Skip to content

GitLab

".github/git@developer.sourcefind.cn:dadigang/Ventoy.git" did not exist on "ba5978d2987cce2545e10696a6ca14c6d54fdc92"
Unverified Commit d5712f7c authored by Thomas Wolf's avatar Thomas Wolf Committed by GitHub
Browse files

Merge branch 'master' into check-link-validity

parents f230d91b 9c58b236
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 2895 additions and 686 deletions
.circleci/config.yml
View file @ d5712f7c
...@@ -70,6 +70,27 @@ jobs: ...@@ -70,6 +70,27 @@ jobs:
- run: sudo pip install pytest codecov pytest-cov - run: sudo pip install pytest codecov pytest-cov
- run: python -m pytest -sv ./transformers/tests/ --cov - run: python -m pytest -sv ./transformers/tests/ --cov
- run: codecov - run: codecov
build_py3_custom_tokenizers:
working_directory: ~/transformers
docker:
- image: circleci/python:3.5
steps:
- checkout
- run: sudo pip install --progress-bar off .
- run: sudo pip install pytest
- run: sudo pip install mecab-python3
- run: RUN_CUSTOM_TOKENIZERS=1 python -m pytest -sv ./transformers/tests/tokenization_bert_japanese_test.py
build_py2_custom_tokenizers:
working_directory: ~/transformers
docker:
- image: circleci/python:2.7
steps:
- checkout
- run: sudo pip install --progress-bar off .
- run: sudo pip install pytest
- run: sudo apt-get -y install libmecab-dev mecab mecab-ipadic-utf8 swig
- run: sudo pip install mecab-python
- run: RUN_CUSTOM_TOKENIZERS=1 python -m pytest -sv ./transformers/tests/tokenization_bert_japanese_test.py
deploy_doc: deploy_doc:
working_directory: ~/transformers working_directory: ~/transformers
docker: docker:
...@@ -102,6 +123,8 @@ workflows: ...@@ -102,6 +123,8 @@ workflows:
build_and_test: build_and_test:
jobs: jobs:
- repository_consistency - repository_consistency
- build_py3_custom_tokenizers
- build_py2_custom_tokenizers
- build_py3_torch_and_tf - build_py3_torch_and_tf
- build_py3_torch - build_py3_torch
- build_py3_tf - build_py3_tf
......
README.md
View file @ d5712f7c
...@@ -101,17 +101,26 @@ pip install [--editable] . ...@@ -101,17 +101,26 @@ pip install [--editable] .
A series of tests are included for the library and the example scripts. Library tests can be found in the [tests folder](https://github.com/huggingface/transformers/tree/master/transformers/tests) and examples tests in the [examples folder](https://github.com/huggingface/transformers/tree/master/examples). A series of tests are included for the library and the example scripts. Library tests can be found in the [tests folder](https://github.com/huggingface/transformers/tree/master/transformers/tests) and examples tests in the [examples folder](https://github.com/huggingface/transformers/tree/master/examples).
These tests can be run using `pytest` (install pytest if needed with `pip install pytest`). These tests can be run using `unittest` or `pytest` (install pytest if needed with `pip install pytest`).
Depending on which framework is installed (TensorFlow 2.0 and/or PyTorch), the irrelevant tests will be skipped. Ensure that both frameworks are installed if you want to execute all tests. Depending on which framework is installed (TensorFlow 2.0 and/or PyTorch), the irrelevant tests will be skipped. Ensure that both frameworks are installed if you want to execute all tests.
You can run the tests from the root of the cloned repository with the commands: You can run the tests from the root of the cloned repository with the commands:
```bash
python -m unittest discover -s transformers/tests -p "*test.py" -t .
python -m unittest discover -s examples -p "*test.py" -t examples
```
or
```bash ```bash
python -m pytest -sv ./transformers/tests/ python -m pytest -sv ./transformers/tests/
python -m pytest -sv ./examples/ python -m pytest -sv ./examples/
``` ```
By default, slow tests are skipped. Set the `RUN_SLOW` environment variable to `yes` to run them.
### Do you want to run a Transformer model on a mobile device? ### Do you want to run a Transformer model on a mobile device?
You should check out our [`swift-coreml-transformers`](https://github.com/huggingface/swift-coreml-transformers) repo. You should check out our [`swift-coreml-transformers`](https://github.com/huggingface/swift-coreml-transformers) repo.
...@@ -134,7 +143,7 @@ At some point in the future, you'll be able to seamlessly move from pre-training ...@@ -134,7 +143,7 @@ At some point in the future, you'll be able to seamlessly move from pre-training
8. **[DistilBERT](https://github.com/huggingface/transformers/tree/master/examples/distillation)** (from HuggingFace), released together with the paper [DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter](https://arxiv.org/abs/1910.01108) by Victor Sanh, Lysandre Debut and Thomas Wolf. The same method has been applied to compress GPT2 into [DistilGPT2](https://github.com/huggingface/transformers/tree/master/examples/distillation), RoBERTa into [DistilRoBERTa](https://github.com/huggingface/transformers/tree/master/examples/distillation), Multilingual BERT into [DistilmBERT](https://github.com/huggingface/transformers/tree/master/examples/distillation) and a German version of DistilBERT. 8. **[DistilBERT](https://github.com/huggingface/transformers/tree/master/examples/distillation)** (from HuggingFace), released together with the paper [DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter](https://arxiv.org/abs/1910.01108) by Victor Sanh, Lysandre Debut and Thomas Wolf. The same method has been applied to compress GPT2 into [DistilGPT2](https://github.com/huggingface/transformers/tree/master/examples/distillation), RoBERTa into [DistilRoBERTa](https://github.com/huggingface/transformers/tree/master/examples/distillation), Multilingual BERT into [DistilmBERT](https://github.com/huggingface/transformers/tree/master/examples/distillation) and a German version of DistilBERT.
9. **[CTRL](https://github.com/salesforce/ctrl/)** (from Salesforce) released with the paper [CTRL: A Conditional Transformer Language Model for Controllable Generation](https://arxiv.org/abs/1909.05858) by Nitish Shirish Keskar*, Bryan McCann*, Lav R. Varshney, Caiming Xiong and Richard Socher. 9. **[CTRL](https://github.com/salesforce/ctrl/)** (from Salesforce) released with the paper [CTRL: A Conditional Transformer Language Model for Controllable Generation](https://arxiv.org/abs/1909.05858) by Nitish Shirish Keskar*, Bryan McCann*, Lav R. Varshney, Caiming Xiong and Richard Socher.
10. **[CamemBERT](https://camembert-model.fr)** (from Inria/Facebook/Sorbonne) released with the paper [CamemBERT: a Tasty French Language Model](https://arxiv.org/abs/1911.03894) by Louis Martin*, Benjamin Muller*, Pedro Javier Ortiz Suárez*, Yoann Dupont, Laurent Romary, Éric Villemonte de la Clergerie, Djamé Seddah and Benoît Sagot. 10. **[CamemBERT](https://camembert-model.fr)** (from Inria/Facebook/Sorbonne) released with the paper [CamemBERT: a Tasty French Language Model](https://arxiv.org/abs/1911.03894) by Louis Martin*, Benjamin Muller*, Pedro Javier Ortiz Suárez*, Yoann Dupont, Laurent Romary, Éric Villemonte de la Clergerie, Djamé Seddah and Benoît Sagot.
11. **[ALBERT](https://github.com/google-research/google-research/tree/master/albert)** (from Google Research and the Toyota Technological Institute at Chicago) released with the paper [ALBERT: A Lite BERT for Self-supervised Learning of Language Representations](https://arxiv.org/abs/1909.11942), by Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut. 11. **[ALBERT](https://github.com/google-research/ALBERT)** (from Google Research and the Toyota Technological Institute at Chicago) released with the paper [ALBERT: A Lite BERT for Self-supervised Learning of Language Representations](https://arxiv.org/abs/1909.11942), by Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut.
11. Want to contribute a new model? We have added a **detailed guide and templates** to guide you in the process of adding a new model. You can find them in the [`templates`](./templates) folder of the repository. Be sure to check the [contributing guidelines](./CONTRIBUTING.md) and contact the maintainers or open an issue to collect feedbacks before starting your PR. 11. Want to contribute a new model? We have added a **detailed guide and templates** to guide you in the process of adding a new model. You can find them in the [`templates`](./templates) folder of the repository. Be sure to check the [contributing guidelines](./CONTRIBUTING.md) and contact the maintainers or open an issue to collect feedbacks before starting your PR.
These implementations have been tested on several datasets (see the example scripts) and should match the performances of the original implementations (e.g. ~93 F1 on SQuAD for BERT Whole-Word-Masking, ~88 F1 on RocStories for OpenAI GPT, ~18.3 perplexity on WikiText 103 for Transformer-XL, ~0.916 Peason R coefficient on STS-B for XLNet). You can find more details on the performances in the Examples section of the [documentation](https://huggingface.co/transformers/examples.html). These implementations have been tested on several datasets (see the example scripts) and should match the performances of the original implementations (e.g. ~93 F1 on SQuAD for BERT Whole-Word-Masking, ~88 F1 on RocStories for OpenAI GPT, ~18.3 perplexity on WikiText 103 for Transformer-XL, ~0.916 Peason R coefficient on STS-B for XLNet). You can find more details on the performances in the Examples section of the [documentation](https://huggingface.co/transformers/examples.html).
......
docs/source/index.rst
View file @ d5712f7c
...@@ -49,7 +49,7 @@ The library currently contains PyTorch and Tensorflow implementations, pre-train ...@@ -49,7 +49,7 @@ The library currently contains PyTorch and Tensorflow implementations, pre-train
8. `DistilBERT <https://huggingface.co/transformers/model_doc/distilbert.html>`_ (from HuggingFace) released together with the paper `DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter <https://arxiv.org/abs/1910.01108>`_ by Victor Sanh, Lysandre Debut and Thomas Wolf. The same method has been applied to compress GPT2 into `DistilGPT2 <https://github.com/huggingface/transformers/tree/master/examples/distillation>`_. 8. `DistilBERT <https://huggingface.co/transformers/model_doc/distilbert.html>`_ (from HuggingFace) released together with the paper `DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter <https://arxiv.org/abs/1910.01108>`_ by Victor Sanh, Lysandre Debut and Thomas Wolf. The same method has been applied to compress GPT2 into `DistilGPT2 <https://github.com/huggingface/transformers/tree/master/examples/distillation>`_.
9. `CTRL <https://github.com/pytorch/fairseq/tree/master/examples/ctrl>`_ (from Salesforce), released together with the paper `CTRL: A Conditional Transformer Language Model for Controllable Generation <https://www.github.com/salesforce/ctrl>`_ by Nitish Shirish Keskar*, Bryan McCann*, Lav R. Varshney, Caiming Xiong and Richard Socher. 9. `CTRL <https://github.com/pytorch/fairseq/tree/master/examples/ctrl>`_ (from Salesforce), released together with the paper `CTRL: A Conditional Transformer Language Model for Controllable Generation <https://www.github.com/salesforce/ctrl>`_ by Nitish Shirish Keskar*, Bryan McCann*, Lav R. Varshney, Caiming Xiong and Richard Socher.
10. `CamemBERT <https://huggingface.co/transformers/model_doc/camembert.html>`_ (from FAIR, Inria, Sorbonne Université) released together with the paper `CamemBERT: a Tasty French Language Model <https://arxiv.org/abs/1911.03894>`_ by Louis Martin, Benjamin Muller, Pedro Javier Ortiz Suarez, Yoann Dupont, Laurent Romary, Eric Villemonte de la Clergerie, Djame Seddah, and Benoît Sagot. 10. `CamemBERT <https://huggingface.co/transformers/model_doc/camembert.html>`_ (from FAIR, Inria, Sorbonne Université) released together with the paper `CamemBERT: a Tasty French Language Model <https://arxiv.org/abs/1911.03894>`_ by Louis Martin, Benjamin Muller, Pedro Javier Ortiz Suarez, Yoann Dupont, Laurent Romary, Eric Villemonte de la Clergerie, Djame Seddah, and Benoît Sagot.
11. `ALBERT <https://github.com/pytorch/fairseq/tree/master/examples/albert>`_ (from Google Research), released together with the paper a `ALBERT: A Lite BERT for Self-supervised Learning of Language Representations <https://arxiv.org/abs/1909.11942>`_ by Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut. 11. `ALBERT <https://github.com/google-research/ALBERT>`_ (from Google Research), released together with the paper a `ALBERT: A Lite BERT for Self-supervised Learning of Language Representations <https://arxiv.org/abs/1909.11942>`_ by Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut.
.. toctree:: .. toctree::
:maxdepth: 2 :maxdepth: 2
......
docs/source/installation.md
View file @ d5712f7c
...@@ -24,15 +24,24 @@ pip install [--editable] . ...@@ -24,15 +24,24 @@ pip install [--editable] .
