Full Code of heartexlabs/label-studio-transformers for AI

Repository: heartexlabs/label-studio-transformers
Branch: master
Commit: de267cbd903e
Files: 8
Total size: 51.2 KB

Directory structure:
gitextract_m_a1r8n7/

├── .gitignore
├── LICENSE
├── README.md
├── models/
│   ├── __init__.py
│   ├── bert_classifier.py
│   ├── ner.py
│   └── utils.py
└── requirements.txt

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

================================================
FILE: .gitignore
================================================
# Byte-compiled / optimized / DLL files
__pycache__/
*.py[cod]
*$py.class

# C extensions
*.so

# Distribution / packaging
.Python
build/
develop-eggs/
dist/
downloads/
eggs/
.eggs/
lib/
lib64/
parts/
sdist/
var/
wheels/
*.egg-info/
.installed.cfg
*.egg
MANIFEST

# PyInstaller
#  Usually these files are written by a python script from a template
#  before PyInstaller builds the exe, so as to inject date/other infos into it.
*.manifest
*.spec

# Installer logs
pip-log.txt
pip-delete-this-directory.txt

# Unit test / coverage reports
htmlcov/
.tox/
.coverage
.coverage.*
.cache
nosetests.xml
coverage.xml
*.cover
.hypothesis/
.pytest_cache/

# Translations
*.mo
*.pot

# Django stuff:
*.log
local_settings.py
db.sqlite3

# Flask stuff:
instance/
.webassets-cache

# Scrapy stuff:
.scrapy

# Sphinx documentation
docs/_build/

# PyBuilder
target/

# Jupyter Notebook
.ipynb_checkpoints

# pyenv
.python-version

# celery beat schedule file
celerybeat-schedule

# SageMath parsed files
*.sage.py

# Environments
.venv
env/
venv/
ENV/
env.bak/
venv.bak/

# Spyder project settings
.spyderproject
.spyproject

# Rope project settings
.ropeproject

# mkdocs documentation
/site

# mypy
.mypy_cache/


================================================
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: README.md
================================================
# Label Studio for Hugging Face's Transformers

[Website](https://labelstud.io/) • [Docs](https://labelstud.io/guide) • [Twitter](https://twitter.com/heartexlabs) • [Join Slack Community <img src="https://app.heartex.ai/docs/images/slack-mini.png" width="18px"/>](https://slack.labelstud.io/?source=github-1)

<br/>

**Transfer learning for NLP models by annotating your textual data without any additional coding.**

This package provides a ready-to-use container that links together:

- [Label Studio](https://github.com/heartexlabs/label-studio) as annotation frontend
- [Hugging Face's transformers](https://github.com/huggingface/transformers) as machine learning backend for NLP

<br/>

[<img src="https://raw.githubusercontent.com/heartexlabs/label-studio-transformers/master/images/codeless.png" height="500">](https://github.com/heartexlabs/label-studio-transformers)

### Quick Usage

#### Install Label Studio and other dependencies

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

##### Create ML backend with BERT classifier
```bash
label-studio-ml init my-ml-backend --script models/bert_classifier.py
cp models/utils.py my-ml-backend/utils.py

# Start ML backend at http://localhost:9090
label-studio-ml start my-ml-backend

# Start Label Studio in the new terminal with the same python environment
label-studio start
```

1. Create a project with `Choices` and `Text` tags in the labeling config.
2. Connect the ML backend in the Project settings with `http://localhost:9090`

##### Create ML backend with BERT named entity recognizer
```bash
label-studio-ml init my-ml-backend --script models/ner.py
cp models/utils.py my-ml-backend/utils.py

# Start ML backend at http://localhost:9090
label-studio-ml start my-ml-backend

# Start Label Studio in the new terminal with the same python environment
label-studio start
```

1. Create a project with `Labels` and `Text` tags in the labeling config.
2. Connect the ML backend in the Project settings with `http://localhost:9090`

#### Training and inference

The browser opens at `http://localhost:8080`. Upload your data on **Import** page then annotate by selecting **Labeling** page.
Once you've annotate sufficient amount of data, go to **Model** page and press **Start Training** button. Once training is finished, model automatically starts serving for inference from Label Studio, and you'll find all model checkpoints inside `my-ml-backend/<ml-backend-id>/` directory.

