Full Code of mattriemer/MER for AI

master 4b7b9c977f31 cached
22 files
88.6 KB
22.2k tokens
80 symbols
1 requests
Download .txt
Repository: mattriemer/MER
Branch: master
Commit: 4b7b9c977f31
Files: 22
Total size: 88.6 KB

Directory structure:
gitextract_jqq8_utk/

├── LICENSE
├── MIT-LICENSE
├── README.md
├── data/
│   └── README.md
├── download.py
├── get_data.py
├── main.py
├── mer_examples.sh
├── metrics/
│   └── metrics.py
├── model/
│   ├── common.py
│   ├── eralg4.py
│   ├── eralg5.py
│   ├── ewc.py
│   ├── gem.py
│   ├── independent.py
│   ├── meralg1.py
│   ├── meralg6.py
│   ├── meralg7.py
│   ├── online.py
│   └── taskinput.py
├── requirements.txt
└── unit_test.sh

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

================================================
FILE: LICENSE
================================================
                                 Apache License
                           Version 2.0, January 2004
                        http://www.apache.org/licenses/

   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION

   1. Definitions.

      "License" shall mean the terms and conditions for use, reproduction,
      and distribution as defined by Sections 1 through 9 of this document.

      "Licensor" shall mean the copyright owner or entity authorized by
      the copyright owner that is granting the License.

      "Legal Entity" shall mean the union of the acting entity and all
      other entities that control, are controlled by, or are under common
      control with that entity. For the purposes of this definition,
      "control" means (i) the power, direct or indirect, to cause the
      direction or management of such entity, whether by contract or
      otherwise, or (ii) ownership of fifty percent (50%) or more of the
      outstanding shares, or (iii) beneficial ownership of such entity.

      "You" (or "Your") shall mean an individual or Legal Entity
      exercising permissions granted by this License.

      "Source" form shall mean the preferred form for making modifications,
      including but not limited to software source code, documentation
      source, and configuration files.

      "Object" form shall mean any form resulting from mechanical
      transformation or translation of a Source form, including but
      not limited to compiled object code, generated documentation,
      and conversions to other media types.

      "Work" shall mean the work of authorship, whether in Source or
      Object form, made available under the License, as indicated by a
      copyright notice that is included in or attached to the work
      (an example is provided in the Appendix below).

      "Derivative Works" shall mean any work, whether in Source or Object
      form, that is based on (or derived from) the Work and for which the
      editorial revisions, annotations, elaborations, or other modifications
      represent, as a whole, an original work of authorship. For the purposes
      of this License, Derivative Works shall not include works that remain
      separable from, or merely link (or bind by name) to the interfaces of,
      the Work and Derivative Works thereof.

      "Contribution" shall mean any work of authorship, including
      the original version of the Work and any modifications or additions
      to that Work or Derivative Works thereof, that is intentionally
      submitted to Licensor for inclusion in the Work by the copyright owner
      or by an individual or Legal Entity authorized to submit on behalf of
      the copyright owner. For the purposes of this definition, "submitted"
      means any form of electronic, verbal, or written communication sent
      to the Licensor or its representatives, including but not limited to
      communication on electronic mailing lists, source code control systems,
      and issue tracking systems that are managed by, or on behalf of, the
      Licensor for the purpose of discussing and improving the Work, but
      excluding communication that is conspicuously marked or otherwise
      designated in writing by the copyright owner as "Not a Contribution."

      "Contributor" shall mean Licensor and any individual or Legal Entity
      on behalf of whom a Contribution has been received by Licensor and
      subsequently incorporated within the Work.

   2. Grant of Copyright License. Subject to the terms and conditions of
      this License, each Contributor hereby grants to You a perpetual,
      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
      copyright license to reproduce, prepare Derivative Works of,
      publicly display, publicly perform, sublicense, and distribute the
      Work and such Derivative Works in Source or Object form.

   3. Grant of Patent License. Subject to the terms and conditions of
      this License, each Contributor hereby grants to You a perpetual,
      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
      (except as stated in this section) patent license to make, have made,
      use, offer to sell, sell, import, and otherwise transfer the Work,
      where such license applies only to those patent claims licensable
      by such Contributor that are necessarily infringed by their
      Contribution(s) alone or by combination of their Contribution(s)
      with the Work to which such Contribution(s) was submitted. If You
      institute patent litigation against any entity (including a
      cross-claim or counterclaim in a lawsuit) alleging that the Work
      or a Contribution incorporated within the Work constitutes direct
      or contributory patent infringement, then any patent licenses
      granted to You under this License for that Work shall terminate
      as of the date such litigation is filed.

   4. Redistribution. You may reproduce and distribute copies of the
      Work or Derivative Works thereof in any medium, with or without
      modifications, and in Source or Object form, provided that You
      meet the following conditions:

      (a) You must give any other recipients of the Work or
          Derivative Works a copy of this License; and

      (b) You must cause any modified files to carry prominent notices
          stating that You changed the files; and

      (c) You must retain, in the Source form of any Derivative Works
          that You distribute, all copyright, patent, trademark, and
          attribution notices from the Source form of the Work,
          excluding those notices that do not pertain to any part of
          the Derivative Works; and

      (d) If the Work includes a "NOTICE" text file as part of its
          distribution, then any Derivative Works that You distribute must
          include a readable copy of the attribution notices contained
          within such NOTICE file, excluding those notices that do not
          pertain to any part of the Derivative Works, in at least one
          of the following places: within a NOTICE text file distributed
          as part of the Derivative Works; within the Source form or
          documentation, if provided along with the Derivative Works; or,
          within a display generated by the Derivative Works, if and
          wherever such third-party notices normally appear. The contents
          of the NOTICE file are for informational purposes only and
          do not modify the License. You may add Your own attribution
          notices within Derivative Works that You distribute, alongside
          or as an addendum to the NOTICE text from the Work, provided
          that such additional attribution notices cannot be construed
          as modifying the License.

      You may add Your own copyright statement to Your modifications and
      may provide additional or different license terms and conditions
      for use, reproduction, or distribution of Your modifications, or
      for any such Derivative Works as a whole, provided Your use,
      reproduction, and distribution of the Work otherwise complies with
      the conditions stated in this License.

   5. Submission of Contributions. Unless You explicitly state otherwise,
      any Contribution intentionally submitted for inclusion in the Work
      by You to the Licensor shall be under the terms and conditions of
      this License, without any additional terms or conditions.
      Notwithstanding the above, nothing herein shall supersede or modify
      the terms of any separate license agreement you may have executed
      with Licensor regarding such Contributions.

   6. Trademarks. This License does not grant permission to use the trade
      names, trademarks, service marks, or product names of the Licensor,
      except as required for reasonable and customary use in describing the
      origin of the Work and reproducing the content of the NOTICE file.

   7. Disclaimer of Warranty. Unless required by applicable law or
      agreed to in writing, Licensor provides the Work (and each
      Contributor provides its Contributions) on an "AS IS" BASIS,
      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
      implied, including, without limitation, any warranties or conditions
      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
      PARTICULAR PURPOSE. You are solely responsible for determining the
      appropriateness of using or redistributing the Work and assume any
      risks associated with Your exercise of permissions under this License.

   8. Limitation of Liability. In no event and under no legal theory,
      whether in tort (including negligence), contract, or otherwise,
      unless required by applicable law (such as deliberate and grossly
      negligent acts) or agreed to in writing, shall any Contributor be
      liable to You for damages, including any direct, indirect, special,
      incidental, or consequential damages of any character arising as a
      result of this License or out of the use or inability to use the
      Work (including but not limited to damages for loss of goodwill,
      work stoppage, computer failure or malfunction, or any and all
      other commercial damages or losses), even if such Contributor
      has been advised of the possibility of such damages.

   9. Accepting Warranty or Additional Liability. While redistributing
      the Work or Derivative Works thereof, You may choose to offer,
      and charge a fee for, acceptance of support, warranty, indemnity,
      or other liability obligations and/or rights consistent with this
      License. However, in accepting such obligations, You may act only
      on Your own behalf and on Your sole responsibility, not on behalf
      of any other Contributor, and only if You agree to indemnify,
      defend, and hold each Contributor harmless for any liability
      incurred by, or claims asserted against, such Contributor by reason
      of your accepting any such warranty or additional liability.

   END OF TERMS AND CONDITIONS

   APPENDIX: How to apply the Apache License to your work.

      To apply the Apache License to your work, attach the following
      boilerplate notice, with the fields enclosed by brackets "[]"
      replaced with your own identifying information. (Don't include
      the brackets!)  The text should be enclosed in the appropriate
      comment syntax for the file format. We also recommend that a
      file or class name and description of purpose be included on the
      same "printed page" as the copyright notice for easier
      identification within third-party archives.

   Copyright [yyyy] [name of copyright owner]

   Licensed under the Apache License, Version 2.0 (the "License");
   you may not use this file except in compliance with the License.
   You may obtain a copy of the License at

       http://www.apache.org/licenses/LICENSE-2.0

   Unless required by applicable law or agreed to in writing, software
   distributed under the License is distributed on an "AS IS" BASIS,
   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
   See the License for the specific language governing permissions and
   limitations under the License.


================================================
FILE: MIT-LICENSE
================================================
Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.


================================================
FILE: README.md
================================================
# Meta-Experience Replay (MER)

Source code for the paper "Learning to Learn without Forgetting by Maximizing Transfer and Minimizing Interference".

Link: https://openreview.net/pdf?id=B1gTShAct7

Reference:
```
@inproceedings{MER,
    title={Learning to Learn without Forgetting by Maximizing Transfer and Minimizing Interference},   
    author={Riemer, Matthew and Cases, Ignacio and Ajemian, Robert and Liu, Miao and Rish, Irina and Tu, Yuhai and Tesauro, Gerald},    
    booktitle={In International Conference on Learning Representations (ICLR)},    
    year={2019}   
}
```

This project is a fork of the GEM project https://github.com/facebookresearch/GradientEpisodicMemory in order to reproduce baselines from their paper. These baselines have been copied into the model/ directory of this repository. Our output and logging mechanisms for all models follow the same format used in the GEM project. 

# Available Datasets

The code in this repository should work on the variants of MNIST used in the experiments. This includes Rotations, Permutations, and Many Permutations. It would need to be slightly extended to be applied to other interesting continual learning benchmarks like CIFAR-100 or Omniglot.

The original MNIST database is available at http://yann.lecun.com/exdb/mnist/ and interface for generating your own MNIST variants is provided as part of the GEM project https://github.com/facebookresearch/GradientEpisodicMemory/tree/master/data. To maximize reproducibility, we have provided an interface for directly downloading the dataset versions used in our experiments.

## Basic Setup (Python 3.5 & torch 0.3.1)

As a first step to get up and running, clone this git repository and navigate into the root directory of your local version of the repository. To get started, please install the requirements inside your environment.

If you don't have an environment, we recommend that you create one (using [conda](http://anaconda.org)). The following instructions will guide you:

Install `conda` and type

```conda create --name mer python=3.5```

This will create a conda environment (an isolated workplace) in which we can install the right versions of the software. Then, activate the environment:

```source activate mer```

or

```conda activate mer```

Within the `mer` environment, install PyTorch and Cython using conda as follows:

```conda install pytorch=0.3.1 -c pytorch```

```conda install cython```

and then install the rest of the requirements using the following command:

```pip install --user -r requirements.txt```


## Basic Setup (Python 3.6+ & torch 1.4+)

As a first step to get up and running, clone this git repository and navigate into the root directory of your local version of the repository. To get started, please install the requirements inside your environment.

