
for example , if the difference between the waist and hips is such that the skirt or trousers do not fit , it is worth to have at least one piece custom-made .
you will definitely not regret the investment , as perfectly fitting clothes are priceless .
in a custom-made skirt , you will feel much better than in one that keeps shifting , rolling up , sticking out in your waist , etc .
boyish figures , on the other hand , can wear clothes with ruffling , appliqué , embroidery , or shirring .
given the abundance of such information and figure types , i am glad to have a whole week to sort it all out in my head .
still , i fear it is too much , with regard to the homework we were given .
each of the course participants was given a celebrity to create two outfits for : day and night .
we must consider their current appearance , their fashion style , and figure type .
i am to dress miranda kerr whom you may know from the victoria &apos;s secret campaigns .
inside i rejoice , as she is so much easier to dress than , say , britney spears , christina aguilera , who has recently gained about a ton ( please , understand i am exaggerating ) , or the very young selena gomez .
we must cut the clothes out of magazines or find them on the internet and print and graphically process them .
moodboard , as the result is called , then expresses our personal style and attitude to graphic depiction .
it is important to think about the graphic part of styling .
a stylist does not just dress physical people , but creates product pages in magazines ( based on specific instructions , for example , trendy purses , coats , etc . ) .
this is what a stylist and designer , mario kameník , was telling us about .
but that was not until sunday .
a stylist needs general knowledge
saturday &apos;s lecture , filled with information on figure types , was livened up by three guests from the field .
a stylist , honza pokorný ; a fashion designer , jakub polanka ; and a fashion photographer , ben renč .
jakub reminded us that fashion is an illusion , and a stylist should be able to use this fact .
a good stylist can work with a single piece of clothing in different ways .
in this case , variability is key .
for example , a trench coat may be worn in the standard way , over jeans and t-shirt , as a coat ; but it may also be worn on naked body , with step-ins , as a dress .
another option is rolling up the sleeves or collar , and the entire look changes completely .
&quot; play with it , &quot; he emphasised .
however , i appreciated the following advice the most : &quot; for your work , you need to know the rules perfectly . &quot;
&apos; then you can start breaking them , &apos; jakub told us .
here i finally understood why we drill all the definitions and rules of what to combine with what and why , while in fashion magazines we often see something else .
not to mention fashion shows .
from ben , the photographer , i learned that a stylist must have more outfit versions prepared than the number he is to take pictures of .
so , for a ten-page fashion story ( equals ten outfits ) , i have to be ready to create at least fifteen of them .
now i also know that in the picture , a piece of clothing often looks and functions in combination with others completely different from reality .
we must therefore take that into consideration when creating an outfit .
honza pokorný confirmed that image and style often are half the success .
just look at madonna , who is no miraculous singer , yet her image and marketing turned her into a worldwide star .
the attitude towards stylist services is still rather cold in the czech republic .
even famous people often consider it redundant ( believing it is sufficient to be good at something , and how they look at it comes second ) .
and if someone finally lets us create their styling , then we should expect them to give us different ( smaller ) size than what they really wear .
then , we are left to google and find out if the tabloids wrote something about a celebrity gaining / losing weight , etc .
on sunday , both honza and mario were stressing the fact that for a stylist it is not enough to know about fashion and be able to create an outfit .
our work chiefly consists of communicating with people we dress ; it is therefore important to be a good psychologist .
to be able to estimate people , not to pander , and stick to one &apos;s opinion .
&quot; if a manager asks you to go through her clothes and add some nice pieces , you cannot throw out her entire wardrobe telling her how awful the clothes are . &quot;
&quot; find at least a few pieces you can praise , and recommend what to wear them with &quot; mario says .
he also showed us what specifically a duty bag must contain .
that is the bag you fasten around your waist to carry all the props you need .
aside from the obvious things , such as scissors , a sticky roller , or safety pins , it should have a nylon thread ( for the stylist to use when taking pictures of products such as purses or hats ) , special clip-ons for holding the clothes in the back in case it is too large ( pins and similar things are forbidden as they damage the clothes ) , or a sewing set .
to have some change from listening , and gain some practical experience , we learned how to properly underlay shoe soles .
even when taking photographs in a studio where the surface is even , the soles might get damaged - not to mention shooting outside .
and so we were sticking away .
you must not touch the shoes with the scissors , to avoid any damage .
this requires real skill .
mario pleased me by praising my job .
the skills mentioned are really not my cup of tea .
this is just a fragment of what we heard during the second weekend course session .
i am beginning to realise that to be a professional stylist is no piece of cake , no matter what people say .
v need for speed : the run takes place across america .
review
an originally carefree ride turned into a routine morning drive to work because of a series of completely superfluous and sloppy mistakes .
the ambitions were met to a certain point , but it does not matter since nobody obviously wanted to do any extra work .
before it was released , the new version subtitled the run created high expectations .
it promised to shift the old need for speed racing series in a new direction ; the demos made it look like a game that might stand up in the action racing genre .
we were curious to see the sections where the main character gets out of the car and runs .
typical for the nfs series though , in the end many things are different from what we expected .
the game represents a possibly decent arcade race , but the whole is burdened by several unnecessary mistakes featured in the series for a number of years .
this means they are not really mistakes , but the expression of an incomprehensible laziness of the developers .
the adventure opens very well .
you are introduced into a story , where the main character , jack , decides to solve some personal issues by participating in an illegal race across america .
you get in the car on the west coast , in san francisco , and your task is to be the first one to reach new york .
the idea brings several interesting features .
the route is divided into ten stages , subdivided into over fifty races .
this means each race takes place on a different route , and given stages are set in varied and interesting environments .
one thus does not feel the game keeps offering the same things .
despite the graphics being on the same level of previous parts of the series , some of the environments are very nice to look at .
still , it is nothing world-shattering .
driving through the desert , autumnal countryside , or high in snow-covered mountains is visually very pleasant though .
in addition , the races are , from time to time , enlivened by an untypical feature .
for example , in the desert , you get into a sandstorm , significantly decreasing the visibility .
on the other hand , in the mountains , you have to be careful not to get under an avalanche .
individual races are ordered fairly well .
there are just a few types , but it is sufficient .
for a while , you try to overtake a number of competitors ( basically a regular race ) ; then you try to beat the time ( standard stopwatch race ) ; or you are in a duel with a single rival .
on top of everything else you sometimes come across the police who make your efforts more difficult .
the parts outside the car are new .
they are just interactive animations , where most of the time jack is running away from someone and you need to help him by pressing the right buttons at the right time .
nothing difficult .
it is a surprise though , that these parts are much less in number than you would expect ( just three to be precise ) .
it actually turns out to be good for the game , because this feature does not get to be boring ; within what was already said , it is one of the things that make the game more entertaining .
the driving model follows the footsteps of the previous parts .
this means that it is nothing special anymore , but it cannot be considered a negative .
if you are looking for a real-to-life simulation , you will have to look elsewhere .
still , this driving model is fully suitable for the action style racing , even though we could find better at the competition .
so far , everything is just right .
however , after some time , problems come .
in contrast with the driving model , the collision model is not that good .
from a game of this type , one does not expect complicated deformations and collisions , but when you have no idea , before crashing into any object , how your car will act , something is not right .
typically , when overtaking or passing cars tightly in common traffic , you sometimes lightly touch the other car . yet sometimes in nearly identical situation your car ends up in wild pirouettes .
or , crashed completely .
this is not such a big deal due to the restart system , but is enough to spoil the fun .
the restarts are done so that if you crash or get far off the track , the game takes you back to the last interrupted point .
nevertheless , there are only five restarts available per each race .
if you use them up , you have to start the race from the beginning .
this makes it an aid in case something does not work out , but you cannot rely on it completely .
sometimes the restarts are not enough and you need to learn a given section of the race by heart .
a significant shortcoming of the game is the fact that the same things often happen in the same places .
for example , some turns have to be driven through in a specific way , otherwise you simply do not fit in .
at other times you meet civilian cars at a given place in the very same configuration , no matter if you come first , fifth , five seconds earlier , or one minute later .
this is especially annoying in situations where you can see how your rival drives through the two cars coming in the opposite direction so easily , while when you get there , the cars are already positioned so that you have to either drive around or slow down .
it makes no difference when you get to them ; everything is set so that the situation always occurs just in front of your nose .
the competitors have other advantages , too .
while playing the game , you find several places , where , regardless of time , the same competitor overtakes you at the very same spot .
this is most obvious at the end of the game , on a long bridge , where you are driving in the opposite direction .
obviously , your rival has to get in front of you no matter what .
the thing is that when he is driving in front of you , he makes your vision worse , which causes you to crash the car a few times in the section .
one can understand the attempt at making the game more interesting and adrenalin-filled , but it is so obvious it gets annoying quickly .
it &apos;s a pity , because despite the simple driving model , the game is entertaining and sometimes even so full of action you do not have time to pay attention to what is happening around you .
it is not important whether the reason you do not get bored is the concept with fifty unique tracks , ordering of races , or something else .
the idea of changing cars during races is unusual .
you have to watch out for gas stations along the way , as that is where you car may change despite the fact that in most of the cases the one you have is fully sufficient .
cars are divided into three categories .
sports cars are suitable for cities ; strong american cars for straight highway sections ; and exotic supersports cars for technical driving , such as along crooked mountain tracks .
the key story mode is closely connected with the challenge mode ; these are the races in individual locations you know from the story .
here , you gain both medals and experience points for your profile , common for both challenges and the story .
various features unlock fir getting a higher driver level : starting with cars and ending with profile backgrounds .
again , the profile is interconnected with an on-line autolog function , which , among other things , compares the results you have reached with those of other players , and ensures community functions .
yet another setback difficult to understand ; is the situation from the last race in the story .
you and your car get into a subway tunnel ( let &apos;s say you do ) , where you speed down the tracks in the darkness ( let &apos;s say you do ) .
your speedometer shows numbers from 150 to 200 km / h ( let &apos;s say it does ) , when suddenly , a subway train comes running down from behind you , turning your car into a pile of scrap metal .
for several days , our editorial staff kept discussing whether when designing the game no one thought for a moment or whether its authors really think we are such morons .
the technical part of the game is nothing extraordinary ; but the overall visual impression is mostly pleasant , sometimes even great .
it is a pity though , that in some dramatic moments , the recording , in some cases rather tragically , decreases .
this especially concerns accidents with large number of participants , where fluent recoding would not help you much anyway .
the sound is all right ; the soundtrack is a matter of taste .
in order to please the widest possible young audience , the soundtrack is &apos; flecked in from all quarters &apos; ; but there probably will not be many people buying the game because of its music .
overall impression of need for speed : the run is quite confused .
if the game avoided those clearly redundant mistakes , we could sincerely say that the run is the best part of nfs for the past several years .
however , the game feels unfinished .
it surely is a nice entertainment , but prepare to be blinded by the mistakes with the strength of airport spotlights .
uk &apos;s choirs will sing at the students for velvet concert
the third year of the festa academia festival offers joining the international students &apos; day and velvet revolution anniversary with a musical event .
it will introduce about twenty high school and university choirs .
the main focus of the festival is on two concerts taking place on november 17 .
&quot; charles university has been greatly involved in the festival . &quot;
&quot; participating choirs include the choir of the pedagogical department and humanities ; václav hampl , the rector , will be in the committee , &quot; jakub čaloun , head of the organisation team says .
the climax will be a free public concert by high school choirs , which takes place at 2 p.m. , on november 17 , at a plaza in front of the national technical library in dejvice .
celebration jazz mass , a piece by an important czech jazzman , karel růžička is on the programme .
in the evening , in the ceremonial national house in vinohrady , there will be a &quot; students for velvet &quot; concert , featuring university choirs , including those from the charles university .
the aim of festa academica is to celebrate the international students &quot; day and commemorate the 22nd anniversary of the velvet revolution with dignity ; as well as to present the public with top quality choir music , and to develop creative cooperation among young people .
the chief organiser is českých pěveckých sborů ( czech choirs ) in cooperation with jan kepler &apos;s grammar school .
the entire event is under the auspices of bohuslav svoboda , lord mayor of prague , and václav hampl , rector of charles university .
the festival takes place from november 16 to 20 , in prague and pardubice .
the programme offers both ceremonial concerts and numerous musical meetings , students &quot; happenings , matinées , and more .
for detailed information please visit festival web pages at : www.festaacademica.cz
eu ambassadress vicenová :
czechs have the reputation of being negative and euro-sceptical
&quot; lisbon treaty caused significant imbalance between individual institutions . &quot;
&quot; i did not expect it to be such a tough war . &quot;
&quot; it will take five or six years before the situation becomes somewhat stable , calm , and before all players get used to their new roles , &quot; eu ambassadress milena vicenová opened her lecture at the law department sharply .
she proved how true she is to her &quot; clearly and openly about europe &quot; motto in the very first part of her lecture , where she described individual eu organs and the relationships between them in such a clear and intelligible , way that many lecturers from the uk law dept. could be jealous .
moreover , vicenová praised the work of new long-term president of the european council , herman van rompuy :
&quot; at first , critics said he was plain , a grey office mouse who cannot fulfil any roles ; i have to say everyone was amazed at how wonderfully he acquitted his role . &quot;
&quot; he is a person who can listen to all twenty-seven member states , communicate with them perfectly , and who has a clear vision of where the eu should be headed . &quot;
she explained her new position of foreign affairs and security policy representative as a reply to a question : &quot; who is the european union ? which phone number should i call ? &quot; ; i.e. as an important step to unification and better clarity of union &apos;s policy towards countries such as china or india .
no one wants to ban our cheese specialties
vicenová further focused on the agenda crucial for the eu in the past , present , and near future .
she described czech presidency from the year before last as a success .
czechs were dealing with the coming financial crisis ; they solved the as crisis ; and organised the east partnership summit .
she said the fall of the government at the time was a great negative causing significant damage to the czech republic .
currently , czech representatives in the eu has been very busy .
as the ambassadress said , both the presidency and the following financial crisis caused an ongoing stirring .
in addition , a new seven-year eu budget needs to be passed , which is very complicated due to the current crisis .
whether the budget will be slim and frugal ; which fields will be given preference ; which countries will be supported more and which less ; all of that is on the coming agenda .
the ambassadress reminded the key slogans she took to brussels : &quot; about europe , clearly . &quot;
&quot; i really do not like all the abbreviations , foreign words , and jargon no one understands .
how are citizens supposed to understand that in brussels , people are not trying to come up with crooked or round cucumbers , different size bananas , or that someone wants to ban our cheese specialties . &apos;
she also wants to use the eu environment better , and get more czechs into european institutions .
multispeed europe is coming
the ambassadress says czechs are said to be euro-sceptics ; as well as people noticing the negative rather than the positive sides of things .
she wishes more czech representatives were involved in european institutions .
she thinks numerous factors are to blame : for example , the modesty of our candidates and their inability to write motivation letters ; together with low support from the czech side .
along with the success of czech presidency , she highlighted the power and cooperation of the visegrad group .
&quot; believe me that when i meet someone from poland and hungary , and a slovak is coming to join us , the frenchman next to me says : &quot; hey , what &apos;s that gathering about , is that the visegrad group again ? &apos; and he is not happy about it in the slightest . &apos;
in the union &apos;s future , vicenová expects mainly new strict rules for the budget policy of member states .
rather than a new european federation , she finds the concept of two-speed , if not multi-speed europe , very realistic .
&quot; we have both countries inside and outside the eurozone . &quot;
&quot; we have schengen area , and we already have an institute of reinforced cooperation . &quot;
&quot; for the first time it was said that the countries who want are to cooperate , while those who are not willing can stand off . &quot;
&quot; the number of such institutes will be growing ; it will be a highly two-speed europe , &quot; vicenová expects .
in this regard , she appreciated the roles of both the czech republic and herman van rompuy in pushing forward the rule based on which , all topics influencing the entire eu should be negotiated by all member states , and not just those belonging into the eurozone .
brand story :
the imb innovator starts the second century of its existence
should someone search for the word that would best express the imb strategy , the closest one would probably be &quot; long-term . &quot;
to the wide public , imb is a symbol of technological revolution ; the company is seen as a highly temporary one , yet it was founded over a hundred years ago .
the company &apos;s durability goes back to its first boss , a visionary , thomas j. watson sr.
he believed in his motto saying that numerous world problems can be easily solved if people are willing to think .
his motto : &apos; think ! &apos; thus become an imprint around which the life of the future giant would unfold ( as well as the name of its internal magazine that imb started to publish in 1935 ) .
moreover , it was the foundation of company culture watson was building carefully .
&quot; for us in imb , long-term thinking means continuous movement towards the future .
imb survived and flourished for a hundred years , because it remained true to its original values , while it was not afraid to change everything around itself .
during our first century , this allowed us to transform the technology , business , and the company .
in the second century , we hope to be able to achieve even more , &quot; says current company director , samuel j. palmisano , on the occasion of the anniversary .
sophisticated system
in the 1930s , watson approached employee care similarly to baťa in the czech republic - people from imb were among the first american employees to have a paid holiday , insurance from their employer , an elegant uniform , and a sophisticated system of bonuses for the best inventors and businessmen .
just as sophisticated is ibm &apos;s long-term relationship with customers .
from the very beginning , the company was brilliant in exploiting talented people . it was one of the first world companies to abolish all discrimination , which rendered it with great selection .
consistent with the anti-discrimination strategy is the current plan expecting a woman , virginia rometty , to take over after palmisano next year .
she has been working in imb for thirty years ; she is currently a director of sales , marketing , and strategy .
in the last few years , imb has been organising jams , on-line brainstorming sessions , from which many future strategies and innovations arise .
it is no surprise that the ideas are plentiful - the company employs nearly half a million people .
institution with a good reputation
company &apos;s strong business position is ensured by its strong internal and external image .
&quot; from the beginning , the concept of ibm was that of an institution rather than a technology company , &apos; a harvard business school professor and author of a book about ibm , rosabeth moss kanter , told the economist .
&quot; ibm is not a technology firm ; it is a company that helps to solve business problems using technology , &quot; george colony of forrester research consultation company adds .
as the company itself says , there is a difference between entering a market and creating it .
during the hundred years , imb , or the &quot; big blue , &quot; as it is called in the us because of its logo , has grown into one of the largest companies in the world .
it seems to be the most universal technology company - developing hardware and software ; infrastructure ; hosting ; and consultation services within a range from servers to nanotechnologies .
it holds the most patents of all american technology companies ; five of its employees were awarded the nobel price .
its inventions include , for example , the cash point , pay card , financial swap , bar code , and floppy disc .
with about 220 billion dollars , based on market capitalisation , ibm has become the second largest technology company ( apple being number one ) this year , overrunning microsoft for the first time since 1996 .
view ahead
not all of its products fitted the taste of both the customers and period . for example , ibm failed with its own operating system os / 2 ; prodigy on-line services ; and ibm pcjr computers .
&quot; if your business is defined by movement forward , you cannot be emotionally attached to the past , &apos; says samuel j. palmisano .
that is why ibm easily builds its foundations on many platforms - if they do not work , it leaves them .
this is how it differs from , for example , its competitor , microsoft , which basically stands and falls on windows , its operating system .
perspective technology
ibm &apos;s revenues chiefly consist of services , about one fifth for each comes from software and hardware .
while in 1990 , with over a half share , it was the strongest hardware representative .
since 2000 , ibm has sold low profitable assets worth 15 billion dollars . during the same time , it has invested 58 billion dollars into purchasing of perspective technologies .
it has continued the investments even after 2008 , when the modern economic crisis first hit the world .
ibm follows the legacy of its first director , thomas j. watson , who increased the amount of investments during the great depression .
so far , imb seems to be getting through the current crisis without any harm .
since the beginning of 2008 , the price of ibm &apos;s shares has grown by remarkable 120 per cent ; company &apos;s net earnings have been growing continuously .
ibm &apos;s company motto is to change everything , except for its values .
we may therefore expect that even if another crisis comes , this flexible giant will manage .
ibm history : from cheese cutter to space travel
ibm ( international business machines ) was founded in new york , a hundred years ago . its original name &quot; computing tabulating recording corporation &quot; was quite complicated .
it took the name of ibm in 1924 . originally , this was a brand used for company &apos;s branches in canada and south america .
foundations for the later ibm were laid by a merger of four companies , the history of which goes back to the 1880s .
they owned technologies such as the legendary &quot; time clock , &quot; i.e. a clock recording the time when employees arrive and leave work .
the architect of the new company was a businessman , charles ranlett flint , who , until 1930 , also presided over the company &apos;s board of trustees .
flint hired a director of a competing company , thomas j. watson , to help him with company management .
the leader then put the ibm &apos;s founder in the shade - watson &apos;s ideas remain the base for invincible company culture and identity .
watson led the company until 1952 .
ibm , as its name suggests , started by producing and selling various business machines - from meat and cheese cutters to cash registers .
during the first four years of watson &apos;s leadership , the revenue doubled to nine million dollars , and company &apos;s vast expansion into all continents commenced .
watson became so essential for the company , that after nearly forty years , his position in the director &apos;s chair was taken by his son , thomas watson junior .
under his leadership , ibm started to develop artificial intelligence and created the first programming language .
in 1960s , both people and computers from ibm helped nasa control space flights - to mercury , saturn , and in 1969 , man &apos;s flight to the moon .
meanwhile , in 1964 , ibm developed the first computer family - the ibm line system / 360 .
the classic pc that defined the computer standards of the period , called ibm 5150 , came to the market in 1981 .
twenty years later , ibm sold computer production to a chinese company , lenovo , while buying the consultant division of pricewaterhousecoopers , and thus commenced its full focus on consulting and high added value services .
in the past three years , ibm has been highly successful with its smarter planet concept .
its goal is both smart exploitation of technologies in areas that have not yet been planned for and improving the situation in traffic or town management .
for commerce , ibm has recently introduced a smarter commerce platform that helps businessmen control the whole commerce cycle .
ibm &apos;s history is strong even in our country .
in former czechoslovakia , as the first country in both central and eastern europe , ibm founded its branch in 1932 .
today , ibm cz also runs a strategic outsourcing centre in brno .
the headquarters for both central and eastern europe resides in prague , where there also is a world level research team focusing on voice recognition .
rychtářová about her unfaithful husband :
iveto , keep the fool , i don &apos;t want him anymore !
darina rychtářová ( 53 ) said it is over !
after her husband , josef rychtář ( 53 ) , became both the bodyguard and lover of iveta bartošová ( 45 ) , darina still hoped he would wake up and come back home .
however , she does not want to wait any longer , and set clear rules for her future life .
&quot; pepa may return home , but no longer as my partner , &quot; rychtářová explained .
when darina rychtářová provided shelter for bartošová in her house in summer , she probably had not the slightest idea of what was going to happen to her marriage .
even though she had iveta move out of říčany 30 days later , pepa would not leave the singer .
moreover , he publically admitted he loved her .
after some time , darina resigned , pretended to accept her husband &apos;s affair ; and now she gave him her final goodbye .
rychytář does not care who his wife sleeps with
even if one day rychtář and bartošová would split up , his wife will not welcome him with open arms .
&quot; he is the biological father to my sons and if he wants to spend his last days in our house , no one will chase him out . &quot;
&quot; however , he will complete his days not as my beloved , lost and found old fool , but as a person i have certain bonds to , &quot; rychtářová said , adding she had already disengaged herself from her husband and only insists on mutual decency .
so far , josef does not mind that his wife closed the hind gate on him .
&quot; i respect my wife &apos;s decision . &quot;
&quot; i will be with iveta , and i don &apos;t mean to change that . &quot;
&quot; my wife decided for marriage on paper , and i respect it . &quot;
&quot; our company still operates ; i am in charge of repairs , garden , the house , and shopping . &quot;
&quot; darina and i communicate in a completely normal way , only we do not share our bed anymore , &quot; rychtář said for blesk .
he claims that if rychtářová found a new love , he would not stand in her way .
&quot; we don &apos;t sleep together , and i don &apos;t care where she goes , with whom , and if she has a boyfriend .
i know it &apos;s my fault , i fell in love with iveta , and i accept full responsibility , &quot; josef concluded .
the worst feminine habits : do they concern you ?
when you ask men to name some typical bad habits women have , they usually come up with all of them in an instant .
dear ladies , do you recognise yourself in the lines to follow , describing several women &apos;s habits that make the men see red ?
shopping
does your other half complain you are addicted to shopping ?
do you need to buy at least one piece of clothing a week ?
do you like shopping , in mass sales ?
if so , then your shopping may be seen as bad habit .
some women recognise they should somewhat control their passion for shopping .
still , who of you could resist walking through clothes shops looking for well fitting clothes and then trying them all ( ideally , in various colour versions ) in the changing rooms ?
shopping may be a hobby for a majority of women , but for men , it may become absolute hell .
just remember how many times you have told your other half to wait for you outside ...
permanent dissatisfaction
&quot; am i too fat ? &quot;
&quot; i need to lose at least five kilos - should i start doing some exercise ? &quot;
&quot; are you still attracted to me ? &quot;
do you find these sentences familiar ?
you do ?
then , you belong to the group of women obsessed by their appearance .
not that you should stop looking after yourself , but many men are literary allergic to their partners &quot; constant complaints about their looks .
the stronger sex also agrees that the higher the frequency of such complaints , the more annoying women are .
order and cleanliness first
it is like a never ending story .
cleaning the windows , vacuuming , wiping the floors , dusting , scrubbing the bathtub - at least twice a week .
do you blame your partner for stepping onto the clean tiles in dirty shoes ; your children for messing up a clean sink and leaving fingerprints over the glass pane in the living room that was so difficult to clean ?
in case you suffer from this &quot; deviation , &quot; try to realise that no extremes are good ; and that within your family , order and cleanliness should not go first .
this harmless desire for cleanliness may threaten your relationships , as well as the entire family .
bossing about
&quot; throw the underwear into the bin . &quot;
&quot; do you plan to wear the dirty socks tomorrow ? &quot;
&quot; wet dishtowels do not belong on the floor . &quot;
if you boss your partner around several times a day , he may find it very annoying .
naturally , you mean well , you just want everything to be perfect .
maybe , if men realised that , women would not have to boss them around so much ...
gossiping
some women love gossiping , offending , and plotting against others .
women are focused on relationships , which is why they like to discuss who , where , who with , how , and why .
every now and then , probably each of us releases a gossip here or there .
let &apos;s be frank , this is probably the nature all women .
however , there are tougher gossips , women who enjoy frequent gossiping and despise everyone . they are , especially within a female work team , very dangerous for their surroundings .
assists of jágr and voráček helped philadelphia &apos;s victory
during monday &apos;s nhl match , hockey forward , jaromír jágr &apos;s two assists helped philadelphia &apos;s victory 5 : 3 on carolina &apos;s ice . again , he leads the productivity of czech players in the game .
in both cases , claude giroux , used the experience of a thirty-nine year old star ; he added an extra pass and was announced best player of the match .
two assists in the match also belonged to jágr &apos;s team-mate , jakub voráček .
after the match , jágr told the overseas journalists he had predicted a great evening for giroux .
&quot; i told him before the match he would score a hat-trick . &quot;
&quot; i felt it . &quot;