[Click here](https://labelstud.io/guide/ml.html) to read more about how to use Machine Learning backend and build Human-in-the-Loop pipelines with Label Studio

## License

This software is licensed under the [Apache 2.0 LICENSE](/LICENSE) © [Heartex](https://www.heartex.com/). 2020

<img src="https://github.com/heartexlabs/label-studio/blob/master/images/opossum_looking.png?raw=true" title="Hey everyone!" height="140" width="140" />


================================================
FILE: models/__init__.py
================================================


================================================
FILE: models/bert_classifier.py
================================================
import torch
import numpy as np
import os

from torch.utils.data import SequentialSampler
from tqdm import tqdm, trange
from collections import deque
from tensorboardX import SummaryWriter
from transformers import BertTokenizer, BertForSequenceClassification
from transformers import AdamW, get_linear_schedule_with_warmup
from torch.utils.data import TensorDataset, DataLoader, RandomSampler

from label_studio_ml.model import LabelStudioMLBase

from utils import prepare_texts, calc_slope


if torch.cuda.is_available():
    device = torch.device("cuda")
    print('There are %d GPU(s) available.' % torch.cuda.device_count())
    print('We will use the GPU:', torch.cuda.get_device_name(0))
else:
    print('No GPU available, using the CPU instead.')
    device = torch.device("cpu")


class BertClassifier(LabelStudioMLBase):

    def __init__(
        self, pretrained_model='bert-base-multilingual-cased', maxlen=64,
        batch_size=32, num_epochs=100, logging_steps=1, train_logs=None, **kwargs
    ):
        super(BertClassifier, self).__init__(**kwargs)
        self.pretrained_model = pretrained_model
        self.maxlen = maxlen
        self.batch_size = batch_size
        self.num_epochs = num_epochs
        self.logging_steps = logging_steps
        self.train_logs = train_logs

        # then collect all keys from config which will be used to extract data from task and to form prediction
        # Parsed label config contains only one output of <Choices> type
        assert len(self.parsed_label_config) == 1
        self.from_name, self.info = list(self.parsed_label_config.items())[0]
        assert self.info['type'] == 'Choices'

        # the model has only one textual input
        assert len(self.info['to_name']) == 1
        assert len(self.info['inputs']) == 1
        assert self.info['inputs'][0]['type'] == 'Text'
        self.to_name = self.info['to_name'][0]
        self.value = self.info['inputs'][0]['value']

        if not self.train_output:
            self.labels = self.info['labels']
            self.reset_model('bert-base-multilingual-cased', cache_dir=None, device='cpu')
            print('Initialized with from_name={from_name}, to_name={to_name}, labels={labels}'.format(
                from_name=self.from_name, to_name=self.to_name, labels=str(self.labels)
            ))
        else:
            self.load(self.train_output)
            print('Loaded from train output with from_name={from_name}, to_name={to_name}, labels={labels}'.format(
                from_name=self.from_name, to_name=self.to_name, labels=str(self.labels)
            ))

    def reset_model(self, pretrained_model, cache_dir, device):
        model = BertForSequenceClassification.from_pretrained(
            pretrained_model,
            num_labels=len(self.labels),
            output_attentions=False,
            output_hidden_states=False,
            cache_dir=cache_dir
        )
        model.to(device)
        return model

    def load(self, train_output):
        pretrained_model = train_output['model_path']
        self.tokenizer = BertTokenizer.from_pretrained(pretrained_model)
        self.model = BertForSequenceClassification.from_pretrained(pretrained_model)
        self.model.to(device)
        self.model.eval()
        self.batch_size = train_output['batch_size']
        self.labels = train_output['labels']
        self.maxlen = train_output['maxlen']

    @property
    def not_trained(self):
        return not hasattr(self, 'tokenizer')

    def predict(self, tasks, **kwargs):
        if self.not_trained:
            print('Can\'t get prediction because model is not trained yet.')
            return []

        texts = [task['data'][self.value] for task in tasks]
        predict_dataloader = prepare_texts(texts, self.tokenizer, self.maxlen, SequentialSampler, self.batch_size)

        pred_labels, pred_scores = [], []
        for batch in predict_dataloader:
            batch = tuple(t.to(device) for t in batch)
            inputs = {
                'input_ids': batch[0],
                'attention_mask': batch[1]
            }
            with torch.no_grad():
                outputs = self.model(**inputs)
                logits = outputs[0]

            batch_preds = logits.detach().cpu().numpy()

            argmax_batch_preds = np.argmax(batch_preds, axis=-1)
            pred_labels.extend(str(self.labels[i]) for i in argmax_batch_preds)

            max_batch_preds = np.max(batch_preds, axis=-1)
            pred_scores.extend(float(s) for s in max_batch_preds)

        predictions = []
        for predicted_label, score in zip(pred_labels, pred_scores):
            result = [{
                'from_name': self.from_name,
                'to_name': self.to_name,
                'type': 'choices',
                'value': {'choices': [predicted_label]}
            }]

            predictions.append({'result': result, 'score': score})
        return predictions

    def fit(self, completions, workdir=None, cache_dir=None, **kwargs):
        input_texts = []
        output_labels, output_labels_idx = [], []
        label2idx = {l: i for i, l in enumerate(self.labels)}
        for completion in completions:
            # get input text from task data

            if completion['annotations'][0].get('skipped'):
                continue

            input_text = completion['data'][self.value]
            input_texts.append(input_text)