If you don't have an environment, we recommend that you create one (using [conda](http://anaconda.org)). The following instructions will guide you:

Install `conda` and type

```conda create --name mer python```

This will create a conda environment (an isolated workplace) in which we can install the right versions of the software. Then, activate the environment:

```source activate mer```

or

```conda activate mer```

Within the `mer` environment, install PyTorch and Cython using conda as follows:

```conda install torch```

```conda install cython```

_For python 3.6+, to install quadprog, first do:_
```sudo apt install gcc build-essential```

and then install the rest of the requirements using the following command:

```pip install --user -r requirements.txt```

## Getting the datasets

The first step is to download and uncompress all three datasets (30 GB of storage needed) execute the following command:

```python get_data.py all```

For just MNIST Rotations (4.1 GB) execute:

```python get_data.py rotations```

For just MNIST Permutations (4.1 GB) execute:

```python get_data.py permutations```

For just MNIST Many Permutations (21 GB) execute:

```python get_data.py manypermutations```

# Getting Started

In mer_examples.sh see examples of how to run variants of MER from the paper and baseline models from the experiments. We make sure first that this script is excutable:

```chmod +x mer_examples.sh```

Now, executing the following command leads to running a full suite of experiments with a random seed of 0:

```
./mer_examples.sh "0"
```

Within the file you can see examples of how to run models from our experiments on each datasets and memory size setting. For instance, we can execute MER from Algorithm 1 in the paper (meralg1) on MNIST Rotations with 5120 memories using the following commands:
```
export ROT="--n_layers 2 --n_hiddens 100 --data_path data/ --save_path results/ --batch_size 1 --log_every 100 --samples_per_task 1000 --data_file mnist_rotations.pt --cuda no --seed 0"
python3 main.py $ROT --model meralg1 --lr 0.03 --beta 0.03 --gamma 1.0 --memories 5120 --replay_batch_size 100 --batches_per_example 10
```

# Available Models

In the model/ directory we have provided various models that were used for experiments on varients of MNIST in the paper. These models include:

- Online Learning (online)

- Task Specific Input Layer (taskinput)

- Independent Models per Task (independent)

- Elastic Weight Consolidation (ewc)

- Gradient Episodic Memory (gem)

- Experience Replay Algorithm 4 (eralg4)

- Experience Replay Algorithm 5 (eralg5)

- Meta-Experience Replay Algorithm 1 (meralg1)

- Meta-Experience Replay Algorithm 6 (meralg6)

- Meta-Experience Replay Algorithm 7 (meralg7)

# System Requirements

The repository has been developed for Python 3.5.2 using PyTorch 0.3.1.

# Reproducing Our Experiments 

We have conducted comprehensive experiments detailed in Appendix M of our paper to make sure that our results are reproducible across runs regardless of the machine and random seed. You should be able to reproduce these experiments using the provided mer_examples.sh script.


================================================
FILE: data/README.md
================================================
This is the directory where data should be stored in order to use main.py for executing continual learning models. If you have not already, populate this folder using the get_data.py file in the root directory.


================================================
FILE: download.py
================================================
########################################################################
#
# Functions for downloading and extracting data-files from the internet.
#
# Implemented in Python 3.5
#
########################################################################
#
# This file is part of the TensorFlow Tutorials available at:
#
# https://github.com/Hvass-Labs/TensorFlow-Tutorials
#
# Published under the MIT License. See the file LICENSE for details.
#
# Copyright 2016 by Magnus Erik Hvass Pedersen
#
########################################################################

import sys
import os
import urllib.request
import tarfile
import zipfile

########################################################################


def _print_download_progress(count, block_size, total_size):
    """
    Function used for printing the download progress.
    Used as a call-back function in maybe_download_and_extract().
    """

    # Percentage completion.
    pct_complete = float(count * block_size) / total_size

    # Limit it because rounding errors may cause it to exceed 100%.
    pct_complete = min(1.0, pct_complete)

    # Status-message. Note the \r which means the line should overwrite itself.
    msg = "\r- Download progress: {0:.1%}".format(pct_complete)

    # Print it.
    sys.stdout.write(msg)
    sys.stdout.flush()


########################################################################

def download(base_url, filename, download_dir):
    """
    Download the given file if it does not already exist in the download_dir.

    :param base_url: The internet URL without the filename.
    :param filename: The filename that will be added to the base_url.
    :param download_dir: Local directory for storing the file.
    :return: Nothing.
    """

    # Path for local file.
    save_path = os.path.join(download_dir, filename)

    # Check if the file already exists, otherwise we need to download it now.
    if not os.path.exists(save_path):
        # Check if the download directory exists, otherwise create it.
        if not os.path.exists(download_dir):
            os.makedirs(download_dir)

        print("Downloading", filename, "...")

        # Download the file from the internet.
        url = base_url + filename
        file_path, _ = urllib.request.urlretrieve(url=url,
                                                  filename=save_path,
                                                  reporthook=_print_download_progress)

        print(" Done!")


def maybe_download_and_extract(url, download_dir):
    """
    Download and extract the data if it doesn't already exist.
    Assumes the url is a tar-ball file.

    :param url:
        Internet URL for the tar-file to download.
        Example: "https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz"

    :param download_dir:
        Directory where the downloaded file is saved.
        Example: "data/CIFAR-10/"

    :return:
        Nothing.
    """

    # Filename for saving the file downloaded from the internet.
    # Use the filename from the URL and add it to the download_dir.
    filename = url.split('/')[-1]
    file_path = os.path.join(download_dir, filename)

    # Check if the file already exists.
    # If it exists then we assume it has also been extracted,
    # otherwise we need to download and extract it now.
    if not os.path.exists(file_path):
        # Check if the download directory exists, otherwise create it.
        if not os.path.exists(download_dir):
            os.makedirs(download_dir)

        # Download the file from the internet.
        file_path, _ = urllib.request.urlretrieve(url=url,
                                                  filename=file_path,
                                                  reporthook=_print_download_progress)

        print()
        print("Download finished. Extracting files.")

        if file_path.endswith(".zip"):
            # Unpack the zip-file.
            zipfile.ZipFile(file=file_path, mode="r").extractall(download_dir)
        elif file_path.endswith((".tar.gz", ".tgz")):
            # Unpack the tar-ball.
            tarfile.open(name=file_path, mode="r:gz").extractall(download_dir)

        print("Done.")
    else:
        print("Data has apparently already been downloaded and unpacked.")


########################################################################


================================================
FILE: get_data.py
================================================
# Copyright 2019-present, IBM Research
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.

import os
import download
import argparse

def get_mnist_data(url, data_dir):
    print("Downloading {} into {}".format(url, data_dir))
    download.maybe_download_and_extract(url, data_dir)

def get_datasets():
    parser = argparse.ArgumentParser()
    parser.add_argument("dataset", help="Either the name of the dataset (rotations, permutations, manypermutations), or `all` to download all datasets")
    args = parser.parse_args()

    # Change dir to the location of this file (repo's root)
    get_data_path = os.path.realpath(__file__)
    os.chdir(os.path.dirname(get_data_path))
    data_dir = os.path.join(os.getcwd(), 'data')

    # get files
    mnist_rotations = "https://nlp.stanford.edu/data/mer/mnist_rotations.tar.gz"
    mnist_permutations = "https://nlp.stanford.edu/data/mer/mnist_permutations.tar.gz"
    mnist_many = "https://nlp.stanford.edu/data/mer/mnist_manypermutations.tar.gz"

    all = {"rotations": mnist_rotations, "permutations": mnist_permutations, "manypermutations": mnist_many}

    if args.dataset == "all":
        for dataset in all.values():
            get_mnist_data(dataset, data_dir)
    else:
        get_mnist_data(all[args.dataset], data_dir)

if __name__ == "__main__":
    get_datasets()


================================================
FILE: main.py
================================================
### This is a modified version of main.py from https://github.com/facebookresearch/GradientEpisodicMemory 
### The most significant changes are to the arguments: 1) allowing for new models and 2) changing the default settings in some cases

# Copyright 2019-present, IBM Research
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.


import importlib
import datetime
import argparse
import random
import uuid
import time
import os

import numpy as np

import torch
from torch.autograd import Variable
from metrics.metrics import confusion_matrix

# continuum iterator #########################################################


def load_datasets(args):
    d_tr, d_te = torch.load(args.data_path + '/' + args.data_file)
    n_inputs = d_tr[0][1].size(1)
    n_outputs = 0
    for i in range(len(d_tr)):
        n_outputs = max(n_outputs, d_tr[i][2].max())
        n_outputs = max(n_outputs, d_te[i][2].max())
    return d_tr, d_te, n_inputs, n_outputs + 1, len(d_tr)


class Continuum:

    def __init__(self, data, args):
        self.data = data
        self.batch_size = args.batch_size
        n_tasks = len(data)
        task_permutation = range(n_tasks)

        if args.shuffle_tasks == 'yes':
            task_permutation = torch.randperm(n_tasks).tolist()

        sample_permutations = []

        for t in range(n_tasks):
            N = data[t][1].size(0)
            if args.samples_per_task <= 0:
                n = N
            else:
                n = min(args.samples_per_task, N)

            p = torch.randperm(N)[0:n]
            sample_permutations.append(p)

        self.permutation = []

        for t in range(n_tasks):
            task_t = task_permutation[t]
            for _ in range(args.n_epochs):
                task_p = [[task_t, i] for i in sample_permutations[task_t]]
                random.shuffle(task_p)
                self.permutation += task_p

        self.length = len(self.permutation)
        self.current = 0

    def __iter__(self):
        return self

    def next(self):
        return self.__next__()

    def __next__(self):
        if self.current >= self.length:
            raise StopIteration
        else:
            ti = self.permutation[self.current][0]
            j = []
            i = 0
            while (((self.current + i) < self.length) and
                   (self.permutation[self.current + i][0] == ti) and
                   (i < self.batch_size)):
                j.append(self.permutation[self.current + i][1])
                i += 1
            self.current += i
            j = torch.LongTensor(j)
            return self.data[ti][1][j], ti, self.data[ti][2][j]

# train handle ###############################################################


def eval_tasks(model, tasks, args):
    model.eval()
    result = []
    for i, task in enumerate(tasks):
        t = i
        x = task[1]
        y = task[2]
        rt = 0
        
        eval_bs = x.size(0)

        with torch.no_grad():  # torch 0.4+
            for b_from in range(0, x.size(0), eval_bs):
                b_to = min(b_from + eval_bs, x.size(0) - 1)
                if b_from == b_to:
                    xb = x[b_from].view(1, -1)
                    yb = torch.LongTensor([y[b_to]]).view(1, -1)
                else:
                    xb = x[b_from:b_to]
                    yb = y[b_from:b_to]
                if args.cuda:
                    xb = xb.cuda()
                # xb = Variable(xb, volatile=True)  # torch 0.4+
                _, pb = torch.max(model(xb, t).data.cpu(), 1, keepdim=False)
                rt += (pb == yb).float().sum()

        result.append(rt / x.size(0))

    return result


def life_experience(model, continuum, x_te, args):
    result_a = []
    result_t = []

    current_task = 0
    time_start = time.time()

    for (i, (x, t, y)) in enumerate(continuum):
        if(((i % args.log_every) == 0) or (t != current_task)):
            result_a.append(eval_tasks(model, x_te, args))
            result_t.append(current_task)
            current_task = t

        v_x = x.view(x.size(0), -1)
        v_y = y.long()

        if args.cuda:
            v_x = v_x.cuda()
            v_y = v_y.cuda()

        model.train()
        model.observe(Variable(v_x), t, Variable(v_y))

    result_a.append(eval_tasks(model, x_te, args))
    result_t.append(current_task)

    time_end = time.time()
    time_spent = time_end - time_start

    return torch.Tensor(result_t), torch.Tensor(result_a), time_spent


if __name__ == "__main__":
    parser = argparse.ArgumentParser(description='Continuum learning')

    # model details
    parser.add_argument('--model', type=str, default='single',
                        help='model to train')
    parser.add_argument('--n_hiddens', type=int, default=100,
                        help='number of hidden neurons at each layer')
    parser.add_argument('--n_layers', type=int, default=2,
                        help='number of hidden layers')
    parser.add_argument('--finetune', default='yes', type=str,help='whether to initialize nets in indep. nets')
    