An extensive test suite is included to test the library behavior and several examples. Library tests can be found in the [tests folder](https://github.com/huggingface/transformers/tree/master/transformers/tests) and examples tests in the [examples folder](https://github.com/huggingface/transformers/tree/master/examples). An extensive test suite is included to test the library behavior and several examples. Library tests can be found in the [tests folder](https://github.com/huggingface/transformers/tree/master/transformers/tests) and examples tests in the [examples folder](https://github.com/huggingface/transformers/tree/master/examples).
Tests can be run using `pytest` (install pytest if needed with `pip install pytest`). Tests can be run using `unittest` or `pytest` (install pytest if needed with `pip install pytest`).
Run all the tests from the root of the cloned repository with the commands: Run all the tests from the root of the cloned repository with the commands:
```bash
python -m unittest discover -s transformers/tests -p "*test.py" -t .
python -m unittest discover -s examples -p "*test.py" -t examples
```
or
``` bash ``` bash
python -m pytest -sv ./transformers/tests/ python -m pytest -sv ./transformers/tests/
python -m pytest -sv ./examples/ python -m pytest -sv ./examples/
``` ```
By default, slow tests are skipped. Set the `RUN_SLOW` environment variable to `yes` to run them.
## OpenAI GPT original tokenization workflow ## OpenAI GPT original tokenization workflow
If you want to reproduce the original tokenization process of the `OpenAI GPT` paper, you will need to install `ftfy` (use version 4.4.3 if you are using Python 2) and `SpaCy`: If you want to reproduce the original tokenization process of the `OpenAI GPT` paper, you will need to install `ftfy` (use version 4.4.3 if you are using Python 2) and `SpaCy`:
......
docs/source/main_classes/optimizer_schedules.rst
View file @ d5712f7c
...@@ -5,6 +5,7 @@ The ``.optimization`` module provides: ...@@ -5,6 +5,7 @@ The ``.optimization`` module provides:
- an optimizer with weight decay fixed that can be used to fine-tuned models, and - an optimizer with weight decay fixed that can be used to fine-tuned models, and
- several schedules in the form of schedule objects that inherit from ``_LRSchedule``: - several schedules in the form of schedule objects that inherit from ``_LRSchedule``:
- a gradient accumulation class to accumulate the gradients of multiple batches
``AdamW`` ``AdamW``
~~~~~~~~~~~~~~~~ ~~~~~~~~~~~~~~~~
...@@ -12,6 +13,15 @@ The ``.optimization`` module provides: ...@@ -12,6 +13,15 @@ The ``.optimization`` module provides:
.. autoclass:: transformers.AdamW .. autoclass:: transformers.AdamW
:members: :members:
``AdamWeightDecay``
~~~~~~~~~~~~~~~~~~~
.. autoclass:: transformers.AdamWeightDecay
:members:
.. autofunction:: transformers.create_optimizer
:members:
Schedules Schedules
---------------------------------------------------- ----------------------------------------------------
...@@ -49,3 +59,17 @@ Learning Rate Schedules ...@@ -49,3 +59,17 @@ Learning Rate Schedules
.. image:: /imgs/warmup_linear_schedule.png .. image:: /imgs/warmup_linear_schedule.png
:target: /imgs/warmup_linear_schedule.png :target: /imgs/warmup_linear_schedule.png
:alt: :alt:
``Warmup``
~~~~~~~~~~~~~~~~
.. autoclass:: transformers.Warmup
:members:
Gradient Strategies
----------------------------------------------------
``GradientAccumulator``
~~~~~~~~~~~~~~~~~~~~~~~
.. autoclass:: transformers.GradientAccumulator
docs/source/main_classes/processors.rst
View file @ d5712f7c
...@@ -54,8 +54,7 @@ Additionally, the following method can be used to load values from a data file ...@@ -54,8 +54,7 @@ Additionally, the following method can be used to load values from a data file
Example usage Example usage
^^^^^^^^^^^^^^^^^^^^^^^^^ ^^^^^^^^^^^^^^^^^^^^^^^^^
An example using these processors is given in the An example using these processors is given in the `run_glue.py <https://github.com/huggingface/pytorch-transformers/blob/master/examples/run_glue.py>`__ script.
`run_glue.py <https://github.com/huggingface/pytorch-transformers/blob/master/examples/run_glue.py>`__ script.
XNLI XNLI
...@@ -74,8 +73,81 @@ This library hosts the processor to load the XNLI data: ...@@ -74,8 +73,81 @@ This library hosts the processor to load the XNLI data:
Please note that since the gold labels are available on the test set, evaluation is performed on the test set. Please note that since the gold labels are available on the test set, evaluation is performed on the test set.
An example using these processors is given in the
`run_xnli.py <https://github.com/huggingface/pytorch-transformers/blob/master/examples/run_xnli.py>`__ script.
SQuAD
~~~~~~~~~~~~~~~~~~~~~
`The Stanford Question Answering Dataset (SQuAD) <https://rajpurkar.github.io/SQuAD-explorer//>`__ is a benchmark that evaluates
the performance of models on question answering. Two versions are available, v1.1 and v2.0. The first version (v1.1) was released together with the paper
`SQuAD: 100,000+ Questions for Machine Comprehension of Text <https://arxiv.org/abs/1606.05250>`__. The second version (v2.0) was released alongside
the paper `Know What You Don't Know: Unanswerable Questions for SQuAD <https://arxiv.org/abs/1806.03822>`__.
This library hosts a processor for each of the two versions:
Processors
^^^^^^^^^^^^^^^^^^^^^^^^^
Those processors are:
- :class:`~transformers.data.processors.utils.SquadV1Processor`
- :class:`~transformers.data.processors.utils.SquadV2Processor`
They both inherit from the abstract class :class:`~transformers.data.processors.utils.SquadProcessor`
.. autoclass:: transformers.data.processors.squad.SquadProcessor
:members:
Additionally, the following method can be used to convert SQuAD examples into :class:`~transformers.data.processors.utils.SquadFeatures`
that can be used as model inputs.
.. automethod:: transformers.data.processors.squad.squad_convert_examples_to_features
These processors as well as the aforementionned method can be used with files containing the data as well as with the `tensorflow_datasets` package.
Examples are given below.
Example usage Example usage
^^^^^^^^^^^^^^^^^^^^^^^^^ ^^^^^^^^^^^^^^^^^^^^^^^^^
Here is an example using the processors as well as the conversion method using data files:
An example using these processors is given in the Example::
`run_xnli.py <https://github.com/huggingface/pytorch-transformers/blob/master/examples/run_xnli.py>`__ script.
\ No newline at end of file # Loading a V2 processor
processor = SquadV2Processor()
examples = processor.get_dev_examples(squad_v2_data_dir)
# Loading a V1 processor
processor = SquadV1Processor()
examples = processor.get_dev_examples(squad_v1_data_dir)
features = squad_convert_examples_to_features(
examples=examples,
tokenizer=tokenizer,
max_seq_length=max_seq_length,
doc_stride=args.doc_stride,
max_query_length=max_query_length,
is_training=not evaluate,
)
Using `tensorflow_datasets` is as easy as using a data file:
Example::
# tensorflow_datasets only handle Squad V1.
tfds_examples = tfds.load("squad")
examples = SquadV1Processor().get_examples_from_dataset(tfds_examples, evaluate=evaluate)
features = squad_convert_examples_to_features(
examples=examples,
tokenizer=tokenizer,
max_seq_length=max_seq_length,
doc_stride=args.doc_stride,
max_query_length=max_query_length,
is_training=not evaluate,
)
Another example using these processors is given in the
`run_squad.py <https://github.com/huggingface/transformers/blob/master/examples/run_squad.py>`__ script.
docs/source/pretrained_models.rst
View file @ d5712f7c
...@@ -61,6 +61,24 @@ Here is the full list of the currently provided pretrained models together with ...@@ -61,6 +61,24 @@ Here is the full list of the currently provided pretrained models together with
| | ``bert-base-german-dbmdz-uncased`` | | 12-layer, 768-hidden, 12-heads, 110M parameters. | | | ``bert-base-german-dbmdz-uncased`` | | 12-layer, 768-hidden, 12-heads, 110M parameters. |
| | | | Trained on uncased German text by DBMDZ | | | | | Trained on uncased German text by DBMDZ |
| | | (see `details on dbmdz repository <https://github.com/dbmdz/german-bert>`__). | | | | (see `details on dbmdz repository <https://github.com/dbmdz/german-bert>`__). |
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
| | ``bert-base-japanese`` | | 12-layer, 768-hidden, 12-heads, 110M parameters. |
| | | | Trained on Japanese text. Text is tokenized with MeCab and WordPiece. |
| | | | `MeCab <https://taku910.github.io/mecab/>`__ is required for tokenization. |
| | | (see `details on cl-tohoku repository <https://github.com/cl-tohoku/bert-japanese>`__). |
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
| | ``bert-base-japanese-whole-word-masking`` | | 12-layer, 768-hidden, 12-heads, 110M parameters. |
| | | | Trained on Japanese text using Whole-Word-Masking. Text is tokenized with MeCab and WordPiece. |
| | | | `MeCab <https://taku910.github.io/mecab/>`__ is required for tokenization. |
| | | (see `details on cl-tohoku repository <https://github.com/cl-tohoku/bert-japanese>`__). |
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
| | ``bert-base-japanese-char`` | | 12-layer, 768-hidden, 12-heads, 110M parameters. |
| | | | Trained on Japanese text. Text is tokenized into characters. |
| | | (see `details on cl-tohoku repository <https://github.com/cl-tohoku/bert-japanese>`__). |
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
| | ``bert-base-japanese-char-whole-word-masking`` | | 12-layer, 768-hidden, 12-heads, 110M parameters. |
| | | | Trained on Japanese text using Whole-Word-Masking. Text is tokenized into characters. |
| | | (see `details on cl-tohoku repository <https://github.com/cl-tohoku/bert-japanese>`__). |
+-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+ +-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
| GPT | ``openai-gpt`` | | 12-layer, 768-hidden, 12-heads, 110M parameters. | | GPT | ``openai-gpt`` | | 12-layer, 768-hidden, 12-heads, 110M parameters. |
| | | | OpenAI GPT English model | | | | | OpenAI GPT English model |
...@@ -169,35 +187,35 @@ Here is the full list of the currently provided pretrained models together with ...@@ -169,35 +187,35 @@ Here is the full list of the currently provided pretrained models together with
+-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+ +-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
| ALBERT | ``albert-base-v1`` | | 12 repeating layers, 128 embedding, 768-hidden, 12-heads, 11M parameters | | ALBERT | ``albert-base-v1`` | | 12 repeating layers, 128 embedding, 768-hidden, 12-heads, 11M parameters |
| | | | ALBERT base model | | | | | ALBERT base model |
| | | (see `details <https://github.com/google-research/google-research/tree/master/albert>`__) | | | | (see `details <https://github.com/google-research/ALBERT>`__) |
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+ | +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
| | ``albert-large-v1`` | | 24 repeating layers, 128 embedding, 1024-hidden, 16-heads, 17M parameters | | | ``albert-large-v1`` | | 24 repeating layers, 128 embedding, 1024-hidden, 16-heads, 17M parameters |
| | | | ALBERT large model | | | | | ALBERT large model |
| | | (see `details <https://github.com/google-research/google-research/tree/master/albert>`__) | | | | (see `details <https://github.com/google-research/ALBERT>`__) |
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+ | +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
| | ``albert-xlarge-v1`` | | 24 repeating layers, 128 embedding, 2048-hidden, 16-heads, 58M parameters | | | ``albert-xlarge-v1`` | | 24 repeating layers, 128 embedding, 2048-hidden, 16-heads, 58M parameters |
| | | | ALBERT xlarge model | | | | | ALBERT xlarge model |
| | | (see `details <https://github.com/google-research/google-research/tree/master/albert>`__) | | | | (see `details <https://github.com/google-research/ALBERT>`__) |
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+ | +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
| | ``albert-xxlarge-v1`` | | 12 repeating layer, 128 embedding, 4096-hidden, 64-heads, 223M parameters | | | ``albert-xxlarge-v1`` | | 12 repeating layer, 128 embedding, 4096-hidden, 64-heads, 223M parameters |
| | | | ALBERT xxlarge model | | | | | ALBERT xxlarge model |
| | | (see `details <https://github.com/google-research/google-research/tree/master/albert>`__) | | | | (see `details <https://github.com/google-research/ALBERT>`__) |
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+ | +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
| | ``albert-base-v2`` | | 12 repeating layers, 128 embedding, 768-hidden, 12-heads, 11M parameters | | | ``albert-base-v2`` | | 12 repeating layers, 128 embedding, 768-hidden, 12-heads, 11M parameters |
| | | | ALBERT base model with no dropout, additional training data and longer training | | | | | ALBERT base model with no dropout, additional training data and longer training |
| | | (see `details <https://github.com/google-research/google-research/tree/master/albert>`__) | | | | (see `details <https://github.com/google-research/ALBERT>`__) |
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+ | +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
| | ``albert-large-v2`` | | 24 repeating layers, 128 embedding, 1024-hidden, 16-heads, 17M parameters | | | ``albert-large-v2`` | | 24 repeating layers, 128 embedding, 1024-hidden, 16-heads, 17M parameters |
| | | | ALBERT large model with no dropout, additional training data and longer training | | | | | ALBERT large model with no dropout, additional training data and longer training |
| | | (see `details <https://github.com/google-research/google-research/tree/master/albert>`__) | | | | (see `details <https://github.com/google-research/ALBERT>`__) |
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+ | +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
| | ``albert-xlarge-v2`` | | 24 repeating layers, 128 embedding, 2048-hidden, 16-heads, 58M parameters | | | ``albert-xlarge-v2`` | | 24 repeating layers, 128 embedding, 2048-hidden, 16-heads, 58M parameters |
| | | | ALBERT xlarge model with no dropout, additional training data and longer training | | | | | ALBERT xlarge model with no dropout, additional training data and longer training |
| | | (see `details <https://github.com/google-research/google-research/tree/master/albert>`__) | | | | (see `details <https://github.com/google-research/ALBERT>`__) |
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+ | +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
| | ``albert-xxlarge-v2`` | | 12 repeating layer, 128 embedding, 4096-hidden, 64-heads, 223M parameters | | | ``albert-xxlarge-v2`` | | 12 repeating layer, 128 embedding, 4096-hidden, 64-heads, 223M parameters |
| | | | ALBERT xxlarge model with no dropout, additional training data and longer training | | | | | ALBERT xxlarge model with no dropout, additional training data and longer training |
| | | (see `details <https://github.com/google-research/google-research/tree/master/albert>`__) | | | | (see `details <https://github.com/google-research/ALBERT>`__) |
+-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+ +-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
......