            # get an annotation
            output_label = completion['annotations'][0]['result'][0]['value']['choices'][0]
            output_labels.append(output_label)
            output_label_idx = label2idx[output_label]
            output_labels_idx.append(output_label_idx)

        new_labels = set(output_labels)
        added_labels = new_labels - set(self.labels)
        if len(added_labels) > 0:
            print('Label set has been changed. Added ones: ' + str(list(added_labels)))
            self.labels = list(sorted(new_labels))
            label2idx = {l: i for i, l in enumerate(self.labels)}
            output_labels_idx = [label2idx[label] for label in output_labels]

        tokenizer = BertTokenizer.from_pretrained(self.pretrained_model, cache_dir=cache_dir)

        train_dataloader = prepare_texts(input_texts, tokenizer, self.maxlen, RandomSampler, self.batch_size, output_labels_idx)
        model = self.reset_model(self.pretrained_model, cache_dir, device)

        total_steps = len(train_dataloader) * self.num_epochs
        optimizer = AdamW(model.parameters(), lr=2e-5, eps=1e-8)
        scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=0, num_training_steps=total_steps)
        global_step = 0
        total_loss, logging_loss = 0.0, 0.0
        model.zero_grad()
        train_iterator = trange(self.num_epochs, desc='Epoch')
        if self.train_logs:
            tb_writer = SummaryWriter(logdir=os.path.join(self.train_logs, os.path.basename(self.output_dir)))
        else:
            tb_writer = None
        loss_queue = deque(maxlen=10)
        for epoch in train_iterator:
            epoch_iterator = tqdm(train_dataloader, desc='Iteration')
            for step, batch in enumerate(epoch_iterator):
                model.train()
                batch = tuple(t.to(device) for t in batch)
                inputs = {'input_ids': batch[0],
                          'attention_mask': batch[1],
                          'labels': batch[2]}
                outputs = model(**inputs)
                loss = outputs[0]
                loss.backward()
                total_loss += loss.item()

                torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
                optimizer.step()
                scheduler.step()
                model.zero_grad()
                global_step += 1
                if global_step % self.logging_steps == 0:
                    last_loss = (total_loss - logging_loss) / self.logging_steps
                    loss_queue.append(last_loss)
                    if tb_writer:
                        tb_writer.add_scalar('lr', scheduler.get_lr()[0], global_step)
                        tb_writer.add_scalar('loss', last_loss, global_step)
                    logging_loss = total_loss

            # slope-based early stopping
            if len(loss_queue) == loss_queue.maxlen:
                slope = calc_slope(loss_queue)
                if tb_writer:
                    tb_writer.add_scalar('slope', slope, global_step)
                if abs(slope) < 1e-2:
                    break

        if tb_writer:
            tb_writer.close()

        model_to_save = model.module if hasattr(model, 'module') else model  # Take care of distributed/parallel training  # noqa
        model_to_save.save_pretrained(workdir)
        tokenizer.save_pretrained(workdir)

        return {
            'model_path': workdir,
            'batch_size': self.batch_size,
            'maxlen': self.maxlen,
            'pretrained_model': self.pretrained_model,
            'labels': self.labels
        }


================================================
FILE: models/ner.py
================================================
import torch
import numpy as np
import re
import os
import io
import logging

from functools import partial
from itertools import groupby
from operator import itemgetter
from torch.nn import CrossEntropyLoss
from torch.utils.data import Dataset, DataLoader
from tqdm import tqdm, trange
from tensorboardX import SummaryWriter
from collections import deque

from transformers import (
    BertTokenizer, BertForTokenClassification, BertConfig,
    RobertaConfig, RobertaForTokenClassification, RobertaTokenizer,
    DistilBertConfig, DistilBertForTokenClassification, DistilBertTokenizer,
    CamembertConfig, CamembertForTokenClassification, CamembertTokenizer,
    AutoConfig, AutoModelForTokenClassification, AutoTokenizer
)
from transformers import AdamW, get_linear_schedule_with_warmup

from label_studio_ml.model import LabelStudioMLBase
from utils import calc_slope


logger = logging.getLogger(__name__)


ALL_MODELS = sum(
    [list(conf.pretrained_config_archive_map.keys()) for conf in (BertConfig, RobertaConfig, DistilBertConfig)],
    [])