    # optimizer parameters influencing all models
    parser.add_argument('--n_epochs', type=int, default=1,
                        help='Number of epochs per task')
    parser.add_argument('--batch_size', type=int, default=1,
                        help='the amount of items received by the algorithm at one time (set to 1 across all experiments). Variable name is from GEM project.')
    parser.add_argument('--lr', type=float, default=1e-3,
                        help='learning rate')

    # memory parameters for GEM baselines
    parser.add_argument('--n_memories', type=int, default=0,
                        help='number of memories per task')
    parser.add_argument('--memory_strength', default=0, type=float,
                        help='memory strength (meaning depends on memory)')

    # parameters specific to models in https://openreview.net/pdf?id=B1gTShAct7 
    
    parser.add_argument('--memories', type=int, default=5120, help='number of total memories stored in a reservoir sampling based buffer')

    parser.add_argument('--gamma', type=float, default=1.0,
                        help='gamma learning rate parameter') #gating net lr in roe 

    parser.add_argument('--batches_per_example', type=float, default=1,
                        help='the number of batch per incoming example')

    parser.add_argument('--s', type=float, default=1,
                        help='current example learning rate multiplier (s)')

    parser.add_argument('--replay_batch_size', type=float, default=20,
                        help='The batch size for experience replay. Denoted as k-1 in the paper.')

    parser.add_argument('--beta', type=float, default=1.0,
                        help='beta learning rate parameter') # exploration factor in roe
    
    # experiment parameters
    parser.add_argument('--cuda', type=str, default='no',
                        help='Use GPU?')
    parser.add_argument('--seed', type=int, default=0,
                        help='random seed of model')
    parser.add_argument('--log_every', type=int, default=100,
                        help='frequency of logs, in minibatches')
    parser.add_argument('--save_path', type=str, default='results/',
                        help='save models at the end of training')

    # data parameters
    parser.add_argument('--data_path', default='data/',
                        help='path where data is located')
    parser.add_argument('--data_file', default='mnist_permutations.pt',
                        help='data file')
    parser.add_argument('--samples_per_task', type=int, default=-1,
                        help='training samples per task (all if negative)')
    parser.add_argument('--shuffle_tasks', type=str, default='no',
                        help='present tasks in order')
    args = parser.parse_args()

    args.cuda = True if args.cuda == 'yes' else False
    args.finetune = True if args.finetune == 'yes' else False

    # taskinput model has one extra layer
    if args.model == 'taskinput':
        args.n_layers -= 1

    # unique identifier
    uid = uuid.uuid4().hex

    # initialize seeds    
    torch.backends.cudnn.enabled = False
    torch.manual_seed(args.seed)
    np.random.seed(args.seed)
    random.seed(args.seed)
    if args.cuda:
        print("Found GPU:", torch.cuda.get_device_name(0))
        torch.cuda.manual_seed_all(args.seed)

    # load data
    x_tr, x_te, n_inputs, n_outputs, n_tasks = load_datasets(args)
    n_outputs = n_outputs.item()  # outputs should not be a tensor, otherwise "TypeError: expected Float (got Long)"

    # set up continuum
    continuum = Continuum(x_tr, args)

    # load model
    Model = importlib.import_module('model.' + args.model)
    model = Model.Net(n_inputs, n_outputs, n_tasks, args)
    if args.cuda:
        try:
            model.cuda()
        except:
            pass 

    # run model on continuum
    result_t, result_a, spent_time = life_experience(
        model, continuum, x_te, args)

    # prepare saving path and file name
    if not os.path.exists(args.save_path):
        os.makedirs(args.save_path)

    fname = args.model + '_' + args.data_file + '_'
    fname += datetime.datetime.now().strftime("%Y_%m_%d_%H_%M_%S")
    fname += '_' + uid
    fname = os.path.join(args.save_path, fname)

    # save confusion matrix and print one line of stats
    stats = confusion_matrix(result_t, result_a, fname + '.txt')
    one_liner = str(vars(args)) + ' # '
    one_liner += ' '.join(["%.3f" % stat for stat in stats])
    print(fname + ': ' + one_liner + ' # ' + str(spent_time))

    # save all results in binary file
    torch.save((result_t, result_a, model.state_dict(),
                stats, one_liner, args), fname + '.pt')


================================================
FILE: mer_examples.sh
================================================
#!/bin/bash

seed=$1
ROT="--n_layers 2 --n_hiddens 100 --data_path data/ --save_path results/ --batch_size 1 --log_every 100 --samples_per_task 1000 --data_file mnist_rotations.pt    --cuda no  --seed"
PERM="--n_layers 2 --n_hiddens 100 --data_path data/ --save_path results/ --batch_size 1 --log_every 100 --samples_per_task 1000 --data_file mnist_permutations.pt --cuda no  --seed"
MANY="--n_layers 2 --n_hiddens 100 --data_path data/ --save_path results/ --batch_size 1 --log_every 100 --samples_per_task 200 --data_file mnist_manypermutations.pt --cuda no  --seed"

echo "Beginning Online Learning" "( seed =" $seed ")"
echo "MNIST Rotations:"
python3 main.py $ROT $seed --model online --lr 0.0003
echo "MNIST Permutations:"
python3 main.py $PERM $seed --model online --lr 0.003
echo "MNIST Many Permutations:"
python3 main.py $MANY $seed --model online --lr 0.003

echo "Beginning Independent Model Per Task" "( seed =" $seed ")"
echo "MNIST Rotations:"
python3 main.py $ROT $seed --model independent --lr 0.01
echo "MNIST Permutations:"
python3 main.py $PERM $seed --model independent --lr 0.01
echo "MNIST Many Permutations:"
python3 main.py $MANY $seed --model independent --lr 0.01

echo "Beginning EWC" "( seed =" $seed ")"
echo "MNIST Rotations:"
python3 main.py $ROT $seed --model ewc --lr 0.001 --n_memories 10 --memory_strength 100.0
echo "MNIST Permutations:"
python3 main.py $PERM $seed --model ewc --lr 0.01 --n_memories 10 --memory_strength 10.0
echo "MNIST Many Permutations:"
python3 main.py $MANY $seed --model ewc --lr 0.003 --n_memories 10 --memory_strength 1.0

echo "Beginning GEM With 5120 Memories" "( seed =" $seed ")"
echo "MNIST Rotations:"
python3 main.py $ROT $seed --model gem --lr 0.01 --n_memories 256 --memory_strength 1.0
echo "MNIST Permutations:"
python3 main.py $PERM $seed --model gem --lr 0.01 --n_memories 256 --memory_strength 1.0
echo "MNIST Many Permutations:"
python3 main.py $MANY $seed --model gem --lr 0.01 --n_memories 51 --memory_strength 0.0

echo "Beginning GEM With 500 Memories" "( seed =" $seed ")"
echo "MNIST Rotations:"
python3 main.py $ROT $seed --model gem --lr 0.01 --n_memories 25 --memory_strength 0.0
echo "MNIST Permutations:"
python3 main.py $PERM $seed --model gem --lr 0.01 --n_memories 25 --memory_strength 1.0
echo "MNIST Many Permutations:"
python3 main.py $MANY $seed --model gem --lr 0.003 --n_memories 5 --memory_strength 0.1

echo "Beginning GEM With 200 Memories" "( seed =" $seed ")"
echo "MNIST Rotations:"
python3 main.py $ROT $seed --model gem --lr 0.01 --n_memories 10 --memory_strength 0.0
echo "MNIST Permutations:"
python3 main.py $PERM $seed --model gem --lr 0.01 --n_memories 10 --memory_strength 0.0

echo "Beginning ER (Algorithm 4) With 5120 Memories" "( seed =" $seed ")"
echo "MNIST Rotations:"
python3 main.py $ROT $seed --model eralg4 --lr 0.1 --memories 5120 --replay_batch_size 25
echo "MNIST Permutations:"
python3 main.py $PERM $seed --model eralg4 --lr 0.1 --memories 5120 --replay_batch_size 25
echo "MNIST Many Permutations:"
python3 main.py $MANY $seed --model eralg4 --lr 0.1 --memories 5120 --replay_batch_size 25

echo "Beginning ER (Algorithm 4) With 500 Memories" "( seed =" $seed ")"
echo "MNIST Rotations:"
python3 main.py $ROT $seed --model eralg4 --lr 0.1 --memories 500 --replay_batch_size 5
echo "MNIST Permutations:"
python3 main.py $PERM $seed --model eralg4 --lr 0.1 --memories 500 --replay_batch_size 10
echo "MNIST Many Permutations:"
python3 main.py $MANY $seed --model eralg4 --lr 0.1 --memories 500 --replay_batch_size 25

echo "Beginning ER (Algorithm 4) With 200 Memories" "( seed =" $seed ")"
echo "MNIST Rotations:"
python3 main.py $ROT $seed --model eralg4 --lr 0.1 --memories 200 --replay_batch_size 10
echo "MNIST Permutations:"
python3 main.py $PERM $seed --model eralg4 --lr 0.1 --memories 200 --replay_batch_size 10  

echo "Beginning ER (Algorithm 5) With 5120 Memories" "( seed =" $seed ")"
echo "MNIST Rotations:"
python3 main.py $ROT $seed --model eralg5 --lr 0.03 --memories 5120 --replay_batch_size 100
echo "MNIST Permutations:"
python3 main.py $PERM $seed --model eralg5 --lr 0.01 --memories 5120 --replay_batch_size 25
echo "MNIST Many Permutations:"
python3 main.py $MANY $seed --model eralg5 --lr 0.003 --memories 5120 --replay_batch_size 10

echo "Beginning ER (Algorithm 5) With 500 Memories" "( seed =" $seed ")"
echo "MNIST Rotations:"
python3 main.py $ROT $seed --model eralg5 --lr 0.01 --memories 500 --replay_batch_size 100
echo "MNIST Permutations:"
python3 main.py $PERM $seed --model eralg5 --lr 0.01 --memories 500 --replay_batch_size 25
echo "MNIST Many Permutations:"
python3 main.py $MANY $seed --model eralg5 --lr 0.01 --memories 500 --replay_batch_size 5

echo "Beginning ER (Algorithm 5) With 200 Memories" "( seed =" $seed ")"
echo "MNIST Rotations:"
python3 main.py $ROT $seed --model eralg5 --lr 0.01 --memories 200 --replay_batch_size 50
echo "MNIST Permutations:"
python3 main.py $PERM $seed --model eralg5 --lr 0.01 --memories 200 --replay_batch_size 10

echo "Beginning MER (Algorithm 1) With 5120 Memories" "( seed =" $seed ")"
echo "MNIST Rotations:"
python3 main.py $ROT $seed --model meralg1 --lr 0.03 --beta 0.03 --gamma 1.0 --memories 5120 --replay_batch_size 100 --batches_per_example 10
echo "MNIST Permutations:"
python3 main.py $PERM $seed --model meralg1 --lr 0.03 --beta 0.03 --gamma 1.0 --memories 5120 --replay_batch_size 100 --batches_per_example 10
echo "MNIST Many Permutations:"
python3 main.py $MANY $seed --model meralg1 --lr 0.1 --beta 0.01 --gamma 1.0 --memories 5120 --replay_batch_size 5 --batches_per_example 10

echo "Beginning MER (Algorithm 1) With 500 Memories" "( seed =" $seed ")"
echo "MNIST Rotations:"
python3 main.py $ROT $seed --model meralg1 --lr 0.1 --beta 0.01 --gamma 1.0 --memories 500 --replay_batch_size 10 --batches_per_example 10
echo "MNIST Permutations:"
python3 main.py $PERM $seed --model meralg1 --lr 0.03 --beta 0.03 --gamma 1.0 --memories 500 --replay_batch_size 25 --batches_per_example 10
echo "MNIST Many Permutations:"
python3 main.py $MANY $seed --model meralg1 --lr 0.03 --beta 0.03 --gamma 1.0 --memories 500 --replay_batch_size 5 --batches_per_example 10