examples/README.md
View file @ d5712f7c
...@@ -24,7 +24,6 @@ pip install -r ./examples/requirements.txt ...@@ -24,7 +24,6 @@ pip install -r ./examples/requirements.txt
| [Multiple Choice](#multiple-choice) | Examples running BERT/XLNet/RoBERTa on the SWAG/RACE/ARC tasks. | [Multiple Choice](#multiple-choice) | Examples running BERT/XLNet/RoBERTa on the SWAG/RACE/ARC tasks.
| [Named Entity Recognition](#named-entity-recognition) | Using BERT for Named Entity Recognition (NER) on the CoNLL 2003 dataset, examples with distributed training. | | [Named Entity Recognition](#named-entity-recognition) | Using BERT for Named Entity Recognition (NER) on the CoNLL 2003 dataset, examples with distributed training. |
| [XNLI](#xnli) | Examples running BERT/XLM on the XNLI benchmark. | | [XNLI](#xnli) | Examples running BERT/XLM on the XNLI benchmark. |
| [Abstractive summarization](#abstractive-summarization) | Fine-tuning the library models for abstractive summarization tasks on the CNN/Daily Mail dataset. |
## TensorFlow 2.0 Bert models on GLUE ## TensorFlow 2.0 Bert models on GLUE
...@@ -467,7 +466,8 @@ Training with the previously defined hyper-parameters yields the following resul ...@@ -467,7 +466,8 @@ Training with the previously defined hyper-parameters yields the following resul
## Named Entity Recognition ## Named Entity Recognition
Based on the script [`run_ner.py`](https://github.com/huggingface/transformers/blob/master/examples/run_ner.py). Based on the scripts [`run_ner.py`](https://github.com/huggingface/transformers/blob/master/examples/run_ner.py) for Pytorch and
[`run_tf_ner.py`(https://github.com/huggingface/transformers/blob/master/examples/run_tf_ner.py)] for Tensorflow 2.
This example fine-tune Bert Multilingual on GermEval 2014 (German NER). This example fine-tune Bert Multilingual on GermEval 2014 (German NER).
Details and results for the fine-tuning provided by @stefan-it. Details and results for the fine-tuning provided by @stefan-it.
...@@ -512,7 +512,7 @@ The GermEval 2014 dataset has much more labels than CoNLL-2002/2003 datasets, so ...@@ -512,7 +512,7 @@ The GermEval 2014 dataset has much more labels than CoNLL-2002/2003 datasets, so
cat train.txt dev.txt test.txt | cut -d " " -f 2 | grep -v "^$"| sort | uniq > labels.txt cat train.txt dev.txt test.txt | cut -d " " -f 2 | grep -v "^$"| sort | uniq > labels.txt
``` ```
### Training ### Prepare the run
Additional environment variables must be set: Additional environment variables must be set:
...@@ -524,6 +524,8 @@ export SAVE_STEPS=750 ...@@ -524,6 +524,8 @@ export SAVE_STEPS=750
export SEED=1 export SEED=1
``` ```
### Run the Pytorch version
To start training, just run: To start training, just run:
```bash ```bash
...@@ -544,7 +546,7 @@ python3 run_ner.py --data_dir ./ \ ...@@ -544,7 +546,7 @@ python3 run_ner.py --data_dir ./ \
If your GPU supports half-precision training, just add the `--fp16` flag. After training, the model will be both evaluated on development and test datasets. If your GPU supports half-precision training, just add the `--fp16` flag. After training, the model will be both evaluated on development and test datasets.
### Evaluation #### Evaluation
Evaluation on development dataset outputs the following for our example: Evaluation on development dataset outputs the following for our example:
...@@ -566,7 +568,7 @@ On the test dataset the following results could be achieved: ...@@ -566,7 +568,7 @@ On the test dataset the following results could be achieved:
10/04/2019 00:42:42 - INFO - __main__ - recall = 0.8624150210424085 10/04/2019 00:42:42 - INFO - __main__ - recall = 0.8624150210424085
``` ```
### Comparing BERT (large, cased), RoBERTa (large, cased) and DistilBERT (base, uncased) #### Comparing BERT (large, cased), RoBERTa (large, cased) and DistilBERT (base, uncased)
Here is a small comparison between BERT (large, cased), RoBERTa (large, cased) and DistilBERT (base, uncased) with the same hyperparameters as specified in the [example documentation](https://huggingface.co/transformers/examples.html#named-entity-recognition) (one run): Here is a small comparison between BERT (large, cased), RoBERTa (large, cased) and DistilBERT (base, uncased) with the same hyperparameters as specified in the [example documentation](https://huggingface.co/transformers/examples.html#named-entity-recognition) (one run):
...@@ -576,32 +578,70 @@ Here is a small comparison between BERT (large, cased), RoBERTa (large, cased) a ...@@ -576,32 +578,70 @@ Here is a small comparison between BERT (large, cased), RoBERTa (large, cased) a
| `roberta-large` | 95.96 | 91.87 | `roberta-large` | 95.96 | 91.87
| `distilbert-base-uncased` | 94.34 | 90.32 | `distilbert-base-uncased` | 94.34 | 90.32
## Abstractive summarization ### Run the Tensorflow 2 version
Based on the script
[`run_summarization_finetuning.py`](https://github.com/huggingface/transformers/blob/master/examples/run_summarization_finetuning.py).
Before running this script you should download **both** CNN and Daily Mail To start training, just run:
datasets from [Kyunghyun Cho's website](https://cs.nyu.edu/~kcho/DMQA/) (the
links next to "Stories") in the same folder. Then uncompress the archives by running:
```bash ```bash
tar -xvf cnn_stories.tgz && tar -xvf dailymail_stories.tgz python3 run_tf_ner.py --data_dir ./ \
--model_type bert \
--labels ./labels.txt \
--model_name_or_path $BERT_MODEL \
--output_dir $OUTPUT_DIR \
--max_seq_length $MAX_LENGTH \
--num_train_epochs $NUM_EPOCHS \
--per_device_train_batch_size $BATCH_SIZE \
--save_steps $SAVE_STEPS \
--seed $SEED \
--do_train \
--do_eval \
--do_predict
``` ```
note that the finetuning script **will not work** if you do not download both Such as the Pytorch version, if your GPU supports half-precision training, just add the `--fp16` flag. After training, the model will be both evaluated on development and test datasets.
datasets. We will refer as `$DATA_PATH` the path to where you uncompressed both
archive. #### Evaluation
Evaluation on development dataset outputs the following for our example:
```bash ```bash
export DATA_PATH=/path/to/dataset/ precision recall f1-score support
python run_summarization_finetuning.py \ LOCderiv 0.7619 0.6154 0.6809 52
--output_dir=output \ PERpart 0.8724 0.8997 0.8858 4057
--model_type=bert2bert \ OTHpart 0.9360 0.9466 0.9413 711
--model_name_or_path=bert2bert \ ORGpart 0.7015 0.6989 0.7002 269
--do_train \ LOCpart 0.7668 0.8488 0.8057 496
--data_path=$DATA_PATH \ LOC 0.8745 0.9191 0.8963 235
ORGderiv 0.7723 0.8571 0.8125 91
OTHderiv 0.4800 0.6667 0.5581 18
OTH 0.5789 0.6875 0.6286 16
PERderiv 0.5385 0.3889 0.4516 18
PER 0.5000 0.5000 0.5000 2
ORG 0.0000 0.0000 0.0000 3
micro avg 0.8574 0.8862 0.8715 5968
macro avg 0.8575 0.8862 0.8713 5968
```
On the test dataset the following results could be achieved:
```bash
precision recall f1-score support
PERpart 0.8847 0.8944 0.8896 9397
OTHpart 0.9376 0.9353 0.9365 1639
ORGpart 0.7307 0.7044 0.7173 697
LOC 0.9133 0.9394 0.9262 561
LOCpart 0.8058 0.8157 0.8107 1150
ORG 0.0000 0.0000 0.0000 8
OTHderiv 0.5882 0.4762 0.5263 42
PERderiv 0.6571 0.5227 0.5823 44
OTH 0.4906 0.6667 0.5652 39
ORGderiv 0.7016 0.7791 0.7383 172
LOCderiv 0.8256 0.6514 0.7282 109
PER 0.0000 0.0000 0.0000 11
micro avg 0.8722 0.8774 0.8748 13869
macro avg 0.8712 0.8774 0.8740 13869
``` ```
## XNLI ## XNLI
......
examples/run_lm_finetuning.py
View file @ d5712f7c
...@@ -188,6 +188,13 @@ def train(args, train_dataset, model, tokenizer): ...@@ -188,6 +188,13 @@ def train(args, train_dataset, model, tokenizer):
] ]
optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon) optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total) scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
# Check if saved optimizer or scheduler states exist
if os.path.isfile(os.path.join(args.model_name_or_path, 'optimizer.pt')) and os.path.isfile(os.path.join(args.model_name_or_path, 'scheduler.pt')):
# Load in optimizer and scheduler states
optimizer.load_state_dict(torch.load(os.path.join(args.model_name_or_path, 'optimizer.pt')))
scheduler.load_state_dict(torch.load(os.path.join(args.model_name_or_path, 'scheduler.pt')))
if args.fp16: if args.fp16:
try: try:
from apex import amp from apex import amp
...@@ -216,17 +223,37 @@ def train(args, train_dataset, model, tokenizer): ...@@ -216,17 +223,37 @@ def train(args, train_dataset, model, tokenizer):
logger.info(" Total optimization steps = %d", t_total) logger.info(" Total optimization steps = %d", t_total)
global_step = 0 global_step = 0
epochs_trained = 0
steps_trained_in_current_epoch = 0
# Check if continuing training from a checkpoint
if os.path.exists(args.model_name_or_path):
# set global_step to gobal_step of last saved checkpoint from model path
global_step = int(args.model_name_or_path.split('-')[-1].split('/')[0])
epochs_trained = global_step // (len(train_dataloader) // args.gradient_accumulation_steps)
steps_trained_in_current_epoch = global_step % (len(train_dataloader) // args.gradient_accumulation_steps)
logger.info(" Continuing training from checkpoint, will skip to saved global_step")
logger.info(" Continuing training from epoch %d", epochs_trained)
logger.info(" Continuing training from global step %d", global_step)
logger.info(" Will skip the first %d steps in the first epoch", steps_trained_in_current_epoch)
tr_loss, logging_loss = 0.0, 0.0 tr_loss, logging_loss = 0.0, 0.0
model_to_resize = model.module if hasattr(model, 'module') else model # Take care of distributed/parallel training model_to_resize = model.module if hasattr(model, 'module') else model # Take care of distributed/parallel training
model_to_resize.resize_token_embeddings(len(tokenizer)) model_to_resize.resize_token_embeddings(len(tokenizer))
model.zero_grad() model.zero_grad()
train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0]) train_iterator = trange(epochs_trained, int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
set_seed(args) # Added here for reproducibility (even between python 2 and 3) set_seed(args) # Added here for reproducibility (even between python 2 and 3)
for _ in train_iterator: for _ in train_iterator:
epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0]) epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
for step, batch in enumerate(epoch_iterator): for step, batch in enumerate(epoch_iterator):
# Skip past any already trained steps if resuming training
if steps_trained_in_current_epoch > 0:
steps_trained_in_current_epoch -= 1
continue
inputs, labels = mask_tokens(batch, tokenizer, args) if args.mlm else (batch, batch) inputs, labels = mask_tokens(batch, tokenizer, args) if args.mlm else (batch, batch)
inputs = inputs.to(args.device) inputs = inputs.to(args.device)
labels = labels.to(args.device) labels = labels.to(args.device)
...@@ -274,11 +301,17 @@ def train(args, train_dataset, model, tokenizer): ...@@ -274,11 +301,17 @@ def train(args, train_dataset, model, tokenizer):
os.makedirs(output_dir) os.makedirs(output_dir)
model_to_save = model.module if hasattr(model, 'module') else model # Take care of distributed/parallel training model_to_save = model.module if hasattr(model, 'module') else model # Take care of distributed/parallel training
model_to_save.save_pretrained(output_dir) model_to_save.save_pretrained(output_dir)
tokenizer.save_pretrained(output_dir)
torch.save(args, os.path.join(output_dir, 'training_args.bin')) torch.save(args, os.path.join(output_dir, 'training_args.bin'))
logger.info("Saving model checkpoint to %s", output_dir) logger.info("Saving model checkpoint to %s", output_dir)
_rotate_checkpoints(args, checkpoint_prefix) _rotate_checkpoints(args, checkpoint_prefix)
torch.save(optimizer.state_dict(), os.path.join(output_dir, 'optimizer.pt'))
torch.save(scheduler.state_dict(), os.path.join(output_dir, 'scheduler.pt'))
logger.info("Saving optimizer and scheduler states to %s", output_dir)
if args.max_steps > 0 and global_step > args.max_steps: if args.max_steps > 0 and global_step > args.max_steps:
epoch_iterator.close() epoch_iterator.close()
break break
......