&quot; you sometimes feel such things , &quot; explained jágr , after whose action giroux scored a goal on the first flyers &quot; shot .
a little bit later , second score for the team bore czech traces too .
the trailing hurricanes &apos; voráček won the puck and presented it to maxim talbot , who scored between the circles .
a second before the first break , a home player , patrick dwyer , narrowed the lead in a power-play , but the beginning of the second part was led by flyers again .
specifically by giroux .
after assist with jágr , he scored the third goal ; consequently , his shot was touched into a fourth goal by wayne simmonds .
&quot; i love playing with him . &quot;
&quot; i don &apos;t ; want to say he &apos;s the best league player , but he &apos;s definitely among the best three , &quot; said jágr , complimenting his center .
&quot; i never thought that at the end of my career i would play with a hockey player like him . &quot;
&quot; it makes me happy . &quot;
&quot; i want to assist him . &quot;
&quot; all those years in nhl , my team-mates tried to assist me , and now its reversing , &quot; jágr added .
carolina made the match more dramatic with tuomo ruutu and dwyer &apos;s goal from a penalty shot , making it a single goal difference .
further complications were avoided by matt read who increased the difference to two points .
&quot; pronger did great in our zone . &quot;
&quot; he poked out the puck , i gained speed , and read went on to the more distant pole . &quot;
&quot; i aimed at him , and i &apos;m glad it worked , &quot; voráček described his view of the last goal of the match .
its victory moved philadelphia onto the top of the eastern conference .
czech montreal defenseman , jaroslav špaček , did not finish the match with buffalo ; he left in the second period due to an upper body injury .
along with the experienced back , the candiens lost their two-goal lead and lost 2 : 3 after a penalty shoot-out .
tomáš plekanec failed to score on one of their raids .
more than lost points , the home trainer was sorry for high sick rate in the back , as the match was finished with just five defenders .
&quot; the situation is bad . &quot;
&quot; we finished the match with just one experienced player , the rest were young men . &quot;
&quot; however , there is nothing we can do about that , we just have to face it , &quot; jacques martin said .
&quot; substitution by substitution , it was overwhelming and we couldn &apos;t resist . &quot;
&quot; we should have made the game simpler , moving the puck out of our zone . &quot;
&quot; that was our weakness , &quot; a defender , p. k. subban added .
with thirty interventions , czech goalie , ondřej pavelec , helped winnipeg &apos;s 5 : 2 victory over tampa bay .
jets &quot; goals , which interrupted a five-match series of losses , were shared by five different scorers .
development of trip prices since 1989
until 1989 , only five state travel agencies were allowed in the former czechoslovakia : čedok , sporturist , ckm , rekrea , and autoturist .
as travelling to the west was strictly limited , travel agencies of the time mostly sold tours to the socialist block countries .
for their holiday , czech tourists thus went to the sea to the german democratic republic ; soviet union ( sochi , crimea near the black sea ) ; bulgaria ; romania ( both to the black sea ) ; and hungary ( balaton ) .
the only exception was the period from 1962 to 1972 , when we had the opportunity to travel to yugoslavia . nevertheless , the yugoslavian regime turned west , which caused significant limitations to trips to yugoslavia for czech tourists .
a great paradox of the socialistic period was that citizens of the socialist block ( i.e. czechoslovaks too ) were not allowed free individual travel within the socialist countries .
if a czechoslovak citizen wanted to travel , as an individual , to , for example , the soviet union or poland , an invitation letter from the country was necessary ; otherwise , individual travel was impossible .
at the time , travel agencies would offer some trips to the west , but one either had to stand in a line overnight , with a sleeping bag or have some &quot; good friends &quot; in the agency .
still , that was not all .
a foreign currency assurance was essential for travels to the west . it was not granted to politically unreliable people , who were not allowed to travel at all .
czechoslovakian currency was not freely convertible ; without the currency assurance , it was not possible to obtain financial means ( western currency ) for travelling west .
most often , czechs spent their holiday at home , which resulted in great number of cottages and summer houses , a habit that survives until now .
since 1989 , czech tourism thus experienced a real revolution .
while until 1989 , a holiday in the gdr or bulgaria was the climax of our travels , last year , czechs spent 4.5 million holidays in countries all over the world .
a great paradox we do not realise anymore is that until 1989 , we could not freely visit the currently most popular croatia , as it was part of former yugoslavia .
right at the beginning of 1990 , visa requirements for neighbouring west european countries were cancelled . this started a huge wave of short-term visits of chiefly vienna and germany .
czechoslovaks could not get enough of travelling to the west . everyone wanted to go there for at least a moment .
for example , in paris , czech tours slept in parks in tents , because their income at the time was not enough to pay for a hotel or a pension .
a typical feature of czech tourism then was bringing their own food ( cans , salami , etc . ) to save money .
another important moment was the introduction of freely convertible czech crown . after decades , our tourists thus finally able to freely buy western currencies .
our living standards and average income were growing , which , around 1997 , resulted in another important moment .
since 1997 , increased demand allowed travel agencies to prepare regular trips by charter airlines . compared to the past , air trips thus became unbelievably cheap due to the ability to order quantity discounts .
by the end of 1990s , czechs did not need visa for visiting nearly all developed world countries , except for the usa . this made our travelling even easier .
the next significant event in the history of after revolution travelling came in 2000 . high demand resulted in the historically first direct charter flight from the czech republic to an exotic country , thailand .
since then , direct charter flights take czech tourists to warm exotic countries . this winter , flights from the czech republic will reach twelve exotic destinations .
currently about 100 thousand czechs visit exotic countries every year !
yet another important event of this period , symbolising our development after the revolution , came in 2008 .
on november 17 , 2008 , the usa removed visa requirements for czech tourists . our country has thus permanently joined the world &apos;s most developed countries .
for the first time in history , in 2008 czechs spent more than a half of their holidays abroad ( 50.3 % ) ; and for the first time in history , they spent over 15 thousand crowns per person per holiday abroad , including their costs at the destination .
general meeting should find a president , mate threat returns
for nearly five months , czech football has had no president . there exists a risk that ivan hašek &apos;s successor , the head of czech republic &apos;s football association ( fačr ) will not be known even after another extraordinary general meeting , which is to take place in nymburk on thursday .
there is still no agreement as to which election rules to follow .
without it , voting for the three candidates will not start .
jindřich rajchl , miroslav pelta , and tomáš paclík are eager to lead czech football .
hašek resigned from his function in june . when leaving , he called to his colleagues to unite and find a strong personality who would lead football in the future .
however , hašek &apos;s era sees football back in quarrels between both association chambers , unable to agree on a common candidate .
this was fully shown on september 16 , when a new president was to be chosen .
general meeting delegates wasted their trip to the national house in prague , smíchov , as the negotiations remained blocked , not just regarding the election method .
no agreement was reached as to who is entitled to vote at the general meeting - whether the statutory representatives only , or also officials given power of attorney .
this single point has been partially settled since , thanks to a recommendation of the institute of state and law , according to which representation based on power of attorney is not possible .
arguments concerning the election rules continue , though .
the executive board has not been of much help either , though at its meetings , both regular and extraordinary , it is mostly preoccupied by this issue .
the view of election options has changed several times , but the key point remains the same .
the question is : how to proceed in the 3rd round to avoid mate ?
&quot; the argument is based on whether to follow the by-laws or to find another model , &quot; says one of the three fačr vice-presidents , dušan svoboda , who represents professional football in the association &apos;s management .
based on the model given by the by-laws , the winner of the 3rd round is a candidate with higher percentage support from the chamber he is the candidate for .
moravia is said to accept a compromise to have a candidate with the highest percentage of total votes from both chambers win the election .
the czech part stresses a condition that in the percentage total , the winner must gain at least two-third superiority .
&quot; however , the two previous rounds would then lose their meaning ; that would completely deny the two-chamber concept , &quot; svoboda agrees with moravia &apos;s arguments . he says it would be best not to bend the rules and vote according to the existing regulations .
somewhat surprisingly tomáš paclík agrees with his opinion even if it is estimated that the election rules defined by the by-laws fits pelta best of the candidates .
&apos; i wish the elections follow the by-laws to the maximum . &apos;
&quot; i want to avoid all kinds of representative nonsense , as at the previous general meeting , &quot; says the owner of viktorie plzeň , who is a candidate for the first time ; and in his statements he chiefly stood against pelta .
pelta remains an optimist .
he believes the new president will be known in nymburk ; he also hopes that in the future , the candidates will cooperate constructively .
&quot; all passionate stirring will calm down with today &apos;s prospective advance of the national team into the european championship . &quot;
&quot; football will be back on the happy wave , people will be more willing to cooperate , &quot; a jablonec functionary says . of the three candidates , he has the most experience with football management on both professional and regional levels .
rajchl is more reserved .
&quot; i hope for an agreement , but the situation is very ambivalent . &quot;
&quot; it is not between people from bohemia and moravia , but there are attempts at purposeful adjustment of regulations . &quot;
&quot; that may cause another blockage of the general meeting . &quot;
&quot; i hope the common sense wins , &quot; rajchl said .
there is not much time left for an agreement between the parties .
some negotiations are said to have taken place directly in montenegro ; but who knows if the football generality will be complete before the match .
an extra flight that was to take , among others , the vice presidents dalibor kučera and rajchl to podgorica did not leave prague in the morning due to a technical failure . an alternative was sought for in order to transport part of the executive board to the scene of the barraged rematch .
finally , a meeting at ministry level , said to happen on wednesday , may decide .
minister of education , youth , and sports , josef dobeš , would like to force the split parties inside the strongest sports association in the country to reach an agreement .
a lot is in the game : government grants ; football trustworthiness ; and the planned čstv general meeting dealing with new direction of czech sports following sazka &apos;s financial problems .
paclík enters the presidential fight as he cannot see a suitable candidate
if , before the extraordinary general meeting , the head of pilsen football players , tomáš paclík , saw a strong candidate , who would meet all of his expectations of a new football association of the czech republic ( fačr ) president , he would not even apply for the top function .
&quot; if there was a good quality candidate , i &apos;d fight for him and not participate in the fight . &quot;
&quot; but no such person has currently been found , &quot; paclík said to čtk .
paclík &apos;s motivation to join the fight with miroslav pelta and jindřich rajchl is to try to prevent pelta , reportedly the favourite of the elections , from entering the top of the fačr .
moreover , paclík emphasizes that should he win , he would only remain the president until 2013 , when a regular general meeting is scheduled .
in the future , he would like to see the position given to an experienced top manager , not excessively connected to football life .
&quot; i &apos;ve spoken to some people that could be considered . &quot;
&quot; i was terrible for me to hear some of them say that at this point , they don &apos;t even want to identify themselves with czech football . &quot;
&quot; i &apos;d like to change this opinion , &quot; he said .
at the moment , paclík &apos;s strongest disagreement with pelta is based on the latter &apos;s connection to roman berbr , president of pilsen &apos;s regional football association . berber is a man of controversial reputation , who is , behind the scenes , known as powerful lobbyist , capable of influencing the votes of general meeting delegates .
paclík therefore finds it unacceptable for pelta to become the president .
&quot; if he wins , current situation will be preserved and nothing in football would move forward . &quot;
&quot; combined with support from berbr , a capable businessman , this would be even more harmful for the football , &quot; paclík said .
he is hinting at the fact that some years ago , pelta &apos;s name was recorded during phone tapping in a corruption affair investigation ; pelta came out without a sentence .
&quot; i don &apos;t know what prestige pelta &apos;s victory would bring to football . &quot;
&quot; and how would we look in the eyes uefa and fifa , &quot; paclík said . he is now more detached from his former statement saying he would &quot; sell the pilsen club if pelta became president . &quot;
in case paclík wins on thursday , he could imagine cooperation with pelta in the future .
this opinion is shared by another candidate , rajchl , whom paclík supported before the september general meeting that ended as a fiasco .
&quot; he ( pelta ) would be the first one i &apos;d go to . &quot;
&apos; he is very capable . &apos;
&quot; i can see him as the head of the representation . &quot;
&apos; but he cannot manage the association . &apos;
&quot; i don &apos;t poke my nose in what i don &apos;t understand either , &apos; said paclík , who gained a lot of attention in september , when he supported the removal of board of referees , led by luděk macela , replaced by dagmar damková .
paclík claims he would dare to manage the association .
he refers to his skills from management of private companies ; since last summer , he has been managing the pilsen team that , with his support , celebrated its first league triumph in may .
&quot; i have the strength to make the football environment change , to support its improvement . &quot;
&quot; to limit the influence of some interest groups that tend to influence the disciplinary and appeal board , leading to disgusting situations , &quot; paclík said . he does not hide his antipathy for , e.g. sparta &apos;s director , daniel křetínský .
one of his priorities would be the strengthening of relations between fačr and the czech olympic committee , with whose support he would lobby at the political top to get more money for sports from the lottery tax .
&quot; that &apos;s one of the key things the new president should attend to . &quot;
&quot; if taxes are imposed on gambling , as the political parties declare , i understand their arguments , but they must clearly say where the money will go , &quot; he said .
paclík does not consider his connection with the pilsen club a burden , though he claims that an fačr president should be a non-party one .
pelta is connected to jablonec , and rajchl to prague dukla .
&quot; it &apos;s not ideal . &quot;
&quot; yet , i don &apos;t see a point in transferring my shares in the pilsen club hypocritically to someone else , while coming back there in six months , &quot; paclík says .
pelta promised new fačr regulations that would avoid election mate
two days before the general meeting of the football association , one of the presidential candidates , miroslav pelta , promised that if he is elected , he would strive for new fačr regulations to avoid future disputes over the election rules .
before the thursday negotiations in nymburk , a jablonec functionary remains an optimist and believes that the pre-election fight would be significantly suppressed by today &apos;s prospective success of our football representation .
prior to the failed election in september , pelta promised that if he was to win , he would focus on both the representation and gaining of financial means .
he decided to participate in the elections before the last general meeting at the last moment . a two-month delay thus could have been useful for him to defend his ideas concerning football management in wider context .
&quot; people at all levels have been in football for long , and they expect the president to be known on thursday . &quot;
&quot; they feel it &apos;s high time someone leads the association , &quot; claims the man chiefly connected with the jablonec club ; but he spent many years in sparta and the association &apos;s executive board .
this gives him the reason to believe he is experienced enough to manage czech football .
he believes current problems with elections are caused by the euphoria there was in the time of ivan hašek &apos;s reign .
&quot; unfortunately , regulations that were not as perfect as we all thought were passed at the general meeting in june . &quot;
&quot; if various representations were not accepted , we would have had a president since september , &quot; pelta said .
he approached the rewriting of the key document of the civil association as a decoy for the delegates .
&quot; if i become president , i will take it as my personal task , and i &apos;m willing to take personal responsibility in case the attempt fails . &quot;
&quot; it is a fundamental issue we cannot exist without , &quot; pelta claims .
other than that , he says he does not try to define himself against the rivals .
in the past , he has already announced he can imagine his cooperation with jindřich rajchl , whom he recognises for his legal education ; he is well aware of the fact that his rival candidate should become his first vice president .
for a long time , he did not attack tomáš paclík , who has constantly criticised pelta .
&quot; he &apos;s been in football for over a year , and gained some success . &quot;
&quot; however , his attacks were not decent , and many people were caught by surprise by the method of his election campaign . &quot;
&quot; this is too , is why some of them now doubt his true character , &quot; pelta said about his rival . pelta himself claims to be true to the proclamations he announced in september .
he sees his strongest weapon in knowledge of football , from the lowest competitions to representation .
the representation should be the shopping window of football .
&quot; í would also try to end all affairs , starting from the bohemians case or the aftermath of last year &apos;s corruption scandal . &quot;
&quot; right after that , i would focus on negotiations with both private and state sectors , to ensure enough money for the whole football , &quot; the forty-six year old candidate told čtk two months ago .
he then promised that his arrival to the head of the football would not mean any personnel housecleaning .
&quot; i &apos;d need to learn about the job descriptions of individual employees . &quot; that was his message for strahov , where the long period of interregnum causes understandable nervousness .
despite current catastrophic scenarios accompanying the general meeting planned , he believes that on thursday , the delegates will not come to nymburk in vain .
&quot; if the only questionable point is how to perform the elections in the third round that is something we can get over . &quot;
&quot; football needs trustworthiness it can only gain by finally having its leader , &quot; says pelta , who , two days before the crucial day of his career of a functionary , remains an optimist .
rajchl promises that under him , the fačr will start working hard
if , on thursday &apos;s general meeting , jindřich rajchl , the existing vice president , is to become the president of czech football association , many people in the association would reportedly not be pleased .
rajchl promises that if he is elected , hard work will come .
he wants to establish system changes and make members of the executive board responsible for individual football areas .
he also promised to solve the bohemians case ; improve relations with both uefa and fifa ; bring more money to football ; and continue the work of the previous president , ivan hašek .
&quot; as president , i &apos;d like to focus on football as a whole . &quot;
&quot; from international contacts with uefa and fifa , to representation , professional football , and performance football for the youth . &quot;
&quot; by this , i don &apos;t mean i &apos;d do everything on my own , &quot; rajchl said in his interview for čtk .
&quot; i &apos;m a team player , and i want to distribute the competences within the executive board so that each member focuses on a certain area . &quot;
&apos; i &apos;ll distribute tasks strictly so that in each of these segments , football makes big steps forward . &quot;
&quot; that would probably not please many people , as it would mean hard work , not just criticizing of the work of others , &quot; he added .
critical atmosphere in football movement supposedly made him doubt his candidature .
&quot; i &apos;m known to work 20 hours a day . &quot;
&quot; but i need to know if it has any sense ; and that i won &apos;t be surrounded by people throwing sticks under my feet and question my every move , &quot; he explained his hesitation .
&quot; however , many people told me i mustn &apos;t give up , otherwise others would follow . &quot;
&quot; that &apos;s a trust i can &apos;t just throw away , &quot; he said .
moreover , he says his motivation is to continue the work he started with hašek .
&quot; a great amount of work was done in those two years , and i don &apos;t want to throw it all away . &quot;
&quot; processes have been started that may bear fruit in a few years , and i don &apos;t want someone else to stop them , &quot; he said , in fear that if the other candidate , miroslav pelta , is elected , situation f the association might return to the period before hašek .
&quot; i believe in systematic approach , which is something i don &apos;t think mirek pelta is competent of . &quot;
&quot; he is chiefly surrounded by people who would do football so that it would only be for the chosen ones . &quot;
&quot; i can be blamed for many things , but certainly not for being subjective . &quot;
&quot; i try to be objective . &quot;
&quot; i want to make football for everyone , because then everyone has a good time , not just the chosen few , &quot; rajchl said .
furthermore , he refutes the allegations he would only be a czech president .
&quot; though some people try to claim the opposite , i plan to do the maximum for moravia . &quot;
&quot; for example , it is high time to consider about splitting the second league into czech and moravian parts , winners of which would continue into the first league . &quot;
&quot; this would prevent the current situation , where there are just three moravian units in the league , which is a problem concerning young players in moravia , &quot; rajchl said .
as the presidential fight has been joined by pilsen club manager , tomáš paclík , rajchl does not dare to estimate his chances to win .
&quot; nevertheless , i can still feel a strong support in the czech chamber that elected me its vice president , and can see i have achieved some clear successes , &quot; rajchl said . among his successes , he includes the saving of dozens of millions of crowns on disadvantageous contracts , negotiating tv contracts , and negotiations about the lottery law .
on the other hand , he considers the fact that the association does not stick together partially as his own fault .
while claiming that pelta has been &apos; disqualified from returning trustworthiness to czech football &apos; , he considers the other rival , paclík , a strong candidate that has things to offer .
&quot; no one can doubt that he did a great job in pilsen this year . &quot;
&quot; if he is elected , i hope he would continue in the things ivan hašek started , without significant personnel changes in the committees . &quot;
&quot; he wouldn &apos;t want to twist things for his own benefit , &quot; rajchl said about his rival .
europe &apos;s most beautiful pools
leap into the water
by julia stanek
paddling pools with universal fibre optics and status pools in budapest . travellers can experience europe &apos;s incredible water - and relax tired limbs in a jacuzzi after a city tour or a ramble .
one book can tell you where the water is at its best .
there was one thing iris meder could not do without on research trips over the last one and a half years : her swimming costume .
lying back on the water , she has looked up at artistically designed domes , has swum towards marble cherubs awaiting her at the edge of the pool and has marvelled at meticulously restored art nouveau swimming pools .
she rushed to add any pool she liked to her notebook - with essential details on the construction and the history of the building .
the architectural historian meder looked at more than 200 swimming pools - and has now compiled here experiences in her recently published book badefreuden ( the joys of bathing ) , covering everything from public swimming baths in munich to historical bathing palaces in the black forest and &quot; hardcore concrete buildings &quot; in taunus .
her &quot; journey to the most unusual baths in central europe &quot; took here through 13 countries . besides germany , austria and switzerland , she visited baths in italy , france , the czech republic , slovakia , slovenia , hungary , romania and poland - and one city each in luxembourg , serbia and croatia .
the 46-year-old is a real bathing culture enthusiast .
she has very little enthusiasm for sport , says meder - but in the water , it &apos;s a different matter .
the only downside is that fans of swimming have to cope with blue glazed tiles and the stench of chlorine while joggers can breathe in the beguiling scent of spruce forests or listen to the bright twittering of birds in the park .
&quot; one day , i asked myself why swimming baths always have to be so ugly , &quot; says vienna-based meder .
&quot; so i started collecting beautiful baths . &quot;
water surface doubles architecture
splashing in the diving pool and ethereal sauna oils - the wellness factor in baths is so great because all the senses are stimulated , according to meder .
and swimming baths can be a real treat for the eyes if the builders have gone to the trouble . &quot; you have the surface of the water in pools instead of a floor . &quot;
&quot; the surface is very transparent , but also reflects the beauty of the entire building . &quot;
everything is mirrored visually : high ceilings , colourful tiles and diving boards .
the illustrated pocketbook the joys of bathing provides valuable information on the architectural and cultural history of historic pleasure and health oases over 190 pages - and makes you want to leap into the water .
pool collector meder discovered some notable examples during research in hungary .
she found nitrate-rich water in a karst cave ( the cave bath in miskolctapolca ) and a thermal bath filled with alkaline water in a bottle-shaped building ( városi termálfürdö in jászberény ) in addition to magnificent therapeutic baths such as the 100-year-old széchenyi baths in budapest .
many of the heated outdoor baths in switzerland and austria , on the other hand , offer spectacular views . you can view all of zurich from a rooftop bath or look out on the snow-capped arlberg from an outdoor pool in st. anton . and there is a window in the steam bath giving a view of the action on the ski piste .
perhaps the most unusual pool design awaits bathers in längenfeld . the aqua dome in ötztal looks like a ufo has just landed in the alps .
it steams during the winter from three open-topped shells dwarfed by the ötztal mountains behind it .
alhambra in the black forest
pool collector meder was not greeted by the smell of chlorine in every swimming pool .
a lot of thermal baths smell of sulphur . the árpád baths in békescsaba , hungary , smell so strongly of oil you may find it difficult to believe you are in a health spa .
the smell comes from the hydrogen carbonate in the alkaline water . the water is almost black with silt that sticks to the skin as soon as you slide into the pool - a wellness experience of a particular kind .
iris meder , however , was at one point disappointed in her search for beautiful bathing places . when she tried to visit the outdoor pool at the czech baths of luhatschowitz , she found the area sealed off .
the baths , designed by the architect dušan jurkovic in 1902 , have been closed .
when meder looked through the weather-beaten windows of the red , white and yellow art nouveau building , she could see weeds growing up through the tiles .
one place where decay has been stopped in its tracks is a magical swimming bath in the black forest - the palais thermal in bad wildbad .
the baths have already gone through three stages of construction . you will not only discover neo-romanesque elements from the period after it was built in 1844 , but also a moorish hall in alhambra style added around 1900 .
following its careful modernisation in 1995 , the palais thermal is the favourite swimming pool of pool connoisseur iris meder - and not just because she comes from the black forest .
&quot; it is a real achievement to harmonise such different architectural styles so brilliantly , &quot; says meder .
a dip in the thermal pool in such a palace is clearly of secondary importance .
agricultural commodities speculation
deutsche bank has commodity trading investigated
by christian teevs
will deutsche bank stop trading agricultural commodities ?
according to spiegel online , josef ackermann has given a working group the task of investigating the consequences of speculation for the poor around the world .
the ceo plans to make changes in january .
hamburg -
for german citizens the situation is clear . according to one study , 84 percent of germans find it unacceptable that banks speculate on agricultural commodities such as wheat and maize .
two thirds of those questioned demand that deutsche bank and other banks cease trading in agricultural commodities , as trading deepens the problems faced by the very poorest people around the world .
the findings were made in a survey conducted by the forsa institute for social research and statistical analysis on behalf of the consumer organisation foodwatch .
the study is based on a report presented by foodwatch in mid-october .
the author of the report , harald schumann , shows that betting on global commodities exchanges drives up prices - and the banks are therefore complicit in global hunger .
the campaign by consumer advocates is directed specifically at the head of deutsche bank , josef ackermann , although large banking houses such as goldman sachs and morgan stanley trade in the same way .
originally , commodity betting was designed as a means of protecting traders against price fluctuations .
however , most experts now agree that the gamblers have left behind the rules of supply and demand - and are profiting at the expense of the poorest .
in contrast to us competitors , ackermann has responded to the criticism . he has promised to investigate the accusations and has said that no trade is &quot; worth putting deutsche bank &apos;s good reputation at risk . &quot;
according to spiegel online , an international working group is currently investigating the accusations contained in the foodwatch report .
the investigation should be complete by the end of the year when the findings will be presented to deutsche bank &apos;s board of managing directors - with recommendations for action .
the head of the bank then plans to announce changes following the investigation at the end of january .
commodities trading may be reduced or stopped completely .
ackermann working on his image
the president of foodwatch , thilo bode , welcomed the news while emphasising his criticism of the financial sector . &quot; betting on rising food prices is an extreme example of how thoughtlessly banks are currently acting against the common interest . &quot;
as top bank lobbyist , ackermann has a particular duty , according to bode .
&quot; deutsche bank must remove all investments from its portfolio that bet money on food prices , &quot; said bode .
&quot; it does not matter whether this is about starving people or the reputation of deutsche bank . &quot;
ackermann &apos;s sensitive response is at first surprising , but then the ceo would like to present a perfect image just a few months before the end of his term in office .
deutsche bank announced on monday that the chairman will not move to the supervisory board as had been planned .
cancelling contentious food commodity speculation could improve his image .
he also need have little concern for the costs - it would be up to his successor to manage the outcome .
one in two people would withdraw commodities investments
foodwatch is striving to keep up the pressure on the manager .
to date , more than 30,000 internet users have joined the organisation &apos;s campaign .
bode published the survey on tuesday . the survey was conducted for the forsa institute for social research and statistical analysis on 7 and 8 november and 1001 people took part .
no more than eleven percent think it is legitimate for deutsche bank to offer investments that bet on the price of foodstuffs .
many banking customers would also take direct action if they discover that their bank is involved in such trading . according to the survey , one in two would withdraw investments that end up in commodities speculation .
43 percent would recommend leaving the bank , 49 percent said they would think about closing their account and switching to a different bank .
21er haus in vienna
the relocated pavilion
by ingeborg wiensowski
an architectural history with highs , lows and a happy-ever-after . austria &apos;s pavilion for the 1958 world &apos;s fair in brussels won prizes , became a museum in vienna , stood empty and was abandoned .
now it is being reopened as an exhibition centre .
austria &apos;s pavilion , designed by karl schwanzer for the 1958 brussels world &apos;s fair , is a famous textbook example of modern post-war architecture with its clear shapes , glass halls , bold strokes and the new construction materials that expressed belief in technical progress .
schwanzer &apos;s pavilion won the grand prix d &apos;architecture in 1958 for its light floating steel-glass structure .
the building is considered a milestone in contemporary architecture .