MODEL_CLASSES = {
    'bert': (BertConfig, BertForTokenClassification, BertTokenizer),
    'roberta': (RobertaConfig, RobertaForTokenClassification, RobertaTokenizer),
    'distilbert': (DistilBertConfig, DistilBertForTokenClassification, DistilBertTokenizer),
    'camembert': (CamembertConfig, CamembertForTokenClassification, CamembertTokenizer),
}


class SpanLabeledTextDataset(Dataset):

    def __init__(
        self, list_of_strings, list_of_spans=None, tokenizer=None, tag_idx_map=None,
        cls_token='[CLS]', sep_token='[SEP]', pad_token_label_id=-1, max_seq_length=128, sep_token_extra=False,
        cls_token_at_end=False, sequence_a_segment_id=0, cls_token_segment_id=1, mask_padding_with_zero=True
    ):
        self.list_of_strings = list_of_strings
        self.list_of_spans = list_of_spans or [[] * len(list_of_strings)]
        self.tokenizer = tokenizer
        self.cls_token = cls_token
        self.sep_token = sep_token
        self.pad_token_label_id = pad_token_label_id
        self.max_seq_length = max_seq_length
        self.sep_token_extra = sep_token_extra
        self.cls_token_at_end = cls_token_at_end
        self.sequence_a_segment_id = sequence_a_segment_id
        self.cls_token_segment_id = cls_token_segment_id
        self.mask_padding_with_zero = mask_padding_with_zero

        (self.original_list_of_tokens, self.original_list_of_tags, tag_idx_map_,
         original_list_of_tokens_start_map) = self._prepare_data()

        if tag_idx_map is None:
            self.tag_idx_map = tag_idx_map_
        else:
            self.tag_idx_map = tag_idx_map

        (self.list_of_tokens, self.list_of_token_ids, self.list_of_labels, self.list_of_label_ids,
         self.list_of_segment_ids, self.list_of_token_start_map) = [], [], [], [], [], []

        for original_tokens, original_tags, original_token_start_map in zip(
            self.original_list_of_tokens,
            self.original_list_of_tags,
            original_list_of_tokens_start_map
        ):
            tokens, token_ids, labels, label_ids, segment_ids, token_start_map = self._convert_to_features(
                original_tokens, original_tags, self.tag_idx_map, original_token_start_map)
            self.list_of_token_ids.append(token_ids)
            self.list_of_tokens.append(tokens)
            self.list_of_labels.append(labels)
            self.list_of_segment_ids.append(segment_ids)
            self.list_of_label_ids.append(label_ids)
            self.list_of_token_start_map.append(token_start_map)

    def get_params_dict(self):
        return {
            'cls_token': self.cls_token,
            'sep_token': self.sep_token,
            'pad_token_label_id': self.pad_token_label_id,
            'max_seq_length': self.max_seq_length,
            'sep_token_extra': self.sep_token_extra,
            'cls_token_at_end': self.cls_token_at_end,
            'sequence_a_segment_id': self.sequence_a_segment_id,
            'cls_token_segment_id': self.cls_token_segment_id,
            'mask_padding_with_zero': self.mask_padding_with_zero
        }

    def dump(self, output_file):
        with io.open(output_file, mode='w') as f:
            for tokens, labels in zip(self.list_of_tokens, self.list_of_labels):
                for token, label in zip(tokens, labels):
                    f.write(f'{token} {label}\n')
                f.write('\n')

    def _convert_to_features(self, words, labels, label_map, list_token_start_map):
        tokens, out_labels, label_ids, tokens_idx_map = [], [], [], []
        for i, (word, label, token_start) in enumerate(zip(words, labels, list_token_start_map)):
            word_tokens = self.tokenizer.tokenize(word)
            tokens.extend(word_tokens)
            tokens_idx_map.extend([token_start] * len(word_tokens))
            # Use the real label id for the first token of the word, and padding ids for the remaining tokens
            label_ids.extend([label_map[label]] + [self.pad_token_label_id] * (len(word_tokens) - 1))
            out_labels.extend([label] + ['X'] * (len(word_tokens) - 1))

        # Account for [CLS] and [SEP] with "- 2" and with "- 3" for RoBERTa.
        special_tokens_count = 3 if self.sep_token_extra else 2
        if len(tokens) > self.max_seq_length - special_tokens_count:
            tokens = tokens[:(self.max_seq_length - special_tokens_count)]
            label_ids = label_ids[:(self.max_seq_length - special_tokens_count)]
            out_labels = out_labels[:(self.max_seq_length - special_tokens_count)]
            tokens_idx_map = tokens_idx_map[:(self.max_seq_length - special_tokens_count)]