examples/run_squad.py
View file @ d5712f7c
...@@ -16,6 +16,8 @@ ...@@ -16,6 +16,8 @@
""" Finetuning the library models for question-answering on SQuAD (DistilBERT, Bert, XLM, XLNet).""" """ Finetuning the library models for question-answering on SQuAD (DistilBERT, Bert, XLM, XLNet)."""
from __future__ import absolute_import, division, print_function from __future__ import absolute_import, division, print_function
from transformers.data.processors.squad import SquadV1Processor, SquadV2Processor, SquadResult
from transformers.data.metrics.squad_metrics import compute_predictions_logits, compute_predictions_log_probs, squad_evaluate
import argparse import argparse
import logging import logging
...@@ -23,11 +25,9 @@ import os ...@@ -23,11 +25,9 @@ import os
import random import random
import glob import glob
import timeit import timeit
import numpy as np import numpy as np
import torch import torch
from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler, from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler, TensorDataset)
TensorDataset)
from torch.utils.data.distributed import DistributedSampler from torch.utils.data.distributed import DistributedSampler
try: try:
...@@ -44,18 +44,11 @@ from transformers import (WEIGHTS_NAME, BertConfig, ...@@ -44,18 +44,11 @@ from transformers import (WEIGHTS_NAME, BertConfig,
XLNetForQuestionAnswering, XLNetForQuestionAnswering,
XLNetTokenizer, XLNetTokenizer,
DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer, DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer,
AlbertConfig, AlbertForQuestionAnswering, AlbertTokenizer) AlbertConfig, AlbertForQuestionAnswering, AlbertTokenizer,
XLMConfig, XLMForQuestionAnswering, XLMTokenizer,
from transformers import AdamW, get_linear_schedule_with_warmup )
from utils_squad import (read_squad_examples, convert_examples_to_features, from transformers import AdamW, get_linear_schedule_with_warmup, squad_convert_examples_to_features
RawResult, write_predictions,
RawResultExtended, write_predictions_extended)
# The follwing import is the official SQuAD evaluation script (2.0).
# You can remove it from the dependencies if you are using this script outside of the library
# We've added it here for automated tests (see examples/test_examples.py file)
from utils_squad_evaluate import EVAL_OPTS, main as evaluate_on_squad
logger = logging.getLogger(__name__) logger = logging.getLogger(__name__)
...@@ -67,7 +60,8 @@ MODEL_CLASSES = { ...@@ -67,7 +60,8 @@ MODEL_CLASSES = {
'xlnet': (XLNetConfig, XLNetForQuestionAnswering, XLNetTokenizer), 'xlnet': (XLNetConfig, XLNetForQuestionAnswering, XLNetTokenizer),
'xlm': (XLMConfig, XLMForQuestionAnswering, XLMTokenizer), 'xlm': (XLMConfig, XLMForQuestionAnswering, XLMTokenizer),
'distilbert': (DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer), 'distilbert': (DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer),
'albert': (AlbertConfig, AlbertForQuestionAnswering, AlbertTokenizer) 'albert': (AlbertConfig, AlbertForQuestionAnswering, AlbertTokenizer),
'xlm': (XLMConfig, XLMForQuestionAnswering, XLMTokenizer)
} }
def set_seed(args): def set_seed(args):
...@@ -100,14 +94,16 @@ def train(args, train_dataset, model, tokenizer): ...@@ -100,14 +94,16 @@ def train(args, train_dataset, model, tokenizer):
optimizer_grouped_parameters = [ 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': args.weight_decay}, {'params': [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)], 'weight_decay': args.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} {'params': [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
] ]
optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon) optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total) scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
if args.fp16: if args.fp16:
try: try:
from apex import amp from apex import amp
except ImportError: except ImportError:
raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.") raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level) model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)
# multi-gpu training (should be after apex fp16 initialization) # multi-gpu training (should be after apex fp16 initialization)
...@@ -135,20 +131,26 @@ def train(args, train_dataset, model, tokenizer): ...@@ -135,20 +131,26 @@ def train(args, train_dataset, model, tokenizer):
model.zero_grad() model.zero_grad()
train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0]) train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
set_seed(args) # Added here for reproductibility (even between python 2 and 3) set_seed(args) # Added here for reproductibility (even between python 2 and 3)
for _ in train_iterator: for _ in train_iterator:
epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0]) epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
for step, batch in enumerate(epoch_iterator): for step, batch in enumerate(epoch_iterator):
model.train() model.train()
batch = tuple(t.to(args.device) for t in batch) batch = tuple(t.to(args.device) for t in batch)
inputs = {'input_ids': batch[0],
'attention_mask': batch[1], inputs = {
'start_positions': batch[3], 'input_ids': batch[0],
'end_positions': batch[4]} 'attention_mask': batch[1],
'start_positions': batch[3],
'end_positions': batch[4]
}
if args.model_type != 'distilbert': if args.model_type != 'distilbert':
inputs['token_type_ids'] = None if args.model_type == 'xlm' else batch[2] inputs['token_type_ids'] = None if args.model_type == 'xlm' else batch[2]
if args.model_type in ['xlnet', 'xlm']: if args.model_type in ['xlnet', 'xlm']:
inputs.update({'cls_index': batch[5], inputs.update({'cls_index': batch[5], 'p_mask': batch[6]})
'p_mask': batch[6]})
outputs = model(**inputs) outputs = model(**inputs)
loss = outputs[0] # model outputs are always tuple in transformers (see doc) loss = outputs[0] # model outputs are always tuple in transformers (see doc)
...@@ -175,8 +177,8 @@ def train(args, train_dataset, model, tokenizer): ...@@ -175,8 +177,8 @@ def train(args, train_dataset, model, tokenizer):
model.zero_grad() model.zero_grad()
global_step += 1 global_step += 1
# Log metrics
if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0: if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
# Log metrics
if args.local_rank == -1 and args.evaluate_during_training: # Only evaluate when single GPU otherwise metrics may not average well if args.local_rank == -1 and args.evaluate_during_training: # Only evaluate when single GPU otherwise metrics may not average well
results = evaluate(args, model, tokenizer) results = evaluate(args, model, tokenizer)
for key, value in results.items(): for key, value in results.items():
...@@ -185,8 +187,8 @@ def train(args, train_dataset, model, tokenizer): ...@@ -185,8 +187,8 @@ def train(args, train_dataset, model, tokenizer):
tb_writer.add_scalar('loss', (tr_loss - logging_loss)/args.logging_steps, global_step) tb_writer.add_scalar('loss', (tr_loss - logging_loss)/args.logging_steps, global_step)
logging_loss = tr_loss logging_loss = tr_loss
# Save model checkpoint
if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0: if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
# Save model checkpoint
output_dir = os.path.join(args.output_dir, 'checkpoint-{}'.format(global_step)) output_dir = os.path.join(args.output_dir, 'checkpoint-{}'.format(global_step))
if not os.path.exists(output_dir): if not os.path.exists(output_dir):
os.makedirs(output_dir) os.makedirs(output_dir)
...@@ -215,6 +217,7 @@ def evaluate(args, model, tokenizer, prefix=""): ...@@ -215,6 +217,7 @@ def evaluate(args, model, tokenizer, prefix=""):
os.makedirs(args.output_dir) os.makedirs(args.output_dir)
args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu) args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
# Note that DistributedSampler samples randomly # Note that DistributedSampler samples randomly
eval_sampler = SequentialSampler(dataset) eval_sampler = SequentialSampler(dataset)
eval_dataloader = DataLoader(dataset, sampler=eval_sampler, batch_size=args.eval_batch_size) eval_dataloader = DataLoader(dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
...@@ -227,38 +230,59 @@ def evaluate(args, model, tokenizer, prefix=""): ...@@ -227,38 +230,59 @@ def evaluate(args, model, tokenizer, prefix=""):
logger.info("***** Running evaluation {} *****".format(prefix)) logger.info("***** Running evaluation {} *****".format(prefix))
logger.info(" Num examples = %d", len(dataset)) logger.info(" Num examples = %d", len(dataset))
logger.info(" Batch size = %d", args.eval_batch_size) logger.info(" Batch size = %d", args.eval_batch_size)
all_results = [] all_results = []
start_time = timeit.default_timer() start_time = timeit.default_timer()
for batch in tqdm(eval_dataloader, desc="Evaluating"): for batch in tqdm(eval_dataloader, desc="Evaluating"):
model.eval() model.eval()
batch = tuple(t.to(args.device) for t in batch) batch = tuple(t.to(args.device) for t in batch)
with torch.no_grad(): with torch.no_grad():
inputs = {'input_ids': batch[0], inputs = {
'attention_mask': batch[1] 'input_ids': batch[0],
} 'attention_mask': batch[1]
}
if args.model_type != 'distilbert': if args.model_type != 'distilbert':
inputs['token_type_ids'] = None if args.model_type == 'xlm' else batch[2] # XLM don't use segment_ids inputs['token_type_ids'] = None if args.model_type == 'xlm' else batch[2] # XLM don't use segment_ids
example_indices = batch[3] example_indices = batch[3]
# XLNet and XLM use more arguments for their predictions
if args.model_type in ['xlnet', 'xlm']: if args.model_type in ['xlnet', 'xlm']:
inputs.update({'cls_index': batch[4], inputs.update({'cls_index': batch[4], 'p_mask': batch[5]})
'p_mask': batch[5]})
outputs = model(**inputs) outputs = model(**inputs)
for i, example_index in enumerate(example_indices): for i, example_index in enumerate(example_indices):
eval_feature = features[example_index.item()] eval_feature = features[example_index.item()]
unique_id = int(eval_feature.unique_id) unique_id = int(eval_feature.unique_id)
if args.model_type in ['xlnet', 'xlm']:
# XLNet uses a more complex post-processing procedure output = [to_list(output[i]) for output in outputs]
result = RawResultExtended(unique_id = unique_id,
start_top_log_probs = to_list(outputs[0][i]), # Some models (XLNet, XLM) use 5 arguments for their predictions, while the other "simpler"
start_top_index = to_list(outputs[1][i]), # models only use two.