echo "Beginning MER (Algorithm 1) With 200 Memories" "( seed =" $seed ")"
echo "MNIST Rotations:"
python3 main.py $ROT $seed --model meralg1 --lr 0.1 --beta 0.01 --gamma 1.0 --memories 200 --replay_batch_size 10 --batches_per_example 5
echo "MNIST Permutations:"
python3 main.py $PERM $seed --model meralg1 --lr 0.03 --beta 0.03 --gamma 1.0 --memories 200 --replay_batch_size 10 --batches_per_example 10

echo "Beginning MER (Algorithm 6) With 5120 Memories" "( seed =" $seed ")"
echo "MNIST Rotations:"
python3 main.py $ROT $seed --model meralg6 --lr 0.03 --gamma 0.1 --memories 5120 --replay_batch_size 100
echo "MNIST Permutations:"
python3 main.py $PERM $seed --model meralg6 --lr 0.03 --gamma 0.1 --memories 5120 --replay_batch_size 50
echo "MNIST Many Permutations:"
python3 main.py $MANY $seed --model meralg6 --lr 0.03 --gamma 0.1 --memories 5120 --replay_batch_size 25

echo "Beginning MER (Algorithm 6) With 500 Memories" "( seed =" $seed ")"
echo "MNIST Rotations:"
python3 main.py $ROT $seed --model meralg6 --lr 0.1 --gamma 0.03 --memories 500 --replay_batch_size 10
echo "MNIST Permutations:"
python3 main.py $PERM $seed --model meralg6 --lr 0.03 --gamma 0.3 --memories 500 --replay_batch_size 10
echo "MNIST Many Permutations:"
python3 main.py $MANY $seed --model meralg6 --lr 0.1 --gamma 0.03 --memories 500 --replay_batch_size 5

echo "Beginning MER (Algorithm 6) With 200 Memories" "( seed =" $seed ")"
echo "MNIST Rotations:"
python3 main.py $ROT $seed --model meralg6 --lr 0.1 --gamma 0.03 --memories 200 --replay_batch_size 25
echo "MNIST Permutations:"
python3 main.py $PERM $seed --model meralg6 --lr 0.1 --gamma 0.03 --memories 200 --replay_batch_size 5


echo "Beginning MER (Algorithm 7) With 5120 Memories" "( seed =" $seed ")"
echo "MNIST Rotations:"
python3 main.py $ROT $seed --model meralg7 --lr 0.03 --gamma 0.03 --memories 5120 --replay_batch_size 100 --s 5
echo "MNIST Permutations:"
python3 main.py $PERM $seed --model meralg7 --lr 0.01 --gamma 0.1 --memories 5120 --replay_batch_size 100 --s 10
echo "MNIST Many Permutations:"
python3 main.py $MANY $seed --model meralg7 --lr 0.03 --gamma 0.03 --memories 5120 --replay_batch_size 100 --s 10

echo "Beginning MER (Algorithm 7) With 500 Memories" "( seed =" $seed ")"
echo "MNIST Rotations:"
python3 main.py $ROT $seed --model meralg7 --lr 0.03 --gamma 0.03 --memories 500 --replay_batch_size 50 --s 5
echo "MNIST Permutations:"
python3 main.py $PERM $seed --model meralg7 --lr 0.01 --gamma 0.1 --memories 500 --replay_batch_size 25 --s 10
echo "MNIST Many Permutations:"
python3 main.py $MANY $seed --model meralg7 --lr 0.03 --gamma 0.03 --memories 500 --replay_batch_size 5 --s 10

echo "Beginning MER (Algorithm 7) With 200 Memories" "( seed =" $seed ")"
echo "MNIST Rotations:"
python3 main.py $ROT $seed --model meralg7 --lr 0.03 --gamma 0.03 --memories 200 --replay_batch_size 50 --s 5
echo "MNIST Permutations:"
python3 main.py $PERM $seed --model meralg7 --lr 0.03 --gamma 0.1 --memories 200 --replay_batch_size 5 --s 2


================================================
FILE: metrics/metrics.py
================================================
### We directly copied the metrics.py model file from the GEM project https://github.com/facebookresearch/GradientEpisodicMemory

# Copyright 2019-present, IBM Research
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.

from __future__ import print_function

import torch


def task_changes(result_t):
    n_tasks = int(result_t.max() + 1)
    changes = []
    current = result_t[0]
    for i, t in enumerate(result_t):
        if t != current:
            changes.append(i)
            current = t

    return n_tasks, changes


def confusion_matrix(result_t, result_a, fname=None):
    nt, changes = task_changes(result_t)

    baseline = result_a[0]
    changes = torch.LongTensor(changes + [result_a.size(0)]) - 1
    result = result_a.index_select(0, torch.LongTensor(changes))  # .index (torch<0.3.1) | .index_select (torch>0.4)

    # acc[t] equals result[t,t]
    acc = result.diag()
    fin = result[nt - 1]
    # bwt[t] equals result[T,t] - acc[t]
    bwt = result[nt - 1] - acc

    # fwt[t] equals result[t-1,t] - baseline[t]
    fwt = torch.zeros(nt)
    for t in range(1, nt):
        fwt[t] = result[t - 1, t] - baseline[t]

    if fname is not None:
        f = open(fname, 'w')

        print(' '.join(['%.4f' % r for r in baseline]), file=f)
        print('|', file=f)
        for row in range(result.size(0)):
            print(' '.join(['%.4f' % r for r in result[row]]), file=f)
        print('', file=f)
        # print('Diagonal Accuracy: %.4f' % acc.mean(), file=f)
        print('Final Accuracy: %.4f' % fin.mean(), file=f)
        print('Backward: %.4f' % bwt.mean(), file=f)
        print('Forward:  %.4f' % fwt.mean(), file=f)
        f.close()

    stats = []
    # stats.append(acc.mean())
    stats.append(fin.mean())
    stats.append(bwt.mean())
    stats.append(fwt.mean())

    return stats


================================================
FILE: model/common.py
================================================
### A copy of common.py from https://github.com/facebookresearch/GradientEpisodicMemory. 
### We leveraged the same architecture and weight initialization for all of our experiments. 

# Copyright 2019-present, IBM Research
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.

import math
import torch
import torch.nn as nn
from torch.nn.functional import relu, avg_pool2d


def Xavier(m):
    if m.__class__.__name__ == 'Linear':
        fan_in, fan_out = m.weight.data.size(1), m.weight.data.size(0)
        std = 1.0 * math.sqrt(2.0 / (fan_in + fan_out))
        a = math.sqrt(3.0) * std
        m.weight.data.uniform_(-a, a)
        m.bias.data.fill_(0.0)


class MLP(nn.Module):
    def __init__(self, sizes):
        super(MLP, self).__init__()
        layers = []

        for i in range(0, len(sizes) - 1):
            layers.append(nn.Linear(sizes[i], sizes[i + 1]))
            if i < (len(sizes) - 2):
                layers.append(nn.ReLU())

        self.net = nn.Sequential(*layers)
        self.net.apply(Xavier)

    def forward(self, x):
        return self.net(x)


def conv3x3(in_planes, out_planes, stride=1):
    return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride,
                     padding=1, bias=False)


class BasicBlock(nn.Module):
    expansion = 1

    def __init__(self, in_planes, planes, stride=1):
        super(BasicBlock, self).__init__()
        self.conv1 = conv3x3(in_planes, planes, stride)
        self.bn1 = nn.BatchNorm2d(planes)
        self.conv2 = conv3x3(planes, planes)
        self.bn2 = nn.BatchNorm2d(planes)

        self.shortcut = nn.Sequential()
        if stride != 1 or in_planes != self.expansion * planes:
            self.shortcut = nn.Sequential(
                nn.Conv2d(in_planes, self.expansion * planes, kernel_size=1,
                          stride=stride, bias=False),
                nn.BatchNorm2d(self.expansion * planes)
            )

    def forward(self, x):
        out = relu(self.bn1(self.conv1(x)))
        out = self.bn2(self.conv2(out))
        out += self.shortcut(x)
        out = relu(out)
        return out


class ResNet(nn.Module):
    def __init__(self, block, num_blocks, num_classes, nf):
        super(ResNet, self).__init__()
        self.in_planes = nf

        self.conv1 = conv3x3(3, nf * 1)
        self.bn1 = nn.BatchNorm2d(nf * 1)
        self.layer1 = self._make_layer(block, nf * 1, num_blocks[0], stride=1)
        self.layer2 = self._make_layer(block, nf * 2, num_blocks[1], stride=2)
        self.layer3 = self._make_layer(block, nf * 4, num_blocks[2], stride=2)
        self.layer4 = self._make_layer(block, nf * 8, num_blocks[3], stride=2)
        self.linear = nn.Linear(nf * 8 * block.expansion, num_classes)

    def _make_layer(self, block, planes, num_blocks, stride):
        strides = [stride] + [1] * (num_blocks - 1)
        layers = []
        for stride in strides:
            layers.append(block(self.in_planes, planes, stride))
            self.in_planes = planes * block.expansion
        return nn.Sequential(*layers)

    def forward(self, x):
        bsz = x.size(0)
        out = relu(self.bn1(self.conv1(x.view(bsz, 3, 32, 32))))
        out = self.layer1(out)
        out = self.layer2(out)
        out = self.layer3(out)
        out = self.layer4(out)
        out = avg_pool2d(out, 4)
        out = out.view(out.size(0), -1)
        out = self.linear(out)
        return out


def ResNet18(nclasses, nf=20):
    return ResNet(BasicBlock, [2, 2, 2, 2], nclasses, nf)


================================================
FILE: model/eralg4.py
================================================
# An implementation of Experience Replay (ER) with reservoir sampling and without using tasks from Algorithm 4 of https://openreview.net/pdf?id=B1gTShAct7

# Copyright 2019-present, IBM Research
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.

import torch
import torch.nn as nn
import torch.optim as optim
from torch.autograd import Variable

import numpy as np

from .common import MLP, ResNet18
import random
from torch.nn.modules.loss import CrossEntropyLoss
from random import shuffle
import sys
import warnings
warnings.filterwarnings("ignore")

class Net(nn.Module):
    def __init__(self,
                 n_inputs,
                 n_outputs,
                 n_tasks,
                 args):
        super(Net, self).__init__()
        nl, nh = args.n_layers, args.n_hiddens
        self.net = MLP([n_inputs] + [nh] * nl + [n_outputs])

        self.bce = CrossEntropyLoss()
        self.n_outputs = n_outputs

        self.opt = optim.SGD(self.parameters(), args.lr)
        self.batchSize = int(args.replay_batch_size)

        self.memories = args.memories

        # allocate buffer
        self.M = []
        self.age = 0
        
        # handle gpus if specified
        self.cuda = args.cuda
        if self.cuda:
            self.net = self.net.cuda()
            

    def forward(self, x, t):
        output = self.net(x)
        return output


    def getBatch(self,x,y,t):
        mxi = np.array(x)
        myi = np.array(y)
        bxs = []
        bys = []
        
        if len(self.M) > 0:
            order = [i for i in range(0,len(self.M))]
            osize = min(self.batchSize,len(self.M))
            for j in range(0,osize):
                shuffle(order)
                k = order[j]
                x,y,t = self.M[k]
                xi = np.array(x)
                yi = np.array(y)
                bxs.append(xi)
                bys.append(yi)

        bxs.append(mxi)
        bys.append(myi)

        bxs = Variable(torch.from_numpy(np.array(bxs))).float().view(-1,784)
        bys = Variable(torch.from_numpy(np.array(bys))).long().view(-1)
        