        # The convention in BERT is:
        # (a) For sequence pairs:
        #  tokens:   [CLS] is this jack ##son ##ville ? [SEP] no it is not . [SEP]
        #  type_ids:   0   0  0    0    0     0       0   0   1  1  1  1   1   1
        # (b) For single sequences:
        #  tokens:   [CLS] the dog is hairy . [SEP]
        #  type_ids:   0   0   0   0  0     0   0
        #
        # Where "type_ids" are used to indicate whether this is the first
        # sequence or the second sequence. The embedding vectors for `type=0` and
        # `type=1` were learned during pre-training and are added to the wordpiece
        # embedding vector (and position vector). This is not *strictly* necessary
        # since the [SEP] token unambiguously separates the sequences, but it makes
        # it easier for the model to learn the concept of sequences.
        #
        # For classification tasks, the first vector (corresponding to [CLS]) is
        # used as as the "sentence vector". Note that this only makes sense because
        # the entire model is fine-tuned.
        tokens += [self.sep_token]
        label_ids += [self.pad_token_label_id]
        out_labels += ['X']
        tokens_idx_map += [-1]
        if self.sep_token_extra:
            # roberta uses an extra separator b/w pairs of sentences
            tokens += [self.sep_token]
            label_ids += [self.pad_token_label_id]
            out_labels += ['X']
            tokens_idx_map += [-1]
        segment_ids = [self.sequence_a_segment_id] * len(tokens)
        if self.cls_token_at_end:
            tokens += [self.cls_token]
            label_ids += [self.pad_token_label_id]
            out_labels += ['X']
            segment_ids += [self.cls_token_segment_id]
            tokens_idx_map += [-1]
        else:
            tokens = [self.cls_token] + tokens
            label_ids = [self.pad_token_label_id] + label_ids
            out_labels = ['X'] + out_labels
            segment_ids = [self.cls_token_segment_id] + segment_ids
            tokens_idx_map = [-1] + tokens_idx_map

        token_ids = self.tokenizer.convert_tokens_to_ids(tokens)

        return tokens, token_ids, out_labels, label_ids, segment_ids, tokens_idx_map

    def _apply_tokenizer(self, original_tokens, original_tags):
        out_tokens, out_tags, out_maps = [], [], []
        for i, (original_token, original_tag) in enumerate(zip(original_tokens, original_tags)):
            tokens = self.tokenizer.tokenize(original_token)
            out_tokens.extend(tokens)
            out_maps.extend([i] * len(tokens))
            start_tag = original_tag.startswith('B-')
            for j in range(len(tokens)):
                if (j == 0 and start_tag) or original_tag == 'O':
                    out_tags.append(original_tag)
                else:
                    out_tags.append(f'I-{original_tag[2:]}')
        return out_tokens, out_tags, out_maps

    def _prepare_data(self):
        list_of_tokens, list_of_tags, list_of_token_idx_maps = [], [], []
        tag_idx_map = {'O': 0}
        for text, spans in zip(self.list_of_strings, self.list_of_spans):
            if not text:
                continue

            tokens = []
            start = 0
            for t in text.split():
                tokens.append((t, start))
                start += len(t) + 1

            if spans:
                spans = list(sorted(spans, key=itemgetter('start')))
                span = spans.pop(0)
                prefix = 'B-'
                tags = []
                for token, token_start in tokens:
                    token_end = token_start + len(token) - 1

                    # token precedes current span
                    if not span or token_end < span['start']:
                        tags.append('O')
                        continue

                    # token jumps over the span (it could happens
                    # when prev label ends with whitespaces, e.g. "cat " "too" or span created for whitespace)
                    if token_start > span['end']:

                        prefix = 'B-'
                        no_more_spans = False
                        while token_start > span['end']:
                            if not len(spans):
                                no_more_spans = True
                                break
                            span = spans.pop(0)

                        if no_more_spans:
                            tags.append('O')
                            span = None
                            continue

                        if token_end < span['start']:
                            tags.append('O')
                            continue

                    label = span['label']
                    if label.startswith(prefix):
                        tag = label
                    else:
                        tag = f'{prefix}{label}'
                    tags.append(tag)
                    if tag not in tag_idx_map:
                        tag_idx_map[tag] = len(tag_idx_map)
                    if span['end'] > token_end:
                        prefix = 'I-'
                    elif len(spans):
                        span = spans.pop(0)
                        prefix = 'B-'
                    else:
                        span = None
            else:
                tags = ['O'] * len(tokens)

            list_of_tokens.append([t[0] for t in tokens])
            list_of_token_idx_maps.append([t[1] for t in tokens])
            list_of_tags.append(tags)

        return list_of_tokens, list_of_tags, tag_idx_map, list_of_token_idx_maps

    def __len__(self):
        return len(self.list_of_token_ids)

    def __getitem__(self, idx):
        return {
            'tokens': self.list_of_token_ids[idx],
            'labels': self.list_of_label_ids[idx],
            'segments': self.list_of_segment_ids[idx],
            'token_start_map': self.list_of_token_start_map[idx],
            'string': self.list_of_strings[idx]
        }

    @property
    def num_labels(self):
        return len(self.tag_idx_map)