end_top_log_probs = to_list(outputs[2][i]), if len(output) >= 5:
end_top_index = to_list(outputs[3][i]), start_logits = output[0]
cls_logits = to_list(outputs[4][i])) start_top_index = output[1]
end_logits = output[2]
end_top_index = output[3]
cls_logits = output[4]
result = SquadResult(
unique_id, start_logits, end_logits,
start_top_index=start_top_index,
end_top_index=end_top_index,
cls_logits=cls_logits
)
else: else:
result = RawResult(unique_id = unique_id, start_logits, end_logits = output
start_logits = to_list(outputs[0][i]), result = SquadResult(
end_logits = to_list(outputs[1][i])) unique_id, start_logits, end_logits
)
all_results.append(result) all_results.append(result)
evalTime = timeit.default_timer() - start_time evalTime = timeit.default_timer() - start_time
...@@ -267,84 +291,84 @@ def evaluate(args, model, tokenizer, prefix=""): ...@@ -267,84 +291,84 @@ def evaluate(args, model, tokenizer, prefix=""):
# Compute predictions # Compute predictions
output_prediction_file = os.path.join(args.output_dir, "predictions_{}.json".format(prefix)) output_prediction_file = os.path.join(args.output_dir, "predictions_{}.json".format(prefix))
output_nbest_file = os.path.join(args.output_dir, "nbest_predictions_{}.json".format(prefix)) output_nbest_file = os.path.join(args.output_dir, "nbest_predictions_{}.json".format(prefix))
if args.version_2_with_negative: if args.version_2_with_negative:
output_null_log_odds_file = os.path.join(args.output_dir, "null_odds_{}.json".format(prefix)) output_null_log_odds_file = os.path.join(args.output_dir, "null_odds_{}.json".format(prefix))
else: else:
output_null_log_odds_file = None output_null_log_odds_file = None
# XLNet and XLM use a more complex post-processing procedure
if args.model_type in ['xlnet', 'xlm']: if args.model_type in ['xlnet', 'xlm']:
# XLNet uses a more complex post-processing procedure start_n_top = model.config.start_n_top if hasattr(model, "config") else model.module.config.start_n_top
write_predictions_extended(examples, features, all_results, args.n_best_size, end_n_top = model.config.end_n_top if hasattr(model, "config") else model.module.config.end_n_top
predictions = compute_predictions_log_probs(examples, features, all_results, args.n_best_size,
args.max_answer_length, output_prediction_file, args.max_answer_length, output_prediction_file,
output_nbest_file, output_null_log_odds_file, args.predict_file, output_nbest_file, output_null_log_odds_file,
model.config.start_n_top, model.config.end_n_top, start_n_top, end_n_top,
args.version_2_with_negative, tokenizer, args.verbose_logging) args.version_2_with_negative, tokenizer, args.verbose_logging)
else: else:
write_predictions(examples, features, all_results, args.n_best_size, predictions = compute_predictions_logits(examples, features, all_results, args.n_best_size,
args.max_answer_length, args.do_lower_case, output_prediction_file, args.max_answer_length, args.do_lower_case, output_prediction_file,
output_nbest_file, output_null_log_odds_file, args.verbose_logging, output_nbest_file, output_null_log_odds_file, args.verbose_logging,
args.version_2_with_negative, args.null_score_diff_threshold) args.version_2_with_negative, args.null_score_diff_threshold)
# Evaluate with the official SQuAD script # Compute the F1 and exact scores.
evaluate_options = EVAL_OPTS(data_file=args.predict_file, results = squad_evaluate(examples, predictions)
pred_file=output_prediction_file,
na_prob_file=output_null_log_odds_file)
results = evaluate_on_squad(evaluate_options)
return results return results
def load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=False): def load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=False):
if args.local_rank not in [-1, 0] and not evaluate: if args.local_rank not in [-1, 0] and not evaluate:
torch.distributed.barrier() # Make sure only the first process in distributed training process the dataset, and the others will use the cache torch.distributed.barrier() # Make sure only the first process in distributed training process the dataset, and the others will use the cache
# Load data features from cache or dataset file # Load data features from cache or dataset file
input_file = args.predict_file if evaluate else args.train_file input_dir = args.data_dir if args.data_dir else "."
cached_features_file = os.path.join(os.path.dirname(input_file), 'cached_{}_{}_{}'.format( cached_features_file = os.path.join(input_dir, 'cached_{}_{}_{}'.format(
'dev' if evaluate else 'train', 'dev' if evaluate else 'train',
list(filter(None, args.model_name_or_path.split('/'))).pop(), list(filter(None, args.model_name_or_path.split('/'))).pop(),
str(args.max_seq_length))) str(args.max_seq_length))
)
# Init features and dataset from cache if it exists
if os.path.exists(cached_features_file) and not args.overwrite_cache and not output_examples: if os.path.exists(cached_features_file) and not args.overwrite_cache and not output_examples:
logger.info("Loading features from cached file %s", cached_features_file) logger.info("Loading features from cached file %s", cached_features_file)
features = torch.load(cached_features_file) features_and_dataset = torch.load(cached_features_file)
features, dataset = features_and_dataset["features"], features_and_dataset["dataset"]
else: else:
logger.info("Creating features from dataset file at %s", input_file) logger.info("Creating features from dataset file at %s", input_dir)
examples = read_squad_examples(input_file=input_file,
is_training=not evaluate, if not args.data_dir:
version_2_with_negative=args.version_2_with_negative) try:
features = convert_examples_to_features(examples=examples, import tensorflow_datasets as tfds
tokenizer=tokenizer, except ImportError:
max_seq_length=args.max_seq_length, raise ImportError("If not data_dir is specified, tensorflow_datasets needs to be installed.")
doc_stride=args.doc_stride,
max_query_length=args.max_query_length, if args.version_2_with_negative:
is_training=not evaluate, logger.warn("tensorflow_datasets does not handle version 2 of SQuAD.")
cls_token_segment_id=2 if args.model_type in ['xlnet'] else 0,
pad_token_segment_id=3 if args.model_type in ['xlnet'] else 0, tfds_examples = tfds.load("squad")
cls_token_at_end=True if args.model_type in ['xlnet'] else False, examples = SquadV1Processor().get_examples_from_dataset(tfds_examples, evaluate=evaluate)
sequence_a_is_doc=True if args.model_type in ['xlnet'] else False) else:
processor = SquadV2Processor() if args.version_2_with_negative else SquadV1Processor()
examples = processor.get_dev_examples(args.data_dir) if evaluate else processor.get_train_examples(args.data_dir)
features, dataset = squad_convert_examples_to_features(
examples=examples,
tokenizer=tokenizer,
max_seq_length=args.max_seq_length,
doc_stride=args.doc_stride,
max_query_length=args.max_query_length,
is_training=not evaluate,
return_dataset='pt'
)
if args.local_rank in [-1, 0]: if args.local_rank in [-1, 0]:
logger.info("Saving features into cached file %s", cached_features_file) logger.info("Saving features into cached file %s", cached_features_file)
torch.save(features, cached_features_file) torch.save({"features": features, "dataset": dataset}, cached_features_file)
if args.local_rank == 0 and not evaluate: if args.local_rank == 0 and not evaluate:
torch.distributed.barrier() # Make sure only the first process in distributed training process the dataset, and the others will use the cache torch.distributed.barrier() # Make sure only the first process in distributed training process the dataset, and the others will use the cache
# Convert to Tensors and build dataset
all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in features], dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in features], dtype=torch.long)
all_cls_index = torch.tensor([f.cls_index for f in features], dtype=torch.long)
all_p_mask = torch.tensor([f.p_mask for f in features], dtype=torch.float)
if evaluate:
all_example_index = torch.arange(all_input_ids.size(0), dtype=torch.long)
dataset = TensorDataset(all_input_ids, all_input_mask, all_segment_ids,
all_example_index, all_cls_index, all_p_mask)
else:
all_start_positions = torch.tensor([f.start_position for f in features], dtype=torch.long)
all_end_positions = torch.tensor([f.end_position for f in features], dtype=torch.long)
dataset = TensorDataset(all_input_ids, all_input_mask, all_segment_ids,
all_start_positions, all_end_positions,
all_cls_index, all_p_mask)
if output_examples: if output_examples:
return dataset, examples, features return dataset, examples, features
return dataset return dataset
...@@ -354,10 +378,6 @@ def main(): ...@@ -354,10 +378,6 @@ def main():
parser = argparse.ArgumentParser() parser = argparse.ArgumentParser()
## Required parameters ## Required parameters
parser.add_argument("--train_file", default=None, type=str, required=True,
help="SQuAD json for training. E.g., train-v1.1.json")
parser.add_argument("--predict_file", default=None, type=str, required=True,
help="SQuAD json for predictions. E.g., dev-v1.1.json or test-v1.1.json")
parser.add_argument("--model_type", default=None, type=str, required=True, parser.add_argument("--model_type", default=None, type=str, required=True,
help="Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys())) help="Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys()))
parser.add_argument("--model_name_or_path", default=None, type=str, required=True, parser.add_argument("--model_name_or_path", default=None, type=str, required=True,
...@@ -366,6 +386,8 @@ def main(): ...@@ -366,6 +386,8 @@ def main():
help="The output directory where the model checkpoints and predictions will be written.") help="The output directory where the model checkpoints and predictions will be written.")
## Other parameters ## Other parameters
parser.add_argument("--data_dir", default=None, type=str,
help="The input data dir. Should contain the .json files for the task. If not specified, will run with tensorflow_datasets.")
parser.add_argument("--config_name", default="", type=str, parser.add_argument("--config_name", default="", type=str,
help="Pretrained config name or path if not the same as model_name") help="Pretrained config name or path if not the same as model_name")
parser.add_argument("--tokenizer_name", default="", type=str, parser.add_argument("--tokenizer_name", default="", type=str,
...@@ -450,6 +472,11 @@ def main(): ...@@ -450,6 +472,11 @@ def main():
parser.add_argument('--server_port', type=str, default='', help="Can be used for distant debugging.") parser.add_argument('--server_port', type=str, default='', help="Can be used for distant debugging.")
args = parser.parse_args() args = parser.parse_args()
args.predict_file = os.path.join(args.output_dir, 'predictions_{}_{}.txt'.format(
list(filter(None, args.model_name_or_path.split('/'))).pop(),
str(args.max_seq_length))
)
if os.path.exists(args.output_dir) and os.listdir(args.output_dir) and args.do_train and not args.overwrite_output_dir: if os.path.exists(args.output_dir) and os.listdir(args.output_dir) and args.do_train and not args.overwrite_output_dir:
raise ValueError("Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(args.output_dir)) raise ValueError("Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(args.output_dir))
......
examples/run_summarization_finetuning.py deleted 100644 → 0
View file @ f230d91b
# coding=utf-8
# Copyright 2019 The HuggingFace Inc. team.
# Copyright (c) 2019 The HuggingFace Inc. All rights reserved.
#
# 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.
""" Finetuning seq2seq models for sequence generation."""
import argparse
import functools
import logging
import os
import random
import sys
import numpy as np
from tqdm import tqdm, trange
import torch
from torch.optim import Adam
from torch.utils.data import DataLoader, RandomSampler, SequentialSampler
from transformers import (
AutoTokenizer,
BertForMaskedLM,
BertConfig,
PreTrainedEncoderDecoder,
Model2Model,
)
from utils_summarization import (
CNNDailyMailDataset,
encode_for_summarization,
fit_to_block_size,
build_lm_labels,
build_mask,
compute_token_type_ids,
)
logger = logging.getLogger(__name__)
logging.basicConfig(stream=sys.stdout, level=logging.INFO)
def set_seed(args):
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
# ------------
# Load dataset
# ------------
def load_and_cache_examples(args, tokenizer):
dataset = CNNDailyMailDataset(tokenizer, data_dir=args.data_dir)
return dataset
def collate(data, tokenizer, block_size):
""" List of tuple as an input. """
# remove the files with empty an story/summary, encode and fit to block
data = filter(lambda x: not (len(x[0]) == 0 or len(x[1]) == 0), data)
data = [
encode_for_summarization(story, summary, tokenizer) for story, summary in data
]
data = [
(
fit_to_block_size(story, block_size, tokenizer.pad_token_id),
fit_to_block_size(summary, block_size, tokenizer.pad_token_id),
)
for story, summary in data
]
stories = torch.tensor([story for story, summary in data])
summaries = torch.tensor([summary for story, summary in data])
encoder_token_type_ids = compute_token_type_ids(stories, tokenizer.cls_token_id)
encoder_mask = build_mask(stories, tokenizer.pad_token_id)
decoder_mask = build_mask(summaries, tokenizer.pad_token_id)
lm_labels = build_lm_labels(summaries, tokenizer.pad_token_id)
return (
stories,
summaries,
encoder_token_type_ids,
encoder_mask,
decoder_mask,
lm_labels,
)
# ----------
# Optimizers
# ----------
class BertSumOptimizer(object):
""" Specific optimizer for BertSum.
As described in [1], the authors fine-tune BertSum for abstractive
summarization using two Adam Optimizers with different warm-up steps and
learning rate. They also use a custom learning rate scheduler.
[1] Liu, Yang, and Mirella Lapata. "Text summarization with pretrained encoders."
arXiv preprint arXiv:1908.08345 (2019).