        # handle gpus if specified
        if self.cuda:
            bxs = bxs.cuda()
            bys = bys.cuda()

 
        return bxs,bys
                

    def observe(self, x, t, y):
        ### step through elements of x
        for i in range(0,x.size()[0]):
            self.age += 1
            xi = x[i].data.cpu().numpy()
            yi = y[i].data.cpu().numpy()
            
            self.net.zero_grad()
            
            # Draw batch from buffer:
            bx,by = self.getBatch(xi,yi,t)
            
            # Update parameters with mini-batch SGD:
            prediction = self.forward(bx,0)
            loss = self.bce(prediction,by)
            loss.backward()
            self.opt.step()
            
            # Reservoir sampling memory update:
            if len(self.M) < self.memories:
                self.M.append([xi,yi,t])

            else:
                p = random.randint(0,self.age)
                if p < self.memories:
                    self.M[p] = [xi,yi,t]
                    




================================================
FILE: model/eralg5.py
================================================
# An implementation of Experience Replay (ER) with tasks from Algorithm 5 in https://openreview.net/pdf?id=B1gTShAct7

# Copyright 2019-present, IBM Research
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.

import torch
import torch.nn as nn
import torch.optim as optim
from torch.autograd import Variable

import numpy as np

from .common import MLP, ResNet18
import random
from torch.nn.modules.loss import CrossEntropyLoss
from random import shuffle
import sys
import warnings
warnings.filterwarnings("ignore")

class Net(nn.Module):
    def __init__(self,
                 n_inputs,
                 n_outputs,
                 n_tasks,
                 args):
        super(Net, self).__init__()
        nl, nh = args.n_layers, args.n_hiddens
        self.net = MLP([n_inputs] + [nh] * nl + [n_outputs])

        self.bce = CrossEntropyLoss()
        self.n_outputs = n_outputs

        self.opt = optim.SGD(self.parameters(), args.lr)
        self.batchSize = int(args.replay_batch_size)

        self.memories = args.memories

        # allocate buffer
        self.M = []
        self.age = 0
        
        # handle gpus if specified
        self.cuda = args.cuda
        if self.cuda:
            self.net = self.net.cuda()

    def forward(self, x, t):
        output = self.net(x)
        return output

    def getBatch(self,x,y,t):
        bx = {}
        by = {}
        
        bx[t] = [x]
        by[t] = [y]
            
        if len(self.M) > 0:
            order = [i for i in range(0,len(self.M))]
            osize = min(self.batchSize,len(self.M))
            for j in range(0,osize):
                shuffle(order)
                k = order[j]
                x,y,t = self.M[k]
                if t in bx:
                    bx[t].append(x)
                    by[t].append(y)
                else:
                    bx[t] = [x]
                    by[t] = [y]

        for task in bx:
            bx[task] = Variable(torch.from_numpy(np.array(bx[task]))).float()
            by[task] = Variable(torch.from_numpy(np.array(by[task]))).long().view(-1)
            
            # handle gpus if specified
            if self.cuda:
                bx[task] = bx[task].cuda()
                by[task] = by[task].cuda()

        return bx,by
                

    def observe(self, x, t, y):
        ### step through elements of x
        for i in range(0,x.size()[0]):
            self.age += 1
            xi = x[i].data.cpu().numpy()
            yi = y[i].data.cpu().numpy()

            self.net.zero_grad()
            
            # Draw batch from buffer:
            bx,by = self.getBatch(xi,yi,t)
            
            # Update parameters with balanced loss across tasks:
            loss = 0.0
            for kz in bx:
                prediction = self.forward(bx[kz],kz)
                loss += self.bce(prediction,by[kz])
            loss.backward()
            self.opt.step()
            
            # Reservoir sampling memory update: 
            
            if len(self.M) < self.memories:
                self.M.append([xi,yi,t])

            else:
                p = random.randint(0,self.age)
                if p < self.memories:
                    self.M[p] = [xi,yi,t]




================================================
FILE: model/ewc.py
================================================
### We use the same version of EwC https://www.pnas.org/content/114/13/3521 originally used in https://github.com/facebookresearch/GradientEpisodicMemory
### We directly copied the ewc.py model file from the GEM project https://github.com/facebookresearch/GradientEpisodicMemory

# Copyright 2019-present, IBM Research
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.

import torch
from torch.autograd import Variable
from .common import MLP, ResNet18


class Net(torch.nn.Module):

    def __init__(self,
                 n_inputs,
                 n_outputs,
                 n_tasks,
                 args):
        super(Net, self).__init__()
        nl, nh = args.n_layers, args.n_hiddens
        self.reg = args.memory_strength

        # setup network
        self.is_cifar = (args.data_file == 'cifar100.pt')
        if self.is_cifar:
            self.net = ResNet18(n_outputs)
        else:
            self.net = MLP([n_inputs] + [nh] * nl + [n_outputs])

        # setup optimizer
        self.opt = torch.optim.SGD(self.net.parameters(), lr=args.lr)

        # setup losses
        self.bce = torch.nn.CrossEntropyLoss()

        # setup memories
        self.current_task = 0
        self.fisher = {}
        self.optpar = {}
        self.memx = None
        self.memy = None

        if self.is_cifar:
            self.nc_per_task = n_outputs / n_tasks
        else:
            self.nc_per_task = n_outputs
        self.n_outputs = n_outputs
        self.n_memories = args.n_memories
        
        # handle gpus if specified
        self.cuda = args.cuda
        if self.cuda:
            self.net = self.net.cuda()

    def compute_offsets(self, task):
        if self.is_cifar:
            offset1 = task * self.nc_per_task
            offset2 = (task + 1) * self.nc_per_task
        else:
            offset1 = 0
            offset2 = self.n_outputs
        return int(offset1), int(offset2)

    def forward(self, x, t):
        output = self.net(x)
        if self.is_cifar:
            # make sure we predict classes within the current task
            offset1, offset2 = self.compute_offsets(t)
            if offset1 > 0:
                output[:, :offset1].data.fill_(-10e10)
            if offset2 < self.n_outputs:
                output[:, int(offset2):self.n_outputs].data.fill_(-10e10)
        return output

    def observe(self, x, t, y):
        self.net.train()

        # next task?
        if t != self.current_task:
            self.net.zero_grad()

            if self.is_cifar:
                offset1, offset2 = self.compute_offsets(self.current_task)
                self.bce((self.net(Variable(self.memx))[:, offset1: offset2]),
                         Variable(self.memy) - offset1).backward()
            else:
                self.bce(self(Variable(self.memx),
                              self.current_task),
                         Variable(self.memy)).backward()
            self.fisher[self.current_task] = []
            self.optpar[self.current_task] = []
            for p in self.net.parameters():
                pd = p.data.clone()
                pg = p.grad.data.clone().pow(2)
                self.optpar[self.current_task].append(pd)
                self.fisher[self.current_task].append(pg)
            self.current_task = t
            self.memx = None
            self.memy = None

        if self.memx is None:
            self.memx = x.data.clone()
            self.memy = y.data.clone()
        else:
            if self.memx.size(0) < self.n_memories:
                self.memx = torch.cat((self.memx, x.data.clone()))
                self.memy = torch.cat((self.memy, y.data.clone()))
                if self.memx.size(0) > self.n_memories:
                    self.memx = self.memx[:self.n_memories]
                    self.memy = self.memy[:self.n_memories]

        self.net.zero_grad()
        if self.is_cifar:
            offset1, offset2 = self.compute_offsets(t)
            loss = self.bce((self.net(x)[:, offset1: offset2]),
                            y - offset1)
        else:
            loss = self.bce(self(x, t), y)
        for tt in range(t):
            for i, p in enumerate(self.net.parameters()):
                l = self.reg * Variable(self.fisher[tt][i])
                l = l * (p - Variable(self.optpar[tt][i])).pow(2)
                loss += l.sum()
        loss.backward()
        self.opt.step()


================================================
FILE: model/gem.py
================================================
### This is a copy of GEM from https://github.com/facebookresearch/GradientEpisodicMemory. 
### In order to ensure complete reproducability, we do not change the file and treat it as a baseline.

# Copyright 2019-present, IBM Research
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.

import torch
import torch.nn as nn
import torch.optim as optim
from torch.autograd import Variable

import numpy as np
import quadprog

from .common import MLP, ResNet18

# Auxiliary functions useful for GEM's inner optimization.

def compute_offsets(task, nc_per_task, is_cifar):
    """
        Compute offsets for cifar to determine which
        outputs to select for a given task.
    """
    if is_cifar:
        offset1 = task * nc_per_task
        offset2 = (task + 1) * nc_per_task
    else:
        offset1 = 0
        offset2 = nc_per_task
    return offset1, offset2


def store_grad(pp, grads, grad_dims, tid):
    """
        This stores parameter gradients of past tasks.
        pp: parameters
        grads: gradients
        grad_dims: list with number of parameters per layers
        tid: task id
    """
    # store the gradients
    grads[:, tid].fill_(0.0)
    cnt = 0
    for param in pp():
        if param.grad is not None:
            beg = 0 if cnt == 0 else sum(grad_dims[:cnt])
            en = sum(grad_dims[:cnt + 1])
            grads[beg: en, tid].copy_(param.grad.data.view(-1))
        cnt += 1


def overwrite_grad(pp, newgrad, grad_dims):
    """
        This is used to overwrite the gradients with a new gradient
        vector, whenever violations occur.
        pp: parameters
        newgrad: corrected gradient
        grad_dims: list storing number of parameters at each layer
    """
    cnt = 0
    for param in pp():
        if param.grad is not None:
            beg = 0 if cnt == 0 else sum(grad_dims[:cnt])
            en = sum(grad_dims[:cnt + 1])
            this_grad = newgrad[beg: en].contiguous().view(
                param.grad.data.size())
            param.grad.data.copy_(this_grad)
        cnt += 1


def project2cone2(gradient, memories, margin=0.5):
    """
        Solves the GEM dual QP described in the paper given a proposed
        gradient "gradient", and a memory of task gradients "memories".
        Overwrites "gradient" with the final projected update.
        input:  gradient, p-vector
        input:  memories, (t * p)-vector
        output: x, p-vector
    """
    memories_np = memories.cpu().t().double().numpy()
    gradient_np = gradient.cpu().contiguous().view(-1).double().numpy()
    t = memories_np.shape[0]
    P = np.dot(memories_np, memories_np.transpose())
    P = 0.5 * (P + P.transpose())
    q = np.dot(memories_np, gradient_np) * -1
    G = np.eye(t)
    h = np.zeros(t) + margin
    v = quadprog.solve_qp(P, q, G, h)[0]
    x = np.dot(v, memories_np) + gradient_np
    gradient.copy_(torch.Tensor(x).view(-1, 1))


class Net(nn.Module):
    def __init__(self,
                 n_inputs,
                 n_outputs,
                 n_tasks,
                 args):
        super(Net, self).__init__()
        nl, nh = args.n_layers, args.n_hiddens
        self.margin = args.memory_strength
        self.is_cifar = (args.data_file == 'cifar100.pt')
        if self.is_cifar:
            self.net = ResNet18(n_outputs)
        else:
            self.net = MLP([n_inputs] + [nh] * nl + [n_outputs])

        self.ce = nn.CrossEntropyLoss()
        self.n_outputs = n_outputs

        self.opt = optim.SGD(self.parameters(), args.lr)

        self.n_memories = args.n_memories
        self.gpu = args.cuda

        # allocate episodic memory
        self.memory_data = torch.FloatTensor(
            n_tasks, self.n_memories, n_inputs)
        self.memory_labs = torch.LongTensor(n_tasks, self.n_memories)
        if args.cuda:
            self.memory_data = self.memory_data.cuda()
            self.memory_labs = self.memory_labs.cuda()