    @classmethod
    def pad_sequences(cls, batch, mask_padding_with_zero, pad_on_left, pad_token, pad_token_segment_id, pad_token_label_id):
        # The mask has 1 for real tokens and 0 for padding tokens. Only real
        # tokens are attended to.
        max_seq_length = max(map(len, (sample['tokens'] for sample in batch)))
        batch_input_ids, batch_label_ids, batch_segment_ids, batch_input_mask, batch_token_start_map = [], [], [], [], []
        batch_strings = []
        for sample in batch:
            input_ids = sample['tokens']
            label_ids = sample['labels']
            segment_ids = sample['segments']
            # The mask has 1 for real tokens and 0 for padding tokens. Only real
            # tokens are attended to.
            input_mask = [1 if mask_padding_with_zero else 0] * len(input_ids)
            padding_length = max_seq_length - len(input_ids)
            if pad_on_left:
                input_ids = ([pad_token] * padding_length) + input_ids
                input_mask = ([0 if mask_padding_with_zero else 1] * padding_length) + input_mask
                segment_ids = ([pad_token_segment_id] * padding_length) + segment_ids
                label_ids = ([pad_token_label_id] * padding_length) + label_ids
            else:
                input_ids += ([pad_token] * padding_length)
                input_mask += ([0 if mask_padding_with_zero else 1] * padding_length)
                segment_ids += ([pad_token_segment_id] * padding_length)
                label_ids += ([pad_token_label_id] * padding_length)
            batch_input_ids.append(input_ids)
            batch_label_ids.append(label_ids)
            batch_segment_ids.append(segment_ids)
            batch_input_mask.append(input_mask)
            batch_token_start_map.append(sample['token_start_map'])
            batch_strings.append(sample['string'])

        return {
            'input_ids': torch.tensor(batch_input_ids, dtype=torch.long),
            'label_ids': torch.tensor(batch_label_ids, dtype=torch.long),
            'segment_ids': torch.tensor(batch_segment_ids, dtype=torch.long),
            'input_mask': torch.tensor(batch_input_mask, dtype=torch.long),
            'token_start_map': batch_token_start_map,
            'strings': batch_strings
        }

    @classmethod
    def get_padding_function(cls, model_type, tokenizer, pad_token_label_id):
        return partial(
            cls.pad_sequences,
            mask_padding_with_zero=True,
            pad_on_left=model_type in ['xlnet'],
            pad_token=tokenizer.convert_tokens_to_ids([tokenizer.pad_token])[0],
            pad_token_segment_id=4 if model_type in ['xlnet'] else 0,
            pad_token_label_id=pad_token_label_id
        )


class TransformersBasedTagger(LabelStudioMLBase):

    def __init__(self, **kwargs):
        super(TransformersBasedTagger, self).__init__(**kwargs)

        assert len(self.parsed_label_config) == 1
        self.from_name, self.info = list(self.parsed_label_config.items())[0]
        assert self.info['type'] == 'Labels'

        # the model has only one textual input
        assert len(self.info['to_name']) == 1
        assert len(self.info['inputs']) == 1
        assert self.info['inputs'][0]['type'] == 'Text'
        self.to_name = self.info['to_name'][0]
        self.value = self.info['inputs'][0]['value']

        if not self.train_output:
            self.labels = self.info['labels']
        else:
            self.load(self.train_output)

    def load(self, train_output):
        pretrained_model = train_output['model_path']
        self._model_type = train_output['model_type']
        _, model_class, tokenizer_class = MODEL_CLASSES[train_output['model_type']]

        self._tokenizer = AutoTokenizer.from_pretrained(pretrained_model)
        self._model = AutoModelForTokenClassification.from_pretrained(pretrained_model)
        self._batch_size = train_output['batch_size']
        self._pad_token = self._tokenizer.convert_tokens_to_ids([self._tokenizer.pad_token])[0]
        self._pad_token_label_id = train_output['pad_token_label_id']
        self._label_map = train_output['label_map']
        self._mask_padding_with_zero = True
        self._dataset_params_dict = train_output['dataset_params_dict']

        self._batch_padding = SpanLabeledTextDataset.get_padding_function(
            self._model_type, self._tokenizer, self._pad_token_label_id)

    def predict(self, tasks, **kwargs):
        texts = [task['data'][self.value] for task in tasks]
        predict_set = SpanLabeledTextDataset(texts, tokenizer=self._tokenizer, **self._dataset_params_dict)
        from_name = self.from_name
        to_name = self.to_name
        predict_loader = DataLoader(
            dataset=predict_set,
            batch_size=self._batch_size,
            collate_fn=self._batch_padding
        )

        results = []
        for batch in tqdm(predict_loader, desc='Prediction'):
            inputs = {
                'input_ids': batch['input_ids'],
                'attention_mask': batch['input_mask'],
                'token_type_ids': batch['segment_ids']
            }
            if self._model_type == 'distilbert':
                inputs.pop('token_type_ids')
            with torch.no_grad():
                model_output = self._model(**inputs)
                logits = model_output[0]

            batch_preds = logits.detach().cpu().numpy()
            argmax_batch_preds = np.argmax(batch_preds, axis=-1)
            max_batch_preds = np.max(batch_preds, axis=-1)
            input_mask = batch['input_mask'].detach().cpu().numpy()
            batch_token_start_map = batch['token_start_map']
            batch_strings = batch['strings']