"""
def __init__(self, model, lr, warmup_steps, beta_1=0.99, beta_2=0.999, eps=1e-8):
self.encoder = model.encoder
self.decoder = model.decoder
self.lr = lr
self.warmup_steps = warmup_steps
self.optimizers = {
"encoder": Adam(
model.encoder.parameters(),
lr=lr["encoder"],
betas=(beta_1, beta_2),
eps=eps,
),
"decoder": Adam(
model.decoder.parameters(),
lr=lr["decoder"],
betas=(beta_1, beta_2),
eps=eps,
),
}
self._step = 0
def _update_rate(self, stack):
return self.lr[stack] * min(
self._step ** (-0.5), self._step * self.warmup_steps[stack] ** (-0.5)
)
def zero_grad(self):
self.optimizer_decoder.zero_grad()
self.optimizer_encoder.zero_grad()
def step(self):
self._step += 1
for stack, optimizer in self.optimizers.items():
new_rate = self._update_rate(stack)
for param_group in optimizer.param_groups:
param_group["lr"] = new_rate
optimizer.step()
# ------------
# Train
# ------------
def train(args, model, tokenizer):
""" Fine-tune the pretrained model on the corpus. """
set_seed(args)
# Load the data
args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
train_dataset = load_and_cache_examples(args, tokenizer)
train_sampler = RandomSampler(train_dataset)
model_collate_fn = functools.partial(collate, tokenizer=tokenizer, block_size=512)
train_dataloader = DataLoader(
train_dataset,
sampler=train_sampler,
batch_size=args.train_batch_size,
collate_fn=model_collate_fn,
)
# Training schedule
if args.max_steps > 0:
t_total = args.max_steps
args.num_train_epochs = t_total // (
len(train_dataloader) // args.gradient_accumulation_steps + 1
)
else:
t_total = (
len(train_dataloader)
// args.gradient_accumulation_steps
* args.num_train_epochs
)
# Prepare the optimizer
lr = {"encoder": 0.002, "decoder": 0.2}
warmup_steps = {"encoder": 20000, "decoder": 10000}
optimizer = BertSumOptimizer(model, lr, warmup_steps)
# Train
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_dataset))
logger.info(" Num Epochs = %d", args.num_train_epochs)
logger.info(
" Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size
)
logger.info(
" Total train batch size (w. parallel, distributed & accumulation) = %d",
args.train_batch_size * args.gradient_accumulation_steps
# * (torch.distributed.get_world_size() if args.local_rank != -1 else 1),
)
logger.info(" Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
logger.info(" Total optimization steps = %d", t_total)
model.zero_grad()
train_iterator = trange(args.num_train_epochs, desc="Epoch", disable=True)
global_step = 0
tr_loss = 0.0
for _ in train_iterator:
epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=True)
for step, batch in enumerate(epoch_iterator):
source, target, encoder_token_type_ids, encoder_mask, decoder_mask, lm_labels = batch
source = source.to(args.device)
target = target.to(args.device)
encoder_token_type_ids = encoder_token_type_ids.to(args.device)
encoder_mask = encoder_mask.to(args.device)
decoder_mask = decoder_mask.to(args.device)
lm_labels = lm_labels.to(args.device)
model.train()
outputs = model(
source,
target,
encoder_token_type_ids=encoder_token_type_ids,
encoder_attention_mask=encoder_mask,
decoder_attention_mask=decoder_mask,
decoder_lm_labels=lm_labels,
)
loss = outputs[0]
print(loss)
if args.gradient_accumulation_steps > 1:
loss /= args.gradient_accumulation_steps
loss.backward()
tr_loss += loss.item()
if (step + 1) % args.gradient_accumulation_steps == 0:
torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
optimizer.step()
model.zero_grad()
global_step += 1
if args.max_steps > 0 and global_step > args.max_steps:
epoch_iterator.close()
break
if args.max_steps > 0 and global_step > args.max_steps:
train_iterator.close()
break
return global_step, tr_loss / global_step
# ------------
# Train
# ------------
def evaluate(args, model, tokenizer, prefix=""):
set_seed(args)
args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
eval_dataset = load_and_cache_examples(args, tokenizer, evaluate=True)
eval_sampler = SequentialSampler(eval_dataset)
eval_dataloader = DataLoader(
eval_dataset, sampler=eval_sampler, batch_size=args.eval_batch_size
)
# multi-gpu evaluate
if args.n_gpu > 1:
model = torch.nn.DataParallel(model)
logger.info("***** Running evaluation {} *****".format(prefix))
logger.info(" Num examples = %d", len(eval_dataset))
logger.info(" Batch size = %d", args.eval_batch_size)
eval_loss = 0.0
nb_eval_steps = 0
model.eval()
for batch in tqdm(eval_dataloader, desc="Evaluating"):
source, target, encoder_token_type_ids, encoder_mask, decoder_mask, lm_labels = batch
source = source.to(args.device)
target = target.to(args.device)
encoder_token_type_ids = encoder_token_type_ids.to(args.device)
encoder_mask = encoder_mask.to(args.device)
decoder_mask = decoder_mask.to(args.device)
lm_labels = lm_labels.to(args.device)
with torch.no_grad():
outputs = model(
source,
target,
encoder_token_type_ids=encoder_token_type_ids,
encoder_attention_mask=encoder_mask,
decoder_attention_mask=decoder_mask,
decoder_lm_labels=lm_labels,
)
lm_loss = outputs[0]
eval_loss += lm_loss.mean().item()
nb_eval_steps += 1
eval_loss = eval_loss / nb_eval_steps
perplexity = torch.exp(torch.tensor(eval_loss))
result = {"perplexity": perplexity}
# Save the evaluation's results
output_eval_file = os.path.join(args.output_dir, "eval_results.txt")
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results {} *****".format(prefix))
for key in sorted(result.keys()):
logger.info(" %s = %s", key, str(result[key]))
writer.write("%s = %s\n" % (key, str(result[key])))
return result
def main():
parser = argparse.ArgumentParser()
# Required parameters
parser.add_argument(
"--data_dir",
default=None,
type=str,
required=True,
help="The input training data file (a text file).",
)
parser.add_argument(
"--output_dir",
default=None,
type=str,
required=True,
help="The output directory where the model predictions and checkpoints will be written.",
)
# Optional parameters
parser.add_argument(
"--gradient_accumulation_steps",
type=int,
default=1,
help="Number of updates steps to accumulate before performing a backward/update pass.",
)
parser.add_argument(
"--do_evaluate",
type=bool,
default=False,
help="Run model evaluation on out-of-sample data.",
)
parser.add_argument("--do_train", type=bool, default=False, help="Run training.")
parser.add_argument(
"--do_overwrite_output_dir",
type=bool,
default=False,
help="Whether to overwrite the output dir.",
)
parser.add_argument(
"--model_name_or_path",
default="bert-base-cased",
type=str,
help="The model checkpoint to initialize the encoder and decoder's weights with.",
)
parser.add_argument(
"--model_type",
default="bert",
type=str,
help="The decoder architecture to be fine-tuned.",
)
parser.add_argument(
"--max_grad_norm", default=1.0, type=float, help="Max gradient norm."
)
parser.add_argument(
"--max_steps",
default=-1,
type=int,
help="If > 0: set total number of training steps to perform. Override num_train_epochs.",
)
parser.add_argument(
"--to_cpu", default=False, type=bool, help="Whether to force training on CPU."
)
parser.add_argument(
"--num_train_epochs",
default=10,
type=int,
help="Total number of training epochs to perform.",
)
parser.add_argument(
"--per_gpu_train_batch_size",
default=4,
type=int,
help="Batch size per GPU/CPU for training.",
)
parser.add_argument("--seed", default=42, type=int)
args = parser.parse_args()
if (
os.path.exists(args.output_dir)
and os.listdir(args.output_dir)
and args.do_train
and not args.do_overwrite_output_dir
):
raise ValueError(
"Output directory ({}) already exists and is not empty. Use --do_overwrite_output_dir to overwrite.".format(
args.output_dir
)
)
# Set up training device
if args.to_cpu or not torch.cuda.is_available():
args.device = torch.device("cpu")
args.n_gpu = 0
else:
args.device = torch.device("cuda")
args.n_gpu = torch.cuda.device_count()
# Load pretrained model and tokenizer. The decoder's weights are randomly initialized.
tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path)
config = BertConfig.from_pretrained(args.model_name_or_path)
decoder_model = BertForMaskedLM(config)
model = Model2Model.from_pretrained(
args.model_name_or_path, decoder_model=decoder_model
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
logger.warning(
"Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
0,
args.device,
args.n_gpu,
False,
False,
)
logger.info("Training/evaluation parameters %s", args)
# Train the model
model.to(args.device)
if args.do_train:
global_step, tr_loss = train(args, model, tokenizer)
logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
logger.info("Saving model checkpoint to %s", args.output_dir)
# Save a trained model, configuration and tokenizer using `save_pretrained()`.
# They can then be reloaded using `from_pretrained()`
model_to_save = (
model.module if hasattr(model, "module") else model
) # Take care of distributed/parallel training
model_to_save.save_pretrained(args.output_dir)
tokenizer.save_pretrained(args.output_dir)
torch.save(args, os.path.join(args.output_dir, "training_arguments.bin"))
# Evaluate the model
results = {}
if args.do_evaluate:
checkpoints = []
logger.info("Evaluate the following checkpoints: %s", checkpoints)
for checkpoint in checkpoints:
encoder_checkpoint = os.path.join(checkpoint, "encoder")
decoder_checkpoint = os.path.join(checkpoint, "decoder")
model = PreTrainedEncoderDecoder.from_pretrained(
encoder_checkpoint, decoder_checkpoint
)
model.to(args.device)
results = "placeholder"
return results
if __name__ == "__main__":
main()
examples/run_tf_ner.py 0 → 100644
View file @ d5712f7c
This diff is collapsed.
examples/summarization/README.md 0 → 100644
View file @ d5712f7c
# Text Summarization with Pretrained Encoders
This folder contains part of the code necessary to reproduce the results on abstractive summarization from the article [Text Summarization with Pretrained Encoders](https://arxiv.org/pdf/1908.08345.pdf) by [Yang Liu](https://nlp-yang.github.io/) and [Mirella Lapata](https://homepages.inf.ed.ac.uk/mlap/). It can also be used to summarize any document.
The original code can be found on the Yang Liu's [github repository](https://github.com/nlpyang/PreSumm).
The model is loaded with the pre-trained weights for the abstractive summarization model trained on the CNN/Daily Mail dataset with an extractive and then abstractive tasks.
## Setup
```
git clone https://github.com/huggingface/transformers && cd transformers
pip install [--editable] .
pip install nltk py-rouge
cd examples/summarization
```
## Reproduce the authors' results on ROUGE
To be able to reproduce the authors' results on the CNN/Daily Mail dataset you first need to download both CNN and Daily Mail datasets [from Kyunghyun Cho's website](https://cs.nyu.edu/~kcho/DMQA/) (the links next to "Stories") in the same folder. Then uncompress the archives by running:
```bash
tar -xvf cnn_stories.tgz && tar -xvf dailymail_stories.tgz
```
And move all the stories to the same folder. We will refer as `$DATA_PATH` the path to where you uncompressed both archive. Then run the following in the same folder as `run_summarization.py`:
```bash
python run_summarization.py \
--documents_dir $DATA_PATH \
--summaries_output_dir $SUMMARIES_PATH \ # optional
--to_cpu false \
--batch_size 4 \
--min_length 50 \
--max_length 200 \
--beam_size 5 \
--alpha 0.95 \
--block_trigram true \
--compute_rouge true
```
The scripts executes on GPU if one is available and if `to_cpu` is not set to `true`. Inference on multiple GPUs is not suported yet. The ROUGE scores will be displayed in the console at the end of evaluation and written in a `rouge_scores.txt` file. The script takes 30 hours to compute with a single Tesla V100 GPU and a batch size of 10 (300,000 texts to summarize).
## Summarize any text
Put the documents that you would like to summarize in a folder (the path to which is referred to as `$DATA_PATH` below) and run the following in the same folder as `run_summarization.py`:
```bash
python run_summarization.py \
--documents_dir $DATA_PATH \
--summaries_output_dir $SUMMARIES_PATH \ # optional
--to_cpu false \
--batch_size 4 \
--min_length 50 \
--max_length 200 \
--beam_size 5 \
--alpha 0.95 \
--block_trigram true \
```
You may want to play around with `min_length`, `max_length` and `alpha` to suit your use case. If you want to compute ROUGE on another dataset you will need to tweak the stories/summaries import in `utils_summarization.py` and tell it where to fetch the reference summaries.
examples/summarization/configuration_bertabs.py 0 → 100644
View file @ d5712f7c
# coding=utf-8
# Copyright 2019 The HuggingFace Inc. team.
# Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
#
# 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.
""" BertAbs configuration """
import json
import logging
import sys
from transformers import PretrainedConfig
logger = logging.getLogger(__name__)
BERTABS_FINETUNED_CONFIG_MAP = {
"bertabs-finetuned-cnndm": "https://s3.amazonaws.com/models.huggingface.co/bert/remi/bertabs-finetuned-cnndm-extractive-abstractive-summarization-config.json",
}
class BertAbsConfig(PretrainedConfig):
r""" Class to store the configuration of the BertAbs model.
Arguments:
max_pos: int
The maximum sequence length that this model will be used with.
enc_layer: int
The numner of hidden layers in the Transformer encoder.
enc_hidden_size: int
The size of the encoder's layers.
enc_heads: int
The number of attention heads for each attention layer in the encoder.
enc_ff_size: int
The size of the encoder's feed-forward layers.
enc_dropout: int
The dropout probabilitiy for all fully connected layers in the
embeddings, layers, pooler and also the attention probabilities in
the encoder.
dec_layer: int
The numner of hidden layers in the decoder.
dec_hidden_size: int
The size of the decoder's layers.
dec_heads: int
The number of attention heads for each attention layer in the decoder.
dec_ff_size: int
The size of the decoder's feed-forward layers.
dec_dropout: int
The dropout probabilitiy for all fully connected layers in the
embeddings, layers, pooler and also the attention probabilities in
the decoder.