it made the architect so famous that he also built austria &apos;s pavilion at the next world &apos;s fair .
he finally shot to fame internationally in 1973 with his legendary four-cylinder bmw tower and its bowl-shaped sister building .
the pavilion , planned as a temporary exhibition building , was relocated to the schweizer garten near vienna &apos;s südbahnhof , converted and opened as the museum of the 20th century in 1962 - and is widely referred to as the &quot; 20er haus &quot; for this reason .
however , the building was not suitable as a museum for technical and spatial reasons - it had no walls to hang artworks and poor air conditioning .
nonetheless , it continued to be used until contemporary art in vienna was given a new home .
the building stood empty from 2001 and fell into a worse and worse state of repair .
the architect adolf krischanitz has now renovated the pavilion , which has been named the &quot; 21er haus &quot; and will be used as a contemporary art museum .
the museum is a branch of the belvedere museum , which will open the first exhibition on tuesday evening - a delight in every respect .
agnes husslein-arco who became the new director of the austrian galerie belvedere in 2007 was instrumental in the new beginning .
husslein-arco , an art history graduate with a remarkable career in her youth as an ice skater , who went on to sotheby &apos;s , the guggenheim museum and to become the founding director of the museum der moderne in salzburg , convinced politicians and conservators to go ahead with renovation , attracted public funding and private sponsorship and decided on the use of the building .
the new museum was to house the fritz wotruba foundation &apos;s permanent collection , a café and a bookshop alongside contemporary art . an office building was also to be built .
strict supervision &amp; limited budget
krischanitz was then the ideal choice to carry out the project .
he was known , of course , for having renovated and expanded joseph olbrich &apos;s secession in vienna , and had won the competition , but he also studied under schwanzer at the vienna university of technology and the pavilion was an &quot; exceptionally important place &quot; for him where had visited exhibitions &quot; almost every weekend . &quot;
he knew the original design and knew the issues the building was facing , a building that had already lost some of its grandeur and lightness when it was relocated to vienna and converted .
the problems facing the latest alterations were much greater . the building was &quot; an energy guzzler , &quot; says krischanitz .
the window profiles of the glass facade were replaced , fibreglass-reinforced safety glass was put in and the glass roof was replaced by high-grade safety glass .
thermal bridge steel girders were strengthened or replaced and asbestos was removed from ceilings .
two lower floors created space for the new use of the building , a sunken light well between the building and the street provides lighting , a bridge leads to the entrance - almost a new building , but still the delicate schwanzer pavilion .
the original external doors have been retained and the building has been painted in the old red-brown colour of rust-proofing .
and the cinema hall is exactly as it used to be .
all this under the strict supervision of the monument authority and with a limited budget .
&quot; we often had to find a material that was good , cheap and met with the approval of the monument authority , &quot; says krischanitz , pointing to the rough flooring in the basement .
the building has remained flexible and light despite the constraints , problems and conversions .
krischanitz stated recently that the pavilion &quot; from brussels &quot; had been tough - a typical &quot; drop in the city &quot; story .
architecture like this is &quot; always in the wrong place &quot; and nonetheless wants to &quot; stimulate a real awareness . &quot;
the 21er haus is very much in the right place , as there has been a stroke of luck in town planning . the completely redesigned &quot; main station quarter &quot; with offices and residential properties is developing around the schweizer garten .
there may also be a second stroke of luck . the 21er haus may be bolstered by another krischanitz building - the collector francesca von habsburg has purchased the temporäre kunsthalle berlin and is reported to be negotiating with the city for a neighbouring location in the schweizer garten .
deutsche bahn plans winter crisis package into the millions
deutsche bahn has come up with a crisis plan to prevent train cancellations in the winter .
the number of employees is to double .
the costs will run into the millions .
according to a newspaper report , deutsche bahn will spend more than 70 million euros more this year in the battle to prevent winter-time cancellations and delays .
referring to an internal crisis plan at deutsche bahn , the bild newspaper has reported that investment of around 300 million euros in total is planned up to 2015 .
investment is intended to improve the availability of locomotives and the rail network , even in the face of extreme weather conditions .
the number of internal and external employees involved in removing snow from rail facilities and platforms is to double to 16,000 .
the aim is to clear platforms of snow and ice by start of business and 90 percent of the critical points within four to five hours .
rüdiger grube , head of deutsche bahn , warned of train cancellations and delays this winter at a summit with government and industry two months ago .
last winter , deutsche bahn had considerable difficulties in its passenger business , as it lacked reserves in the face of difficult weather conditions .
numerous new regional trains have been waiting for years for approval from the federal railway office .
there is also some delay in the delivery of ice trains .
deutsche bahn has also been forced to have the axles of ice trains checked significantly more frequently for many years , after an ice train &apos;s axle broke in cologne &apos;s main station .
deutsche bank pays multi-million pound fine in the usa
fresh trouble is following hard on the heels of the uproar around josef ackermann , ceo of deutsche bank : deutsche bank faces a multi-million pound fine in the usa .
as if the problems at home were not already enough for deutsche bank , the industry giant &apos;s past in the usa is now also catching up with it . frankfurt-based deutsche bank will pay 145 million dollars ( 106 million euros ) to settle disputes arising from the bankruptcy of five major credit unions during the financial crisis .
the issue centres on the sale of mortgage securities .
the national credit union administration ( ncua ) , a financial market regulator , accuses a string of major banks of having used false promises to mislead credit unions into buying the financial products .
the value of the securities dropped dramatically during the financial crisis and brought down the credit unions .
&quot; we are happy to have resolved the issue without the parties going to court , &quot; said a deutsche bank spokesman in new york .
the bank did not admit culpability in the settlement .
the same is true of citigroup , which agreed to pay 20.5 million dollars .
the ncua chairman , debbie matz , welcomed the responsiveness of the two banks .
the national credit union administration ( ncua ) is responsible for the us credit unions and intervenes in bankruptcies to protect customer deposits .
the ncua is fighting for compensation for losses that run into the billions .
the current settlements are the first of their kind .
the regulator had also approached other major banks and sued jpmorgan chase , the royal bank of scotland and goldman sachs in the summer .
the dubious mortgage securities are so-called mortgage-backed securities .
they are underpinned by home loans .
this sealed the fate of a number of financial companies when the us property bubble exploded in 2007 .
the us investment bank , lehman brothers , went belly up in september 2008 at the height of the financial crisis .
the banks battling against a strong wind in the usa several years later .
investors and regulators have brought a vast number of claims for compensation or to punish misconduct .
the federal housing finance agency ( fhfa ) has been behind the largest wave of claims .
it accuses 18 major international banks of having taken in us lenders fannie mae and freddie mac on mortgage transactions valued at around 200 billion dollars .
deutsche bank also faces legal action by the fhfa covering multiple transactions with a total value of 14.2 billion dollars from 2005 to 2007 .
the regulator is demanding that the frankfurt-based bank take responsibility for &quot; substantial losses , &quot; but has not yet specified the exact sum .
deutsche bank has rejected the claims as baseless and has declared its intention to fend off the claims .
warren buffett pumps 10 billion into ibm
financial guru buffett is investing heavily in the it company , ibm .
since march , he has bought up shares totalling more than ten billion dollars .
until now , the major american investor , warren buffett , had always claimed not to invest in it and computing companies , as it was difficult to forecast the long-term development .
however , ibm &apos;s annual reports changed his mind on investing in the industry .
according to buffett , he should have realised much earlier that ibm primarily offers services and purchases computer systems for the it departments of other companies .
investment by warren buffett is considered a knighthood in the financial world .
the it veteran ibm has now earned the title of &quot; sir ibm . &quot;
buffett revealed that he had purchased ibm shares totalling 10.7 billion dollars since march on the us business news broadcaster cnbc .
his investment holding company , berkshire hathaway , has become one of the largest shareholders in ibm with a 5.5 percent stake .
buffett said that even ibm knew nothing of his involvement until recently .
he praised the management , which has managed to bring in profits even during the economic crisis .
&quot; they have done a great job , &quot; said buffett , referring to the management strategy .
for a fairly long time , ibm has focused on lucrative it services such as data centres .
the computer pioneer - now more than 100 years old - also provides software and consultancy services , as well as powerful business computers .
buffett had avoided technology firms to date .
he claimed only to invest in companies whose business he understands
and invested in a freight railway , for example , a lubricant manufacturer and a machine manufacturer .
his holding company , berkshire hathaway , owns shares in a long list of major companies such as coca-cola and munich re ( formerly münchener rück ) alongside around 80 of its own subsidiaries .
the 81-year-old took over the small textile company berkshire hathaway in the 1960s and built it up , with clever investments , into one of the most valuable companies in the world .
his lifestyle , however , has remained humble ,
making him a cult figure for countless investors around the world .
his almost infallible instinct for making money has earned him a nickname as the &quot; oracle from omaha . &quot;
the ibm share price rose by one percent before the markets opened following the announcement of buffett &apos;s investment .
&quot; big blue , &quot; as ibm is also known , is worth more than 220 billion dollars on the stock exchange , making it one of the most valuable technology companies in the world alongside apple and microsoft .
buffett , however , has said he is not planning to invest in microsoft .
the founder , bill gates , is a close friend .
imminent danger
following the discovery of the &quot; zwickau cell , &quot; turkish media express doubt about the rule of law in germany .
one newspaper speaks of a &quot; bloody ideology &quot; that is now resurfacing .
the murder of eight turkish and one greek businessman between 2000 and 2006 was the result of right-wing extremist ideology .
fear and concern has been the reaction to this report in turkey and within the german-turkish community in germany .
for many german turks , the news brings back to life a long-dead spectre . there has been no comparable right-wing extremist violence against turks since the arson attacks in mölln in november 1992 and in solingen in may 1993 .
there was never any question , however , that xenophobia does exist in germany .
the right-wing extremist network that has now been discovered , however , is on a scale that has not yet been fully understood .
&quot; is this a revival of the bloody ideology ? &quot; was the headline of the internet edition of the turkish newspaper haberturk .
the sabah newspaper , on the other hand , points out that one of the perpetrators was a confidential informant for the intelligence services and expresses concern that the german authorities may be caught up in the right-wing extremist quagmire .
the turkish community of germany ( tgd ) reacted calmly , remembering the victims of racial violence during a silent vigil in front of the brandenburg gate at the weekend .
just two weeks earlier , the turkish community in berlin had commemorated the fiftieth anniversary of the german-turkish bilateral agreement on labour recruitment in a ceremony with german and turkish politicians .
cooperation between the two countries was presented in a positive light - as a success story .
the coverage by some turkish media has been like a slap in the face . &quot; germany &apos;s present on the fiftieth anniversary &quot; is the title of one article on the haber x turkish internet platform , which comments on the circumstances around the series of murders .
the german edition of hürriyet also does not miss the chance to make this connection .
&quot; it is beginning to stink , &quot; is the title of the column by ahmet külahci .
he points out that people have been murdered who had paid taxes and made an important contribution to germany &apos;s redevelopment with their work .
the behaviour of the police and the subsequently successful manhunt is heavily criticised on german-turkish internet forums .
the authors are asking themselves why the german authorities did not think earlier to look for suspects in right-wing extremist circles .
some writers even go as far as some strong rhetoric . &quot; if things continue to be done so cack-handedly , migrants have no option but armed self-defence . &quot;
&quot; that will be the best self-defence . &quot;
&quot; peace in germany depends very strongly on how the entire network is uncovered and on the appropriateness of punishment , &quot; writes &quot; selen . &quot;
the attacks on mölln and solingen in the 90s are an example of what it means for peace in germany to be at risk .
the photo of the burnt out ruin of solingen is deeply engrained in german and turkish memory .
no other event has shown the turks living here more clearly that there is imminent danger and shown the germans that mistrust and distance can culminate in real violence .
the attacks , however , increased the sense of turkish unity .
do not repeat the mistakes of the past
the turkish press at the time complained mainly of the failure of german policy following the attacks . helmut kohl declined to visit the survivors in mölln .
he sent a telegram of commiseration to the turkish president after the arson attack in solingen , but it was his foreign minister , klaus kinkel , who took his place at the memorial ceremony in cologne - kinkel summed up in his speech the taxes turks contribute in germany ,
just as if he were calculating the value of a person by the taxes he pays .
the coverage by the turkish media at the time again shows how deep the mistrust of many german turks in the current german state has become since .
straight talking is what is needed now , alongside efforts to identify and reassess right-wing extremist terrorism in germany .
the lake zurich banking scandal
the most risk-taking bank in europe is not in italy or france , but in switzerland .
ubs almost went bankrupt during the financial crisis and has been struggling from one crisis to the next ever since .
a german is now to bring it back in line .
axel weber certainly had not imagined that would be his role .
after his spectacular departure from the head of the federal bank , it was announced in july that he would be starting an almost equally spectacular new job in 2013 : chairman of the board of directors of the major swiss bank ubs - an unexpected and lucrative career move .
axel weber was clearly delighted , praising switzerland for its beauty and ubs for its newly won solidity and making some friendly remarks about his future colleagues .
&quot; ubs has managed to turn around . &quot;
&quot; i am looking forward to working with kaspar villiger and oswald grübel . &quot;
but by september it was already clear that everything would be very different .
a ubs trader in london had bypassed all the bank &apos;s security systems and gambled away around 1.7 billion euros .
ubs ceo , oswald grübel - a banker not only famous in switzerland - had to go .
the ubs chairman has been considered ailing ever since - and everyone is asking : when will axel weber finally arrive ?
weber is to advise the bank from february 2012 .
and he will take on the role of chairman of the board in may , and not in 2013 , the bank announced on tuesday .
that will be tough for axel weber .
he is an economics professor and central banker , not a conventional banker , and clearly would need some time to adjust .
all at once , he is taking on responsibility for the strategy of a bank that has once again proven that it has long deserved the title of &quot; europe &apos;s biggest banking scandal . &quot;
ubs is both huge - it was the biggest asset manager in the world until recently - and has been struggling from one scandal to the next for years .
the proud swiss bank , it seems , is to be spared nothing .
it began in the financial crisis .
ubs had an excellent reputation until 2007 .
the bank combined the image of the secretive and solid , but equally savvy asset manager with ambitious targets , aiming to become a major player in investment banking .
as a result , in 2007 ubs was managing 1.6 billion , thereby managing more more money from wealthy clients than any other bank in the world ,
and had risen to the league of the world &apos;s biggest banks as an investment bank .
that made an impression .
the former banker and long-standing banking professor at the swiss institute of banking of the university of zurich , hans geiger , explains : &quot; i always presented ubs to my students as the model of a good bank . &quot;
&quot; from what we know today , you have to ask how i could be so wrong . &quot;
he was very wrong indeed .
it only became clear in 2007 that something was really going very wrong at ubs .
in the middle of the year , ceo peter wuffli - considered a highly intellectual and equally brilliant banker - suddenly stepped down .
some months later we could guess at why he had left .
the investment division of the bank had gambled on subprime securities and was forced to write off billions .
ubs announced a capital increase , believe it or not , of 13 billion swiss francs to face the crisis .
the bank actually managed to find investors , but investors many bankers would have respectfully declined a short while earlier - a less than transparent asian sovereign wealth fund from singapore and an arabic sovereign wealth fund that was so secretive , even its name was not disclosed .
that was just the first blow .
it got a lot worse in april 2008 .
the bank was forced to announce new losses running into the billions and chairman of the board of directors , marcel ospel - the second best paid chairman in switzerland in 2006 - resigned .
catastrophe followed in october 2008 . ubs faced collapse from a lack of investment during the financial crisis and had to be saved by the state .
the national bank bought up toxic securities from ubs for untold billions and the swiss confederation paid six billion swiss francs for a mandatory convertible bond to become part owner of the bank .
the whole of switzerland was shocked - the country recognised the dramatic consequences for a small country of being home to such large banks .
after all , the balance sheet total for ubs in 2007 was more than five times the economic output of switzerland ( see chart ) .
&quot; we would not be able to save the ubs for a second time , &quot; says banking researcher geiger .
the second catastrophe followed closely on the heels of the first .
a few months later , ubs capitulated in the tax dispute with the usa and declared that it was prepared to hand over the data of wealthy clients who had potentially avoided paying tax with the help of the bank .
ubs also agreed to pay 780 million dollars consisting of repayment of unjustified profits from the business and a penalty .
ubs therefore more or less admitted having been an accessory to tax evasion .
at the same time , ubs lost something that had always been its advantage when competing for rich clients - the fog of secrecy .
swiss banking secrecy was not the same after the scandal .
the ceo and the chairman then resigned and
the exodus of client money began .
the bank had lost its credibility in both of its core business divisions within a few months - in investment banking thanks to bad speculation and in asset management by aiding and abetting tax evasion .
and just as things had calmed down and ubs looked like it would start to turn things around , along came a disastrous rogue trader . a single trader gambled away 1.7 billion euros before the bank noticed something was wrong .
one thing is clear - risk management in the investment banking division is still not working .
&quot; the adoboli incident clearly showed that risk control had changed too little , &quot; says martin janssen , economic and professor of finance at the swiss banking institute of the university of zurich .
the cfo of one major german bank also has no doubts : &quot; the problem facing ubs is that it does not having a handle on risk control . &quot;
it was precisely the concept of caution that had long ago made swiss bankers the guardians of the world &apos;s major assets - no excessive risk , but security for money , whether the world fell apart outside the alpine republic or not .
the safes of swiss banks had become a symbol of unshakeable reliability , just as lindt and sprüngli had come to symbolise sophisticated chocolate .
the ubs headquarters in zurich still fit the cliché of the dignified swiss private bank - just a bit bigger .
the location alone is classic .
the bank has its head office on zurich &apos;s bahnhofstrasse that leads from the main station directly to zurich lake , surrounded by the snow-capped peaks of the swiss mountains .
the buildings are ostentatious and the shop rates are among the highest in europe .
jewellers such as tiffany &apos;s are to be found here between traditional swiss shops such as confiserie sprüngli .
the swiss banks have resided here for hundreds of years - from the national bank to the palace of credit suisse and elite private banks such as julius bär .
a heavy revolving door opens onto the ubs reception .
the walls within are clad in marble behind vast leather armchairs . there is a large sign to the &quot; safe &quot; next to the bank &apos;s logo .
ubs knows that its clients are looking for - security .
but both the major swiss banks , ubs and credit suisse , at some point ousted the swiss banking concept - or betrayed it , as the head of one small swiss private bank believes .
he would prefer not to see his name in the newspaper - he too , like all swiss bankers , does business with the two major banks .
he believes ubs has grown too big - particularly in investment banking where the anglo-saxons are also very active .
in his opinion , the swiss have nothing to add .
&quot; and if they do , they have adapted to the customs of the anglo-saxons long ago . &quot;
it was in fact anglo-saxon investment banking that changed ubs , brought in its first big profits - and then the debacle . no other european financial institution has lost as much money in the subprime jungle as the major swiss bank .
the catastrophe became evident on 15 october 2008 .
&quot; that wednesday , the government coalition of a small country with a population of seven million undertook the largest rescue operation in the history of the nation , &quot; writes author lukas hässig in his book the ubs crash .
the politicians approved up to 68 billion swiss francs ( around 45 billion euros ) to save the bank from bankruptcy .
how could it have come to that ?
talking to swiss banks , they see the issue mainly in the size of ubs .
two of the three major swiss banks consolidated into a single , vast group in june 1998 .
a bank was created that was already too big for small switzerland when the merger was agreed - and a bank that continued to grow .
the fact that the two banks that went to form ubs complemented each other helped . one was a wealthy but sedate bank specialising in asset management in zurich and the other was less wealthy , but much more internationalised .
the internationalised investment bankers suddenly had access to a lot of money that they could use to turn an almighty wheel .
two names are connected with the rise of ubs : peter wuffli , group ceo of ubs from 2001 to 2007 , and marcel ospel , chairman of the board of directors .
&quot; i always thought that wuffli and ospel were a dream team , &quot; says the banking research , hans geiger .
&quot; wuffli , highly intellectual , modest , with a deep understanding of models and statistics , and ospel , with gut instinct who had worked his way up from being a banker &apos;s apprentice . &quot;
but ospel in the end relied on wuffli .
and wuffli ultimately may have made a critical error . &quot; he thought his models were reality , &quot; says geiger .
&quot; that is tragic . &quot;
tragic it may be .
pride comes before a fall .
employees of ubs clearly remember the management meetings where ceo wuffli always repeated the same slogan - dethrone goldman sachs .
the american bank was the world &apos;s most powerful investment bank .
it led the &quot; league tables &quot; that documented the success of the banks at that time .
ubs wanted to get to the very top of the league .
it managed some of the way .
ubs was a major player in foreign exchange and stocks .
commissioned management consultants , however , pointed out a weakness . in terms of what in banking terminology is referred to as &quot; fixed income , &quot; ubs was far behind the very largest banks .
ubs had gaps in its us mortgage products .
to plug the hole , ubs entered on an adventure that almost cost it its existence .
hungry for growth , the bank itself and a hedge fund founded by the bank invested in us subprime securities .
it developed into a vast machine that packaged up securities based on us mortgage credit and passed them on - while itself continuing to bear risks .
even when other banks had long been looking for a way out , ubs carried on .
&quot; ubs was buying as late as the summer of 2007 when the us housing market was in flames , &quot; says banking expert geiger .
the greater the rise , the greater the fall .
one manager after the next had to go as the attempts to clean up proved more difficult than anticipated .
posts at ubs became a game of musical chairs and remain uncertain to this day .
any head of the major bank now has an almost impossible job .
he must find a strategy for a bank that has turned its back on all existing business models .
banking expert geiger already has an idea - quit investment banking , focus on asset management and turn back to swiss roots .
the bank , however , is not ( yet ) prepared to go that far .
it is clear , nonetheless , that investment banking must shrink significantly .
the new head of the bank , sergio ermotti , plans to present his future strategy to the world on tuesday .
the 51-year-old italo-swiss banker was actually only imagined as a transitional solution when the bank parted ways unexpectedly with his predecessor grübel .
since then , he has proved his mettle and is now to head ubs permanently , the zurich-based bank announced on tuesday .
ermotti plans to take the investment bank back to where it was in the mid-90s .
he is cutting 3500 jobs .
but that will not be enough .
ubs needs a vision .
what is ubs ?
how does it plan to make money ?
those are the questions axel weber is definitely already asking himself .
publicly , he has only so far that he and his wife will look for a flat in zurich .
next year .
fear of recession in europe depresses us stock exchange
the impending recession in europe has depressed the mood of us investors at the start of the week .
the change of government in athens and rome had little impact .
curbed industrial production on the continent nipped hopes that the debt crisis might stabilise in the bud .
the general relief following the change of government in rome and athens was not sustained .
the euro fell against the dollar .
the italian and greek glass was still considered half full last week , not half empty - now the problems are coming to the fore once again , said mark luschini of janney montgomery scott .
financial stocks in particular were on the sales receipts .
the dow jones industrial average closed business at 12,078 with a fall of 0.6 percent .
the market barometer fluctuated between 12,027 and 12,170 points in the course of trading .
the broader s &amp; p 500 index closed at 1251 point , down one percent .
the nasdaq technology index lost 0.8 percent and closed trading at 2657 points .
in frankfurt , the dax fell 1.2 percent closing at 5985 points .
industry in the 17 euro countries has significantly reduced production and is adjusting to the end of the upturn .
the eurostat statistics office announced that , in september , companies produced two percent less than in the previous month .
the announcement made us investors sceptical . &quot; we are not an island . we are dependent , &quot; said steve goldman of goldman management .
he also claimed that europe is not likely to shake off a recession quickly and that there are further signs of weakness in the banks .
the investment legend , warren buffett , could only support such caution . in a cnbc interview , he said that it was not yet clear whether europe is strong enough to do all that is needed to end the crisis .
in his opinion , it is therefore too early to invest in european government bonds or banks .
instead , buffett threw a decades-old principle out of the window to invest in the us technology industry .
his investment of more than ten billion dollars in ibm boosted the veteran company &apos;s share price by almost one percent against the trend during the course of trading .
the shares closed almost unchanged at 187.35 dollars .
buffett claims to have paid 170 dollars on average per share and now has a 5.5 percent stake .
the share price of his investment company berkshire hathaway fell 1.3 percent .
bank of america share certificates fell significantly more strongly than the market by 2.6 percent .
the bank is converting almost all its remaining shares in the china construction bank into cash , bringing in 6.6 billion dollars .
the bank primarily plans to bolster its capital quota with the net profit from the sale to meet tougher regulations .
the aircraft manufacturer boeing gained , limiting losses on the dow .
the share price went up 1.5 percent after the corporation picked up the biggest order in its history .
the emirates airline ordered 50 long-range type 777 aeroplanes on sunday with an order value of 18 billion dollars .
including options to buy 20 more planes , the total volume is in fact 26 billion dollars .
around 710 million shares changed hands on the new york stock exchange .
719 shares rose , 2281 fell and 79 remained unchanged .
630 shares finished up , 1913 down and 82 unchanged on the nasdaq on turnover of 1.38 billion shares .
intelligent crows are spreading in cities
rooks are considered problem birds in many places . experts now say the highly intelligent animals can no longer be driven from cities .
some like the birds , others would like to get rid of them as fast as possible . crows have been part of the cityscape for hundreds of years , but residents complaining about the filth and noise the animals create are increasing in number in many communities .
a solution to the problem birds , however , is far from within reach .
experts at the first crow symposium in the east frisian town of leer have said that rooks can no longer be forced out of cities .
specialists from all over germany discussed possible solutions .
looking for alternatives in the cities
conservationists stated that migration of specially protected rooks was the root of the problem in cities , pointing out that the use of chemicals in agriculture was increasingly destroying traditional habitats .
hunting , increasing development and felling of typical breeding trees was making life hard for the birds .
crows are protected .
there are large colonies with more than 600 breeding pairs in the westphalian town of soest and in the north-west in jever , diepholz , achim bei bremen and leer .
a controversial project to cull crows in leer was behind negative headlines in 2005 .
critics described the capture of the birds in bird traps and how they were beaten to death with clubs as particularly brutal .
&quot; normal displacement and deterrence does nothing , &quot; said town and country planner werner klöver from leer .
spraying off nests with water or felling trees often has undesired consequences , said environment expert manfred kaiser from lahr in baden-württemberg .
&quot; the animals disperse and large colonies split up and spread to other areas . &quot;
deterrence would only be conceivable as an exception in hospitals , retirement homes and schools .
&quot; we have to live with them . getting rid of them is not an option . &quot;
without the surrounding area , the problem in the cities will not be solved , said leer &apos;s mayor , wolfgang kellner .
a combination of deterring birds from sensitive residential areas and offering incentives for them to move back to other areas would be an option .
&quot; information boards instead of chain saws &quot; was the slogan proposed by the psychologist and ethnologist uta maria jürgens , who argued for a totally different approach to the birds . &quot; we should have the self-confidence to live with this problem . &quot;
jürgens was a co-founder of the ascherberg crow educational trail in schleswig-holstein , where notice boards provide information about the life of the highly intelligent rook .
the project is already attracting interest from tourists .
jürgens also expressed her opinion on the criticism of the noise the birds make . &quot; the noise conveys a lot of information . they have a lot to say to each other . &quot;
the intelligence and skill of the birds are legendary . rooks often do not just use tools to get hold of a choice morsel , but create tools themselves .
behavioural researchers at the university of cambridge recently tested the extent of the birds &quot; intelligence on four rooks .
in one experiment , the birds quickly learnt that they had to knock over a platform to get to a moth larva .
they recognised the size and shape of stone needed to collapse the platform without much training .
in a second experiment , the canny crows had to solve a trickier problem .
the researchers put a small bucket with a moth larva in a vertical cylinder .
the birds could not reach their prey with their beaks .
however , the researchers had provided pieces of wire .
all four birds used the wire to form a hook and used the hook to fish the bucket out of the tube by its handle .
three of the four rooks even managed it at the first attempt .
ronaldo rockets portugal to euro 2012
the hopes of bosnians and turks for a last minute win were short-lived .