            for max_preds, argmax_preds, mask_tokens, token_start_map, string in zip(
                max_batch_preds, argmax_batch_preds, input_mask, batch_token_start_map, batch_strings
            ):
                preds, scores, starts = [], [], []
                for max_pred, argmax_pred, mask_token, token_start in zip(max_preds, argmax_preds, mask_tokens, token_start_map):
                    if token_start != -1:
                        preds.append(self._label_map[str(argmax_pred)])
                        scores.append(max_pred)
                        starts.append(token_start)
                mean_score = np.mean(scores) if len(scores) > 0 else 0

                result = []

                for label, group in groupby(zip(preds, starts, scores), key=lambda i: re.sub('^(B-|I-)', '', i[0])):
                    _, group_start, _ = list(group)[0]
                    if len(result) > 0:
                        if group_start == 0:
                            result.pop(-1)
                        else:
                            result[-1]['value']['end'] = group_start - 1
                    if label != 'O':
                        result.append({
                            'from_name': from_name,
                            'to_name': to_name,
                            'type': 'labels',
                            'value': {
                                'labels': [label],
                                'start': group_start,
                                'end': None, 
                                'text': '...'
                            }
                        })
                if result and result[-1]['value']['end'] is None:
                    result[-1]['value']['end'] = len(string)
                results.append({
                    'result': result,
                    'score': float(mean_score),
                    'cluster': None
                })
        return results

    def get_spans(self, completion):
        spans = []
        for r in completion['result']:
            if r['from_name'] == self.from_name and r['to_name'] == self.to_name:
                labels = r['value'].get('labels')
                if not isinstance(labels, list) or len(labels) == 0:
                    logger.warning(f'Error while parsing {r}: list type expected for "labels"')
                    continue
                label = labels[0]
                start, end = r['value'].get('start'), r['value'].get('end')
                if start is None or end is None:
                    logger.warning(f'Error while parsing {r}: "labels" should contain "start" and "end" fields')
                spans.append({
                    'label': label,
                    'start': start,
                    'end': end
                })
        return spans

    def fit(
        self, completions, workdir=None, model_type='bert', pretrained_model='bert-base-uncased',
        batch_size=32, learning_rate=5e-5, adam_epsilon=1e-8, num_train_epochs=100, weight_decay=0.0, logging_steps=1,
        warmup_steps=0, save_steps=50, dump_dataset=True, cache_dir='~/.heartex/cache', train_logs=None,
        **kwargs
    ):
        train_logs = train_logs or os.path.join(workdir, 'train_logs')
        os.makedirs(train_logs, exist_ok=True)
        logger.debug('Prepare models')
        cache_dir = os.path.expanduser(cache_dir)
        os.makedirs(cache_dir, exist_ok=True)

        model_type = model_type.lower()
        # assert model_type in MODEL_CLASSES.keys(), f'Input model type {model_type} not in {MODEL_CLASSES.keys()}'
        # assert pretrained_model in ALL_MODELS, f'Pretrained model {pretrained_model} not in {ALL_MODELS}'

        tokenizer = AutoTokenizer.from_pretrained(pretrained_model, cache_dir=cache_dir)

        logger.debug('Read data')
        # read input data stream
        texts, list_of_spans = [], []
        for item in completions:
            texts.append(item['data'][self.value])
            list_of_spans.append(self.get_spans(item['annotations'][0]))

        logger.debug('Prepare dataset')
        pad_token_label_id = CrossEntropyLoss().ignore_index
        train_set = SpanLabeledTextDataset(
            texts, list_of_spans, tokenizer,
            cls_token_at_end=model_type in ['xlnet'],
            cls_token_segment_id=2 if model_type in ['xlnet'] else 0,
            sep_token_extra=model_type in ['roberta'],
            pad_token_label_id=pad_token_label_id
        )

        if dump_dataset:
            dataset_file = os.path.join(workdir, 'train_set.txt')
            train_set.dump(dataset_file)

        # config = config_class.from_pretrained(pretrained_model, num_labels=train_set.num_labels, cache_dir=cache_dir)
        config = AutoConfig.from_pretrained(pretrained_model, num_labels=train_set.num_labels, cache_dir=cache_dir)
        # model = model_class.from_pretrained(pretrained_model, config=config, cache_dir=cache_dir)
        model = AutoModelForTokenClassification.from_pretrained(pretrained_model, config=config, cache_dir=cache_dir)

        batch_padding = SpanLabeledTextDataset.get_padding_function(model_type, tokenizer, pad_token_label_id)

        train_loader = DataLoader(
            dataset=train_set,
            batch_size=batch_size,
            shuffle=True,
            collate_fn=batch_padding
        )

        no_decay = ['bias', 'LayerNorm.weight']
        optimizer_grouped_parameters = [
            {'params': [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)],
             'weight_decay': weight_decay},
            {'params': [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
        ]