        # allocate temporary synaptic memory
        self.grad_dims = []
        for param in self.parameters():
            self.grad_dims.append(param.data.numel())
        self.grads = torch.Tensor(sum(self.grad_dims), n_tasks)
        if args.cuda:
            self.grads = self.grads.cuda()

        # allocate counters
        self.observed_tasks = []
        self.old_task = -1
        self.mem_cnt = 0
        if self.is_cifar:
            self.nc_per_task = int(n_outputs / n_tasks)
        else:
            self.nc_per_task = n_outputs
        
        if args.cuda:
            self.cuda()

    def forward(self, x, t):
        output = self.net(x)
        if self.is_cifar:
            # make sure we predict classes within the current task
            offset1 = int(t * self.nc_per_task)
            offset2 = int((t + 1) * self.nc_per_task)
            if offset1 > 0:
                output[:, :offset1].data.fill_(-10e10)
            if offset2 < self.n_outputs:
                output[:, offset2:self.n_outputs].data.fill_(-10e10)
        return output

    def observe(self, x, t, y):
        # update memory
        if t != self.old_task:
            self.observed_tasks.append(t)
            self.old_task = t

        # Update ring buffer storing examples from current task
        bsz = y.data.size(0)
        endcnt = min(self.mem_cnt + bsz, self.n_memories)
        effbsz = endcnt - self.mem_cnt
        self.memory_data[t, self.mem_cnt: endcnt].copy_(
            x.data[: effbsz])
        if bsz == 1:
            self.memory_labs[t, self.mem_cnt] = y.data[0]
        else:
            self.memory_labs[t, self.mem_cnt: endcnt].copy_(
                y.data[: effbsz])
        self.mem_cnt += effbsz
        if self.mem_cnt == self.n_memories:
            self.mem_cnt = 0

        # compute gradient on previous tasks
        if len(self.observed_tasks) > 1:
            for tt in range(len(self.observed_tasks) - 1):
                self.zero_grad()
                # fwd/bwd on the examples in the memory
                past_task = self.observed_tasks[tt]

                offset1, offset2 = compute_offsets(past_task, self.nc_per_task,
                                                   self.is_cifar)
                ptloss = self.ce(
                    self.forward(
                        Variable(self.memory_data[past_task]),
                        past_task)[:, offset1: offset2],
                    Variable(self.memory_labs[past_task] - offset1))
                ptloss.backward()
                store_grad(self.parameters, self.grads, self.grad_dims,
                           past_task)

        # now compute the grad on the current minibatch
        self.zero_grad()

        offset1, offset2 = compute_offsets(t, self.nc_per_task, self.is_cifar)
        loss = self.ce(self.forward(x, t)[:, offset1: offset2], y - offset1)
        loss.backward()

        # check if gradient violates constraints
        if len(self.observed_tasks) > 1:
            # copy gradient
            store_grad(self.parameters, self.grads, self.grad_dims, t)
            indx = torch.cuda.LongTensor(self.observed_tasks[:-1]) if self.gpu \
                else torch.LongTensor(self.observed_tasks[:-1])
            dotp = torch.mm(self.grads[:, t].unsqueeze(0),
                            self.grads.index_select(1, indx))
            if (dotp < 0).sum() != 0:
                project2cone2(self.grads[:, t].unsqueeze(1),
                              self.grads.index_select(1, indx), self.margin)
                # copy gradients back
                overwrite_grad(self.parameters, self.grads[:, t],
                               self.grad_dims)

        self.opt.step()


================================================
FILE: model/independent.py
================================================
### code for this basline is copied directly from independent.py in https://github.com/facebookresearch/GradientEpisodicMemory

# Copyright 2019-present, IBM Research
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.

import torch
from torch.autograd import Variable
from .common import MLP, ResNet18


class Net(torch.nn.Module):

    def __init__(self,
                 n_inputs,
                 n_outputs,
                 n_tasks,
                 args):
        super(Net, self).__init__()
        nl, nh = args.n_layers, args.n_hiddens
        self.nets = torch.nn.ModuleList()
        self.opts = []

        self.is_cifar = (args.data_file == 'cifar100.pt')
        if self.is_cifar:
            self.nc_per_task = n_outputs / n_tasks
        self.n_outputs = n_outputs

        # setup network
        for _ in range(n_tasks):
            if self.is_cifar:
                self.nets.append(
                    ResNet18(int(n_outputs / n_tasks), int(20 / n_tasks)))
            else:
                self.nets.append(
                    MLP([n_inputs] + [int(nh / n_tasks)] * nl + [n_outputs]))

        # setup optimizer
        for t in range(n_tasks):
            self.opts.append(torch.optim.SGD(self.nets[t].parameters(),
                                             lr=args.lr))

        # setup loss
        self.bce = torch.nn.CrossEntropyLoss()

        self.finetune = args.finetune
        self.gpu = args.cuda
        self.old_task = 0
        

    def forward(self, x, t):
        output = self.nets[t](x)
        if self.is_cifar:
            bigoutput = torch.Tensor(x.size(0), self.n_outputs)
            if self.gpu:
                bigoutput = bigoutput.cuda()
            bigoutput.fill_(-10e10)
            bigoutput[:, int(t * self.nc_per_task): int((t + 1) * self.nc_per_task)].copy_(
                output.data)
            return Variable(bigoutput)
        else:
            return output

    def observe(self, x, t, y):
        # detect beginning of a new task
        if self.finetune and t > 0 and t != self.old_task:
            # initialize current network like the previous one
            for ppold, ppnew in zip(self.nets[self.old_task].parameters(),
                                    self.nets[t].parameters()):
                ppnew.data.copy_(ppold.data)
            self.old_task = t

        self.train()
        self.zero_grad()
        if self.is_cifar:
            self.bce(self.nets[t](x), y - int(t * self.nc_per_task)).backward()
        else:
            self.bce(self(x, t), y).backward()
        self.opts[t].step()


================================================
FILE: model/meralg1.py
================================================
# An implementation of MER Algorithm 1 from https://openreview.net/pdf?id=B1gTShAct7

# Copyright 2019-present, IBM Research
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.

import torch
import torch.nn as nn
import torch.optim as optim
from torch.autograd import Variable

import numpy as np

from .common import MLP, ResNet18
import random
from torch.nn.modules.loss import CrossEntropyLoss
from random import shuffle
import sys
from copy import deepcopy
import warnings
warnings.filterwarnings("ignore")

class Net(nn.Module):
    def __init__(self,
                 n_inputs,
                 n_outputs,
                 n_tasks,
                 args):
        super(Net, self).__init__()
        
        nl, nh = args.n_layers, args.n_hiddens
        
        self.net = MLP([n_inputs] + [nh] * nl + [n_outputs])

        self.bce = CrossEntropyLoss()
        self.n_outputs = n_outputs

        self.opt = optim.SGD(self.parameters(), args.lr)
        self.batchSize = int(args.replay_batch_size)

        self.memories = args.memories
        self.steps = int(args.batches_per_example)
        self.beta = args.beta
        self.gamma = args.gamma

        # allocate buffer
        self.M = []
        self.age = 0
        
        # handle gpus if specified
        self.cuda = args.cuda
        if self.cuda:
            self.net = self.net.cuda()


    def forward(self, x, t):
        output = self.net(x)
        return output

    def getBatch(self,x,y,t):
        xi = Variable(torch.from_numpy(np.array(x))).float().view(1,-1)
        yi = Variable(torch.from_numpy(np.array(y))).long().view(1)
        if self.cuda:
            xi = xi.cuda()
            yi = yi.cuda()
        bxs = [xi]
        bys = [yi]
        
        if len(self.M) > 0:
            order = [i for i in range(0,len(self.M))]
            osize = min(self.batchSize,len(self.M))
            for j in range(0,osize):
                shuffle(order)
                k = order[j]
                x,y,t = self.M[k]
                xi = Variable(torch.from_numpy(np.array(x))).float().view(1,-1)
                yi = Variable(torch.from_numpy(np.array(y))).long().view(1)
                # handle gpus if specified
                if self.cuda:
                    xi = xi.cuda()
                    yi = yi.cuda()
                bxs.append(xi)
                bys.append(yi)
                
                

 
        return bxs,bys
                

    def observe(self, x, t, y):
        ### step through elements of x
        for i in range(0,x.size()[0]):
            self.age += 1
            xi = x[i].data.cpu().numpy()
            yi = y[i].data.cpu().numpy()
            self.net.zero_grad()

            before = deepcopy(self.net.state_dict())
            for step in range(0,self.steps):                
                weights_before = deepcopy(self.net.state_dict())
                # Draw batch from buffer:
                bxs,bys = self.getBatch(xi,yi,t)
                loss = 0.0
                for idx in range(len(bxs)):
                    self.net.zero_grad()
                    bx = bxs[idx]
                    by = bys[idx] 
                    prediction = self.forward(bx,0)
                    loss = self.bce(prediction,by)
                    loss.backward()
                    self.opt.step()
                
                weights_after = self.net.state_dict()
                
                # Within batch Reptile meta-update:
                self.net.load_state_dict({name : weights_before[name] + ((weights_after[name] - weights_before[name]) * self.beta) for name in weights_before})
            
            after = self.net.state_dict()
            
            # Across batch Reptile meta-update:
            self.net.load_state_dict({name : before[name] + ((after[name] - before[name]) * self.gamma) for name in before})

                    

            # Reservoir sampling memory update: 
            
            if len(self.M) < self.memories:
                self.M.append([xi,yi,t])

            else:
                p = random.randint(0,self.age)
                if p < self.memories:
                    self.M[p] = [xi,yi,t]




================================================
FILE: model/meralg6.py
================================================
# An implementation of Meta-Experience Replay (MER) Algorithm 6 from https://openreview.net/pdf?id=B1gTShAct7 

# Copyright 2019-present, IBM Research
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.

import torch
import torch.nn as nn
import torch.optim as optim
from torch.autograd import Variable

import numpy as np

from .common import MLP, ResNet18
import random
from torch.nn.modules.loss import CrossEntropyLoss
from random import shuffle
import sys
from copy import deepcopy
import warnings
from random import shuffle
warnings.filterwarnings("ignore")

class Net(nn.Module):
    def __init__(self,
                 n_inputs,
                 n_outputs,
                 n_tasks,
                 args):
        super(Net, self).__init__()
        nl, nh = args.n_layers, args.n_hiddens
        self.net = MLP([n_inputs] + [nh] * nl + [n_outputs])

        self.bce = CrossEntropyLoss()
        self.n_outputs = n_outputs

        self.opt = optim.SGD(self.parameters(), args.lr)
        self.batchSize = int(args.replay_batch_size)

        self.memories = args.memories
        self.steps = int(args.batches_per_example)
        self.gamma = args.gamma

        # allocate buffer
        self.M = []
        self.age = 0
        
        # handle gpus if specified
        self.cuda = args.cuda
        if self.cuda:
            self.net = self.net.cuda()


    def forward(self, x, t):
        output = self.net(x)
        return output

    def getBatch(self,x,y,t):
        mxi = Variable(torch.from_numpy(np.array(x))).float().view(1,-1)
        myi = Variable(torch.from_numpy(np.array(y))).long().view(1)
        if self.cuda:
            mxi = mxi.cuda()
            myi = myi.cuda()
            
        bxs = []
        bys = []
        
        if len(self.M) > 0:
            order = [i for i in range(0,len(self.M))]
            osize = min(self.batchSize,len(self.M))
            for j in range(0,osize):
                shuffle(order)
                k = order[j]
                x,y,t = self.M[k]
                xi = Variable(torch.from_numpy(np.array(x))).float().view(1,-1)
                yi = Variable(torch.from_numpy(np.array(y))).long().view(1)
                # handle gpus if specified
                if self.cuda:
                    xi = xi.cuda()
                    yi = yi.cuda()
                bxs.append(xi)
                bys.append(yi)

        bxs.append(mxi)
        bys.append(myi)

 
        return bxs,bys
                

    def observe(self, x, t, y):
        ### step through elements of x
        for i in range(0,x.size()[0]):
            self.age += 1
            xi = x[i].data.cpu().numpy()
            yi = y[i].data.cpu().numpy()

            bx,by = self.getBatch(xi,yi,t)
            self.net.zero_grad()


            for step in range(0,self.steps):                
                weights_before = deepcopy(self.net.state_dict())
                
                # Draw batch from buffer:
                bxs,bys = self.getBatch(xi,yi,t)
                
                loss = 0.0
                for idx in range(len(bxs)):
                    self.net.zero_grad()
                    bx = bxs[idx]
                    by = bys[idx] 
                    prediction = self.forward(bx,0)
                    loss = self.bce(prediction,by)
                    loss.backward()
                    self.opt.step()
                
                weights_after = self.net.state_dict()
                # Reptile meta-update:
                self.net.load_state_dict({name : weights_before[name] + ((weights_after[name] - weights_before[name]) * self.gamma) for name in weights_before})
            
                

            sys.stdout.flush()
            
            # Reservoir sampling memory update:
                
            if len(self.M) < self.memories:
                self.M.append([xi,yi,t])

            else:
                p = random.randint(0,self.age)
                if p < self.memories:
                    self.M[p] = [xi,yi,t]




================================================
FILE: model/meralg7.py
================================================
# An implementation of Meta-Experience Replay (MER) Algorithm 7 from https://openreview.net/pdf?id=B1gTShAct7

# Copyright 2019-present, IBM Research
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.

import torch
import torch.nn as nn
import torch.optim as optim
from torch.autograd import Variable

import numpy as np

from .common import MLP, ResNet18
import random
from torch.nn.modules.loss import CrossEntropyLoss
from random import shuffle
import sys
from copy import deepcopy
import warnings
from random import shuffle

warnings.filterwarnings("ignore")


class Net(nn.Module):
    def __init__(self,
                 n_inputs,
                 n_outputs,
                 n_tasks,
                 args):
        super(Net, self).__init__()
        nl, nh = args.n_layers, args.n_hiddens
        self.net = MLP([n_inputs] + [nh] * nl + [n_outputs])

        self.bce = CrossEntropyLoss()
        self.n_outputs = n_outputs

        self.opt = optim.SGD(self.parameters(), args.lr)
        self.batchSize = int(args.replay_batch_size)
        

        self.memories = args.memories
        self.s = float(args.s)
        self.gamma = args.gamma



        # allocate buffer
        self.M = []
        self.age = 0
        
        # handle gpus if specified
        self.cuda = args.cuda
        if self.cuda:
            self.net = self.net.cuda()


    def forward(self, x, t):
        output = self.net(x)
        return output

    def getBatch(self,x,y,t):
        mxi = Variable(torch.from_numpy(np.array(x))).float().view(1,-1)
        myi = Variable(torch.from_numpy(np.array(y))).long().view(1)
        if self.cuda:
            mxi = mxi.cuda()
            myi = myi.cuda()
        bxs = [mxi]
        bys = [myi]
        
        if len(self.M) > 0:
            order = [i for i in range(0,len(self.M))]
            osize = min(self.batchSize,len(self.M))
            for j in range(0,osize):
                shuffle(order)
                k = order[j]
                x,y,t = self.M[k]
                xi = Variable(torch.from_numpy(np.array(x))).float().view(1,-1)
                yi = Variable(torch.from_numpy(np.array(y))).long().view(1)
                
                if self.cuda:
                    xi = xi.cuda()
                    yi = yi.cuda()
                    
                bxs.append(xi)
                bys.append(yi)

        bx2 = []
        by2 = []
        indexes = [ind for ind in range(len(bxs))]
        shuffle(indexes)
        for index in indexes:
            if index == 0:
                myindex = len(bx2)
            bx2.append(bxs[index])
            by2.append(bys[index])
 
        return bx2,by2,myindex
                

    def observe(self, x, t, y):
        ### step through elements of x
        for i in range(0,x.size()[0]):
            self.age += 1
            xi = x[i].data.cpu().numpy()
            yi = y[i].data.cpu().numpy()
            if self.age > 1:
                self.net.zero_grad()


   
                weights_before = deepcopy(self.net.state_dict())
    
                # Draw batch from buffer:
                bxs,bys,myind = self.getBatch(xi,yi,t)
                
                # SGD on individual samples from batch:
                loss = 0.0
                for idx in range(len(bxs)):
                    self.net.zero_grad()
                    bx = bxs[idx]
                    by = bys[idx] 
                    prediction = self.forward(bx,0)
                    if myind == idx:
                        # High learning rate SGD on current example:
                        loss = self.s*self.bce(prediction,by)
                    else:
                        loss = self.bce(prediction,by)
                    loss.backward()
                    self.opt.step()
                
                weights_after = self.net.state_dict()
                # Reptile meta-update:
                self.net.load_state_dict({name : weights_before[name] + ((weights_after[name] - weights_before[name]) * self.gamma) for name in weights_before})
            
                    

            sys.stdout.flush()
            
            # Reservoir sampling memory update:
            if len(self.M) < self.memories:
                self.M.append([xi,yi,t])

            else:
                p = random.randint(0,self.age)
                if p < self.memories:
                    self.M[p] = [xi,yi,t]




================================================
FILE: model/online.py
================================================
### code for this basline is copied directly from single.py in https://github.com/facebookresearch/GradientEpisodicMemory

# Copyright 2019-present, IBM Research
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.

import torch
from .common import MLP, ResNet18


class Net(torch.nn.Module):

    def __init__(self,
                 n_inputs,
                 n_outputs,
                 n_tasks,
                 args):
        super(Net, self).__init__()
        nl, nh = args.n_layers, args.n_hiddens

        # setup network
        self.is_cifar = (args.data_file == 'cifar100.pt')
        if self.is_cifar:
            self.net = ResNet18(n_outputs)
        else:
            self.net = MLP([n_inputs] + [nh] * nl + [n_outputs])

        # setup optimizer
        self.opt = torch.optim.SGD(self.parameters(), lr=args.lr)

        # setup losses
        self.bce = torch.nn.CrossEntropyLoss()

        if self.is_cifar:
            self.nc_per_task = n_outputs / n_tasks
        else:
            self.nc_per_task = n_outputs
        self.n_outputs = n_outputs
        # handle gpus if specified
        self.cuda = args.cuda
        if self.cuda:
            self.net = self.net.cuda()


    def compute_offsets(self, task):
        if self.is_cifar:
            offset1 = task * self.nc_per_task
            offset2 = (task + 1) * self.nc_per_task
        else:
            offset1 = 0
            offset2 = self.n_outputs
        return int(offset1), int(offset2)

    def forward(self, x, t):
        output = self.net(x)
        if self.is_cifar:
            # make sure we predict classes within the current task
            offset1, offset2 = self.compute_offsets(t)
            if offset1 > 0:
                output[:, :offset1].data.fill_(-10e10)
            if offset2 < self.n_outputs:
                output[:, offset2:self.n_outputs].data.fill_(-10e10)
        return output

    def observe(self, x, t, y):
        self.train()
        self.zero_grad()
        if self.is_cifar:
            offset1, offset2 = self.compute_offsets(t)
            self.bce((self.net(x)[:, offset1: offset2]),
                     y - offset1).backward()
        else:
            self.bce(self(x, t), y).backward()
        self.opt.step()


================================================
FILE: model/taskinput.py
================================================
### Code for this basline is copied directly from multimodal.py in https://github.com/facebookresearch/GradientEpisodicMemory

# Copyright 2019-present, IBM Research
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.

import torch
import torch.nn as nn


def reset_bias(m):
    m.bias.data.fill_(0.0)


class Net(nn.Module):
    def __init__(self,
                 n_inputs,
                 n_outputs,
                 n_tasks,
                 args):
        super(Net, self).__init__()

        self.i_layer = nn.ModuleList()
        self.h_layer = nn.ModuleList()
        self.o_layer = nn.ModuleList()

        self.n_layers = args.n_layers
        nh = args.n_hiddens

        if self.n_layers > 0:
            # dedicated input layer
            for _ in range(n_tasks):
                self.i_layer += [nn.Linear(n_inputs, nh)]
                reset_bias(self.i_layer[-1])

            # shared hidden layer
            self.h_layer += [nn.ModuleList()]
            for _ in range(self.n_layers):
                self.h_layer[0] += [nn.Linear(nh, nh)]
                reset_bias(self.h_layer[0][0])

            # shared output layer
            self.o_layer += [nn.Linear(nh, n_outputs)]
            reset_bias(self.o_layer[-1])

        # linear model falls back to independent models
        else:
            self.i_layer += [nn.Linear(n_inputs, n_outputs)]
            reset_bias(self.i_layer[-1])

        self.relu = nn.ReLU()
        self.soft = nn.LogSoftmax()
        self.loss = nn.NLLLoss()
        self.optimizer = torch.optim.SGD(self.parameters(), args.lr)
        

    def forward(self, x, t):
        h = x

        if self.n_layers == 0:
            y = self.soft(self.i_layer[t if isinstance(t, int) else t[0]](h))
        else:
            # task-specific input
            h = self.relu(self.i_layer[t if isinstance(t, int) else t[0]](h))
            # shared hiddens
            for l in range(self.n_layers):
                h = self.relu(self.h_layer[0][l](h))
            # shared output
            y = self.soft(self.o_layer[0](h))

        return y

    def observe(self, x, t, y):
        self.zero_grad()
        self.loss(self.forward(x, t), y).backward()
        self.optimizer.step()


================================================
FILE: requirements.txt
================================================
argparse
datetime
matplotlib
numpy
pillow
quadprog
torchvision
urllib3
uuid


================================================
FILE: unit_test.sh
================================================
#!/bin/bash

seed=0
# Test on a small number of samples per task
ROT="--n_layers 2 --n_hiddens 100 --data_path data/ --save_path results/ --batch_size 1 --log_every 100 --samples_per_task 10 --data_file mnist_rotations.pt    --cuda no  --seed"

echo "Begin Unit Test with seed =" $seed " on MNIST Rotations"
echo "Testing Online"
python3 main.py $ROT $seed --model online --lr 0.0003
if [ $? -eq 0 ]
then
  echo "Online finished successfully"
else
  echo "Exited with error." >&2  # Redirect stdout from echo command to stderr.
fi

echo "Testing Independent"
python3 main.py $ROT $seed --model independent --lr 0.01
if [ $? -eq 0 ]
then
  echo "Independent finished successfully"
else
  echo "Exited with error." >&2  # Redirect stdout from echo command to stderr.
fi

echo "Testing EWC"
python3 main.py $ROT $seed --model ewc --lr 0.001 --n_memories 10 --memory_strength 100.0
if [ $? -eq 0 ]
then
  echo "EWC finished successfully"
else
  echo "Exited with error." >&2  # Redirect stdout from echo command to stderr.
fi

echo "Testing GEM"
python3 main.py $ROT $seed --model gem --lr 0.01 --n_memories 256 --memory_strength 1.0
if [ $? -eq 0 ]
then
  echo "GEM finished successfully"
else
  echo "Exited with error." >&2  # Redirect stdout from echo command to stderr.
fi


echo "Testing eralg4"
python3 main.py $ROT $seed --model eralg4 --lr 0.1 --memories 5120 --replay_batch_size 25
if [ $? -eq 0 ]
then
  echo "eralg4 finished successfully"
else
  echo "Exited with error." >&2  # Redirect stdout from echo command to stderr.
fi


echo "Testing eralg5"
python3 main.py $ROT $seed --model eralg5 --lr 0.03 --memories 5120 --replay_batch_size 100
if [ $? -eq 0 ]
then
  echo "eralg5 finished successfully"
else
  echo "Exited with error." >&2  # Redirect stdout from echo command to stderr.
fi


echo "Testing meralg1"
python3 main.py $ROT $seed --model meralg1 --lr 0.03 --beta 0.03 --gamma 1.0 --memories 5120 --replay_batch_size 100 --batches_per_example 10
if [ $? -eq 0 ]
then
  echo "meralg1 finished successfully"
else
  echo "Exited with error." >&2  # Redirect stdout from echo command to stderr.
fi


echo "Testing meralg6"
python3 main.py $ROT $seed --model meralg6 --lr 0.03 --gamma 0.1 --memories 5120 --replay_batch_size 100
if [ $? -eq 0 ]
then
  echo "meralg6 finished successfully"
else
  echo "Exited with error." >&2  # Redirect stdout from echo command to stderr.
fi


echo "Testing meralg7"
python3 main.py $ROT $seed --model meralg7 --lr 0.03 --gamma 0.03 --memories 500 --replay_batch_size 50 --s 5
if [ $? -eq 0 ]
then
  echo "meralg7 finished successfully"
else
  echo "Exited with error." >&2  # Redirect stdout from echo command to stderr.
fi


echo "Finished"
Download .txt
gitextract_jqq8_utk/

├── LICENSE
├── MIT-LICENSE
├── README.md
├── data/
│   └── README.md
├── download.py
├── get_data.py
├── main.py
├── mer_examples.sh
├── metrics/
│   └── metrics.py
├── model/
│   ├── common.py
│   ├── eralg4.py
│   ├── eralg5.py
│   ├── ewc.py
│   ├── gem.py
│   ├── independent.py
│   ├── meralg1.py
│   ├── meralg6.py
│   ├── meralg7.py
│   ├── online.py
│   └── taskinput.py
├── requirements.txt
└── unit_test.sh
Download .txt
SYMBOL INDEX (80 symbols across 15 files)

FILE: download.py
  function _print_download_progress (line 28) | def _print_download_progress(count, block_size, total_size):
  function download (line 50) | def download(base_url, filename, download_dir):
  function maybe_download_and_extract (line 80) | def maybe_download_and_extract(url, download_dir):

FILE: get_data.py
  function get_mnist_data (line 11) | def get_mnist_data(url, data_dir):
  function get_datasets (line 15) | def get_datasets():

FILE: main.py
  function load_datasets (line 28) | def load_datasets(args):
  class Continuum (line 38) | class Continuum:
    method __init__ (line 40) | def __init__(self, data, args):
    method __iter__ (line 73) | def __iter__(self):
    method next (line 76) | def next(self):
    method __next__ (line 79) | def __next__(self):
  function eval_tasks (line 98) | def eval_tasks(model, tasks, args):
  function life_experience (line 129) | def life_experience(model, continuum, x_te, args):

FILE: metrics/metrics.py
  function task_changes (line 14) | def task_changes(result_t):
  function confusion_matrix (line 26) | def confusion_matrix(result_t, result_a, fname=None):

FILE: model/common.py
  function Xavier (line 16) | def Xavier(m):
  class MLP (line 25) | class MLP(nn.Module):
    method __init__ (line 26) | def __init__(self, sizes):
    method forward (line 38) | def forward(self, x):
  function conv3x3 (line 42) | def conv3x3(in_planes, out_planes, stride=1):
  class BasicBlock (line 47) | class BasicBlock(nn.Module):
    method __init__ (line 50) | def __init__(self, in_planes, planes, stride=1):
    method forward (line 65) | def forward(self, x):
  class ResNet (line 73) | class ResNet(nn.Module):
    method __init__ (line 74) | def __init__(self, block, num_blocks, num_classes, nf):
    method _make_layer (line 86) | def _make_layer(self, block, planes, num_blocks, stride):
    method forward (line 94) | def forward(self, x):
  function ResNet18 (line 107) | def ResNet18(nclasses, nf=20):

FILE: model/eralg4.py
  class Net (line 24) | class Net(nn.Module):
    method __init__ (line 25) | def __init__(self,
    method forward (line 52) | def forward(self, x, t):
    method getBatch (line 57) | def getBatch(self,x,y,t):
    method observe (line 90) | def observe(self, x, t, y):

FILE: model/eralg5.py
  class Net (line 24) | class Net(nn.Module):
    method __init__ (line 25) | def __init__(self,
    method forward (line 51) | def forward(self, x, t):
    method getBatch (line 55) | def getBatch(self,x,y,t):
    method observe (line 88) | def observe(self, x, t, y):

FILE: model/ewc.py
  class Net (line 15) | class Net(torch.nn.Module):
    method __init__ (line 17) | def __init__(self,
    method compute_offsets (line 58) | def compute_offsets(self, task):
    method forward (line 67) | def forward(self, x, t):
    method observe (line 78) | def observe(self, x, t, y):

FILE: model/gem.py
  function compute_offsets (line 22) | def compute_offsets(task, nc_per_task, is_cifar):
  function store_grad (line 36) | def store_grad(pp, grads, grad_dims, tid):
  function overwrite_grad (line 55) | def overwrite_grad(pp, newgrad, grad_dims):
  function project2cone2 (line 74) | def project2cone2(gradient, memories, margin=0.5):
  class Net (line 96) | class Net(nn.Module):
    method __init__ (line 97) | def __init__(self,
    method forward (line 147) | def forward(self, x, t):
    method observe (line 159) | def observe(self, x, t, y):

FILE: model/independent.py
  class Net (line 14) | class Net(torch.nn.Module):
    method __init__ (line 16) | def __init__(self,
    method forward (line 53) | def forward(self, x, t):
    method observe (line 66) | def observe(self, x, t, y):

FILE: model/meralg1.py
  class Net (line 25) | class Net(nn.Module):
    method __init__ (line 26) | def __init__(self,
    method forward (line 58) | def forward(self, x, t):
    method getBatch (line 62) | def getBatch(self,x,y,t):
    method observe (line 93) | def observe(self, x, t, y):

FILE: model/meralg6.py
  class Net (line 26) | class Net(nn.Module):
    method __init__ (line 27) | def __init__(self,
    method forward (line 56) | def forward(self, x, t):
    method getBatch (line 60) | def getBatch(self,x,y,t):
    method observe (line 93) | def observe(self, x, t, y):

FILE: model/meralg7.py
  class Net (line 28) | class Net(nn.Module):
    method __init__ (line 29) | def __init__(self,
    method forward (line 61) | def forward(self, x, t):
    method getBatch (line 65) | def getBatch(self,x,y,t):
    method observe (line 104) | def observe(self, x, t, y):

FILE: model/online.py
  class Net (line 13) | class Net(torch.nn.Module):
    method __init__ (line 15) | def __init__(self,
    method compute_offsets (line 47) | def compute_offsets(self, task):
    method forward (line 56) | def forward(self, x, t):
    method observe (line 67) | def observe(self, x, t, y):

FILE: model/taskinput.py
  function reset_bias (line 13) | def reset_bias(m):
  class Net (line 17) | class Net(nn.Module):
    method __init__ (line 18) | def __init__(self,
    method forward (line 59) | def forward(self, x, t):
    method observe (line 75) | def observe(self, x, t, y):
Condensed preview — 22 files, each showing path, character count, and a content snippet. Download the .json file or copy for the full structured content (95K chars).
[
  {
    "path": "LICENSE",
    "chars": 11357,
    "preview": "                                 Apache License\n                           Version 2.0, January 2004\n                   "
  },
  {
    "path": "MIT-LICENSE",
    "chars": 1023,
    "preview": "Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentati"
  },
  {
    "path": "README.md",
    "chars": 5973,
    "preview": "# Meta-Experience Replay (MER)\n\nSource code for the paper \"Learning to Learn without Forgetting by Maximizing Transfer a"
  },
  {
    "path": "data/README.md",
    "chars": 211,
    "preview": "This is the directory where data should be stored in order to use main.py for executing continual learning models. If yo"
  },
  {
    "path": "download.py",
    "chars": 4353,
    "preview": "########################################################################\n#\n# Functions for downloading and extracting da"
  },
  {
    "path": "get_data.py",
    "chars": 1423,
    "preview": "# Copyright 2019-present, IBM Research\n# All rights reserved.\n#\n# This source code is licensed under the license found i"
  },
  {
    "path": "main.py",
    "chars": 9909,
    "preview": "### This is a modified version of main.py from https://github.com/facebookresearch/GradientEpisodicMemory \n### The most "
  },
  {
    "path": "mer_examples.sh",
    "chars": 9195,
    "preview": "#!/bin/bash\n\nseed=$1\nROT=\"--n_layers 2 --n_hiddens 100 --data_path data/ --save_path results/ --batch_size 1 --log_every"
  },
  {
    "path": "metrics/metrics.py",
    "chars": 1920,
    "preview": "### We directly copied the metrics.py model file from the GEM project https://github.com/facebookresearch/GradientEpisod"
  },
  {
    "path": "model/common.py",
    "chars": 3608,
    "preview": "### A copy of common.py from https://github.com/facebookresearch/GradientEpisodicMemory. \n### We leveraged the same arch"
  },
  {
    "path": "model/eralg4.py",
    "chars": 3233,
    "preview": "# An implementation of Experience Replay (ER) with reservoir sampling and without using tasks from Algorithm 4 of https:"
  },
  {
    "path": "model/eralg5.py",
    "chars": 3325,
    "preview": "# An implementation of Experience Replay (ER) with tasks from Algorithm 5 in https://openreview.net/pdf?id=B1gTShAct7\n\n#"
  },
  {
    "path": "model/ewc.py",
    "chars": 4494,
    "preview": "### We use the same version of EwC https://www.pnas.org/content/114/13/3521 originally used in https://github.com/facebo"
  },
  {
    "path": "model/gem.py",
    "chars": 7721,
    "preview": "### This is a copy of GEM from https://github.com/facebookresearch/GradientEpisodicMemory. \n### In order to ensure compl"
  },
  {
    "path": "model/independent.py",
    "chars": 2674,
    "preview": "### code for this basline is copied directly from independent.py in https://github.com/facebookresearch/GradientEpisodic"
  },
  {
    "path": "model/meralg1.py",
    "chars": 4274,
    "preview": "# An implementation of MER Algorithm 1 from https://openreview.net/pdf?id=B1gTShAct7\n\n# Copyright 2019-present, IBM Rese"
  },
  {
    "path": "model/meralg6.py",
    "chars": 4120,
    "preview": "# An implementation of Meta-Experience Replay (MER) Algorithm 6 from https://openreview.net/pdf?id=B1gTShAct7 \n\n# Copyri"
  },
  {
    "path": "model/meralg7.py",
    "chars": 4498,
    "preview": "# An implementation of Meta-Experience Replay (MER) Algorithm 7 from https://openreview.net/pdf?id=B1gTShAct7\n\n# Copyrig"
  },
  {
    "path": "model/online.py",
    "chars": 2335,
    "preview": "### code for this basline is copied directly from single.py in https://github.com/facebookresearch/GradientEpisodicMemor"
  },
  {
    "path": "model/taskinput.py",
    "chars": 2320,
    "preview": "### Code for this basline is copied directly from multimodal.py in https://github.com/facebookresearch/GradientEpisodicM"
  },
  {
    "path": "requirements.txt",
    "chars": 76,
    "preview": "argparse\ndatetime\nmatplotlib\nnumpy\npillow\nquadprog\ntorchvision\nurllib3\nuuid\n"
  },
  {
    "path": "unit_test.sh",
    "chars": 2692,
    "preview": "#!/bin/bash\n\nseed=0\n# Test on a small number of samples per task\nROT=\"--n_layers 2 --n_hiddens 100 --data_path data/ --s"
  }
]

About this extraction

This page contains the full source code of the mattriemer/MER GitHub repository, extracted and formatted as plain text for AI agents and large language models (LLMs). The extraction includes 22 files (88.6 KB), approximately 22.2k tokens, and a symbol index with 80 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.

Copied to clipboard!