"""
pretrained_config_archive_map = BERTABS_FINETUNED_CONFIG_MAP
def __init__(
self,
vocab_size_or_config_json_file=30522,
max_pos=512,
enc_layers=6,
enc_hidden_size=512,
enc_heads=8,
enc_ff_size=512,
enc_dropout=0.2,
dec_layers=6,
dec_hidden_size=768,
dec_heads=8,
dec_ff_size=2048,
dec_dropout=0.2,
**kwargs,
):
super(BertAbsConfig, self).__init__(**kwargs)
if self._input_is_path_to_json(vocab_size_or_config_json_file):
path_to_json = vocab_size_or_config_json_file
with open(path_to_json, "r", encoding="utf-8") as reader:
json_config = json.loads(reader.read())
for key, value in json_config.items():
self.__dict__[key] = value
elif isinstance(vocab_size_or_config_json_file, int):
self.vocab_size = vocab_size_or_config_json_file
self.max_pos = max_pos
self.enc_layers = enc_layers
self.enc_hidden_size = enc_hidden_size
self.enc_heads = enc_heads
self.enc_ff_size = enc_ff_size
self.enc_dropout = enc_dropout
self.dec_layers = dec_layers
self.dec_hidden_size = dec_hidden_size
self.dec_heads = dec_heads
self.dec_ff_size = dec_ff_size
self.dec_dropout = dec_dropout
else:
raise ValueError(
"First argument must be either a vocabulary size (int)"
"or the path to a pretrained model config file (str)"
)
def _input_is_path_to_json(self, first_argument):
""" Checks whether the first argument passed to config
is the path to a JSON file that contains the config.
"""
is_python_2 = sys.version_info[0] == 2
if is_python_2:
return isinstance(first_argument, unicode)
else:
return isinstance(first_argument, str)
examples/summarization/convert_bertabs_original_pytorch_checkpoint.py 0 → 100644
View file @ d5712f7c
# coding=utf-8
# Copyright 2018 The HuggingFace Inc. team.
#
# 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.
""" Convert BertExtAbs's checkpoints.
The script looks like it is doing something trivial but it is not. The "weights"
proposed by the authors are actually the entire model pickled. We need to load
the model within the original codebase to be able to only save its `state_dict`.
"""
import argparse
from collections import namedtuple
import logging
import torch
from models.model_builder import AbsSummarizer # The authors' implementation
from model_bertabs import BertAbsSummarizer
from transformers import BertTokenizer
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)
SAMPLE_TEXT = 'Hello world! cécé herlolip'
BertAbsConfig = namedtuple(
"BertAbsConfig",
["temp_dir", "large", "use_bert_emb", "finetune_bert", "encoder", "share_emb", "max_pos", "enc_layers", "enc_hidden_size", "enc_heads", "enc_ff_size", "enc_dropout", "dec_layers", "dec_hidden_size", "dec_heads", "dec_ff_size", "dec_dropout"],
)
def convert_bertabs_checkpoints(path_to_checkpoints, dump_path):
""" Copy/paste and tweak the pre-trained weights provided by the creators
of BertAbs for the internal architecture.
"""
# Instantiate the authors' model with the pre-trained weights
config = BertAbsConfig(
temp_dir=".",
finetune_bert=False,
large=False,
share_emb=True,
use_bert_emb=False,
encoder="bert",
max_pos=512,
enc_layers=6,
enc_hidden_size=512,
enc_heads=8,
enc_ff_size=512,
enc_dropout=0.2,
dec_layers=6,
dec_hidden_size=768,
dec_heads=8,
dec_ff_size=2048,
dec_dropout=0.2,
)
checkpoints = torch.load(path_to_checkpoints, lambda storage, loc: storage)
original = AbsSummarizer(config, torch.device("cpu"), checkpoints)
original.eval()
new_model = BertAbsSummarizer(config, torch.device("cpu"))
new_model.eval()
# -------------------
# Convert the weights
# -------------------
logging.info("convert the model")
new_model.bert.load_state_dict(original.bert.state_dict())
new_model.decoder.load_state_dict(original.decoder.state_dict())
new_model.generator.load_state_dict(original.generator.state_dict())
# ----------------------------------
# Make sure the outpus are identical
# ----------------------------------
logging.info("Make sure that the models' outputs are identical")
tokenizer = BertTokenizer.from_pretrained("bert-base-uncased")
# prepare the model inputs
encoder_input_ids = tokenizer.encode("This is sample éàalj'-.")
encoder_input_ids.extend([tokenizer.pad_token_id] * (512 - len(encoder_input_ids)))
encoder_input_ids = torch.tensor(encoder_input_ids).unsqueeze(0)
decoder_input_ids = tokenizer.encode("This is sample 3 éàalj'-.")
decoder_input_ids.extend([tokenizer.pad_token_id] * (512 - len(decoder_input_ids)))
decoder_input_ids = torch.tensor(decoder_input_ids).unsqueeze(0)
# failsafe to make sure the weights reset does not affect the
# loaded weights.
assert torch.max(torch.abs(original.generator[0].weight - new_model.generator[0].weight)) == 0
# forward pass
src = encoder_input_ids
tgt = decoder_input_ids
segs = token_type_ids = None
clss = None
mask_src = encoder_attention_mask = None
mask_tgt = decoder_attention_mask = None
mask_cls = None
# The original model does not apply the geneator layer immediatly but rather in
# the beam search (where it combines softmax + linear layer). Since we already
# apply the softmax in our generation process we only apply the linear layer here.
# We make sure that the outputs of the full stack are identical
output_original_model = original(src, tgt, segs, clss, mask_src, mask_tgt, mask_cls)[0]
output_original_generator = original.generator(output_original_model)
output_converted_model = new_model(encoder_input_ids, decoder_input_ids, token_type_ids, encoder_attention_mask, decoder_attention_mask)[0]
output_converted_generator = new_model.generator(output_converted_model)
maximum_absolute_difference = torch.max(torch.abs(output_converted_model - output_original_model)).item()
print("Maximum absolute difference beween weights: {:.2f}".format(maximum_absolute_difference))
maximum_absolute_difference = torch.max(torch.abs(output_converted_generator - output_original_generator)).item()
print("Maximum absolute difference beween weights: {:.2f}".format(maximum_absolute_difference))
are_identical = torch.allclose(output_converted_model, output_original_model, atol=1e-3)
if are_identical:
logging.info("all weights are equal up to 1e-3")
else:
raise ValueError("the weights are different. The new model is likely different from the original one.")
# The model has been saved with torch.save(model) and this is bound to the exact
# directory structure. We save the state_dict instead.
logging.info("saving the model's state dictionary")
torch.save(new_model.state_dict(), "bertabs-finetuned-cnndm-extractive-abstractive-summarization-pytorch_model.bin")
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument(
"--bertabs_checkpoint_path",
default=None,
type=str,
required=True,
help="Path the official PyTorch dump.",
)
parser.add_argument(
"--pytorch_dump_folder_path",
default=None,
type=str,
required=True,
help="Path to the output PyTorch model.",
)
args = parser.parse_args()
convert_bertabs_checkpoints(
args.bertabs_checkpoint_path,
args.pytorch_dump_folder_path,
)
examples/summarization/modeling_bertabs.py 0 → 100644
View file @ d5712f7c
This diff is collapsed.
examples/summarization/requirements.txt 0 → 100644
View file @ d5712f7c
# progress bars in model download and training scripts
tqdm
# Accessing files from S3 directly.
boto3
# Used for downloading models over HTTP
requests
# For ROUGE
nltk
py-rouge
examples/summarization/run_summarization.py 0 → 100644
View file @ d5712f7c
#! /usr/bin/python3
import argparse
from collections import namedtuple
import logging
import os
import sys
import torch
from torch.utils.data import DataLoader, SequentialSampler
from tqdm import tqdm
from transformers import BertTokenizer
from modeling_bertabs import BertAbs, build_predictor
from utils_summarization import (
SummarizationDataset,
encode_for_summarization,
build_mask,
fit_to_block_size,
compute_token_type_ids,
)
logger = logging.getLogger(__name__)
logging.basicConfig(stream=sys.stdout, level=logging.INFO)
Batch = namedtuple(
"Batch", ["document_names", "batch_size", "src", "segs", "mask_src", "tgt_str"]
)
def evaluate(args):
tokenizer = BertTokenizer.from_pretrained("bert-base-uncased", do_lower_case=True)
model = BertAbs.from_pretrained("bertabs-finetuned-cnndm")
model.to(args.device)
model.eval()
symbols = {
"BOS": tokenizer.vocab["[unused0]"],
"EOS": tokenizer.vocab["[unused1]"],
"PAD": tokenizer.vocab["[PAD]"],
}
if args.compute_rouge:
reference_summaries = []
generated_summaries = []
import rouge
import nltk
nltk.download('punkt')
rouge_evaluator = rouge.Rouge(
metrics=['rouge-n', 'rouge-l'],
max_n=2,
limit_length=True,
length_limit=args.beam_size,
length_limit_type='words',
apply_avg=True,
apply_best=False,
alpha=0.5, # Default F1_score
weight_factor=1.2,
stemming=True,
)
# these (unused) arguments are defined to keep the compatibility
# with the legacy code and will be deleted in a next iteration.
args.result_path = ""
args.temp_dir = ""
data_iterator = build_data_iterator(args, tokenizer)
predictor = build_predictor(args, tokenizer, symbols, model)
logger.info("***** Running evaluation *****")
logger.info(" Number examples = %d", len(data_iterator.dataset))
logger.info(" Batch size = %d", args.batch_size)
logger.info("")
logger.info("***** Beam Search parameters *****")
logger.info(" Beam size = %d", args.beam_size)
logger.info(" Minimum length = %d", args.min_length)
logger.info(" Maximum length = %d", args.max_length)
logger.info(" Alpha (length penalty) = %.2f", args.alpha)
logger.info(" Trigrams %s be blocked", ("will" if args.block_trigram else "will NOT"))
for batch in tqdm(data_iterator):
batch_data = predictor.translate_batch(batch)
translations = predictor.from_batch(batch_data)
summaries = [format_summary(t) for t in translations]
save_summaries(summaries, args.summaries_output_dir, batch.document_names)
if args.compute_rouge:
reference_summaries += batch.tgt_str
generated_summaries += summaries
if args.compute_rouge:
scores = rouge_evaluator.get_scores(generated_summaries, reference_summaries)
str_scores = format_rouge_scores(scores)
save_rouge_scores(str_scores)
print(str_scores)
def save_summaries(summaries, path, original_document_name):
""" Write the summaries in fies that are prefixed by the original
files' name with the `_summary` appended.
Attributes:
original_document_names: List[string]
Name of the document that was summarized.
path: string
Path were the summaries will be written
summaries: List[string]
The summaries that we produced.
"""
for summary, document_name in zip(summaries, original_document_name):
# Prepare the summary file's name
if "." in document_name:
bare_document_name = ".".join(document_name.split(".")[:-1])
extension = document_name.split(".")[-1]
name = bare_document_name + "_summary." + extension
else:
name = document_name + "_summary"
file_path = os.path.join(path, name)
with open(file_path, "w") as output:
output.write(summary)
def format_summary(translation):
""" Transforms the output of the `from_batch` function
into nicely formatted summaries.
"""
raw_summary, _, _ = translation
summary = (
raw_summary.replace("[unused0]", "")
.replace("[unused3]", "")
.replace("[PAD]", "")
.replace("[unused1]", "")
.replace(r" +", " ")
.replace(" [unused2] ", ". ")
.replace("[unused2]", "")
.strip()
)
return summary
def format_rouge_scores(scores):
return """\n
****** ROUGE SCORES ******
** ROUGE 1
F1 >> {:.3f}
Precision >> {:.3f}
Recall >> {:.3f}
** ROUGE 2
F1 >> {:.3f}
Precision >> {:.3f}
Recall >> {:.3f}
** ROUGE L
F1 >> {:.3f}
Precision >> {:.3f}
Recall >> {:.3f}""".format(
scores['rouge-1']['f'],
scores['rouge-1']['p'],
scores['rouge-1']['r'],
scores['rouge-2']['f'],
scores['rouge-2']['p'],
scores['rouge-2']['r'],
scores['rouge-l']['f'],
scores['rouge-l']['p'],
scores['rouge-l']['r'],
)
def save_rouge_scores(str_scores):
with open("rouge_scores.txt", "w") as output:
output.write(str_scores)
#
# LOAD the dataset
#
def build_data_iterator(args, tokenizer):
dataset = load_and_cache_examples(args, tokenizer)
sampler = SequentialSampler(dataset)
collate_fn = lambda data: collate(data, tokenizer, block_size=512, device=args.device)
iterator = DataLoader(
dataset, sampler=sampler, batch_size=args.batch_size, collate_fn=collate_fn,
)
return iterator
def load_and_cache_examples(args, tokenizer):
dataset = SummarizationDataset(args.documents_dir)
return dataset
def collate(data, tokenizer, block_size, device):
""" Collate formats the data passed to the data loader.