        num_training_steps = len(train_loader) * num_train_epochs
        optimizer = AdamW(optimizer_grouped_parameters, lr=learning_rate, eps=adam_epsilon)
        scheduler = get_linear_schedule_with_warmup(
            optimizer, num_warmup_steps=warmup_steps, num_training_steps=num_training_steps)

        tr_loss, logging_loss = 0, 0
        global_step = 0
        if train_logs:
            tb_writer = SummaryWriter(logdir=os.path.join(train_logs, os.path.basename(workdir)))
        epoch_iterator = trange(num_train_epochs, desc='Epoch')
        loss_queue = deque(maxlen=10)
        for _ in epoch_iterator:
            batch_iterator = tqdm(train_loader, desc='Batch')
            for step, batch in enumerate(batch_iterator):

                model.train()
                inputs = {
                    'input_ids': batch['input_ids'],
                    'attention_mask': batch['input_mask'],
                    'labels': batch['label_ids'],
                    'token_type_ids': batch['segment_ids']
                }
                if model_type == 'distilbert':
                    inputs.pop('token_type_ids')

                model_output = model(**inputs)
                loss = model_output[0]
                loss.backward()
                tr_loss += loss.item()
                optimizer.step()
                scheduler.step()
                model.zero_grad()
                global_step += 1
                if global_step % logging_steps == 0:
                    last_loss = (tr_loss - logging_loss) / logging_steps
                    loss_queue.append(last_loss)
                    if train_logs:
                        tb_writer.add_scalar('lr', scheduler.get_lr()[0], global_step)
                        tb_writer.add_scalar('loss', last_loss, global_step)
                    logging_loss = tr_loss

            # slope-based early stopping
            if len(loss_queue) == loss_queue.maxlen:
                slope = calc_slope(loss_queue)
                if train_logs:
                    tb_writer.add_scalar('slope', slope, global_step)
                if abs(slope) < 1e-2:
                    break

        if train_logs:
            tb_writer.close()

        model_to_save = model.module if hasattr(model, "module") else model  # Take care of distributed/parallel training
        model_to_save.save_pretrained(workdir)
        tokenizer.save_pretrained(workdir)
        label_map = {i: t for t, i in train_set.tag_idx_map.items()}

        return {
            'model_path': workdir,
            'batch_size': batch_size,
            'pad_token_label_id': pad_token_label_id,
            'dataset_params_dict': train_set.get_params_dict(),
            'model_type': model_type,
            'pretrained_model': pretrained_model,
            'label_map': label_map
        }


================================================
FILE: models/utils.py
================================================
import torch
import numpy as np

from torch.utils.data import TensorDataset, DataLoader


def pad_sequences(input_ids, maxlen):
    padded_ids = []
    for ids in input_ids:
        nonpad = min(len(ids), maxlen)
        pids = [ids[i] for i in range(nonpad)]
        for i in range(nonpad, maxlen):
            pids.append(0)
        padded_ids.append(pids)
    return padded_ids


def prepare_texts(texts, tokenizer, maxlen, sampler_class, batch_size, choices_ids=None):
    # create input token indices
    input_ids = []
    for text in texts:
        input_ids.append(tokenizer.encode(text, add_special_tokens=True))
    # input_ids = pad_sequences(input_ids, maxlen=maxlen, dtype='long', value=0, truncating='post', padding='post')
    input_ids = pad_sequences(input_ids, maxlen)
    # Create attention masks
    attention_masks = []
    for sent in input_ids:
        attention_masks.append([int(token_id > 0) for token_id in sent])

    if choices_ids is not None:
        dataset = TensorDataset(torch.tensor(input_ids, dtype=torch.long), torch.tensor(attention_masks, dtype=torch.long), torch.tensor(choices_ids, dtype=torch.long))
    else:
        dataset = TensorDataset(torch.tensor(input_ids, dtype=torch.long), torch.tensor(attention_masks, dtype=torch.long))
    sampler = sampler_class(dataset)
    dataloader = DataLoader(dataset, sampler=sampler, batch_size=batch_size)
    return dataloader


def calc_slope(y):
    n = len(y)
    if n == 1:
        raise ValueError('Can\'t compute slope for array of length=1')
    x_mean = (n + 1) / 2
    x2_mean = (n + 1) * (2 * n + 1) / 6
    xy_mean = np.average(y, weights=np.arange(1, n + 1))
    y_mean = np.mean(y)
    slope = (xy_mean - x_mean * y_mean) / (x2_mean - x_mean * x_mean)
    return slope

================================================
FILE: requirements.txt
================================================
torch==1.13.1
transformers==4.4.2
tensorboardX==1.9
label-studio>=1.0.0
git+git://github.com/heartexlabs/label-studio-ml-backend@master#egg=label-studio-ml