the favourites portugal , croatia and the czech republic , plus ireland , will be at the 2012 european championship .
portugal , the czech republic , croatia and ireland have got the final spots for the euro 2012 championship in poland and ukraine .
led by superstar cristiano ronaldo who scored in the 8th and 53rd minutes , portugal beat bosnia-herzegovina 6 : 2 ( 2 : 1 ) , rounding off their fifth euro appearance perfectly after 0 : 0 in the first leg .
the czechs only needed 1 : 0 ( 0 : 0 ) in the playoff second leg against outsider montenegro after a clear win in the first leg ( 2 : 0 ) .
croatia did not miss a trick with 0 : 0 against turkey and ireland managed an easy 1 : 1 ( 1 : 0 ) against estonia .
nani ( 24th minute ) , helder postiga ( 72nd and 82nd ) and miguel veloso ( 80th ) scored the other goals for portugal .
the former wolfsburg bundesliga professional zvjezdan misimovic and captain emir spahic evened the score with a converted 41st-minute penalty and a goal in the 65th for bosnia .
sena lulic got a second yellow card in the 54th minute .
bosnia had already been knocked out of the world cup 2010 playoffs in south africa against portugal .
portugal got off to a perfect start in lisbon &apos;s estadio da luz .
ronaldo pounded a free kick into the nets from about 30 meters to take an early lead .
a good quarter of an hour later , nani displayed his long-range abilities , raising the lead to 2 : 0 from the 25 metre line .
the bosnians , with no opening in the first half hour , had their first opportunity with goal scorer edin dzeko .
the former wolfsburg player from the english championship leader manchester city put a header onto the bottom edge of the crossbar in the 33rd minute , but the ball did not go over the goal line .
jiracek scores for the czech republic
the referee wolfgang stark from ergolding then garnered some attention .
following an attack on helder postiga in the bosnian penalty area in the 36th minute , stark decided the portuguese player had simulated and gave him a yellow card .
stark also awarded a controversial penalty for handball after a foul by fabio coentrao - misimovic converted .
after half-time , the vice-european champions of 2004 settled the score with four more goals .
petr jiracek cleared up any remaining doubt for the czechs in the 81st minute .
the european champions of 1976 are going to the euro championship for the fifth time in a row .
luck also smiled on trainer michal bilek &apos;s team and bundesliga legionaries michael kadlec ( bayer leverkusen ) and tomas pekhart ( 1st fc nuremberg ) with a potentially catastrophic error early on .
goalie petr cech prevented a potential own goal from tomas sivok in the 9th minute with a split-second reflex .
montenegro had its best opportunities when the stevan jovetic and dejan damjanovic duo were on the field .
damjanovic took a shot from the edge of the penalty area that just missed cech &apos;s goal in the 40th minute .
after half-time , the crossbar saved the czechs &quot; already battered world-class goalie after the forward from fc seoul retreated in the 50th minute .
the winning goal was scored just before full-time .
turkey , meanwhile , was lacking football genius in zagreb .
four days after losing 0 : 3 in istanbul , the turks could do no better than a 0 : 0 draw in zagreb .
while it is now virtually a certainty that national trainer guus hiddink must now give up his place on the bench of the team that reached third place in 2002 , the croatians are through to the final round of a european championship for the fourth time .
the turks almost managed the perfect start in front of 31,000 spectators at the maksimir stadium .
selcuk inan &apos;s long-range shot hit the post in the seventh minute and striker kazim kazim was unable to put the rebound into the net .
croatia then woke up , taking more and more control of the game .
the croatians continued in fine form after half-time .
the budesliga legionaries danijel pranjic ( bayern munich ) and gordon schildenfeld ( eintracht frankfurt ) were on the field alongside mario mandzukic from vfl wolfsburg , who made space for ivan perisic from dortmund from the 62nd minute .
trapattoni going to euro 2012
the irish in dublin were celebrating their first appearance at the european championship since 1988 in germany .
trainer giovanni trapattoni &apos;s team - already looking assured of a place following the 4 : 0 victory in the first leg - took the lead in the 32nd minute thanks to a goal from stephen ward .
konstantin vassiljev equalised for the away team in the 57th minute .
ireland kept up the pace , but had to wait until the 23rd minute when robbie keane from los angeles had the first clean shot on goal .
the estonian rearguard prevented him at the last moment from taking ireland into the lead .
the goal came in the 32nd minute .
after a failed defence by goalie pavel londak , ward was on the spot and pushed the ball over the line to bring the score to 1 : 0 .
woman critically injured after falling over in tram
a 52-year-old woman was critically injured in berlin-lichtenberg on monday after falling over in a tram .
the police announced in a statement on tuesday that she had undergone an emergency operation in hospital .
the tram was forced to break abruptly on herzbergallee when a vehicle stopped suddenly in front of it .
the car then turned off and drove away - the driver is being traced .
the other passengers on the tram were unhurt .
armed man holds up chemist &apos;s in friedenau
an armed man held up a chemist &apos;s on bundesallee in berlin-friedenau on monday evening .
the police announced in a statement that he had threatened a 25-year-old employee with a firearm and demanded money be handed over .
after the cash draw was handed over , the robber fled unrecognised and with an unknown amount of money .
the female employee suffered from shock .
drunk driver seriously injured
a 21-year-old drunk driver was seriously injured early on tuesday morning on märkische allee in berlin-marzahn .
he attempted to swerve out of the way of a van with trailer that shot out onto the road from a petrol station .
his vehicle skidded , crossed the centre strip into oncoming traffic and crashed into several trees .
the 21-year-old was taken to hospital .
the 40-year-old van driver was unhurt .
cyclist knocked down and seriously injured
a 61-year-old female cyclist was knocked down and seriously injured by a car in berlin-mitte on monday .
the police announced in a statement on tuesday that she had attempted to cross the road on her bike at the pedestrian crossing on the karl-liebknecht-strasse / alexanderstrasse crossing .
a 72-year-old driver drove into her as he was turning off .
the woman was knocked down and taken to hospital for inpatient care in a critical state .
on the run : car thief turns ghost driver
a man was driving a stolen car in charlottenburg at night .
he rammed a parked vehicle and pushed it up against two other cars and three bikes when turning around .
a police pursuit began .
the fugitive sped onto the motorway -
against oncoming traffic !
he scraped against a concrete crash barrier and lost parts of the car .
the car came to a halt and the man fled on foot .
pedestrian shot with an air rifle from a balcony
a female pedestrian in schwedt has been shot with an air rifle from a balcony .
the police announced in a statement that the 32-year-old was out and about on sunday in the vicinity of her flat when she was shot in the back .
she was taken to hospital to treat a laceration .
the police identified four suspects - a 19-year-old is the primary suspect .
a police spokesman said there is no connection between the woman and the suspects .
25-year-old seriously injured in a car accident
a male driver has been seriously injured in the potsdam-mittelmark district in a head-on collision with a lorry .
the police announced in a statement that the 25-year-old mini van driver came off the road on monday afternoon on an a road near wiesenburg and skidded into a crash barrier . the cause of the accident is not yet known .
the car was hurled onto the opposite side of the road by the violence of the impact and collided with a lorry .
the 25-year-old had to be cut out of his vehicle by the fire service .
he was taken to hospital in a rescue helicopter .
fire in cinema on alexanderplatz
a fire broke out in kino cubix on alexanderplatz on sunday night .
the police announced in a statement that around 70 cinema-goers at a late night viewing had to leave the cinema on rathausstrasse shortly after midnight .
no one was hurt .
the fire was probably caused by a technical failure in the electricity system according to the investigations so far .
the power failed at the same time and powerful fumes developed .
woman seriously injured by attempted handbag robbery
a woman has been seriously injured by an attempted handbag robbery in niedergörsdorf in the teltow-fläming district .
the police announced in a statement that the hooded perpetrator surprised the woman on friday at a parking space in front of a shop and tried to grab her handbag .
when the victim defended herself , he struck her hands and head with a truncheon .
he then fled empty-handed .
the woman was admitted to hospital with serious injuries .
grave-tenders robbed
elderly people have repeatedly been robbed in recent weeks at a cemetery in fürstenwalde .
the police announced in a statement that bags , money , mobile phones , bank cards and bunches of keys have been stolen in a total of twelve thefts since mid-october .
the victims were mainly elderly women tending the plots of relatives .
they had left their bicycles or zimmer frames with valuables nearby .
the most recent theft of several hundred euros in cash occurred last friday .
six graffiti sprayers arrested
the police have caught six graffiti sprayers .
the police announced in a statement that the six youths were arrested on sunday in marzahn .
the men , ranging in age from 15 to 25 , had sprayed a railway carriage at hohenschönhausener-grenzgraben .
they had used a ladder to climb onto the carriage and had smashed a window .
a witness notified the police .
the offenders initially fled when officers arrived .
however , they were arrested shortly after at a stop on landsberger allee .
wheelchair user hit by a car and seriously injured
a male wheelchair user has been seriously injured crossing a road in schwedt .
the police announced in a statement that the man attempted to cross a street at the pedestrian crossing on saturday and was not spotted by a female driver .
the vehicle hit the wheelchair and the man fell onto the road and suffered head injuries .
he had to be taken to hospital for inpatient treatment .
seven-year-old hit and injured by a car in wedding
a seven-year-old boy was knocked down and seriously injured by a car on tegeler strasse in berlin-wedding on sunday evening .
the police announced in a statement that he was taken to hospital with injuries to the face and leg .
the child ran out onto the road in front of a car driven by a 53-year-old woman .
the driver was not able to brake in time and the car hit the boy .
the boy was not accompanied by adults .
cyclist seriously injured in kreuzberg
a 46-year-old male cyclist was knocked down and seriously injured by a car in berlin-kreuzberg late on sunday evening .
the police announced in a statement on monday that he had been taken to hospital with head injuries and could not be questioned .
the cause of the collision with the vehicle driven by a 22-year-old man is not yet clear .
burnt out car in köpenick
a car was burnt out early on monday morning on dregerhoffstrasse in berlin-köpenick .
a police spokeswoman said a 34-year-old man attempted to put the fire out .
he was unsuccessful and the fire brigade had to bring the fire under control .
the vehicle was an older model .
arson is suspected , but without a political motive .
hit and run after accident in lichtenberg
a driver fled after a serious accident in lichtenberg on saturday night without helping his wounded passenger .
the police announced in a statement that the man had skidded on möllendorfstrasse , crashed into a tramway pylon and had landed on the trackbed .
a female tram driver was able to brake in time and saw the driver flee .
the fire brigade was informed and had to free a seriously injured 23-year-old from the car wreck and take him to hospital .
the owner of the vehicle later stated that her car had been stolen some time earlier .
fire in a cellar in tempelhof
there was a fire in a part of a cellar on monday night in a house in tempelhof .
the police announced in a statement that tenants of the building on marienfelder allee noticed smoke on the staircase and contacted the fire brigade .
an 18-year-old who had shown conspicuous interest in the fire-fighting operation was initially arrested .
however , he was released as suspicion was not confirmed .
injured people at punch-up in fast food restaurant
two people at a fast food restaurant in pankow were injured in a punch-up on saturday night .
the police announced in a statement that four men aged between 18 and 21 began arguing with two restaurant workers on prenzlauer promenade at around 01 : 20 .
two 19-year-olds attempted to help , but were immediately punched and kicked by the four men .
both suffered head injuries .
the four assailants were arrested in the restaurant .
courageous bystander attacked at underground station - offender arrested
a courageous bystander leapt to help girls being bothered at gesundbrunnen underground station and was beaten up .
three hooligans attacked the 41-year-old man on saturday and sprayed pepper spray into his face after punching him .
the three then fled the scene .
the police announced in a statement that one 19-year-old has been arrested subsequently on suspicion .
a female passer-by called the police from her mobile .
officers found the pepper spray in a bin near the underground station .
the courageous bystander was taken to hospital for outpatient treatment .
construction pit in flames in the embassy quarter
a fire broke out in a construction pit in the embassy quarter of tiergarten on saturday .
the police announced in a statement that construction materials set on fire at around 12 : 00 during welding work on tiergartenstrasse .
the fire caused a large cloud of fumes that could be seen clearly from some distance .
the fire brigade arrived with four crews and was able to bring the fire under control quickly .
a district heating pipe and a power supply line were also affected by the fire .
the structural stability of the construction pit would now also have to be checked , said a fire brigade spokesman .
no one was hurt during the fire .
tiergartenstrasse was closed to car and bus traffic during the fire-fighting operation .
40-year-old seriously injured during robbery
unknown assailants robbed and seriously injured a 40-year-old man on friday evening in berlin-mitte .
the police announced in a statement on saturday that the man was out and about on martin-opitz-strasse at around 23 : 30 when he was approached by the three offenders and asked for the time .
the men then followed him , punched and kicked their unsuspecting victim and fled with his wallet and mobile .
police officers found the seriously injured man lying on the street and called casualty .
the man was taken to hospital .
father attacks tram driver - child looks on
a drunken father verbally abused and then hit a female tram driver on friday evening in front of his four-year-old son .
the police announced in a statement on saturday that the 49-year-old berlin transport authority ( bvg ) worker had to stop on seestrasse in wedding when she was suddenly attacked by the 30-year-old man .
the man punched her in the face while his son looked on .
police officers quickly arrested the drunken man . the boy had to be picked up by his grandfather .
the tram driver was slightly injured .
rubbish bins set alight in wedding
unknown persons set rubbish bins on fire in a back yard in wedding on friday night .
the police announced in a statement that a resident noticed the fire at around midnight and also saw two darkly dressed men fleeing across the adjacent cemetery wall on steegerstrasse .
the fire brigade was notified and quickly put out the fire .
hungry labrador turns on hotplate - apartment fire
a hungry dog is believed to have started a fire in an apartment in brandenburg / havel .
the police announced in a statement on friday that the labrador is thought to have leapt onto the cooker to get to its food .
the dog then must have pushed a switch and turned on the hotplate .
paper on the cooker caught fire .
the fourth floor flat was set alight on thursday afternoon .
the dog died of smoke intoxication .
the 18-year-old tenant who was not in the flat at the time of the fire is being investigated for negligent arson .
star-studded twilight world premiere in hollywood
a flurry of flashing cameras , screaming fans and beaming twilight stars . thousands of curious onlookers gathered on the black-coloured carpet in front of the nokia theatre in los angeles at the world premiere of breaking dawn part 1 .
the procession of stars lasted several hours .
the fans of the vampire saga were not disappointed . vampire actor robert pattinson , his werewolf rival taylor lautner and &quot; bella &quot; beauty kristen stewart signed autographs and posed with fans .
bella finally gets to marry her true love in part four .
sap plans to double workforce in china
europe &apos;s largest software producer , sap , plans to invest billions and almost double its workforce in china in the next few years .
the dax-listed company announced on tuesday that around two billion dollars ( around 1.5 billion euros ) is to be invested by 2015 .
&quot; we will create more research and development facilities and hire thousands of people , &quot; bill mcdermott , co-chief executive of sap , said in a statement at a consumer show in peking .
the 2500-strong workforce is to grow to 4500 .
the world market leader in enterprise software plans to double its number of offices from five to ten or eleven .
sap is currently located in beijing , shanghai , guanzhou , chengdu and dalian .
around 4000 sap customers are in the boom country compared to the company &apos;s more than 176,000 customers worldwide .
the plans for growth in china are also an important pillar of the dax-listed company &apos;s 2015 strategy .
sap plans to increase turnover to 20 billion euros by 2015 .
the company from walldorf made around 12.5 billion euros last year .
business for the software provider is looking better than ever , with 55,000 employees worldwide .
sap has been in china for 20 years .
&quot; we now want to scale our operations , &quot; mcdermott said .
&quot; our aim is to drive sustainable growth through informisation in china . &quot;
mass slaughter on a personal level
peter englund &apos;s intense and bighearted new book , &quot; the beauty and the sorrow : an intimate history of the first world war , &quot; begins with a long dramatis personae of the sort that can make your heart sink .
here is advice for proceeding : gently excise this page and make it your bookmark .
you will be getting to know these people very well in mr. englund &apos;s novelistic telling , and this dramatis personae will function as your gps , a beacon during those few moments when , like one of his men and women , you are confused and bereft in the fog of war .
mr. englund is a swedish historian and journalist .
he is also the new permanent secretary of the swedish academy , which awards the nobel prize in literature .
what he has written here is an unusual book , one he describes , not inaccurately , as &quot; a work of anti-history . &quot;
it contains few big names , major treaties or famous battles ; there are almost no ambassadors , dashing journalists or discussions of tactics and materiel .
it &apos;s not so much a book about what happened , he explains , as &quot; a book about what it was like . &quot; it &apos;s about &quot; feelings , impressions , experiences and moods . &quot;
&quot; the beauty and the sorrow &quot; threads together the wartime experiences of 20 more or less unremarkable men and women , on both sides of the war , from schoolgirls and botanists to mountain climbers , doctors , ambulance drivers and clerks .
a few of these people will become heroes .
a few will become prisoners of war , or lose limbs , go mad or die .
their lives flicker here like votive candles lighted in a church ; new ones are added to the mix while some wink out .
mr. englund &apos;s book is a deviation from standard history books .
it is a corrective too to the notion that world war i was only about the dire trench warfare on the western front .
&quot; the beauty and the sorrow &quot; expertly pans across other theaters of war : the alps , the balkans , the eastern front , mesopotamia , east africa .
soldiers in this book have beehives fall on them ; one has christmas in egypt under the pyramids ; tsetse flies are an intractable problem .
this is a moving book , almost from the start .
war floods these people &apos;s lives like a natural catastrophe , a hurricane katrina that reeks of cordite .
when cannon fire is heard in the distance , and you are a woman at home with your children , do you stay or flee ?
who is coming , anyway ?
almost no one understands what &apos;s happening , even why this war is to be fought .
&quot; lack of facts , &quot; mr. englund observes , &quot; has been padded out with guesses , suppositions , hopes , fears , idées fixes , conspiracy theories , dreams , nightmares and rumors . &quot;
&quot; the beauty and the sorrow &quot; follows individuals like florence farmborough , an english nurse in the russian army , and richard stumpf , a young german high-seas fleet seaman .
their stories are mostly taken from memoirs , letters and other already published material .
the accounts of their lives can be terrifying or stirring , but are most fully alive in mr. englund &apos;s accumulation of small moments , stray details .
many of these are about deprivation and making do .
in telling a german schoolgirl &apos;s story he says : &quot; ersatz , everywhere ersatz . &quot;
&quot; substitute coffee , fake aluminum , imitation rubber , paper bandages , wooden buttons . &quot;
he goes on about the ersatz food in germany .
he describes &quot; meat made of pressed rice boiled in mutton fat ( and finished off with a fake bone made of wood ) ; tobacco made of dried roots and dried potato peel ; shoes soled with wood . &quot;
there are , he notes , &quot; 837 registered meat substitutes permissible in the production of sausages , 511 registered coffee substitutes . &quot;
some stories are about honor and bravery .
one american recognizes his own impulse toward savagery and declares about war : &quot; you feel that , after all , this is what men were intended for rather than to sit in easy chairs with a cigarette and whiskey , the evening paper or the best seller , and to pretend that such a veneer means civilization and that there is no barbarian behind your starched and studded shirt front . &quot;
conversely a british soldier realizes that he will most likely die , and that no one will notice or care .
&quot; once one is resigned to the thought of sacrificing oneself , one would like to think that it might happen in front of an audience , &quot; he writes .
instead , it is horribly like &quot; a condemned man being strangled secretly . &quot;
other observations are about an old europe falling away , and about new types of terror .
&quot; the conflict has increasingly become an economic competition , &quot; mr. englund writes , &quot; a war between factories . &quot;
he catches the arrival of what he calls &quot; a new species in the bestiary of the young century : the articulate and ideologically convinced mass murderer in well-cut clothes who performs his butchery while sitting behind a desk . &quot;
people behave in unanticipated ways ; there is as much base behavior as heroism .
mr. englund discussed the soldiers who actively tried to catch a venereal disease from prostitutes as a way to evade service at the front .
&quot; the most grotesque expression of this can be seen in the trade of gonococcal pus , which soldiers buy and smear into their genitals in the hope of ending up in hospital , &quot; he writes .
&quot; those who are really desperate rub it into their eyes , which often results in lifelong blindness . &quot;
in this translation from the swedish by peter graves , mr. englund &apos;s prose is supple but unshowy , perfectly suited to his humane task .
in dozens of small scenes he catches the way the war has &quot; unleashed uniquely uncontrollable forces : extreme nationalism , social revolution , religious hatred . &quot;
people begin to ask why their leaders are making them fight .
the best books about world war i have often been oblique , like paul fussell &apos;s &quot; great war and modern memory , &quot; or novels , like erich maria remarque &apos;s &quot; all quiet on the western front , &quot; rather than comprehensive histories .
mr. englund &apos;s volume joins an unconventional pantheon .
his book has the most devastating ending i can remember in a piece of nonfiction .
i won &apos;t give it away .
but it &apos;s as if he has reached from his book , snatched that dramatis personae page from your fingers , and lighted it with a match .
after an earlier misstep , a minutely planned raid
hundreds of police officers were involved , some of them wearing riot helmets .
the overnight hours of monday into tuesday were chosen because zuccotti park would be at its emptiest .
the operation was kept secret from all but a few high-ranking officers , with others initially being told that they were embarking on an exercise .
police commissioner raymond w. kelly was at the center , his presence underscoring how the operation was fraught with challenges for the police department .
there could be no repeat of episodes in recent weeks , like the pepper-spraying of protesters , that violated department rules and created a firestorm of public sympathy for the squatters .
and so the police operation to clear zuccotti park of protesters unfolded after two weeks of planning and training .
officials had prepared by watching how occupations in other cities played out .
a major disaster drill was held on randalls island , with an eye toward zuccotti .
officials increased so-called disorder training - counterterrorism measures that involve moving large numbers of police officers quickly - to focus on lower manhattan .
the last training session was on monday night , on the manhattan side of the east river .
the orders to move into the park came down at the &quot; last minute , &quot; said someone familiar with the orders , which referred to the assignment only as &quot; an exercise . &quot;
&quot; the few cops that i know that were called into this thing , they were not told it was for going into zuccotti park , &quot; said the person , who spoke on the condition of anonymity .
&quot; the only people who were aware of them going into zuccotti park were at the very highest levels of the department . &quot;
one reason for the secrecy was a lesson learned by the city .
on oct . 14 , officials wanted to clear the park , but then backed off as hundreds of protesters streamed in ahead of time after hearing of the plans .
the operation on tuesday involved officers from various police units , including boroughwide task forces - scores of mobile officers who are usually used to flood high-crime neighborhoods .
mr. kelly said many people , almost like commuters , had been coming and going from the park during the day , making 1 a.m. a good time to move in .
&quot; it was appropriate to do it when the smallest number of people were in the park , &quot; he said .
emergency service unit trucks with klieg lights and loudspeakers gathered at pike slip and the franklin d. roosevelt drive , near the manhattan bridge , before moving out .
the lights and prerecorded messages booming from the loudspeakers seemed to cow many protesters .
as the community affairs officers moved into the park in their light-blue windbreakers , many protesters simply gathered their belongings and left .
no tents were touched until 1 : 45 a.m. , the police said , giving the protesters time to gather their belongings .
other teams of officers were seen gathering on the perimeter to move in if arrests were needed in the park .
reporters in the park were forced to leave .
paul j. browne , the police department &apos;s chief spokesman , said it was for their safety .
but many journalists said that they had been prevented from seeing the police take action in the park , and that they had been roughly handled by officers .
mr. browne said television camera trucks on church street , along the park &apos;s western border , were able to capture images .
as the police moved west through the dense tangle of protesters &quot; personal belongings , including luggage and plastic lawn and leaf bags stuffed with clothing , crews from the sanitation department followed , scooping up what was left behind .
some protesters who refused to move were dragged out , the images appearing on the internet soon after .
a core group of protesters took up positions close to the encampment &apos;s kitchen area , near the center of the park .
some made a barrier , and the police moved in to methodically arrest them .
about 10 people in the epicenter of the encampment locked themselves together by their arms .
and two people chained themselves to trees , mr. kelly said .
emergency service officers were called in to cut the locks .
no arrests were made in the park until about 3 : 30 a.m. , mr. kelly said .
the clearing operation was complete about 75 minutes later , the police said .
mr. browne said about 142 people were arrested in the park .
most of the arrests were for &quot; disorderly conduct and resisting arrest , &quot; he said .
the highest-ranking officer on the scene was joseph j. esposito , the chief of department , the top uniformed officer in the department .
phil t. pulaski , the department &apos;s chief of detectives , was also there .
mr. kelly , while present , was &quot; not directing operations , &quot; mr. browne said .
a second group of officers waited along the eastern side of broadway , between liberty and cedar streets , in case they had to move in .
barricades were set up at cortlandt street , a block north of the park , and at pine street , a block south .
from those positions , the police were seen expanding the perimeter even farther outward from zuccotti park .
at one point , several officers , many with shields , were seen pushing people farther out .
about 28 people were on the northern perimeter .
some of the rowdiest action of the night took place south of the park .
around 5 a.m. , south of pine street , one protester jumped on the hood of a police car , and others were seen releasing the air from the tires of a police van .
at one point , a piece of plywood came flying from the crowd .
in the end , one officer and one protester were hospitalized .
support workers &apos; union will sue city over layoffs
the union representing hundreds of school support workers who lost their jobs last month plans to sue the city on wednesday , claiming the layoffs were unnecessary and discriminatory because of their disproportionate impact on schools that serve poor students .
the lawsuit is possibly the last weapon available to the union , district council 37 , in its effort to reverse the layoffs .
six hundred seventy-two school aides , parent coordinators and other employees who were among the city &apos;s lowest paid lost their jobs on oct . 7 , in a move union leaders have described as political payback for their refusal to let the city access a health care fund run by labor groups to close its budget gap last spring .
city officials have categorically denied that accusation .
at a city council hearing last month , schools chancellor dennis m. walcott said the layoffs were a part of the budget they signed off on in june .
one council member fired back , telling him that nothing in the budget specified that school support staff members would be laid off .
district council 37 officials said they filed a notice of claim , warning the department of education that it would be sued ; it will hold a news conference on wednesday to announce the lawsuit &apos;s filing formally , in state supreme court in manhattan .
the suit will list mr. walcott and the city &apos;s department of education as defendants , and counts eight laid-off workers as plaintiffs , all of them black or latino women , as were most of the aides who lost their jobs .
elizabeth thomas , a spokeswoman for the city &apos;s law department , said the city had not seen the lawsuit and therefore could not comment on its specific allegations .
according to a draft of the complaint obtained by the new york times , the union has zeroed in on the number of layoffs at schools in poor parts of the city to support a claim of inequitable treatment , saying poor schools lost more aides because they had been chronically underfinanced .
it points out that while there were no layoffs on staten island and few layoffs in schools in the upper east side or tribeca , 17 school aides lost their jobs in district 23 in brownsville , brooklyn , and 46 were let go in districts 8 and 9 in the south bronx .
noting the loss of five out of eight school aides employed by public school 36 in harlem , where 68 percent of the students live in poverty , the lawsuit says , &quot; this is not a neighborhood of well-heeled parents who can fund-raise to make up the budget shortfalls . &quot;
p.s. 36 &quot; s budget was cut by 3.26 percent this year , and it also lost money when the city changed the way it calculates the extra financing for each poor student schools enroll .
the union also claims that high school principals may have been misled by the language in a memo authorizing the dismissal of parent coordinators .
the memo used the word &quot; excess &quot; to describe the action , even though &quot; excess &quot; more commonly describes the act of moving a teacher from a school &apos;s payroll into a pool of available workers whose salaries are paid by the central office .
&quot; sixty-six parent coordinators were laid off , &quot; the draft complaint says , &quot; and not merely excessed . &quot;
mission over , congress ready to agree on libya
it finally looks like republicans and democrats on capitol hill are ready to agree on something related to the uprising in libya .
all it took was for the revolt to succeed , longtime dictator moammar gaddafi to be captured and killed and american involvement to end .
on tuesday , the senate foreign relations committee is to consider a resolution , sponsored by chairman john kerry ( d-mass . ) , along with sens. john mccain ( r-ariz . ) and joseph i. lieberman ( i-conn . ) , applauding libyans for their successful rebellion and u.s. troops for their &quot; bravery . &quot;
&quot; this resolution is about honoring the courageous people of libya as they begin to rebuild their country , &quot; kerry said .
&quot; it &apos;s an affirmation of bipartisan support for their democratic aspirations . &quot;
the language may sound innocuous , but nothing about the libya debate has been easy so far .
this summer , lawmakers from both parties squabbled with each other and the white house over the wording of a series of resolutions to approve or disapprove of the u.s. military &apos;s role in the nato mission in libya .
in june , kerry and mccain introduced a resolution authorizing the limited use of u.s. military forces in libya .
the bill was defeated in the house by a wide bipartisan margin , and it never came up for a vote in the senate .
but on the same day that bill failed in the house , the chamber also voted down a measure that would have cut funding for american operations in libya .
that left the u.s. mission in a form of legislative purgatory - with congress upset enough to criticize obama &apos;s handling of the mission but not so angry as to actually pull the rug out from under it .
then congress

Download .txt

gitextract_7rhl_yau/

├── .gitattributes
├── .gitignore
├── README.md
├── README_XING.md
├── executable/
│   ├── ZOPH_RNN_1
│   ├── ZOPH_RNN_2
│   ├── ZOPH_RNN_3
│   ├── ZOPH_RNN_4
│   └── ZOPH_RNN_XING
├── sample_data/
│   ├── dev_english.txt.tok.lc
│   ├── dev_french.txt.tok.lc
│   ├── ptb.train.txt
│   ├── ptb.valid.txt
│   ├── test_english.tok.lc
│   ├── test_french.tok.lc
│   ├── train_english.txt.tok.lc.10k
│   └── train_french.txt.tok.lc.10k
├── scripts/
│   ├── att_unk_rep.py
│   ├── berk_aligner/
│   │   ├── berkeleyaligner.jar
│   │   ├── run_aligner.sh
│   │   └── unk_replace.conf
│   ├── bleu_format.py
│   ├── bleu_format_valid.py
│   ├── compile.sh
│   ├── compile.xing.sh
│   ├── create_vocab_mapping_file.py
│   ├── create_vocab_mapping_file_preinit.py
│   ├── fsa/
│   │   ├── convert.py
│   │   ├── demo.sh
│   │   ├── enc1.txt
│   │   ├── enc2.txt
│   │   ├── fsa.txt
│   │   ├── generate_fsa.py
│   │   ├── source.singleline.txt
│   │   ├── source.test.txt
│   │   ├── source.train.txt
│   │   ├── source.valid.txt
│   │   ├── target.test.txt
│   │   ├── target.train.txt
│   │   └── target.valid.txt
│   ├── generate_train_decode.py
│   ├── load_lstm.py
│   ├── load_model.py
│   ├── multi-bleu.perl
│   ├── pretrain.py
│   ├── read_hpc_output.py
│   ├── translate/
│   │   ├── f2e_decode.sh
│   │   └── f2e_train.sh
│   └── unk_format.py
└── src/
    ├── BZ_CUDA_UTIL.h
    ├── Eigen_Util.h
    ├── Full_Input_To_Hidden_Layer.h
    ├── Hidden_To_Hidden_Layer.h
    ├── Hidden_To_Hidden_Layer.hpp
    ├── Input_To_Hidden_Layer.h
    ├── Input_To_Hidden_Layer.hpp
    ├── LSH_WTA.h
    ├── LSTM.h
    ├── LSTM.hpp
    ├── LSTM_HH.h
    ├── LSTM_HH.hpp
    ├── NCE.h
    ├── NCE.hpp
    ├── NCE_node.h
    ├── add_model_info.h
    ├── attention_combiner.h
    ├── attention_combiner.hpp
    ├── attention_combiner_node.h
    ├── attention_combiner_node.hpp
    ├── attention_layer.h
    ├── attention_layer.hpp
    ├── attention_node.h
    ├── attention_node.hpp
    ├── base_layer.h
    ├── base_layer.hpp
    ├── base_loss.h
    ├── bi_encoder.h
    ├── bi_encoder.hpp
    ├── charCNN_node.h
    ├── char_file_helper.h
    ├── conv_char.h
    ├── conv_char.hpp
    ├── custom_kernels.h
    ├── decoder.h
    ├── decoder_model_wrapper.h
    ├── decoder_model_wrapper.hpp
    ├── encoder_multi_source.h
    ├── encoder_multi_source.hpp
    ├── ensemble_factory.h
    ├── ensemble_factory.hpp
    ├── fileHelper.h
    ├── fileHelper_source.h
    ├── file_helper_char_decoder.h
    ├── file_helper_decoder.h
    ├── format.cc
    ├── format.h
    ├── fsa.h
    ├── fsa.hpp
    ├── global_params.h
    ├── gpu_info_struct.h
    ├── highway_network.h
    ├── highway_network.hpp
    ├── highway_node.h
    ├── input_file_prep.h
    ├── input_file_prep.hpp
    ├── logger.h
    ├── main.cu
    ├── memory_util.h
    ├── model.h
    ├── model.hpp
    ├── multinomial.h
    ├── prev_states.h
    ├── softmax.h
    ├── softmax.hpp
    ├── softmax_node.h
    ├── transfer_layer.h
    ├── tree_LSTM.h
    ├── tree_LSTM.hpp
    └── trunc_softmax.h

Download .txt

SYMBOL INDEX (528 symbols across 47 files)

FILE: scripts/fsa/generate_fsa.py
  function generate (line 3) | def generate(word,start,end,i,f):
  function generate_fsa (line 17) | def generate_fsa():
  function generate_pair (line 28) | def generate_pair():
  function generate_fn_pairs (line 40) | def generate_fn_pairs(n,fn1,fn2):
  function generate_train_valid (line 50) | def generate_train_valid():

FILE: scripts/generate_train_decode.py
  function main (line 47) | def main():

FILE: scripts/load_lstm.py
  function parse (line 6) | def parse(fn,key = "all", side = 0, the_layer = 0):
  function split_into_sentece (line 94) | def split_into_sentece(res):
  function convert_to_hdf (line 113) | def convert_to_hdf(lstm_path,en_path,hdf_path):

FILE: scripts/load_model.py
  function read_matrix (line 8) | def read_matrix(row,column,f):
  function write_matrix (line 22) | def write_matrix(m,f):
  function same_matrix (line 34) | def same_matrix(m1,m2):
  function random_matrix (line 43) | def random_matrix(row,column,upper = 0.08):
  class IH (line 47) | class IH:
    method __init__ (line 48) | def __init__(self,LSTM_size, vocab_size):
    method parse (line 53) | def parse(self,f):
    method dump (line 69) | def dump(self,f):
    method random_weight (line 86) | def random_weight(self):
    method is_same (line 102) | def is_same(self,other):
  class HH (line 116) | class HH:
    method __init__ (line 117) | def __init__(self,LSTM_size):
    method parse (line 121) | def parse(self,f):
    method random_weight (line 136) | def random_weight(self):
    method dump (line 151) | def dump(self,f):
    method is_same (line 166) | def is_same(self,other):
  class Softmax (line 178) | class Softmax:
    method __init__ (line 179) | def __init__(self,LSTM_size, vocab_size):
    method parse (line 184) | def parse(self,f):
    method random_weight (line 188) | def random_weight(self):
    method dump (line 192) | def dump(self,f):
    method is_same (line 196) | def is_same(self,other):
  class Model (line 208) | class Model:
    method __init__ (line 209) | def __init__(self):
    method diff (line 221) | def diff(self,other):
    method random_model (line 232) | def random_model(self,num_layers,LSTM_size, source_size, target_size):
    method load_vocab (line 260) | def load_vocab(self,side,f):
    method write_vocab (line 277) | def write_vocab(self,side,f):
    method parse (line 289) | def parse(self,fn):
    method dump (line 336) | def dump(self,fn):

FILE: scripts/pretrain.py
  function prepfile (line 27) | def prepfile(fh, code):
  class py34FileType (line 41) | class py34FileType(object):
    method __init__ (line 58) | def __init__(self, mode='r', bufsize=-1, encoding=None, errors=None):
    method __call__ (line 64) | def __call__(self, string):
    method __repr__ (line 83) | def __repr__(self):
  function main (line 92) | def main():

FILE: scripts/read_hpc_output.py
  function parse_arg (line 7) | def parse_arg():
  function process_file (line 12) | def process_file(path):
  function main (line 30) | def main():

FILE: src/BZ_CUDA_UTIL.h
  function namespace (line 31) | namespace deniz {
  function namespace (line 52) | namespace BZ_STATS {
  function namespace (line 59) | namespace BZ_CUDA {
  function CUDA_ERROR_WRAPPER (line 193) | void CUDA_ERROR_WRAPPER(cudaError_t cudaStat,std::string error_message) {
  function CUBLAS_ERROR_WRAPPER (line 239) | void CUBLAS_ERROR_WRAPPER(cublasStatus_t cudaStat,std::string error_mess...
  function CUDA_GET_LAST_ERROR (line 252) | void CUDA_GET_LAST_ERROR() {
  function CUDA_GET_LAST_ERROR (line 262) | void CUDA_GET_LAST_ERROR(std::string msg) {
  function initialize_vector_vocab (line 401) | void initialize_vector_vocab(int *h_vector,int rows,int vocab_size) {
  function initialize_vector_vocab_01 (line 408) | void initialize_vector_vocab_01(int *h_vector,int rows) {
  function full_vector_setup_vocab (line 415) | void full_vector_setup_vocab(int **h_vector,int **d_vector,int rows,int ...
  function full_vector_setup_vocab_01 (line 424) | void full_vector_setup_vocab_01(int **h_vector,int **d_vector,int rows) {
  function cublasStatus_t (line 691) | inline cublasStatus_t cublas_gemm_wrapper(cublasHandle_t handle, cublasO...
  function cublasStatus_t (line 698) | inline cublasStatus_t cublas_gemm_wrapper(cublasHandle_t handle, cublasO...
  function cublasStatus_t (line 708) | inline cublasStatus_t cublas_geam_wrapper(cublasHandle_t handle, cublasO...
  function cublasStatus_t (line 716) | inline cublasStatus_t cublas_geam_wrapper(cublasHandle_t handle, cublasO...
  function cublasStatus_t (line 727) | inline cublasStatus_t cublas_gemv_wrapper(cublasHandle_t handle, cublasO...
  function cublasStatus_t (line 734) | inline cublasStatus_t cublas_gemv_wrapper(cublasHandle_t handle, cublasO...
  function cublasStatus_t (line 745) | inline cublasStatus_t cublas_dgmm_wrapper(cublasHandle_t handle, cublasS...
  function cublasStatus_t (line 751) | inline cublasStatus_t cublas_dgmm_wrapper(cublasHandle_t handle, cublasS...
  function __device__ (line 761) | __device__ double atomicAddDouble(double* address, double val)
  function curandGenerateUniform_wrapper (line 793) | void curandGenerateUniform_wrapper(float *d_mask,int size,curandGenerato...
  function curandGenerateUniform_wrapper (line 797) | void curandGenerateUniform_wrapper(double *d_mask,int size,curandGenerat...
  function __device__ (line 802) | __device__
  function __device__ (line 807) | __device__
  function __device__ (line 812) | __device__
  function __device__ (line 817) | __device__
  function __device__ (line 823) | __device__
  function __device__ (line 828) | __device__
  function __device__ (line 834) | __device__
  function __device__ (line 840) | __device__
  function __device__ (line 846) | __device__
  function __device__ (line 851) | __device__
  function __device__ (line 856) | __device__
  function __device__ (line 861) | __device__
  function __device__ (line 866) | __device__
  function __device__ (line 871) | __device__
  function __device__ (line 876) | __device__
  function __device__ (line 881) | __device__
  function __device__ (line 886) | __device__
  function __device__ (line 893) | __device__
  function __device__ (line 898) | __device__
  function namespace (line 960) | namespace gpu_info {
  function devSynchAll (line 977) | void devSynchAll() {
  function nan_check_kernel (line 992) | void nan_check_kernel(dType *d_ptr,int rows,int cols,bool *d_check) {
  function zero_check (line 1025) | void zero_check(dType *d_mat, int size) {
  function check_nonseg (line 1035) | void check_nonseg(dType *d_ptr,int rows,int cols) {
  function printIntroMessage (line 1040) | void printIntroMessage(global_params &params) {

FILE: src/Eigen_Util.h
  type sigmoid_functor (line 17) | struct sigmoid_functor {
  type tanh_functor (line 22) | struct tanh_functor {
  type tanh_sq_functor (line 27) | struct tanh_sq_functor {
  type exp_functor (line 33) | struct exp_functor {
  function get_file_stats (line 202) | void get_file_stats(int &num_lines,int &num_words,std::ifstream &input,i...
  function get_file_stats_source (line 243) | void get_file_stats_source(int &num_lines,std::ifstream &input) {
  function extract_char_info (line 257) | void extract_char_info(int &longest_word,int &num_unique_chars_source,in...
  function extract_charCNN (line 295) | void extract_charCNN(std::unordered_map<int,std::vector<int>> &word_to_c...
  function create_char_vocab (line 332) | void create_char_vocab(int *h_word_indicies,int num_words,int longest_wo...

FILE: src/Full_Input_To_Hidden_Layer.h
  type Full_Input_To_Hidden_Layer (line 6) | struct Full_Input_To_Hidden_Layer {

FILE: src/Hidden_To_Hidden_Layer.h
  type neuralMT_model (line 260) | struct neuralMT_model
  type neuralMT_model (line 265) | struct neuralMT_model

FILE: src/Hidden_To_Hidden_Layer.hpp
  type neuralMT_model<precision> (line 6) | struct neuralMT_model<precision>
  type neuralMT_model<precision> (line 127) | struct neuralMT_model<precision>

FILE: src/Input_To_Hidden_Layer.h
  type neuralMT_model (line 304) | struct neuralMT_model
  type neuralMT_model (line 309) | struct neuralMT_model

FILE: src/Input_To_Hidden_Layer.hpp
  type neuralMT_model<precision> (line 5) | struct neuralMT_model<precision>
  type neuralMT_model<precision> (line 173) | struct neuralMT_model<precision>

FILE: src/LSH_WTA.h
  function set_to_n (line 108) | void set_to_n(int *data, int size, int n){
  function get_permute (line 264) | void get_permute(int n, int k, int *data){
  function topm_cpu (line 278) | void topm_cpu(dType *d_outputdist, dType *d_h_t, int batch_size){
  function get_top_m_cpu (line 618) | void get_top_m_cpu(int *h_codes, int * h_top_ids, int m, int batch_index...
  function create_hash_cpu (line 658) | void create_hash_cpu(){

FILE: src/LSTM.h
  type Input_To_Hidden_Layer (line 81) | struct Input_To_Hidden_Layer

FILE: src/LSTM.hpp
  type Input_To_Hidden_Layer<dType> (line 4) | struct Input_To_Hidden_Layer<dType>
  type Input_To_Hidden_Layer<dType> (line 23) | struct Input_To_Hidden_Layer<dType>

FILE: src/LSTM_HH.h
  type Hidden_To_Hidden_Layer (line 77) | struct Hidden_To_Hidden_Layer

FILE: src/LSTM_HH.hpp
  type Hidden_To_Hidden_Layer<dType> (line 4) | struct Hidden_To_Hidden_Layer<dType>
  type Hidden_To_Hidden_Layer<dType> (line 21) | struct Hidden_To_Hidden_Layer<dType>

FILE: src/NCE.h
  type neuralMT_model (line 101) | struct neuralMT_model

FILE: src/NCE.hpp
  type neuralMT_model<precision> (line 3) | struct neuralMT_model<precision>
  function softmax_layer_gpu_info (line 727) | softmax_layer_gpu_info NCE_layer<dType>::gpu_init(int device_number) {
  function dType (line 738) | dType *NCE_layer<dType>::get_ht_ptr(int index) {
  function cudaEvent_t (line 748) | cudaEvent_t NCE_layer<dType>::get_ERR_ht_event() {
  function dType (line 753) | dType *NCE_layer<dType>::get_dist_ptr() {
  function __global__ (line 811) | __global__

FILE: src/add_model_info.h
  function add_model_info (line 7) | void add_model_info(int num_layers,int LSTM_size,int target_vocab_size,i...

FILE: src/base_loss.h
  type neuralMT_model (line 19) | struct neuralMT_model

FILE: src/bi_encoder.h
  type model_type_t (line 10) | enum model_type_t {SEND_REV,COMBINE}

FILE: src/char_file_helper.h
  function read_minibatch (line 35) | void read_minibatch() {
  function reset_file (line 85) | void reset_file() {

FILE: src/conv_char.hpp
  function __global__ (line 520) | __global__
  function __global__ (line 529) | __global__
  function __device__ (line 537) | __device__
  function __global__ (line 549) | __global__
  function __global__ (line 577) | __global__

FILE: src/custom_kernels.h
  function __global__ (line 9) | __global__
  function __global__ (line 22) | __global__
  function vocab_softmax (line 38) | void vocab_softmax(int *d_vocab_indicies,int *d_vocab_indicies_01,dType ...
  function __global__ (line 57) | __global__
  function __global__ (line 67) | __global__
  function forward_sigmoid_kernel_small (line 80) | void forward_sigmoid_kernel_small(dType *d_final,dType *temp1,dType *d_b...
  function forward_sigmoid_kernel_feed (line 92) | void forward_sigmoid_kernel_feed(dType *d_final,dType *temp1,dType *temp...
  function __global__ (line 105) | __global__
  function __global__ (line 115) | __global__
  function forward_tanh_kernel_feed (line 128) | void forward_tanh_kernel_feed(dType *d_final,dType *temp1,dType *temp2,d...
  function forward_c_t_kernel (line 141) | void forward_c_t_kernel(dType *d_c_t,dType *d_f_t, dType *d_c_t_prev,dTy...
  function forward_c_t_kernel_tree (line 151) | void forward_c_t_kernel_tree(dType *d_c_t,dType *d_f_t_1,dType *d_c_t_pr...
  function __global__ (line 161) | __global__
  function __global__ (line 170) | __global__
  function zero_c_t_and_h_t (line 184) | void zero_c_t_and_h_t(dType *d_h_t,dType *d_c_t,int *d_vocab_indices_01,...
  function __global__ (line 210) | __global__
  function __global__ (line 222) | __global__
  function __global__ (line 236) | __global__
  function __global__ (line 247) | __global__
  function d_ERRnTOt_ft_it_kernel (line 262) | void d_ERRnTOt_ft_it_kernel(dType *d_d_ERRnTOt,dType *d_d_ERRnTOt_ct,dTy...
  function d_ERRnTOt_tanhcpt_kernel (line 274) | void d_ERRnTOt_tanhcpt_kernel(dType *d_d_ERRnTOt_tanhcpt,dType *d_d_ERRn...
  function zero_columns_kernel (line 286) | void zero_columns_kernel(int hiddenstate_size, dType *d_mat,int *d_vec,d...
  function add_four_matrices_kernel (line 298) | void add_four_matrices_kernel(dType *d_final,dType *d_mat1,dType *d_mat2...
  function elementwise_mult_kernel (line 310) | void elementwise_mult_kernel(dType *d_mat1,dType *d_mat2,dType *d_final,...
  function elementwise_mult_kernel_add (line 320) | void elementwise_mult_kernel_add(dType *d_mat1,dType *d_mat2,dType *d_fi...
  function sparse_lookup_kernel (line 342) | void sparse_lookup_kernel(dType *d_lookup, dType *d_W,int *d_vocab_indic...
  function __global__ (line 353) | __global__
  function __global__ (line 366) | __global__
  function __global__ (line 381) | __global__
  function __global__ (line 394) | __global__
  function W_small_gradient_kernel (line 409) | void W_small_gradient_kernel(dType *d_small_W_grad,int *d_reverse_unique...
  function W_small_dropout_gradient_kernel (line 428) | void W_small_dropout_gradient_kernel(dType *d_small_W_grad,int *d_revers...
  function warpReduceSum (line 456) | void warpReduceSum(volatile dType* sdata, int tid) {
  function warpReduceMax (line 467) | void warpReduceMax(volatile dType* sdata, int tid) {
  function train_perplexity_kernel (line 479) | void train_perplexity_kernel(int *d_output_vocab_indices_single,int *d_o...
  function outputdist_overflow_prevention_kernel (line 503) | void outputdist_overflow_prevention_kernel(dType *output, dType *input, ...
  function outputdist_perplexity_kernel (line 583) | void outputdist_perplexity_kernel(dType2 *output, dType *input, int dim,...
  function outputdist_perplexity_kernel_NCE (line 665) | void outputdist_perplexity_kernel_NCE(dType2 *output, dType *input, int ...
  function nce_score_dot (line 748) | void nce_score_dot(double *d_outputdist_perp,dType *d_h_t,dType *d_D,dTy...
  function matrix_bias_kernel (line 787) | void matrix_bias_kernel(int hiddenstate_size, dType *d_mat,dType *d_vec,...
  type exp_functor_gpu (line 796) | struct exp_functor_gpu {
  type exp_functor_eigen (line 805) | struct exp_functor_eigen {
  type inv_functor_gpu (line 811) | struct inv_functor_gpu {
  function zero_columns_kernel_128 (line 820) | void zero_columns_kernel_128(int hiddenstate_size, dType *d_mat,int *d_v...
  function matrix_row_to_matrix_column_kernel (line 834) | void matrix_row_to_matrix_column_kernel(dType *d_mat_final,dType *d_mat_...
  function copy_matrix_float_to_int_kernel (line 845) | void copy_matrix_float_to_int_kernel(int *d_source,dType *d_destination,...
  function __global__ (line 857) | __global__
  function __global__ (line 865) | __global__
  function matrix_column_to_matrix_row_kernel_2 (line 877) | void matrix_column_to_matrix_row_kernel_2(dType *d_mat_final,dType *d_ma...
  function __global__ (line 892) | __global__
  function __global__ (line 900) | __global__
  type scale_functor (line 913) | struct scale_functor {
  type square (line 927) | struct square {
  function __device__ (line 933) | __device__
  function basic_compute_norm_p1 (line 956) | void basic_compute_norm_p1(dType *d_gradient,int size,dType *result) {
  function basic_compute_norm_p2 (line 990) | void basic_compute_norm_p2(dType *temp_result,dType *final_result) {
  function print_norm_function_softmax (line 1019) | void print_norm_function_softmax(dType *d_mat,int size,int index,dType *...
  function clip_individual (line 1139) | void clip_individual(dType *d_mat, int size,dType threshold) {
  function scale_W_gradient (line 1150) | void scale_W_gradient(dType *d_W_gradient,int *d_vocab_indicies_m1,int h...
  function indv_clip_W_gradient (line 1164) | void indv_clip_W_gradient(dType *d_W_gradient,int *d_vocab_indicies_m1,i...
  function norm_W_compute_p1 (line 1184) | void norm_W_compute_p1(dType *d_W_gradient,dType *global_tempsum,int *d_...
  function norm_W_compute_p2 (line 1230) | void norm_W_compute_p2(dType *global_tempsum) {
  function zero_W_gradient (line 1347) | void zero_W_gradient(dType *d_W_gradient,int *d_vocab_indicies_m1,int hi...
  function update_W_gradient (line 1359) | void update_W_gradient(dType *d_W, dType *d_W_gradient,int *d_vocab_indi...
  function add_grad_vecs (line 1371) | void add_grad_vecs(dType *vec1,dType *vec2,dType learning_rate,int size) {
  function ones_mat (line 1384) | void ones_mat(dType *mat,int size) {
  function trunc_D_grad_nonshort (line 1397) | void trunc_D_grad_nonshort(dType *d_subset_D_grad,dType *d_D,int *d_voca...
  function trunc_D_grad_short (line 1409) | void trunc_D_grad_short(dType *d_subset_D_grad,dType *d_subset_D,int hid...
  function trunc_set_D (line 1421) | void trunc_set_D(dType *d_D,dType *d_subset_D,int trunc_size,int output_...
  function load_shortlist_D (line 1446) | void load_shortlist_D(dType *d_subset_D,dType *d_D,int hiddenstate_size,...
  function outputdist_truncated_kernel (line 1459) | void outputdist_truncated_kernel(dType *output, dType *input, int dim,dT...
  function dropout_kernel (line 1556) | void dropout_kernel(dType *d_dropout_mask,dType rate,dType *d_final, int...
  function __global__ (line 1566) | __global__
  function __global__ (line 1573) | __global__
  function sigmoid_kernel (line 1583) | void sigmoid_kernel(dType *d_in,dType *d_out,int total_length) {
  function alignment_pos_kernel (line 1598) | void alignment_pos_kernel(dType *d_in,dType *d_out,int total_length,int ...
  function lower_upper_kernel (line 1608) | void lower_upper_kernel(dType *d_p_t,int *d_lower_upper,int D,int *d_bat...
  function get_viterbi_alignment_kernel (line 1623) | void get_viterbi_alignment_kernel(dType *d_alignments,int *d_indicies,in...
  function __global__ (line 1649) | __global__
  function load_in_hs_kernel (line 1694) | void load_in_hs_kernel(dType **d_total_hs_mat, int D,dType *d_hs_mat,int...
  function exp_mask_kernel (line 1729) | void exp_mask_kernel(int *d_indicies,dType *d_alignments,int minibatch_s...
  function alignment_reduction_kernel (line 1752) | void alignment_reduction_kernel(dType *d_alignments, int LSTM_size,int m...
  function create_c_t_kernel (line 1804) | void create_c_t_kernel(dType *d_alignments,dType *d_hs_mat,dType *d_c_t,...
  function add_two_mats_kernel (line 1818) | void add_two_mats_kernel(dType *d_mat1,dType *d_mat2,int size) {
  function add_two_mats_into_third_kernel (line 1826) | void add_two_mats_into_third_kernel(dType *d_mat1,dType *d_mat2,dType *d...
  function tanh_grad_kernel (line 1835) | void tanh_grad_kernel(dType *d_output,dType *d_input_Error,dType *d_tanh...
  function tanh_att_forward_kernel (line 1844) | void tanh_att_forward_kernel(dType *d_output,dType *d_in1,dType *d_in2,d...
  function elem_reduce_kernel (line 1855) | void elem_reduce_kernel(dType *d_h_t,dType *d_Wa_hs_temp, dType *d_align...
  function elem_reduce_kernel_large (line 1892) | void elem_reduce_kernel_large(dType *d_h_t,dType *d_Wa_hs_temp, dType *d...
  function error_alignments_kernel (line 1932) | void error_alignments_kernel(dType *d_ERRnTOt_ct,dType *d_hs_mat, dType ...
  function error_alignments_kernel_large (line 1969) | void error_alignments_kernel_large(dType *d_ERRnTOt_ct,dType *d_hs_mat, ...
  function error_pt_kernel (line 2010) | void error_pt_kernel(dType *d_ERRnTOt_pt,dType *d_ERRnTOt_as,int D,dType...
  function att_vp_error (line 2045) | void att_vp_error(dType *d_sigma,dType *d_tanh,dType *d_temp_grad,dType ...
  function grad_W_p_kernel (line 2056) | void grad_W_p_kernel(dType *d_v_p,dType *d_temp,dType *d_sigma,dType *d_...
  function get_ht_scalings_Wa_grad_kernel (line 2100) | void get_ht_scalings_Wa_grad_kernel(dType *d_scalings,dType *d_ERRnTOt_a...
  function scale_ht_kernel (line 2121) | void scale_ht_kernel(dType *d_scalings,dType *d_temp1,dType *d_h_t,int L...
  function scale_ht_kernel_large (line 2132) | void scale_ht_kernel_large(dType *d_hs_sum,dType *d_hs_mat,dType *d_scal...
  function copy_errors_source (line 2147) | void copy_errors_source(dType **d_total_hs_error,dType *d_temp_error,int...
  function error_hs_ct_kernel (line 2166) | void error_hs_ct_kernel(dType *d_ERRnTOt_ct, dType *d_alignments,int *d_...
  function error_hs_ct_kernel_large (line 2183) | void error_hs_ct_kernel_large(dType *d_ERRnTOt_ct, dType *d_alignments,i...
  function gradient_update_mats (line 2202) | void gradient_update_mats(dType *d_mat,dType *d_mat_grad,dType learning_...
  function zero_h_t (line 2212) | void zero_h_t(dType *d_h_t, int *d_01_mask,int LSTM_size,int minibatch_s...
  function clip_mat_kernel (line 2222) | void clip_mat_kernel(dType *d_mat,dType threshold,int size) {
  function load_in_embeddings (line 2242) | void load_in_embeddings(dType *d_temp_embeddings,dType *d_D,int *d_sampl...
  function nce_dot_product_SPARSE (line 2258) | void nce_dot_product_SPARSE(dType *d_dot_products,dType *d_D,dType *d_h_...
  function dType (line 2301) | inline dType log_add_exp(dType x,dType y) {
  function calc_p_true_kernel (line 2311) | void calc_p_true_kernel(dType *d_p_true,dType *d_dot_products,dType *d_s...
  function calc_p_true_kernel_nonshare (line 2339) | void calc_p_true_kernel_nonshare(dType *d_p_true,dType *d_dot_products,d...
  function objective_val_p1_NCE_kernel (line 2360) | void objective_val_p1_NCE_kernel(dType *d_p_true,double *d_OBJ_val_temp,...
  function objective_val_p2_NCE_kernel (line 2403) | void objective_val_p2_NCE_kernel(dType *d_p_true,double *d_final_NCE_OBJ...
  function objective_val_p2_NCE_kernel_nonshare (line 2423) | void objective_val_p2_NCE_kernel_nonshare(dType *d_p_true,double *d_fina...
  function zero_err_ht (line 2438) | void zero_err_ht(dType *d_err_ht,int *d_vocab_indicies_01,int LSTM_size,...
  function error_ht_positive_kernel (line 2451) | void error_ht_positive_kernel(dType *d_d_ERRt_ht,dType *d_p_true,dType *...
  function backprop_ht_SPARSE (line 2463) | void backprop_ht_SPARSE(dType *d_d_ERRt_ht,dType *d_D,dType *d_p_true,in...
  function embedding_gradient_sparse (line 2531) | void embedding_gradient_sparse(dType *d_D_grad,dType *d_h_t,dType *d_p_t...
  function bias_gradient_sparse (line 2561) | void bias_gradient_sparse(dType *d_b_d_grad,dType *d_p_true,int *d_sampl...
  function positive_embedding_NCE (line 2581) | void positive_embedding_NCE(dType *d_h_t,dType *d_small_D_grad,dType *d_...
  function negative_embedding_NCE (line 2601) | void negative_embedding_NCE(dType *d_temp_D_grad,dType *d_small_D_grad,i...
  function __global__ (line 2617) | __global__
  function update_sparse_grad (line 2629) | void update_sparse_grad(dType *d_mat,dType *d_small_grad,int *d_unique_i...
  function negative_bias_NCE (line 2647) | void negative_bias_NCE(dType *d_temp_b_d_grad,dType *d_b_d_grad,int *d_s...
  function positive_bias_NCE (line 2656) | void positive_bias_NCE(dType *d_b_d_grad,dType *d_p_true,int *d_samples,...
  function load_in_embeddings_trunc (line 2671) | void load_in_embeddings_trunc(dType *d_temp_embeddings,dType *d_D,int *d...
  function tanh_bi_forward_kernel (line 2695) | void tanh_bi_forward_kernel(dType *d_mat,dType *d_bias,int LSTM_size,int...
  function add_to_errors (line 2705) | void add_to_errors(dType *d_error_ht,dType *d_additional_error_ht,int LS...
  function add_two_mats (line 2716) | void add_two_mats(dType *d_final,dType *d_mat1,dType *d_mat2,int size) {
  function check_elems_kernel (line 2725) | void check_elems_kernel(dType *d_mat,int size) {
  function add_four_matrices_kernel_stride (line 2735) | void add_four_matrices_kernel_stride(dType *d_final,dType *d_mat1,dType ...
  function add_features (line 2748) | void add_features(dType *probs, dType * sentence_scores, dType * encourage,
  function add_feature_repeat (line 2776) | void add_feature_repeat(dType *probs,
  function top_k (line 2797) | void top_k(dType *probs, dType *results, int* pointers, int* dict, int *...
  function top_k (line 2814) | void top_k(dType *probs, dType *results, int* dict, int dict_size) {
  function hash_code_kernel (line 2829) | void hash_code_kernel(int *d_codes, dType *d_vectors, int * d_permutes, ...
  function hash_code_kernel_T (line 2860) | void hash_code_kernel_T(int *d_codes, dType *d_vectors, int * d_permutes...
  function pad_h_t_T (line 2890) | void pad_h_t_T(dType * d_h_t_pad, dType *d_h_t, int LSTM_size, int beam_...
  function pad_h_t (line 2908) | void pad_h_t(dType * d_h_t_pad, dType *d_h_t, int LSTM_size, int beam_si...
  function sparse_dot_product (line 2929) | void sparse_dot_product(dType *d_outputdist, dType *d_results, dType *d_...
  function sparse_dot_product_2 (line 2966) | void sparse_dot_product_2(dType *d_outputdist, dType *d_Db, dType *d_h_t...
  function sparse_dot_product_3 (line 3006) | void sparse_dot_product_3(dType *d_outputdist, dType *d_Db, dType *d_h_t...
  function hash_func_1 (line 3027) | int hash_func_1(int a){
  function hash_func_2 (line 3038) | int hash_func_2(int key){
  function prepare_Db (line 3053) | void prepare_Db(dType * d_Db, dType * d_D, dType * d_b, int vocab_size, ...
  function cuckoo_lookup (line 3069) | void cuckoo_lookup(int *d_codes, dType *d_outputdist,int batch_size, int...
  function __global__ (line 3098) | __global__ void inc_range(float *d_outputdist, int * d_bands_index, int ...
  function cuckoo_lookup_T (line 3111) | void cuckoo_lookup_T(int *d_codes, dType *d_outputdist,int batch_size, i...
  function cuckoo_lookup_T_2 (line 3158) | void cuckoo_lookup_T_2(int *d_codes, dType *d_outputdist,int batch_size,...
  function shrink_vocab (line 3219) | void shrink_vocab(dType *d_D_shrink, dType *d_D, dType *d_b_shrink, dTyp...
  function __global__ (line 3232) | __global__ void dense2array(float *matrix, int vocab_size, int batch_siz...
  function __global__ (line 3279) | __global__
  type non_negative (line 3291) | struct non_negative
  function fill_number (line 3303) | void fill_number(dType *d_dist, int vocab_size, int batch_size, dType val){
  function __global__ (line 3314) | __global__
  function compact (line 3323) | void compact(int* h_input, int* h_output, int *d_input, int *d_output, i...

FILE: src/decoder.h
  function operator (line 70) | bool operator==(const dec_obj &other) const
  function namespace (line 80) | namespace std {
  function compare_pq (line 140) | bool compare_pq(dec_obj<float> &a,dec_obj<float> &b)
  type pq_compare_functor (line 146) | struct pq_compare_functor {
  type pq_global_compare_functor (line 151) | struct pq_global_compare_functor {
  type k_best_compare_functor (line 156) | struct k_best_compare_functor {
  function init_encourage_lists (line 462) | void init_encourage_lists(std::vector<std::string> fns, std::vector<floa...
  function init_encourage_list (line 485) | void init_encourage_list(std::string fn, float weight){
  function init_fsa_inner (line 518) | void init_fsa_inner(global_params &params){
  function init_fsa (line 528) | void init_fsa(fsa *fsa_model, std::unordered_map<std::string,int> &tgt_m...
  function empty_queue_pq (line 595) | void empty_queue_pq() {
  function empty_queue_global (line 605) | void empty_queue_global(std::priority_queue<dec_global_obj<dType>,std::v...
  function empty_queue_global (line 615) | void empty_queue_global() {
  function expand_pq_global_gpu (line 956) | void expand_pq_global_gpu(
  function print_current_hypotheses (line 1357) | void print_current_hypotheses() {

FILE: src/fileHelper.h
  type char_cnn_params (line 16) | struct char_cnn_params
  function zero_bitmaps (line 19) | struct file_helper {
  function preprocess_output_truncated_softmax (line 119) | void preprocess_output_truncated_softmax() {
  function preprocess_input_Wgrad (line 187) | void preprocess_input_Wgrad() {
  function read_minibatch (line 367) | bool read_minibatch() {

FILE: src/fileHelper_source.h
  function zero_bitmaps (line 17) | struct file_helper_source {
  function preprocess_input_Wgrad (line 83) | void preprocess_input_Wgrad() {
  function init_file_helper_source (line 133) | void init_file_helper_source(std::string fn,int ms,int max_sent_len,int ...
  function read_minibatch (line 158) | void read_minibatch() {

FILE: src/file_helper_char_decoder.h
  function read_minibatch (line 7) | struct file_helper_char_decoder {
  function reset_file (line 42) | void reset_file() {

FILE: src/file_helper_decoder.h
  function file_helper_decoder (line 15) | struct file_helper_decoder {
  function read_sentence (line 75) | bool read_sentence() {

FILE: src/format.cc
  function fmt_snprintf (line 93) | inline int fmt_snprintf(char *buffer, size_t size, const char *format, ....
  type IntChecker (line 106) | struct IntChecker {
    method fits_in_int (line 108) | static bool fits_in_int(T value) {
  type IntChecker<true> (line 115) | struct IntChecker<true> {
    method fits_in_int (line 117) | static bool fits_in_int(T value) {
  function safe_strerror (line 135) | int safe_strerror(
  function format_error_code (line 169) | void format_error_code(fmt::Writer &out, int error_code,
  function report_error (line 187) | void report_error(FormatFunc func,
  class IsZeroInt (line 198) | class IsZeroInt : public fmt::internal::ArgVisitor<IsZeroInt, bool> {
    method visit_any_int (line 201) | bool visit_any_int(T value) { return value == 0; }
  function parse_nonnegative_int (line 207) | int parse_nonnegative_int(const Char *&s) {
  function require_numeric_argument (line 224) | inline void require_numeric_argument(const Arg &arg, char spec) {
  function check_sign (line 233) | void check_sign(const Char *&s, const Arg &arg) {
  class WidthHandler (line 245) | class WidthHandler : public fmt::internal::ArgVisitor<WidthHandler, unsi...
    method WidthHandler (line 252) | explicit WidthHandler(fmt::FormatSpec &spec) : spec_(spec) {}
    method visit_unhandled_arg (line 254) | unsigned visit_unhandled_arg() {
    method visit_any_int (line 260) | unsigned visit_any_int(T value) {
  class PrecisionHandler (line 273) | class PrecisionHandler :
    method visit_unhandled_arg (line 276) | unsigned visit_unhandled_arg() {
    method visit_any_int (line 282) | int visit_any_int(T value) {
  class ArgConverter (line 291) | class ArgConverter : public fmt::internal::ArgVisitor<ArgConverter<T>, v...
    method ArgConverter (line 299) | ArgConverter(fmt::internal::Arg &arg, wchar_t type)
    method visit_any_int (line 303) | void visit_any_int(U value) {
  class CharConverter (line 331) | class CharConverter : public fmt::internal::ArgVisitor<CharConverter, vo...
    method CharConverter (line 338) | explicit CharConverter(fmt::internal::Arg &arg) : arg_(arg) {}
    method visit_any_int (line 341) | void visit_any_int(T value) {
  function ignore_incompatible_str (line 354) | inline Arg::StringValue<char> ignore_incompatible_str(
  function ignore_incompatible_str (line 358) | inline Arg::StringValue<wchar_t> ignore_incompatible_str(
  function FMT_FUNC (line 362) | FMT_FUNC void fmt::SystemError::init(
  function FMT_FUNC (line 433) | FMT_FUNC void fmt::internal::report_unknown_type(char code, const char *...
  function FMT_FUNC (line 445) | FMT_FUNC fmt::internal::UTF8ToUTF16::UTF8ToUTF16(fmt::StringRef s) {
  function FMT_FUNC (line 458) | FMT_FUNC fmt::internal::UTF16ToUTF8::UTF16ToUTF8(fmt::WStringRef s) {
  function FMT_FUNC (line 465) | FMT_FUNC int fmt::internal::UTF16ToUTF8::convert(fmt::WStringRef s) {
  function FMT_FUNC (line 477) | FMT_FUNC void fmt::WindowsError::init(
  function FMT_FUNC (line 510) | FMT_FUNC void fmt::internal::format_windows_error(
  function FMT_CATCH (line 535) | FMT_CATCH(...) {}
  class fmt::internal::ArgFormatter (line 542) | class fmt::internal::ArgFormatter :
    method ArgFormatter (line 553) | ArgFormatter(
    method visit_any_int (line 558) | void visit_any_int(T value) { writer_.write_int(value, spec_); }
    method visit_any_double (line 561) | void visit_any_double(T value) { writer_.write_double(value, spec_); }
    method visit_char (line 563) | void visit_char(int value) {
    method visit_string (line 594) | void visit_string(Arg::StringValue<char> value) {
    method visit_wstring (line 597) | void visit_wstring(Arg::StringValue<wchar_t> value) {
    method visit_pointer (line 601) | void visit_pointer(const void *value) {
    method visit_custom (line 609) | void visit_custom(Arg::CustomValue c) {
  function Arg (line 642) | inline Arg fmt::BasicFormatter<Char>::parse_arg_index(const Char *&s) {
  function FMT_FUNC (line 653) | FMT_FUNC Arg fmt::internal::FormatterBase::do_get_arg(
  function Arg (line 661) | inline Arg fmt::internal::FormatterBase::next_arg(const char *&error) {
  function Arg (line 668) | inline Arg fmt::internal::FormatterBase::get_arg(
  function Arg (line 706) | Arg fmt::internal::PrintfFormatter<Char>::get_arg(
  function Char (line 918) | const Char *fmt::BasicFormatter<Char>::format(
  function FMT_FUNC (line 1077) | FMT_FUNC void fmt::report_system_error(
  function FMT_FUNC (line 1083) | FMT_FUNC void fmt::report_windows_error(
  function FMT_FUNC (line 1089) | FMT_FUNC void fmt::print(std::FILE *f, StringRef format_str, ArgList arg...
  function FMT_FUNC (line 1095) | FMT_FUNC void fmt::print(StringRef format_str, ArgList args) {
  function FMT_FUNC (line 1099) | FMT_FUNC void fmt::print(std::ostream &os, StringRef format_str, ArgList...
  function FMT_FUNC (line 1105) | FMT_FUNC void fmt::print_colored(Color c, StringRef format, ArgList args) {
  function FMT_FUNC (line 1113) | FMT_FUNC int fmt::fprintf(std::FILE *f, StringRef format, ArgList args) {

FILE: src/format.h
  function namespace (line 50) | namespace fmt {
  function namespace (line 174) | namespace fmt {
  type BasicStringRef (line 272) | typedef BasicStringRef<char> StringRef;
  type BasicStringRef (line 273) | typedef BasicStringRef<wchar_t> WStringRef;
  function class (line 278) | class FormatError : public std::runtime_error {
  function namespace (line 284) | namespace internal {
  function wchar_t (line 515) | static wchar_t convert(wchar_t value) { return value; }
  function is_negative (line 526) | bool is_negative(T value) { return value < 0; }
  type SignChecker (line 530) | struct SignChecker
  function is_negative (line 532) | bool is_negative(T) { return false; }
  function is_negative (line 538) | bool is_negative(T value) {
  type Type (line 544) | typedef uint32_t Type;
  type TypeSelector (line 547) | struct TypeSelector
  type Type (line 547) | typedef uint64_t Type;
  type typename (line 553) | typedef typename
  type T (line 559) | typedef T Type;
  type BasicData (line 583) | typedef BasicData<> Data;
  function count_digits (line 596) | inline unsigned count_digits(uint64_t n) {
  function count_digits (line 604) | inline unsigned count_digits(uint64_t n) {
  function count_digits (line 622) | inline unsigned count_digits(uint32_t n) {
  function format_decimal (line 630) | void format_decimal(Char *buffer, UInt value, unsigned num_digits) {
  function class (line 654) | class UTF8ToUTF16 {
  function class (line 668) | class UTF16ToUTF8 {
  type NonZero (line 703) | struct NonZero
  type Value (line 709) | struct Value {
  type CustomValue (line 719) | struct CustomValue {
  function Value (line 740) | struct Arg : Value {
  type None (line 759) | typedef None<T> Supported;
  type T (line 760) | typedef T Unsupported;
  type T (line 765) | typedef T Supported;
  type None (line 766) | typedef None<T> Unsupported;
  function set_string (line 793) | void set_string(StringRef str) {
  function set_string (line 798) | void set_string(WStringRef str) {
  function format_custom_arg (line 805) | void format_custom_arg(
  function type (line 833) | static uint64_t type(long) {
  function type (line 843) | static uint64_t type(unsigned long) {
  function type (line 860) | static uint64_t type(wchar_t) { return Arg::CHAR; }
  function type (line 877) | static uint64_t type(Type) { return Arg::TYPE; }
  function FMT_MAKE_WSTR_VALUE (line 879) | FMT_MAKE_WSTR_VALUE(wchar_t *, WSTRING)
  function type (line 893) | uint64_t type(const T &) { return Arg::CUSTOM; }
  function Result (line 923) | Result visit_int(int value) {
  function Result (line 926) | Result visit_long_long(LongLong value) {
  function Result (line 929) | Result visit_uint(unsigned value) {
  function Result (line 932) | Result visit_ulong_long(ULongLong value) {
  function Result (line 935) | Result visit_char(int value) {
  function visit_any_int (line 939) | visit_any_int(T) {
  function Result (line 943) | Result visit_double(double value) {
  function Result (line 946) | Result visit_long_double(long double value) {
  function visit_any_double (line 950) | visit_any_double(T) {
  function Result (line 954) | Result visit_string(Arg::StringValue<char>) {
  function Result (line 957) | Result visit_wstring(Arg::StringValue<wchar_t>) {
  function Result (line 960) | Result visit_pointer(const void *) {
  function Result (line 963) | Result visit_custom(Arg::CustomValue) {
  function Result (line 967) | Result visit(const Arg &arg) {
  function class (line 1003) | class RuntimeError : public std::runtime_error {
  function class (line 1015) | class ArgList {
  type FormatSpec (line 1051) | struct FormatSpec
  function namespace (line 1053) | namespace internal {
  function writer_ (line 1116) | writer_(w) {}
  type Alignment (line 1125) | enum Alignment {
  type EmptySpec (line 1136) | struct EmptySpec {}
  function flag (line 1144) | bool flag(unsigned) const { return false; }
  function fill (line 1150) | struct WidthSpec {
  function flag (line 1179) | bool flag(unsigned) const { return false; }
  function AlignSpec (line 1184) | struct FormatSpec : AlignSpec {
  function namespace (line 1362) | namespace internal {
  function class (line 1497) | class SystemError : public internal::RuntimeError {
  type typename (line 1568) | typedef typename internal::CharTraits<Char>::CharPtr CharPtr;
  function Char (line 1572) | static Char *get(CharPtr p) { return p.base(); }
  function Char (line 1574) | static Char *get(Char *p) { return p; }
  function CharPtr (line 1584) | CharPtr grow_buffer(std::size_t n) {
  function CharPtr (line 1591) | CharPtr prepare_int_buffer(unsigned num_digits,
  function append_float_length (line 1630) | void append_float_length(Char *&format_ptr, long double) {
  function Char (line 1661) | const Char *data() const FMT_NOEXCEPT { return &buffer_[0]; }
  function Char (line 1667) | const Char *c_str() const {
  function write (line 1706) | void write(BasicStringRef<Char> format, ArgList args) {
  function clear (line 1786) | void clear() FMT_NOEXCEPT { buffer_.clear(); }
  type typename (line 1894) | typedef typename internal::IntTraits<T>::MainType UnsignedType;
  type BasicMemoryWriter (line 2189) | typedef BasicMemoryWriter<char> MemoryWriter;
  type BasicMemoryWriter (line 2190) | typedef BasicMemoryWriter<wchar_t> WMemoryWriter;
  type BasicArrayWriter (line 2239) | typedef BasicArrayWriter<char> ArrayWriter;
  type BasicArrayWriter (line 2240) | typedef BasicArrayWriter<wchar_t> WArrayWriter;
  function class (line 2263) | class WindowsError : public SystemError {
  type Color (line 2308) | enum Color { BLACK, RED, GREEN, YELLOW, BLUE, MAGENTA, CYAN, WHITE }
  function std (line 2327) | inline std::string format(StringRef format_str, ArgList args) {
  function std (line 2333) | inline std::wstring format(WStringRef format_str, ArgList args) {
  function std (line 2386) | inline std::string sprintf(StringRef format, ArgList args) {
  function printf (line 2412) | inline int printf(StringRef format, ArgList args) {
  function class (line 2419) | class FormatInt {
  function explicit (line 2461) | explicit FormatInt(long value) { FormatSigned(value); }
  function explicit (line 2462) | explicit FormatInt(LongLong value) { FormatSigned(value); }
  function explicit (line 2463) | explicit FormatInt(unsigned value) : str_(format_decimal(value)) {}
  function explicit (line 2464) | explicit FormatInt(unsigned long value) : str_(format_decimal(value)) {}
  function explicit (line 2465) | explicit FormatInt(ULongLong value) : str_(format_decimal(value)) {}
  function namespace (line 2624) | namespace fmt {

FILE: src/fsa.h
  type sw (line 22) | struct sw
  function class (line 24) | class state{
  type sw (line 66) | struct sw{
  function namespace (line 72) | namespace std{

FILE: src/global_params.h
  type precision (line 3) | typedef float precision;
  type attention_params (line 9) | struct attention_params {
  type bi_directional_params (line 18) | struct bi_directional_params {
  type global_params (line 61) | struct global_params {

FILE: src/gpu_info_struct.h
  type layer_gpu_info (line 4) | struct layer_gpu_info {
  type softmax_layer_gpu_info (line 161) | struct softmax_layer_gpu_info {
  type bi_layer_info (line 191) | struct bi_layer_info {
  type attention_layer_gpu_info (line 208) | struct attention_layer_gpu_info {

FILE: src/highway_network.hpp
  function __global__ (line 3) | __global__
  function __global__ (line 13) | __global__
  function __global__ (line 27) | __global__
  function __global__ (line 35) | __global__
  function __global__ (line 47) | __global__
  function __global__ (line 55) | __global__

FILE: src/input_file_prep.h
  type comb_sent_info (line 14) | struct comb_sent_info {
  function prep_files_train_nonLM_multi_source (line 542) | bool prep_files_train_nonLM_multi_source(int minibatch_size,int max_sent...
  function prep_files_train_nonLM_ensemble (line 1044) | bool prep_files_train_nonLM_ensemble(int minibatch_size,int max_sent_cut...
  function prep_files_train_LM (line 1432) | bool prep_files_train_LM(int minibatch_size,int max_sent_cutoff,
  function prep_files_train_LM_ensemble (line 1696) | bool prep_files_train_LM_ensemble(int minibatch_size,int max_sent_cutoff,
  function integerize_file_nonLM (line 1958) | void integerize_file_nonLM(std::string output_weights_name,std::string s...
  function integerize_file_LM_carve (line 2550) | void integerize_file_LM_carve(std::string output_weights_name,std::strin...
  function integerize_file_LM (line 2743) | void integerize_file_LM(std::string output_weights_name,std::string targ...
  function unint_file (line 3137) | void unint_file(std::string output_weights_name,std::string unint_file,s...
  function load_word_index_mapping (line 3207) | void load_word_index_mapping(std::string output_weights_name,bool LM,boo...
  function unint_alignments (line 3255) | void unint_alignments(std::string output_weights_name,std::string int_al...

FILE: src/logger.h
  function class (line 5) | class OutputLogger {

FILE: src/memory_util.h
  function process_mem_usage (line 20) | void process_mem_usage(double& vm_usage, double& resident_set)
  function print_CPU_Info (line 57) | void print_CPU_Info()
  function class (line 83) | class Timer {

FILE: src/model.h
  type file_helper (line 29) | struct file_helper
  function namespace (line 31) | namespace debug_flag {
  type attention_params (line 132) | struct attention_params
  type file_helper (line 196) | struct file_helper

FILE: src/model.hpp
  type file_helper (line 2244) | struct file_helper

FILE: src/multinomial.h
  function estimate (line 26) | void estimate(const std::vector<Count>& counts)
  function dType (line 44) | dType prob(int i) const { return m_prob[i]; }
  function dType (line 45) | dType logprob(int i) const { return m_logprob[i]; }

FILE: src/softmax.h
  type neuralMT_model (line 137) | struct neuralMT_model
  type neuralMT_model (line 140) | struct neuralMT_model

FILE: src/softmax.hpp
  type neuralMT_model<precision> (line 4) | struct neuralMT_model<precision>
  type neuralMT_model<precision> (line 31) | struct neuralMT_model<precision>
  function softmax_layer_gpu_info (line 1034) | softmax_layer_gpu_info softmax_layer<dType>::gpu_init(int device_number) {
  function dType (line 1045) | dType *softmax_layer<dType>::get_ht_ptr(int index) {
  function cudaEvent_t (line 1055) | cudaEvent_t softmax_layer<dType>::get_ERR_ht_event() {
  function dType (line 1060) | dType *softmax_layer<dType>::get_dist_ptr() {

FILE: src/transfer_layer.h
  function init_upper_transfer_layer (line 27) | void init_upper_transfer_layer(bool upper_softmax,bool copy_h_t,bool sou...
  function init_lower_transfer_layer (line 48) | void init_lower_transfer_layer(bool lower_input,bool copy_d_Err_ht,Input...

FILE: src/trunc_softmax.h
  function zero_bitmap (line 41) | void zero_bitmap() {
  function prep_GPU_vocab_indices (line 45) | void prep_GPU_vocab_indices(int *h_output_vocab_indicies_target,int curr...

Copy disabled (too large) Download .json

Condensed preview — 119 files, each showing path, character count, and a content snippet. Download the .json file for the full structured content (12,135K chars).

[
  {
    "path": ".gitattributes",
    "chars": 61,
    "preview": "executable/ZOPH_RNN_XING filter=lfs diff=lfs merge=lfs -text\n"
  },
  {
    "path": ".gitignore",
    "chars": 48,
    "preview": ".DS_Store\n*~\n[#].[#]\n.[#]*\n*[#]\n\n*pyc\n.history\n\n"
  },
  {
    "path": "README.md",
    "chars": 30484,
    "preview": "# Zoph\\_RNN: A C++/CUDA toolkit for training sequence and sequence-to-sequence models across multiple GPUs\n\nThis is [Bar"
  },
  {
    "path": "README_XING.md",
    "chars": 9153,
    "preview": "# Decoding with Finate State Acceptor (FSA), Locality Sensitive Hashing (LSH) and Word Alignemnt (WA)\n\nThis is the descr"
  },
  {
    "path": "executable/ZOPH_RNN_XING",
    "chars": 128,
    "preview": "version https://git-lfs.github.com/spec/v1\noid sha256:9331caaa8c6bf6ac7eb07a2d59e0706ad70f610ead632953de49015229cb0b57\ns"
  },
  {
    "path": "sample_data/dev_english.txt.tok.lc",
    "chars": 749313,
    "preview": "parliament does not support amendment freeing tymoshenko\ntoday , the ukraine parliament dismissed , within the code of c"
  },
  {
    "path": "sample_data/dev_french.txt.tok.lc",
    "chars": 886897,
    "preview": "le parlement n&apos; a pas ratifié l&apos; amendement pour la libération de tymosenko\nle parlement ukrainien a refusé , "
  },
  {
    "path": "sample_data/ptb.train.txt",
    "chars": 5101618,
    "preview": " aer banknote berlitz calloway centrust cluett fromstein gitano guterman hydro-quebec ipo kia memotec mlx nahb punts rak"
  },
  {
    "path": "sample_data/ptb.valid.txt",
    "chars": 399782,
    "preview": " consumers may want to move their telephones a little closer to the tv set \n <unk> <unk> watching abc 's monday night fo"
  },
  {
    "path": "sample_data/test_english.tok.lc",
    "chars": 393811,
    "preview": "jet makers feud over seat width with big orders at stake\na row has flared up between leading plane makers over the width"
  },
  {
    "path": "sample_data/test_french.tok.lc",
    "chars": 463901,
    "preview": "les avionneurs se querellent au sujet de la largeur des sièges alors que de grosses commandes sont en jeu\nla dispute fai"
  },
  {
    "path": "sample_data/train_english.txt.tok.lc.10k",
    "chars": 1145746,
    "preview": "milla jovovich , john malkovich , faye dunaway , and dustin hoffman .\nel colegio de méxico .\nthese medications include a"
  },
  {
    "path": "sample_data/train_french.txt.tok.lc.10k",
    "chars": 1198403,
    "preview": "milla jovovich , john malkovich , faye dunaway et dustin hoffman .\nel colegio de mexico .\ncette classe de médicaments in"
  },
  {
    "path": "scripts/att_unk_rep.py",
    "chars": 2410,
    "preview": "#Run this with the output of the neural MT program\nimport sys\nimport codecs\nimport re\n\nsrc_file_name = str(sys.argv[1])\n"
  },
  {
    "path": "scripts/berk_aligner/run_aligner.sh",
    "chars": 28,
    "preview": "bash align unk_replace.conf\n"
  },
  {
    "path": "scripts/berk_aligner/unk_replace.conf",
    "chars": 1014,
    "preview": "##########################################\n# Training: Defines the training regimen #\n##################################"
  },
  {
    "path": "scripts/bleu_format.py",
    "chars": 548,
    "preview": "#pass the output of the kbest from the RNN toolkit into this to strip it of the extra padding\n\nimport codecs\nimport sys\n"
  },
  {
    "path": "scripts/bleu_format_valid.py",
    "chars": 1272,
    "preview": "#pass the output of the kbest from the RNN toolkit into this to strip it of the extra padding\n\n# The difference between "
  },
  {
    "path": "scripts/compile.sh",
    "chars": 1955,
    "preview": "#Written by Barret Zoph, for questions email barretzoph@gmail.com\n\n#Compilation script for compiling ZOPH_RNN\n#The follo"
  },
  {
    "path": "scripts/compile.xing.sh",
    "chars": 2225,
    "preview": "#Written by Barret Zoph, for questions email barretzoph@gmail.com\n#Futher edit by Xing Shi (shixing19910105@gmail.com)\n\n"
  },
  {
    "path": "scripts/create_vocab_mapping_file.py",
    "chars": 1954,
    "preview": "import codecs\nimport sys\nfrom itertools import izip\nfrom collections import defaultdict as dd\n\nif len(sys.argv)!=5:\n\tpri"
  },
  {
    "path": "scripts/create_vocab_mapping_file_preinit.py",
    "chars": 2464,
    "preview": "# -*- coding: utf-8 -*-\nimport codecs\nimport sys\nfrom itertools import izip\nfrom collections import defaultdict as dd\n\ni"
  },
  {
    "path": "scripts/fsa/convert.py",
    "chars": 120,
    "preview": "import sys\nfor line in sys.stdin:\n    ll = line.split()\n    for letter in ll:\n        print ord(letter) % 10,\n    print\n"
  },
  {
    "path": "scripts/fsa/demo.sh",
    "chars": 4135,
    "preview": "# we have a simple task, translate number into letters\n\n# first run generate_fsa.py to generate the fsa file and trainin"
  },
  {
    "path": "scripts/fsa/enc1.txt",
    "chars": 6,
    "preview": "l\ns\nt\n"
  },
  {
    "path": "scripts/fsa/enc2.txt",
    "chars": 13,
    "preview": "i 0.3\nm -0.5\n"
  },
  {
    "path": "scripts/fsa/fsa.txt",
    "chars": 574,
    "preview": "E\n(S (1 l))\n(1 (2 s))\n(2 (3 t))\n(3 (E m))\n(S (5 i))\n(5 (E s))\n(S (7 g))\n(7 (8 r))\n(8 (9 e))\n(9 (10 a))\n(10 (E t))\n(S (12"
  },
  {
    "path": "scripts/fsa/generate_fsa.py",
    "chars": 1317,
    "preview": "import random\n\ndef generate(word,start,end,i,f):\n    s = start\n    n = i\n    for l, w in enumerate(word):\n        if l ="
  },
  {
    "path": "scripts/fsa/source.singleline.txt",
    "chars": 20,
    "preview": "8 5 6 9 5 5 8 7 9 7\n"
  },
  {
    "path": "scripts/fsa/source.test.txt",
    "chars": 34,
    "preview": "11 6 11 12 11 4 7 2 10 6 7 0 12 0\n"
  },
  {
    "path": "scripts/fsa/source.train.txt",
    "chars": 27371,
    "preview": "4 11 12 5 1 12 1 12\n2 10 11 11 6 4 0 8 3 1 4 4\n2 10 12 10 10 6 12 2 1 12 0\n1 12 6 6 9 4 0 8 3 1 4 4\n10 6 7 0 12 0 4 0 8 "
  },
  {
    "path": "scripts/fsa/source.valid.txt",
    "chars": 2802,
    "preview": "6 6 9 12 10 10 6 12 12 10 10 6 12\n2 1 12 0 11 6 11 12 11 4 7 2\n6 6 9 1 11 11 6 11 12\n2 1 12 0 1 11 11 4 7 2\n10 6 7 0 12 "
  },
  {
    "path": "scripts/fsa/target.test.txt",
    "chars": 28,
    "preview": "f a s t s l o w e n o u g h\n"
  },
  {
    "path": "scripts/fsa/target.train.txt",
    "chars": 22952,
    "preview": "l s t m i t i t\nw e f s a l u c k i l y\nw e g r e a t w i t h\ni t a n d l u c k i l y\ne n o u g h l u c k i l y f s a\nm "
  },
  {
    "path": "scripts/fsa/target.valid.txt",
    "chars": 2374,
    "preview": "a n d g r e a t g r e a t\nw i t h f a s t s l o w\na n d i s f a s t\nw i t h i s s l o w\ne n o u g h f a s t l s t m\ni t "
  },
  {
    "path": "scripts/generate_train_decode.py",
    "chars": 2904,
    "preview": "import sys\nimport os\n\nhead = \"\"\"\n#!/bin/bash\n#PBS -q isi\n#PBS -l walltime=300:00:00\n#PBS -l gpus=2\n\nPREFIX=__PREFIX__\nmo"
  },
  {
    "path": "scripts/load_lstm.py",
    "chars": 3327,
    "preview": "import sys\nimport re\nimport numpy as np\nimport pandas as pd\n\ndef parse(fn,key = \"all\", side = 0, the_layer = 0):\n\n    f "
  },
  {
    "path": "scripts/load_model.py",
    "chars": 11798,
    "preview": "#load the non attentional model \n\nimport numpy as np\nfrom cStringIO import StringIO\nfrom datetime import datetime\nfrom n"
  },
  {
    "path": "scripts/multi-bleu.perl",
    "chars": 4826,
    "preview": "#!/usr/bin/env perl\n#\n# This file is part of moses.  Its use is licensed under the GNU Lesser General\n# Public License v"
  },
  {
    "path": "scripts/pretrain.py",
    "chars": 9117,
    "preview": "#!/usr/bin/env python3\n# code by Jon May [jonmay@isi.edu]. Port of code by Deniz Yuret with some interface improvements\n"
  },
  {
    "path": "scripts/read_hpc_output.py",
    "chars": 1057,
    "preview": "# python read_hpc_output.py folder \n\nimport sys\nimport os\nimport pandas as pd\n\ndef parse_arg():\n    d = {}\n    d['folder"
  },
  {
    "path": "scripts/translate/f2e_decode.sh",
    "chars": 1154,
    "preview": "#!/usr/bin/env bash\n#PBS -q isi\n#PBS -l walltime=1:00:00\n#PBS -l gpus=2\n\nSCRIPTDIR=$( cd \"$( dirname \"${BASH_SOURCE[0]}\""
  },
  {
    "path": "scripts/translate/f2e_train.sh",
    "chars": 1061,
    "preview": "#!/usr/bin/env bash\n#PBS -q isi80\n#PBS -l walltime=300:00:00\n#PBS -l gpus=4\n\nSCRIPTDIR=$( cd \"$( dirname \"${BASH_SOURCE["
  },
  {
    "path": "scripts/unk_format.py",
    "chars": 475,
    "preview": "import codecs\nimport sys\nimport re\n\nif len(sys.argv) != 3:\n    print(\"format: <input file> <output file name>\")\n    sys."
  },
  {
    "path": "src/BZ_CUDA_UTIL.h",
    "chars": 32848,
    "preview": "//CUDA utilility for LSTM RNN\n\n#ifndef BZ_CUDA_UTIL_H\n#define BZ_CUDA_UTIL_H\n\n#include <stdlib.h>\n#include <boost/random"
  },
  {
    "path": "src/Eigen_Util.h",
    "chars": 11167,
    "preview": "\n#ifndef EIGEN_UTIL_H\n#define EIGEN_UTIL_H\n\n#include <fstream>\n//#include <boost/random/uniform_01.hpp>\n#include <boost/"
  },
  {
    "path": "src/Full_Input_To_Hidden_Layer.h",
    "chars": 186,
    "preview": "//This contains the full input to hidden layer for the model \n\n#ifndef LSTM_FULL_INPUT_TO_HIDDEN_H\n#define LSTM_FULL_INP"
  },
  {
    "path": "src/Hidden_To_Hidden_Layer.h",
    "chars": 9435,
    "preview": "//LSTM layer that connects input to hidden\n#ifndef LSTM_HIDDEN_TO_HIDDEN_H\n#define LSTM_HIDDEN_TO_HIDDEN_H\n\n#include \"LS"
  },
  {
    "path": "src/Hidden_To_Hidden_Layer.hpp",
    "chars": 33516,
    "preview": "\n\ntemplate<typename dType>\nvoid Hidden_To_Hidden_Layer<dType>::init_Hidden_To_Hidden_Layer_GPU(int LSTM_size,int minibat"
  },
  {
    "path": "src/Input_To_Hidden_Layer.h",
    "chars": 11371,
    "preview": "//LSTM layer that connects input to hidden\n#ifndef LSTM_INPUT_TO_HIDDEN_H\n#define LSTM_INPUT_TO_HIDDEN_H\n\ntemplate<typen"
  },
  {
    "path": "src/Input_To_Hidden_Layer.hpp",
    "chars": 54035,
    "preview": "\ntemplate<typename dType>\nvoid Input_To_Hidden_Layer<dType>::init_Input_To_Hidden_Layer_GPU(int LSTM_size,int minibatch_"
  },
  {
    "path": "src/LSH_WTA.h",
    "chars": 31584,
    "preview": "//\n//  LSH_WTA.h\n//  lstm_github\n//\n//  Created by Xing Shi on 1/3/17.\n//  Copyright © 2017 Xing Shi. All rights reserve"
  },
  {
    "path": "src/LSTM.h",
    "chars": 2769,
    "preview": " //The LSTM file that contains all the info for the LSTM that is needed for forward and backward propagation for gradien"
  },
  {
    "path": "src/LSTM.hpp",
    "chars": 59566,
    "preview": "\n//Constructor\ntemplate<typename dType>\nLSTM_IH_Node<dType>::LSTM_IH_Node(int LSTM_size,int minibatch_size,int vocab_siz"
  },
  {
    "path": "src/LSTM_HH.h",
    "chars": 2515,
    "preview": " //The LSTM file that contains all the info for the LSTM that is needed for forward and backward propagation for gradien"
  },
  {
    "path": "src/LSTM_HH.hpp",
    "chars": 43096,
    "preview": "\n//Constructor\ntemplate<typename dType>\nLSTM_HH_Node<dType>::LSTM_HH_Node(int LSTM_size,int minibatch_size,struct Hidden"
  },
  {
    "path": "src/NCE.h",
    "chars": 4344,
    "preview": "#ifndef NCE_H\n#define NCE_H\n\n#include \"multinomial.h\"\n#include <algorithm> \n#include <fstream>\n#include <unordered_map>\n"
  },
  {
    "path": "src/NCE.hpp",
    "chars": 43342,
    "preview": "\ntemplate<typename dType>\nvoid NCE_layer<dType>::init_loss_layer(struct neuralMT_model<precision> *model,global_params &"
  },
  {
    "path": "src/NCE_node.h",
    "chars": 1488,
    "preview": "\n#ifndef NCE_NODE_H\n#define NCE_NODE_H\n\ntemplate<typename dType>\nstruct NCE_Node {\n\n\t//each node stores the unnormalized"
  },
  {
    "path": "src/add_model_info.h",
    "chars": 1163,
    "preview": "//function for adding in the model information\n#ifndef ADD_MODEL_INFO_H\n#define ADD_MODEL_INFO_H\n\n#include <fstream>\n#in"
  },
  {
    "path": "src/attention_combiner.h",
    "chars": 1416,
    "preview": "//for combining two attention layers\n\n#ifndef ATT_COMBINER_H\n#define ATT_COMBINER_H\n\n#include \"attention_combiner_node.h"
  },
  {
    "path": "src/attention_combiner.hpp",
    "chars": 9361,
    "preview": "//hpp\n\ntemplate<typename dType>\nattention_combiner_layer<dType>::attention_combiner_layer(global_params &params,int devi"
  },
  {
    "path": "src/attention_combiner_node.h",
    "chars": 953,
    "preview": "#ifndef ATT_COMBINER_NODE_H\n#define ATT_COMBINER_NODE_H\n\ntemplate<typename dType>\nclass attention_combiner_layer;\n\ntempl"
  },
  {
    "path": "src/attention_combiner_node.hpp",
    "chars": 5553,
    "preview": "\n\n\n\ntemplate<typename dType>\nattention_combiner_node<dType>::attention_combiner_node(global_params &params,attention_com"
  },
  {
    "path": "src/attention_layer.h",
    "chars": 3873,
    "preview": "#ifndef ATTENTION_LAYER_H\n#define ATTENTION_LAYER_H\n\n\ntemplate<typename dType>\nclass neuralMT_model;\n\ntemplate<typename "
  },
  {
    "path": "src/attention_layer.hpp",
    "chars": 28923,
    "preview": "\ntemplate<typename dType>\nattention_layer<dType>::attention_layer(int LSTM_size,int minibatch_size, int device_number, i"
  },
  {
    "path": "src/attention_node.h",
    "chars": 2520,
    "preview": "#ifndef ATTENTION_NODE_H\n#define ATTENTION_NODE_H\n\n\ntemplate<typename dType>\nclass attention_layer;\n\n\ntemplate<typename "
  },
  {
    "path": "src/attention_node.hpp",
    "chars": 35331,
    "preview": "\ntemplate<typename dType>\nattention_node<dType>::attention_node(int LSTM_size,int minibatch_size,int device_number,int D"
  },
  {
    "path": "src/base_layer.h",
    "chars": 243,
    "preview": "#ifndef BASE_LAYER_H\n#define BASE_LAYER_H\n\n#include \"model.h\"\n#include \"Eigen_Util.h\"\n\ntemplate<typename dType>\nclass ne"
  },
  {
    "path": "src/base_layer.hpp",
    "chars": 0,
    "preview": ""
  },
  {
    "path": "src/base_loss.h",
    "chars": 1786,
    "preview": "\n//base loss class, so MLE,NCE, etc ...\n#ifndef BASE_LOSS_H\n#define BASE_LOSS_H\n\ntemplate<typename dType>\nclass Hidden_T"
  },
  {
    "path": "src/bi_encoder.h",
    "chars": 3704,
    "preview": "//Bidirectional source encoder for MT\n#ifndef BI_ENCODER_H\n#define BI_ENCODER_H\n\n#include \"gpu_info_struct.h\"\n\ntemplate<"
  },
  {
    "path": "src/bi_encoder.hpp",
    "chars": 38567,
    "preview": "template<typename dType>\nbi_encoder<dType>::bi_encoder() {\n\n}\n\n\ntemplate<typename dType>\nvoid bi_encoder<dType>::init_la"
  },
  {
    "path": "src/charCNN_node.h",
    "chars": 1904,
    "preview": "#ifndef CHARCNN_NODE_H\n#define CHARCNN_NODE_H\n\n//charCNN node\n\ntemplate<typename dType>\nstruct charCNN_node {\n\n\tdType *d"
  },
  {
    "path": "src/char_file_helper.h",
    "chars": 2713,
    "preview": "#ifndef FILE_INPUT_CHAR\n#define FILE_INPUT_CHAR\n\n//file helper for character stuff\n\n\nstruct file_helper_char {\n\n\tint num"
  },
  {
    "path": "src/conv_char.h",
    "chars": 4228,
    "preview": "#ifndef CONV_CHAR_H\n#define CONV_CHAR_H\n\ntemplate<typename dType>\nclass neuralMT_model;\n\n#include \"highway_network.h\"\n#i"
  },
  {
    "path": "src/conv_char.hpp",
    "chars": 30438,
    "preview": "template<typename dType>\nvoid conv_char_layer<dType>::prep_vocab_indicies(int *h_vocab_indicies_full,int curr_sent_len,\n"
  },
  {
    "path": "src/custom_kernels.h",
    "chars": 114935,
    "preview": "//Custom Kernels\n#ifndef CUSTOM_KERNELS_H\n#define CUSTOM_KERNELS_H\n#include <thrust/transform_reduce.h>\n#include <assert"
  },
  {
    "path": "src/decoder.h",
    "chars": 48454,
    "preview": "#ifndef DECODER_H\n#define DECODER_H\n\n#include <queue>\n#include <vector>\n#include <fstream>\n#include <utility> \n#include "
  },
  {
    "path": "src/decoder_model_wrapper.h",
    "chars": 4207,
    "preview": "#ifndef DECODER_MODEL_WRAPPER_H\n#define DECODER_MODEL_WRAPPER_H\n\n#include \"file_helper_decoder.h\"\n\ntemplate<typename dTy"
  },
  {
    "path": "src/decoder_model_wrapper.hpp",
    "chars": 18006,
    "preview": "template<typename dType>\ndecoder_model_wrapper<dType>::decoder_model_wrapper(int gpu_num,int beam_size,\n\t\t\t\t\t\t    std::s"
  },
  {
    "path": "src/encoder_multi_source.h",
    "chars": 2830,
    "preview": "//Bidirectional source encoder for MT\n#ifndef ENCODER_MULTI_SOURCE_H\n#define ENCODER_MULTI_SOURCE_H\n\n#include \"gpu_info_"
  },
  {
    "path": "src/encoder_multi_source.hpp",
    "chars": 23438,
    "preview": "template<typename dType>\nencoder_multi_source<dType>::encoder_multi_source() {\n\n}\n\n\ntemplate<typename dType>\nvoid encode"
  },
  {
    "path": "src/ensemble_factory.h",
    "chars": 1896,
    "preview": "#ifndef ENSEMBLE_FACTORY_H\n#define ENSEMBLE_FACTORY_H\n\n#include \"file_helper_decoder.h\"\n#include \"decoder.h\"\n#include \"m"
  },
  {
    "path": "src/ensemble_factory.hpp",
    "chars": 33189,
    "preview": "\ntemplate<typename dType>\nensemble_factory<dType>::ensemble_factory(std::vector<std::string> weight_file_names,int num_h"
  },
  {
    "path": "src/fileHelper.h",
    "chars": 17943,
    "preview": "//Load in the training examples from the file\n\n#ifndef FILE_INPUT\n#define FILE_INPUT\n\n#include <string>\n#include <vector"
  },
  {
    "path": "src/fileHelper_source.h",
    "chars": 8140,
    "preview": "//file helper for other language\n\n//Load in the training examples from the file\n\n#ifndef FILE_INPUT_SOURCE\n#define FILE_"
  },
  {
    "path": "src/file_helper_char_decoder.h",
    "chars": 1103,
    "preview": "#ifndef FILE_INPUT_CHAR_DECODER\n#define FILE_INPUT_CHAR_DECODER\n\n//file helper for character stuff\n\n\nstruct file_helper_"
  },
  {
    "path": "src/file_helper_decoder.h",
    "chars": 4433,
    "preview": "//This is the file reader for the beam decoder\n#ifndef FILE_INPUT_DECODER\n#define FILE_INPUT_DECODER\n\n#include <string>\n"
  },
  {
    "path": "src/format.cc",
    "chars": 35032,
    "preview": "/*\n Formatting library for C++\n\n Copyright (c) 2012 - 2015, Victor Zverovich\n All rights reserved.\n\n Redistribution and "
  },
  {
    "path": "src/format.h",
    "chars": 79948,
    "preview": "/*\n Formatting library for C++\n\n Copyright (c) 2012 - 2015, Victor Zverovich\n All rights reserved.\n\n Redistribution and "
  },
  {
    "path": "src/fsa.h",
    "chars": 2952,
    "preview": "#ifndef FSA_H\n#define FSA_H\n#include <unordered_map>\n#include <unordered_set>\n#include <fstream>\n#include <vector>\n#incl"
  },
  {
    "path": "src/fsa.hpp",
    "chars": 11326,
    "preview": "#ifndef FSA_HPP\n#define FSA_HPP\n\n#include \"fsa.h\"\n\n//------------------- state ----------------\n\n\n// the copy constructe"
  },
  {
    "path": "src/global_params.h",
    "chars": 9893,
    "preview": "//Global parameter file that needs to be specified for learning\n\ntypedef float precision;\n#define NDEBUG\n//#define REMOV"
  },
  {
    "path": "src/gpu_info_struct.h",
    "chars": 6631,
    "preview": "#ifndef GPU_INFO_STRUCT_H\n#define GPU_INFO_STRUCT_H\n\nstruct layer_gpu_info {\n\tint device_number = 0;//Input layer always"
  },
  {
    "path": "src/highway_network.h",
    "chars": 1777,
    "preview": "#ifndef HIGHWAY_NETWORK_H\n#define HIGHWAY_NETWORK_H\n\n#include \"highway_node.h\"\n\n//highway network layer\ntemplate<typenam"
  },
  {
    "path": "src/highway_network.hpp",
    "chars": 13955,
    "preview": "//cuda kernels\ntemplate<typename dType>\n__global__\nvoid sigmoid_bias_kernel(dType *d_final,dType *d_bias,int state_size,"
  },
  {
    "path": "src/highway_node.h",
    "chars": 982,
    "preview": "template<typename dType>\nstruct highway_node {\n\n\t//each node stores the unnormalized probabilities, plus the h_t\n\tdType "
  },
  {
    "path": "src/input_file_prep.h",
    "chars": 107552,
    "preview": "#ifndef INPUT_FILE_PREP_H\n#define INPUT_FILE_PREP_H\n\n#include <fstream>\n#include <unordered_map>\n#include <vector>\n#incl"
  },
  {
    "path": "src/input_file_prep.hpp",
    "chars": 11463,
    "preview": "//additional ugly file stuff\n\nbool input_file_prep::prep_files_train_nonLM_multi_source_ensemble(int minibatch_size,int "
  },
  {
    "path": "src/logger.h",
    "chars": 1267,
    "preview": "//logger for writing to both \n#ifndef OUTPUT_LOGGER_H\n#define OUTPUT_LOGGER_H\n\nclass OutputLogger {\npublic:\n\tbool log_ou"
  },
  {
    "path": "src/main.cu",
    "chars": 81646,
    "preview": "\n#include <iostream>\n#include <vector>\n#include <time.h>\n#include <cmath>\n#include <chrono>\n#include <iomanip>\n#include "
  },
  {
    "path": "src/memory_util.h",
    "chars": 3326,
    "preview": "#ifndef MEMORY_UTIL_H\n#define MEMORY_UTIL_H\n\n#include <unistd.h>\n#include <ios>\n#include <iostream>\n#include <fstream>\n#"
  },
  {
    "path": "src/model.h",
    "chars": 8041,
    "preview": "//Model file that contains the parameters for the model\n\n#ifndef MODEL_H\n#define MODEL_H\n\n#include <vector>\n#include <Ei"
  },
  {
    "path": "src/model.hpp",
    "chars": 111998,
    "preview": "//Model.hpp file that contains implementations for the model class\ntemplate<typename dType>\nvoid neuralMT_model<dType>::"
  },
  {
    "path": "src/multinomial.h",
    "chars": 3192,
    "preview": "#ifndef MULTINOMIAL_H\n#define MULTINOMIAL_H\n\n#include <vector>\n#include <set>\n#include <cassert>\n#include <cmath>\n\n#incl"
  },
  {
    "path": "src/prev_states.h",
    "chars": 586,
    "preview": "\n\n//for decoding multilayer models\ntemplate<typename dType>\nstruct prev_source_state {\n\tdType *d_h_t_prev;\n\tdType *d_c_t"
  },
  {
    "path": "src/softmax.h",
    "chars": 6410,
    "preview": "#ifndef SOFTMAX_H\n#define SOFTMAX_H\n\n#include <Eigen/Dense>\n#include \"Eigen_Util.h\"\n#include \"gpu_info_struct.h\"\n#includ"
  },
  {
    "path": "src/softmax.hpp",
    "chars": 40249,
    "preview": "\n\ntemplate<typename dType>\nvoid softmax_layer<dType>::init_loss_layer(struct neuralMT_model<precision> *model,global_par"
  },
  {
    "path": "src/softmax_node.h",
    "chars": 989,
    "preview": "#ifndef SOFTMAX_NODE_H\n#define SOFTMAX_NODE_H\n\n//for multigpu training\ntemplate<typename dType>\nstruct softmax_node {\n\n\t"
  },
  {
    "path": "src/transfer_layer.h",
    "chars": 1635,
    "preview": "#ifndef TRANSFER_LAYER_H\n#define TRANSFER_LAYER_H\n\n#include \"base_loss.h\"\n\ntemplate<typename dType>\nclass softmax_layer;"
  },
  {
    "path": "src/tree_LSTM.h",
    "chars": 2507,
    "preview": "//tree LSTM for 2 children\n#ifndef TREE_LSTM_H\n#define TREE_LSTM_H\n\n\ntemplate<typename dType>\nclass encoder_multi_source"
  },
  {
    "path": "src/tree_LSTM.hpp",
    "chars": 40180,
    "preview": "//tree LSTM implementation\n\n\n\ntemplate<typename dType>\ntree_LSTM<dType>::tree_LSTM(global_params &params,int device_numb"
  },
  {
    "path": "src/trunc_softmax.h",
    "chars": 2865,
    "preview": "#ifndef TRUNC_SOFTMAX_H\n#define TRUNC_SOFTMAX_H\n\n\n\n\ntemplate<typename dType>\nclass trunc_softmax {\npublic:\n\n\tint output_"
  }
]

// ... and 5 more files (download for full content)

About this extraction

This page contains the full source code of the isi-nlp/Zoph_RNN GitHub repository, extracted and formatted as plain text for AI agents and large language models (LLMs). The extraction includes 119 files (125.4 MB), approximately 3.0M tokens, and a symbol index with 528 extracted functions, classes, methods, constants, and types. Use this with OpenClaw, Claude, ChatGPT, Cursor, Windsurf, or any other AI tool that accepts text input. You can copy the full output to your clipboard or download it as a .txt file.

Extracted by GitExtract — free GitHub repo to text converter for AI. Built by Nikandr Surkov.

Extract another repo