In particular we tokenize the data batch after batch to avoid keeping them
all in memory. We output the data as a namedtuple to fit the original BertAbs's
API.
"""
data = [x for x in data if not len(x[1]) == 0] # remove empty_files
names = [name for name, _, _ in data]
summaries = [" ".join(summary_list) for _, _, summary_list in data]
encoded_text = [
encode_for_summarization(story, summary, tokenizer) for _, story, summary in data
]
encoded_stories = torch.tensor(
[
fit_to_block_size(story, block_size, tokenizer.pad_token_id)
for story, _ in encoded_text
]
)
encoder_token_type_ids = compute_token_type_ids(encoded_stories, tokenizer.cls_token_id)
encoder_mask = build_mask(encoded_stories, tokenizer.pad_token_id)
batch = Batch(
document_names=names,
batch_size=len(encoded_stories),
src=encoded_stories.to(device),
segs=encoder_token_type_ids.to(device),
mask_src=encoder_mask.to(device),
tgt_str=summaries,
)
return batch
def decode_summary(summary_tokens, tokenizer):
""" Decode the summary and return it in a format
suitable for evaluation.
"""
summary_tokens = summary_tokens.to("cpu").numpy()
summary = tokenizer.decode(summary_tokens)
sentences = summary.split(".")
sentences = [s + "." for s in sentences]
return sentences
def main():
""" The main function defines the interface with the users.
"""
parser = argparse.ArgumentParser()
parser.add_argument(
"--documents_dir",
default=None,
type=str,
required=True,
help="The folder where the documents to summarize are located.",
)
parser.add_argument(
"--summaries_output_dir",
default=None,
type=str,
required=False,
help="The folder in wich the summaries should be written. Defaults to the folder where the documents are",
)
parser.add_argument(
"--compute_rouge",
default=False,
type=bool,
required=False,
help="Compute the ROUGE metrics during evaluation. Only available for the CNN/DailyMail dataset.",
)
# EVALUATION options
parser.add_argument(
"--no_cuda",
default=False,
type=bool,
help="Whether to force the execution on CPU.",
)
parser.add_argument(
"--batch_size", default=4, type=int, help="Batch size per GPU/CPU for training.",
)
# BEAM SEARCH arguments
parser.add_argument(
"--min_length",
default=50,
type=int,
help="Minimum number of tokens for the summaries.",
)
parser.add_argument(
"--max_length",
default=200,
type=int,
help="Maixmum number of tokens for the summaries.",
)
parser.add_argument(
"--beam_size",
default=5,
type=int,
help="The number of beams to start with for each example.",
)
parser.add_argument(
"--alpha",
default=0.95,
type=float,
help="The value of alpha for the length penalty in the beam search.",
)
parser.add_argument(
"--block_trigram",
default=True,
type=bool,
help="Whether to block the existence of repeating trigrams in the text generated by beam search.",
)
args = parser.parse_args()
# Select device (distibuted not available)
args.device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
# Check the existence of directories
if not args.summaries_output_dir:
args.summaries_output_dir = args.documents_dir
if not documents_dir_is_valid(args.documents_dir):
raise FileNotFoundError(
"We could not find the directory you specified for the documents to summarize, or it was empty. Please specify a valid path."
)
os.makedirs(args.summaries_output_dir, exist_ok=True)
evaluate(args)
def documents_dir_is_valid(path):
if not os.path.exists(path):
return False
file_list = os.listdir(path)
if len(file_list) == 0:
return False
return True
if __name__ == "__main__":
main()
examples/utils_summarization.py → examples/summarization/utils_summarization.py
View file @ d5712f7c
...@@ -10,9 +10,14 @@ from torch.utils.data import Dataset ...@@ -10,9 +10,14 @@ from torch.utils.data import Dataset
# ------------ # ------------
class CNNDailyMailDataset(Dataset): class SummarizationDataset(Dataset):
""" Abstracts the dataset used to train seq2seq models. """ Abstracts the dataset used to train seq2seq models.
The class will process the documents that are located in the specified
folder. The preprocessing will work on any document that is reasonably
formatted. On the CNN/DailyMail dataset it will extract both the story
and the summary.
CNN/Daily News: CNN/Daily News:
The CNN/Daily News raw datasets are downloaded from [1]. The stories are The CNN/Daily News raw datasets are downloaded from [1]. The stories are
...@@ -25,33 +30,33 @@ class CNNDailyMailDataset(Dataset): ...@@ -25,33 +30,33 @@ class CNNDailyMailDataset(Dataset):
[2] https://github.com/abisee/cnn-dailymail/ [2] https://github.com/abisee/cnn-dailymail/
""" """
def __init__(self, tokenizer, prefix="train", data_dir=""): def __init__(self, path="", prefix="train"):
assert os.path.isdir(data_dir) """ We initialize the class by listing all the documents to summarize.
self.tokenizer = tokenizer Files are not read in memory due to the size of some datasets (like CNN/DailyMail).
"""
# We initialize the class by listing all the files that contain assert os.path.isdir(path)
# stories and summaries. Files are not read in memory given
# the size of the corpus. self.documents = []
self.stories_path = [] story_filenames_list = os.listdir(path)
datasets = ("cnn", "dailymail") for story_filename in story_filenames_list:
for dataset in datasets: if "summary" in story_filename:
path_to_stories = os.path.join(data_dir, dataset, "stories") continue
story_filenames_list = os.listdir(path_to_stories) path_to_story = os.path.join(path, story_filename)
for story_filename in story_filenames_list: if not os.path.isfile(path_to_story):
path_to_story = os.path.join(path_to_stories, story_filename) continue
if not os.path.isfile(path_to_story): self.documents.append(path_to_story)
continue
self.stories_path.append(path_to_story)
def __len__(self): def __len__(self):
return len(self.stories_path) """ Returns the number of documents. """
return len(self.documents)
def __getitem__(self, idx): def __getitem__(self, idx):
story_path = self.stories_path[idx] document_path = self.documents[idx]
with open(story_path, encoding="utf-8") as source: document_name = document_path.split("/")[-1]
with open(document_path, encoding="utf-8") as source:
raw_story = source.read() raw_story = source.read()
story_lines, summary_lines = process_story(raw_story) story_lines, summary_lines = process_story(raw_story)
return story_lines, summary_lines return document_name, story_lines, summary_lines
def process_story(raw_story): def process_story(raw_story):
...@@ -81,7 +86,7 @@ def process_story(raw_story): ...@@ -81,7 +86,7 @@ def process_story(raw_story):
story_lines.append(element) story_lines.append(element)
except IndexError: except IndexError:
# if "@highlight" is absent from the file we pop # if "@highlight" is absent from the file we pop
# all elements until there is None. # all elements until there is None, raising an exception.
return story_lines, [] return story_lines, []
# gather summary lines # gather summary lines
...@@ -104,31 +109,22 @@ def _add_missing_period(line): ...@@ -104,31 +109,22 @@ def _add_missing_period(line):
# -------------------------- # --------------------------
def fit_to_block_size(sequence, block_size, pad_token): def fit_to_block_size(sequence, block_size, pad_token_id):
""" Adapt the source and target sequences' lengths to the block size. """ Adapt the source and target sequences' lengths to the block size.
If the sequence is shorter than the block size we pad it with -1 ids If the sequence is shorter we append padding token to the right of the sequence.
which correspond to padding tokens.
""" """
if len(sequence) > block_size: if len(sequence) > block_size:
return sequence[:block_size] return sequence[:block_size]
else: else:
sequence.extend([pad_token] * (block_size - len(sequence))) sequence.extend([pad_token_id] * (block_size - len(sequence)))
return sequence return sequence
def build_lm_labels(sequence, pad_token): def build_mask(sequence, pad_token_id):
""" Padding token, encoded as 0, are represented by the value -1 so they
are not taken into account in the loss computation. """
padded = sequence.clone()
padded[padded == pad_token] = -1
return padded
def build_mask(sequence, pad_token):
""" Builds the mask. The attention mechanism will only attend to positions """ Builds the mask. The attention mechanism will only attend to positions
with value 1. """ with value 1. """
mask = torch.ones_like(sequence) mask = torch.ones_like(sequence)
idx_pad_tokens = sequence == pad_token idx_pad_tokens = sequence == pad_token_id
mask[idx_pad_tokens] = 0 mask[idx_pad_tokens] = 0
return mask return mask
...@@ -138,18 +134,11 @@ def encode_for_summarization(story_lines, summary_lines, tokenizer): ...@@ -138,18 +134,11 @@ def encode_for_summarization(story_lines, summary_lines, tokenizer):
as specified in [1] by using `[SEP] [CLS]` tokens to separate as specified in [1] by using `[SEP] [CLS]` tokens to separate
sentences. sentences.
""" """
story_lines_token_ids = [ story_lines_token_ids = [tokenizer.encode(line) for line in story_lines]
tokenizer.add_special_tokens_single_sequence(tokenizer.encode(line))
for line in story_lines
]
summary_lines_token_ids = [
tokenizer.add_special_tokens_single_sequence(tokenizer.encode(line))
for line in summary_lines
]
story_token_ids = [ story_token_ids = [
token for sentence in story_lines_token_ids for token in sentence token for sentence in story_lines_token_ids for token in sentence
] ]
summary_lines_token_ids = [tokenizer.encode(line) for line in summary_lines]
summary_token_ids = [ summary_token_ids = [
token for sentence in summary_lines_token_ids for token in sentence token for sentence in summary_lines_token_ids for token in sentence
] ]
...@@ -174,7 +163,7 @@ def compute_token_type_ids(batch, separator_token_id): ...@@ -174,7 +163,7 @@ def compute_token_type_ids(batch, separator_token_id):
""" """
batch_embeddings = [] batch_embeddings = []
for sequence in batch: for sequence in batch:
sentence_num = 0 sentence_num = -1
embeddings = [] embeddings = []
for s in sequence: for s in sequence:
if s == separator_token_id: if s == separator_token_id:
......
examples/utils_summarization_test.py → examples/summarization/utils_summarization_test.py
View file @ d5712f7c
...@@ -21,7 +21,6 @@ from utils_summarization import ( ...@@ -21,7 +21,6 @@ from utils_summarization import (
compute_token_type_ids, compute_token_type_ids,
fit_to_block_size, fit_to_block_size,
build_mask, build_mask,
build_lm_labels,
process_story, process_story,
) )
...@@ -88,20 +87,6 @@ class SummarizationDataProcessingTest(unittest.TestCase): ...@@ -88,20 +87,6 @@ class SummarizationDataProcessingTest(unittest.TestCase):
expected_summary_lines = ["It was the best of times."] expected_summary_lines = ["It was the best of times."]
self.assertEqual(expected_summary_lines, summary_lines) self.assertEqual(expected_summary_lines, summary_lines)
def test_build_lm_labels_no_padding(self):
sequence = torch.tensor([1, 2, 3, 4])
expected = sequence
np.testing.assert_array_equal(
build_lm_labels(sequence, 0).numpy(), expected.numpy()
)
def test_build_lm_labels(self):
sequence = torch.tensor([1, 2, 3, 4, 0, 0, 0])
expected = torch.tensor([1, 2, 3, 4, -1, -1, -1])
np.testing.assert_array_equal(
build_lm_labels(sequence, 0).numpy(), expected.numpy()
)
def test_build_mask_no_padding(self): def test_build_mask_no_padding(self):
sequence = torch.tensor([1, 2, 3, 4]) sequence = torch.tensor([1, 2, 3, 4])
expected = torch.tensor([1, 1, 1, 1]) expected = torch.tensor([1, 1, 1, 1])
...@@ -125,7 +110,7 @@ class SummarizationDataProcessingTest(unittest.TestCase): ...@@ -125,7 +110,7 @@ class SummarizationDataProcessingTest(unittest.TestCase):
[[1, 2, 3, 4, 5, 6], [1, 2, 3, 101, 5, 6], [1, 101, 3, 4, 101, 6]] [[1, 2, 3, 4, 5, 6], [1, 2, 3, 101, 5, 6], [1, 101, 3, 4, 101, 6]]
) )
expected = torch.tensor( expected = torch.tensor(
[[0, 0, 0, 0, 0, 0], [0, 0, 0, 1, 1, 1], [0, 1, 1, 1, 0, 0]] [[1, 1, 1, 1, 1, 1], [1, 1, 1, 0, 0, 0], [1, 0, 0, 0, 1, 1]]
) )
result = compute_token_type_ids(batch, separator) result = compute_token_type_ids(batch, separator)
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment