Merge pull request #1127 from huggingface/dilbert

DilBERT

README.md

@@ -12,7 +12,9 @@ The library currently contains PyTorch implementations, pre-trained model weight
 4. **[Transformer-XL](https://github.com/kimiyoung/transformer-xl)** (from Google/CMU) released with the paper [Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context](https://arxiv.org/abs/1901.02860) by Zihang Dai*, Zhilin Yang*, Yiming Yang, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov.
 5. **[XLNet](https://github.com/zihangdai/xlnet/)** (from Google/CMU) released with the paper [​XLNet: Generalized Autoregressive Pretraining for Language Understanding](https://arxiv.org/abs/1906.08237) by Zhilin Yang*, Zihang Dai*, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, Quoc V. Le.
 6. **[XLM](https://github.com/facebookresearch/XLM/)** (from Facebook) released together with the paper [Cross-lingual Language Model Pretraining](https://arxiv.org/abs/1901.07291) by Guillaume Lample and Alexis Conneau.
-7. **[RoBERTa](https://github.com/pytorch/fairseq/tree/master/examples/roberta)** (from Facebook), a [Robustly Optimized BERT Pretraining Approach](https://arxiv.org/abs/1907.11692) by Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du et al.
+7. **[RoBERTa](https://github.com/pytorch/fairseq/tree/master/examples/roberta)** (from Facebook), released together with the paper a [Robustly Optimized BERT Pretraining Approach](https://arxiv.org/abs/1907.11692) by Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, Veselin Stoyanov.
+8. **[DistilBERT](https://github.com/huggingface/pytorch-transformers/tree/master/examples/distillation)** (from HuggingFace), released together with the blogpost [Smaller, faster, cheaper, lighter: Introducing DistilBERT, a distilled version of BERT](https://medium.com/huggingface/distilbert-8cf3380435b5
+) by Victor Sanh, Lysandre Debut and Thomas Wolf.
 
 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/pytorch-transformers/examples.html).
 

docs/source/index.rst

@@ -48,3 +48,4 @@ The library currently contains PyTorch implementations, pre-trained model weight
     model_doc/xlm
     model_doc/xlnet
     model_doc/roberta
+    model_doc/distilbert

docs/source/model_doc/distilbert.rst

+DistilBERT
+----------------------------------------------------
+
+``DistilBertConfig``
+~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.DistilBertConfig
+    :members:
+
+
+``DistilBertTokenizer``
+~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.DistilBertTokenizer
+    :members:
+
+
+``DistilBertModel``
+~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.DistilBertModel
+    :members:
+
+
+``DistilBertForMaskedLM``
+~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.DistilBertForMaskedLM
+    :members:
+
+
+``DistilBertForSequenceClassification``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.DistilBertForSequenceClassification
+    :members:
+
+
+``DistilBertForQuestionAnswering``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.DistilBertForQuestionAnswering
+    :members:

docs/source/pretrained_models.rst

@@ -111,5 +111,13 @@ Here is the full list of the currently provided pretrained models together with
 |                   |                                                            | | ``roberta-large`` fine-tuned on `MNLI <http://www.nyu.edu/projects/bowman/multinli/>`__.                                            |
 |                   |                                                            | (see `details <https://github.com/pytorch/fairseq/tree/master/examples/roberta>`__)                                                   |
 +-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+| DistilBERT        | ``distilbert-base-uncased``                                | | 6-layer, 768-hidden, 12-heads, 66M parameters                                                                                       |
+|                   |                                                            | | The DistilBERT model distilled from the BERT model `bert-base-uncased` checkpoint                                                   |
+|                   |                                                            | (see `details <https://medium.com/huggingface/distilbert-8cf3380435b5>`__)                                                                       |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``distilbert-base-uncased-distilled-squad``                | | 6-layer, 768-hidden, 12-heads, 66M parameters                                                                                       |
+|                   |                                                            | | The DistilBERT model distilled from the BERT model `bert-base-uncased` checkpoint, with an additional linear layer.                 |
+|                   |                                                            | (see `details <https://medium.com/huggingface/distilbert-8cf3380435b5>`__)                                                                       |
++-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 
 .. <https://huggingface.co/pytorch-transformers/examples.html>`__
\ No newline at end of file

examples/distillation/README.md

+# DistilBERT
+
+This folder contains the original code used to train DistilBERT as well as examples showcasing how to use DistilBERT.
+
+## What is DistilBERT
+
+DistilBERT stands for Distillated-BERT. DistilBERT is a small, fast, cheap and light Transformer model based on Bert architecture. It has 40% less parameters than `bert-base-uncased`, runs 60% faster while preserving over 95% of Bert's performances as measured on the GLUE language understanding benchmark. DistilBERT is trained using knowledge distillation, a technique to compress a large model called the teacher into a smaller model called the student. By distillating Bert, we obtain a smaller Transformer model that bears a lot of similarities with the original BERT model while being lighter, smaller and faster to run. DistilBERT is thus an interesting option to put large-scaled trained Transformer model into production.
+
+For more information on DistilBERT, please refer to our [detailed blog post](https://medium.com/huggingface/smaller-faster-cheaper-lighter-introducing-distilbert-a-distilled-version-of-bert-8cf3380435b5
+).
+
+## How to use DistilBERT
+
+PyTorch-Transformers includes two pre-trained DistilBERT models, currently only provided for English (we are investigating the possibility to train and release a multilingual version of DistilBERT):
+
+- `distilbert-base-uncased`: DistilBERT English language model pretrained on the same data used to pretrain Bert (concatenation of the Toronto Book Corpus and full English Wikipedia) using distillation with the supervision of the `bert-base-uncased` version of Bert. The model has 6 layers, 768 dimension and 12 heads, totalizing 66M parameters.
+- `distilbert-base-uncased-distilled-squad`: A finetuned version of `distilbert-base-uncased` finetuned using (a second step of) knwoledge distillation on SQuAD 1.0. This model reaches a F1 score of 86.2 on the dev set (for comparison, Bert `bert-base-uncased` version reaches a 88.5 F1 score).
+
+Using DistilBERT is very similar to using BERT. DistilBERT share the same tokenizer as BERT's `bert-base-uncased` even though we provide a link to this tokenizer under the `DistilBertTokenizer` name to have a consistent naming between the library models.
+
+```python
+tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-uncased')
+model = DistilBertModel.from_pretrained('distilbert-base-uncased')
+
+input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)
+outputs = model(input_ids)
+last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
+```
+
+## How to train DistilBERT
+
+In the following, we will explain how you can train your own compressed model.
+
+### A. Preparing the data
+
+The weights we release are trained using a concatenation of Toronto Book Corpus and English Wikipedia (same training data as the English version of BERT).
+
+To avoid processing the data several time, we do it once and for all before the training. From now on, will suppose that you have a text file `dump.txt` which contains one sequence per line (a sequence being composed of one of several coherent sentences).
+
+First, we will binarize the data, i.e. tokenize the data and convert each token in an index in our model's vocabulary.
+
+```bash
+python scripts/binarized_data.py \
+    --file_path data/dump.txt \
+    --bert_tokenizer bert-base-uncased \
+    --dump_file data/binarized_text
+```
+
+Our implementation of masked language modeling loss follows [XLM](https://github.com/facebookresearch/XLM)'s one and smoothes the probability of masking with a factor that put more emphasis on rare words. Thus we count the occurences of each tokens in the data:
+
+```bash
+python scripts/token_counts.py \
+    --data_file data/binarized_text.bert-base-uncased.pickle \
+    --token_counts_dump data/token_counts.bert-base-uncased.pickle
+```
+
+### B. Training
+
+Training with distillation is really simple once you have pre-processed the data:
+
+```bash
+python train.py \
+    --dump_path serialization_dir/my_first_training \
+    --data_file data/binarized_text.bert-base-uncased.pickle \
+    --token_counts data/token_counts.bert-base-uncased.pickle \
+    --force # overwrites the `dump_path` if it already exists.
+```
+
+By default, this will launch a training on a single GPU (even if more are available on the cluster). Other parameters are available in the command line, please look in `train.py` or run `python train.py --help` to list them.
+
+We highly encourage you to distributed training for training DistilBert as the training corpus is quite large. Here's an example that runs a distributed training on a single node having 4 GPUs:
+
+```bash
+export NODE_RANK=0
+export N_NODES=1
+
+export N_GPU_NODE=4
+export WORLD_SIZE=4
+export MASTER_PORT=<AN_OPEN_PORT>
+export MASTER_ADDR=<I.P.>
+
+pkill -f 'python -u train.py'
+
+python -m torch.distributed.launch \
+    --nproc_per_node=$N_GPU_NODE \
+    --nnodes=$N_NODES \
+    --node_rank $NODE_RANK \
+    --master_addr $MASTER_ADDR \
+    --master_port $MASTER_PORT \
+    train.py \
+        --force \
+        --n_gpu $WORLD_SIZE \
+        --data_file data/dump_concat_wiki_toronto_bk.bert-base-uncased.pickle \
+        --token_counts data/token_counts_concat_wiki_toronto_bk.bert-base-uncased.pickle \
+        --dump_path serialization_dir/with_transform/last_word
+```
+
+**Tips** Starting distillated training with good initialization of the model weights is crucial to reach decent performance. In our experiments, we initialized our model from a few layers of the teacher (Bert) itself! Please refer to `scripts/extract_for_distil.py` to create a valid initialization checkpoint and use `--from_pretrained_weights` and `--from_pretrained_config` arguments to use this initialization for the distilled training!
+
+Happy distillation!

examples/distillation/dataset.py

+# coding=utf-8
+# Copyright 2019-present, the HuggingFace Inc. team and Facebook, Inc.
+#
+# 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.
+""" Dataloaders to train DistilBERT
+    adapted in part from Facebook, Inc XLM model (https://github.com/facebookresearch/XLM)
+"""
+from typing import List
+import math
+from itertools import chain
+from collections import Counter
+import numpy as np
+import torch
+
+from utils import logger
+
+class Dataset:
+    def __init__(self,
+                 params,
+                 data):
+        self.params = params
+        self.tokens_per_batch = params.tokens_per_batch
+        self.batch_size = params.batch_size
+        self.shuffle = params.shuffle
+        self.group_by_size = params.group_by_size
+
+        self.token_ids = np.array(data)
+        self.lengths = np.uint16([len(t) for t in data])
+
+        self.check()
+        self.remove_long_sequences()
+        self.remove_empty_sequences()
+        self.check()
+        self.print_statistics()
+
+    def __len__(self):
+        return len(self.lengths)
+
+    def check(self):
+        """
+        Some sanity checks
+        """
+        assert len(self.token_ids) == len(self.lengths)
+
+    def remove_long_sequences(self):
+        """
+        Sequences that are too long are splitted by chunk of max_position_embeddings.
+        """
+        indices = self.lengths >= self.params.max_position_embeddings
+        logger.info(f'Splitting {sum(indices)} too long sequences.')
+
+        def divide_chunks(l, n):
+            return [l[i:i + n] for i in range(0, len(l), n)]
+
+        new_tok_ids = []
+        new_lengths = []
+        cls_id, sep_id = self.params.special_tok_ids['cls_token'], self.params.special_tok_ids['sep_token']
+        max_len = self.params.max_position_embeddings
+
+        for seq_, len_ in zip(self.token_ids, self.lengths):
+            if len_ <= max_len:
+                new_tok_ids.append(seq_)
+                new_lengths.append(len_)
+            else:
+                sub_seqs = []
+                for sub_s in divide_chunks(seq_, max_len-2):
+                    if sub_s[0] != cls_id:
+                        sub_s = np.insert(sub_s, 0, cls_id)
+                    if sub_s[-1] != sep_id:
+                        sub_s = np.insert(sub_s, len(sub_s), cls_id)
+                    assert len(sub_s) <= max_len
+                    sub_seqs.append(sub_s)
+
+                new_tok_ids.extend(sub_seqs)
+                new_lengths.extend([len(l) for l in sub_seqs])
+
+        self.token_ids = np.array(new_tok_ids)
+        self.lengths = np.array(new_lengths)
+
+    def remove_empty_sequences(self):
+        """
+        Too short sequences are simply removed. This could be tunedd.
+        """
+        init_size = len(self)
+        indices = self.lengths > 5
+        self.token_ids = self.token_ids[indices]
+        self.lengths = self.lengths[indices]
+        new_size = len(self)
+        logger.info(f'Remove {init_size - new_size} too short (<=5 tokens) sequences.')
+
+    def print_statistics(self):
+        """
+        Print some statistics on the corpus. Only the master process.
+        """
+        if not self.params.is_master:
+            return
+        logger.info(f'{len(self)} sequences')
+        # data_len = sum(self.lengths)
+        # nb_unique_tokens = len(Counter(list(chain(*self.token_ids))))
+        # logger.info(f'{data_len} tokens ({nb_unique_tokens} unique)')
+
+        # unk_idx = self.params.special_tok_ids['unk_token']
+        # nb_unkown = sum([(t==unk_idx).sum() for t in self.token_ids])
+        # logger.info(f'{nb_unkown} unknown tokens (covering {100*nb_unkown/data_len:.2f}% of the data)')
+
+    def select_data(self, a: int, b: int):
+        """
+        Select a subportion of the data.
+        """
+        n_sequences = len(self)
+        assert 0 <= a < b <= n_sequences, ValueError(f'`0 <= a < b <= n_sequences` is not met with a={a} and b={b}')
+
+        logger.info(f'Selecting sequences from {a} to {b} (excluded).')
+        self.token_ids = self.token_ids[a:b]
+        self.lengths = self.lengths[a:b]
+
+        self.check()
+
+    def split(self):
+        """
+        Distributed training: split the data accross the processes.
+        """
+        assert self.params.n_gpu > 1
+        logger.info('Splitting the data accross the processuses.')
+        n_seq = len(self)
+        n_seq_per_procesus = n_seq // self.params.world_size
+        a = n_seq_per_procesus * self.params.global_rank
+        b = a + n_seq_per_procesus
+        self.select_data(a=a, b=b)
+
+    def batch_sequences(self,
+                        token_ids: List[List[int]],
+                        lengths: List[int]):
+        """
+        Do the padding and transform into torch.tensor.
+        """
+        assert len(token_ids) == len(lengths)
+
+        # Max for paddings
+        max_seq_len_ = max(lengths)
+
+        # Pad token ids
+        pad_idx = self.params.special_tok_ids['pad_token']
+        tk_ = [list(t.astype(int)) + [pad_idx]*(max_seq_len_-len(t)) for t in token_ids]
+        assert len(tk_) == len(token_ids)
+        assert all(len(t) == max_seq_len_ for t in tk_)
+
+        tk_t = torch.tensor(tk_)                  # (bs, max_seq_len_)
+        lg_t = torch.tensor(lengths.astype(int))  # (bs)
+        return tk_t, lg_t
+
+    def get_batches_iterator(self,
+                             batches):
+        """
+        Return an iterator over batches.
+        """
+        for sequences_ids in batches:
+            token_ids, lengths = self.batch_sequences(self.token_ids[sequences_ids],
+                                                    self.lengths[sequences_ids])
+            yield (token_ids, lengths)
+
+    def get_iterator(self,
+                     seed: int = None):
+        """
+        Return a data iterator.
+        """
+        rng = np.random.RandomState(seed)
+
+        n_sequences = len(self)
+        indices = np.arange(n_sequences)
+
+        if self.group_by_size:
+            indices = indices[np.argsort(self.lengths[indices], kind='mergesort')]
+
+        if self.tokens_per_batch == -1:
+            batches = np.array_split(indices, math.ceil(len(indices) * 1. / self.batch_size))
+        else:
+            assert self.tokens_per_batch > 0
+            batch_ids = np.cumsum(self.lengths[indices]) // self.tokens_per_batch
+            _, bounds = np.unique(batch_ids, return_index=True)
+            batches = [indices[bounds[i]:bounds[i + 1]] for i in range(len(bounds) - 1)]
+            if bounds[-1] < len(indices):
+                batches.append(indices[bounds[-1]:])
+
+        if self.shuffle:
+            rng.shuffle(batches)
+
+        assert n_sequences == sum([len(x) for x in batches])
+        assert self.lengths[indices].sum() == sum([self.lengths[x].sum() for x in batches])
+
+        return self.get_batches_iterator(batches=batches)

examples/distillation/requirements.txt

+gitpython==3.0.2

examples/distillation/scripts/binarized_data.py

+# coding=utf-8
+# Copyright 2019-present, 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.
+"""
+Preprocessing script before training DistilBERT.
+"""
+import argparse
+import pickle
+import random
+import time
+import numpy as np
+from pytorch_transformers import BertTokenizer
+
+from ..utils import logger
+
+def main():
+    parser = argparse.ArgumentParser(description="Preprocess the data to avoid re-doing it several times by (tokenization + token_to_ids).")
+    parser.add_argument('--file_path', type=str, default='data/dump.txt',
+                        help='The path to the data.')
+    parser.add_argument('--bert_tokenizer', type=str, default='bert-base-uncased',
+                        help="The tokenizer to use.")
+    parser.add_argument('--dump_file', type=str, default='data/dump',
+                        help='The dump file prefix.')
+    args = parser.parse_args()
+
+
+    logger.info(f'Loading Tokenizer ({args.bert_tokenizer})')
+    bert_tokenizer = BertTokenizer.from_pretrained(args.bert_tokenizer)
+
+
+    logger.info(f'Loading text from {args.file_path}')
+    with open(args.file_path, 'r', encoding='utf8') as fp:
+        data = fp.readlines()
+
+
+    logger.info(f'Start encoding')
+    logger.info(f'{len(data)} examples to process.')
+
+    rslt = []
+    iter = 0
+    interval = 10000
+    start = time.time()
+    for text in data:
+        text = f'[CLS] {text.strip()} [SEP]'
+        token_ids = bert_tokenizer.encode(text)
+        rslt.append(token_ids)
+
+        iter += 1
+        if iter % interval == 0:
+            end = time.time()
+            logger.info(f'{iter} examples processed. - {(end-start)/interval:.2f}s/expl')
+            start = time.time()
+    logger.info('Finished binarization')
+    logger.info(f'{len(data)} examples processed.')
+
+
+    dp_file = f'{args.dump_file}.{args.bert_tokenizer}.pickle'
+    rslt_ = [np.uint16(d) for d in rslt]
+    random.shuffle(rslt_)
+    logger.info(f'Dump to {dp_file}')
+    with open(dp_file, 'wb') as handle:
+        pickle.dump(rslt_, handle, protocol=pickle.HIGHEST_PROTOCOL)
+
+
+if __name__ == "__main__":
+    main()
\ No newline at end of file

examples/distillation/scripts/extract_for_distil.py

+# coding=utf-8
+# Copyright 2019-present, 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.
+"""
+Preprocessing script before training DistilBERT.
+"""
+from pytorch_transformers import BertForPreTraining
+import torch
+import argparse
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description="Extraction some layers of the full BertForPreTraining for Transfer Learned Distillation")
+    parser.add_argument("--bert_model", default='bert-base-uncased', type=str)
+    parser.add_argument("--dump_checkpoint", default='serialization_dir/transfer_learning_checkpoint_0247911.pth', type=str)
+    parser.add_argument("--vocab_transform", action='store_true')
+    args = parser.parse_args()
+
+
+    model = BertForPreTraining.from_pretrained(args.bert_model)
+
+    state_dict = model.state_dict()
+    compressed_sd = {}
+
+    for w in ['word_embeddings', 'position_embeddings']:
+        compressed_sd[f'distilbert.embeddings.{w}.weight'] = \
+            state_dict[f'bert.embeddings.{w}.weight']
+    for w in ['weight', 'bias']:
+        compressed_sd[f'distilbert.embeddings.LayerNorm.{w}'] = \
+            state_dict[f'bert.embeddings.LayerNorm.{w}']
+
+    std_idx = 0
+    for teacher_idx in [0, 2, 4, 7, 9, 11]:
+        for w in ['weight', 'bias']:
+            compressed_sd[f'distilbert.transformer.layer.{std_idx}.attention.q_lin.{w}'] = \
+                state_dict[f'bert.encoder.layer.{teacher_idx}.attention.self.query.{w}']
+            compressed_sd[f'distilbert.transformer.layer.{std_idx}.attention.k_lin.{w}'] = \
+                state_dict[f'bert.encoder.layer.{teacher_idx}.attention.self.key.{w}']
+            compressed_sd[f'distilbert.transformer.layer.{std_idx}.attention.v_lin.{w}'] = \
+                state_dict[f'bert.encoder.layer.{teacher_idx}.attention.self.value.{w}']
+
+            compressed_sd[f'distilbert.transformer.layer.{std_idx}.attention.out_lin.{w}'] = \
+                state_dict[f'bert.encoder.layer.{teacher_idx}.attention.output.dense.{w}']
+            compressed_sd[f'distilbert.transformer.layer.{std_idx}.sa_layer_norm.{w}'] = \
+                state_dict[f'bert.encoder.layer.{teacher_idx}.attention.output.LayerNorm.{w}']
+
+            compressed_sd[f'distilbert.transformer.layer.{std_idx}.ffn.lin1.{w}'] = \
+                state_dict[f'bert.encoder.layer.{teacher_idx}.intermediate.dense.{w}']
+            compressed_sd[f'distilbert.transformer.layer.{std_idx}.ffn.lin2.{w}'] = \
+                state_dict[f'bert.encoder.layer.{teacher_idx}.output.dense.{w}']
+            compressed_sd[f'distilbert.transformer.layer.{std_idx}.output_layer_norm.{w}'] = \
+                state_dict[f'bert.encoder.layer.{teacher_idx}.output.LayerNorm.{w}']
+        std_idx += 1
+
+    compressed_sd[f'vocab_projector.weight'] = state_dict[f'cls.predictions.decoder.weight']
+    compressed_sd[f'vocab_projector.bias'] = state_dict[f'cls.predictions.bias']
+    if args.vocab_transform:
+        for w in ['weight', 'bias']:
+            compressed_sd[f'vocab_transform.{w}'] = state_dict[f'cls.predictions.transform.dense.{w}']
+            compressed_sd[f'vocab_layer_norm.{w}'] = state_dict[f'cls.predictions.transform.LayerNorm.{w}']
+
+    print(f'N layers selected for distillation: {std_idx}')
+    print(f'Number of params transfered for distillation: {len(compressed_sd.keys())}')
+
+    print(f'Save transfered checkpoint to {args.dump_checkpoint}.')
+    torch.save(compressed_sd, args.dump_checkpoint)

examples/distillation/scripts/token_counts.py

+# coding=utf-8
+# Copyright 2019-present, 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.
+"""
+Preprocessing script before training DistilBERT.
+"""
+from collections import Counter
+import argparse
+import pickle
+
+from utils import logger
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description="Token Counts for smoothing the masking probabilities in MLM (cf XLM/word2vec)")
+    parser.add_argument("--data_file", type=str, default="data/dump.bert-base-uncased.pickle",
+                        help="The binarized dataset."
+    parser.add_argument("--token_counts_dump", type=str, default="data/token_counts.bert-base-uncased.pickle",
+                        help="The dump file.")
+    parser.add_argument("--vocab_size", default=30522, type=int)
+    args = parser.parse_args()
+
+    logger.info(f'Loading data from {args.data_file}')
+    with open(args.data_file, 'rb') as fp:
+        data = pickle.load(fp)
+
+    logger.info('Counting occurences for MLM.')
+    counter = Counter()
+    for tk_ids in data:
+        counter.update(tk_ids)
+    counts = [0]*args.vocab_size
+    for k, v in counter.items():
+        counts[k] = v
+
+    logger.info(f'Dump to {args.token_counts_dump}')
+    with open(args.token_counts_dump, 'wb') as handle:
+        pickle.dump(counts, handle, protocol=pickle.HIGHEST_PROTOCOL)

examples/distillation/train.py

+# coding=utf-8
+# Copyright 2019-present, 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.
+"""
+Training DistilBERT.
+"""
+import os
+import argparse
+import pickle
+import json
+import shutil
+import numpy as np
+import torch
+
+from pytorch_transformers import BertTokenizer, BertForMaskedLM
+from pytorch_transformers import DistilBertForMaskedLM, DistilBertConfig
+
+from distiller import Distiller
+from utils import git_log, logger, init_gpu_params, set_seed
+from dataset import Dataset
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Training")
+
+    parser.add_argument("--dump_path", type=str, required=True,
+                        help="The output directory (log, checkpoints, parameters, etc.)")
+    parser.add_argument("--data_file", type=str, required=True,
+                        help="The binarized file (tokenized + tokens_to_ids) and grouped by sequence.")
+    parser.add_argument("--token_counts", type=str, required=True,
+                        help="The token counts in the data_file for MLM.")
+    parser.add_argument("--force", action='store_true',
+                        help="Overwrite dump_path if it already exists.")
+
+    parser.add_argument("--vocab_size", default=30522, type=int,
+                        help="The vocabulary size.")
+    parser.add_argument("--max_position_embeddings", default=512, type=int,
+                        help="Maximum sequence length we can model (including [CLS] and [SEP]).")
+    parser.add_argument("--sinusoidal_pos_embds", action='store_false',
+                        help="If true, the position embeddings are simply fixed with sinusoidal embeddings.")
+    parser.add_argument("--n_layers", default=6, type=int,
+                        help="Number of Transformer blocks.")
+    parser.add_argument("--n_heads", default=12, type=int,
+                        help="Number of heads in the self-attention module.")
+    parser.add_argument("--dim", default=768, type=int,
+                        help="Dimension through the network. Must be divisible by n_heads")
+    parser.add_argument("--hidden_dim", default=3072, type=int,
+                        help="Intermediate dimension in the FFN.")
+    parser.add_argument("--dropout", default=0.1, type=float,
+                        help="Dropout.")
+    parser.add_argument("--attention_dropout", default=0.1, type=float,
+                        help="Dropout in self-attention.")
+    parser.add_argument("--activation", default='gelu', type=str,
+                        help="Activation to use in self-attention")
+    parser.add_argument("--tie_weights_", action='store_false',
+                        help="If true, we tie the embeddings matrix with the projection over the vocabulary matrix. Default is true.")
+
+    parser.add_argument("--from_pretrained_weights", default=None, type=str,
+                        help="Load student initialization checkpoint.")
+    parser.add_argument("--from_pretrained_config", default=None, type=str,
+                        help="Load student initialization architecture config.")
+    parser.add_argument("--bert_model", default='bert-base-uncased', type=str,
+                        help="The teacher BERT model.")
+
+    parser.add_argument("--temperature", default=2., type=float,
+                        help="Temperature for the softmax temperature.")
+    parser.add_argument("--alpha_ce", default=0.5, type=float,
+                        help="Linear weight for the distillation loss. Must be >=0.")
+    parser.add_argument("--alpha_mlm", default=0.5, type=float,
+                        help="Linear weight for the MLM loss. Must be >=0.")
+    parser.add_argument("--alpha_mse", default=0.0, type=float,
+                        help="Linear weight of the MSE loss. Must be >=0.")
+    parser.add_argument("--mlm_mask_prop", default=0.15, type=float,
+                        help="Proportion of tokens for which we need to make a prediction.")
+    parser.add_argument("--word_mask", default=0.8, type=float,
+                        help="Proportion of tokens to mask out.")
+    parser.add_argument("--word_keep", default=0.1, type=float,
+                        help="Proportion of tokens to keep.")
+    parser.add_argument("--word_rand", default=0.1, type=float,
+                        help="Proportion of tokens to randomly replace.")
+    parser.add_argument("--mlm_smoothing", default=0.7, type=float,
+                        help="Smoothing parameter to emphasize more rare tokens (see XLM, similar to word2vec).")
+    parser.add_argument("--restrict_ce_to_mask", action='store_true',
+                        help="If true, compute the distilation loss only the [MLM] prediction distribution.")
+
+    parser.add_argument("--n_epoch", type=int, default=3,
+                        help="Number of pass on the whole dataset.")
+    parser.add_argument("--batch_size", type=int, default=5,
+                        help="Batch size (for each process).")
+    parser.add_argument("--tokens_per_batch", type=int, default=-1,
+                        help="If specified, modify the batches so that they have approximately this number of tokens.")
+    parser.add_argument("--shuffle", action='store_false',
+                        help="If true, shuffle the sequence order. Default is true.")
+    parser.add_argument("--group_by_size", action='store_false',
+                        help="If true, group sequences that have similar length into the same batch. Default is true.")
+
+    parser.add_argument("--gradient_accumulation_steps", type=int, default=50,
+                        help="Gradient accumulation for larger training batches.")
+    parser.add_argument("--warmup_prop", default=0.05, type=float,
+                        help="Linear warmup proportion.")
+    parser.add_argument("--weight_decay", default=0.0, type=float,
+                        help="Weight deay if we apply some.")
+    parser.add_argument("--learning_rate", default=5e-4, type=float,
+                        help="The initial learning rate for Adam.")
+    parser.add_argument("--adam_epsilon", default=1e-6, type=float,
+                        help="Epsilon for Adam optimizer.")
+    parser.add_argument("--max_grad_norm", default=5.0, type=float,
+                        help="Max gradient norm.")
+    parser.add_argument("--initializer_range", default=0.02, type=float,
+                        help="Random initialization range.")
+
+    parser.add_argument('--fp16', action='store_true',
+                        help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit")
+    parser.add_argument('--fp16_opt_level', type=str, default='O1',
+                        help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
+                             "See details at https://nvidia.github.io/apex/amp.html")
+    parser.add_argument("--n_gpu", type=int, default=1,
+                        help="Number of GPUs in the node.")
+    parser.add_argument("--local_rank", type=int, default=-1,
+                        help="Distributed training - Local rank")
+    parser.add_argument("--seed", type=int, default=56,
+                        help="Random seed")
+
+    parser.add_argument("--log_interval", type=int, default=500,
+                        help="Tensorboard logging interval.")
+    parser.add_argument("--checkpoint_interval", type=int, default=4000,
+                        help="Checkpoint interval.")
+    args = parser.parse_args()
+
+
+    ## ARGS ##
+    init_gpu_params(args)
+    set_seed(args)
+    if args.is_master:
+        if os.path.exists(args.dump_path):
+            if not args.force:
+                raise ValueError(f'Serialization dir {args.dump_path} already exists, but you have not precised wheter to overwrite it'
+                                   'Use `--force` if you want to overwrite it')
+            else:
+                shutil.rmtree(args.dump_path)
+
+        if not os.path.exists(args.dump_path):
+            os.makedirs(args.dump_path)
+        logger.info(f'Experiment will be dumped and logged in {args.dump_path}')
+
+
+        ### SAVE PARAMS ###
+        logger.info(f'Param: {args}')
+        with open(os.path.join(args.dump_path, 'parameters.json'), 'w') as f:
+            json.dump(vars(args), f, indent=4)
+        git_log(args.dump_path)
+    assert (args.from_pretrained_weights is None and args.from_pretrained_config is None) or \
+           (args.from_pretrained_weights is not None and args.from_pretrained_config is not None)
+
+
+    ### TOKENIZER ###
+    bert_tokenizer = BertTokenizer.from_pretrained(args.bert_model)
+    special_tok_ids = {}
+    for tok_name, tok_symbol in bert_tokenizer.special_tokens_map.items():
+        idx = bert_tokenizer.all_special_tokens.index(tok_symbol)
+        special_tok_ids[tok_name] = bert_tokenizer.all_special_ids[idx]
+    logger.info(f'Special tokens {special_tok_ids}')
+    args.special_tok_ids = special_tok_ids
+
+
+    ## DATA LOADER ##
+    logger.info(f'Loading data from {args.data_file}')
+    with open(args.data_file, 'rb') as fp:
+        data = pickle.load(fp)
+
+
+    assert os.path.isfile(args.token_counts)
+    logger.info(f'Loading token counts from {args.token_counts} (already pre-computed)')
+    with open(args.token_counts, 'rb') as fp:
+        counts = pickle.load(fp)
+        assert len(counts) == args.vocab_size
+    token_probs = np.maximum(counts, 1) ** -args.mlm_smoothing
+    for idx in special_tok_ids.values():
+        token_probs[idx] = 0.  # do not predict special tokens
+    token_probs = torch.from_numpy(token_probs)
+
+
+    train_dataloader = Dataset(params=args, data=data)
+    logger.info(f'Data loader created.')
+
+
+    ## STUDENT ##
+    if args.from_pretrained_weights is not None:
+        assert os.path.isfile(os.path.join(args.from_pretrained_weights))
+        assert os.path.isfile(os.path.join(args.from_pretrained_config))
+        logger.info(f'Loading pretrained weights from {args.from_pretrained_weights}')
+        logger.info(f'Loading pretrained config from {args.from_pretrained_config}')
+        stu_architecture_config = DistilBertConfig.from_json_file(args.from_pretrained_config)
+        student = DistilBertForMaskedLM.from_pretrained(args.from_pretrained_weights,
+                                                     config=stu_architecture_config)
+    else:
+        args.vocab_size_or_config_json_file = args.vocab_size
+        stu_architecture_config = DistilBertConfig(**vars(args))
+        student = DistilBertForMaskedLM(stu_architecture_config)
+
+
+    if args.n_gpu > 0:
+        student.to(f'cuda:{args.local_rank}')
+    logger.info(f'Student loaded.')
+
+
+    ## TEACHER ##
+    teacher = BertForMaskedLM.from_pretrained(args.bert_model)
+    if args.n_gpu > 0:
+        teacher.to(f'cuda:{args.local_rank}')
+    logger.info(f'Teacher loaded from {args.bert_model}.')
+
+    ## DISTILLER ##
+    torch.cuda.empty_cache()
+    distiller = Distiller(params=args,
+                          dataloader=train_dataloader,
+                          token_probs=token_probs,
+                          student=student,
+                          teacher=teacher)
+    distiller.train()
+    logger.info("Let's go get some drinks.")
+
+
+if __name__ == "__main__":
+    main()

examples/distillation/utils.py

+# coding=utf-8
+# Copyright 2019-present, the HuggingFace Inc. team and Facebook, Inc.
+#
+# 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.
+""" Utils to train DistilBERT
+    adapted in part from Facebook, Inc XLM model (https://github.com/facebookresearch/XLM)
+"""
+import git
+import json
+import os
+import socket
+import torch
+import numpy as np
+
+import logging
+logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s - PID: %(process)d -  %(message)s',
+                    datefmt = '%m/%d/%Y %H:%M:%S',
+                    level = logging.INFO)
+logger = logging.getLogger(__name__)
+
+
+def git_log(folder_path: str):
+    """
+    Log commit info.
+    """
+    repo = git.Repo(search_parent_directories=True)
+    repo_infos = {
+        'repo_id': str(repo),
+        'repo_sha': str(repo.head.object.hexsha),
+        'repo_branch': str(repo.active_branch)
+    }
+
+    with open(os.path.join(folder_path, 'git_log.json'), 'w') as f:
+        json.dump(repo_infos, f, indent=4)
+
+
+def init_gpu_params(params):
+    """
+    Handle single and multi-GPU / multi-node.
+    """
+    if params.n_gpu <= 0:
+        params.local_rank = 0
+        params.master_port = -1
+        params.is_master = True
+        params.multi_gpu = False
+        return
+
+    assert torch.cuda.is_available()
+
+    logger.info('Initializing GPUs')
+    if params.n_gpu > 1:
+        assert params.local_rank != -1
+
+        params.world_size = int(os.environ['WORLD_SIZE'])
+        params.n_gpu_per_node = int(os.environ['N_GPU_NODE'])
+        params.global_rank = int(os.environ['RANK'])
+
+        # number of nodes / node ID
+        params.n_nodes = params.world_size // params.n_gpu_per_node
+        params.node_id = params.global_rank // params.n_gpu_per_node
+        params.multi_gpu = True
+
+        assert params.n_nodes == int(os.environ['N_NODES'])
+        assert params.node_id == int(os.environ['NODE_RANK'])
+
+    # local job (single GPU)
+    else:
+        assert params.local_rank == -1
+
+        params.n_nodes = 1
+        params.node_id = 0
+        params.local_rank = 0
+        params.global_rank = 0
+        params.world_size = 1
+        params.n_gpu_per_node = 1
+        params.multi_gpu = False
+
+    # sanity checks
+    assert params.n_nodes >= 1
+    assert 0 <= params.node_id < params.n_nodes
+    assert 0 <= params.local_rank <= params.global_rank < params.world_size
+    assert params.world_size == params.n_nodes * params.n_gpu_per_node
+
+    # define whether this is the master process / if we are in multi-node distributed mode
+    params.is_master = params.node_id == 0 and params.local_rank == 0
+    params.multi_node = params.n_nodes > 1
+
+    # summary
+    PREFIX = f"--- Global rank: {params.global_rank} - "
+    logger.info(PREFIX + "Number of nodes: %i" % params.n_nodes)
+    logger.info(PREFIX + "Node ID        : %i" % params.node_id)
+    logger.info(PREFIX + "Local rank     : %i" % params.local_rank)
+    logger.info(PREFIX + "World size     : %i" % params.world_size)
+    logger.info(PREFIX + "GPUs per node  : %i" % params.n_gpu_per_node)
+    logger.info(PREFIX + "Master         : %s" % str(params.is_master))
+    logger.info(PREFIX + "Multi-node     : %s" % str(params.multi_node))
+    logger.info(PREFIX + "Multi-GPU      : %s" % str(params.multi_gpu))
+    logger.info(PREFIX + "Hostname       : %s" % socket.gethostname())
+
+    # set GPU device
+    torch.cuda.set_device(params.local_rank)
+
+    # initialize multi-GPU
+    if params.multi_gpu:
+        logger.info("Initializing PyTorch distributed")
+        torch.distributed.init_process_group(
+            init_method='env://',
+            backend='nccl',
+        )
+
+
+def set_seed(args):
+    """
+    Set the random seed.
+    """
+    np.random.seed(args.seed)
+    torch.manual_seed(args.seed)
+    if args.n_gpu > 0:
+        torch.cuda.manual_seed_all(args.seed)

pytorch_transformers/__init__.py

@@ -7,6 +7,7 @@ from .tokenization_gpt2 import GPT2Tokenizer
 from .tokenization_xlnet import XLNetTokenizer, SPIECE_UNDERLINE
 from .tokenization_xlm import XLMTokenizer
 from .tokenization_roberta import RobertaTokenizer
+from .tokenization_distilbert import DistilBertTokenizer
 
 from .tokenization_utils import (PreTrainedTokenizer)
 
@@ -40,6 +41,9 @@ from .modeling_xlm import (XLMConfig, XLMPreTrainedModel , XLMModel,
                            XLM_PRETRAINED_MODEL_ARCHIVE_MAP)
 from .modeling_roberta import (RobertaConfig, RobertaForMaskedLM, RobertaModel, RobertaForSequenceClassification,
                                ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP, ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP)
+from .modeling_distilbert import (DistilBertConfig, DistilBertForMaskedLM, DistilBertModel,
+                               DistilBertForSequenceClassification, DistilBertForQuestionAnswering,
+                               DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP)
 from .modeling_utils import (WEIGHTS_NAME, CONFIG_NAME, TF_WEIGHTS_NAME,
                           PretrainedConfig, PreTrainedModel, prune_layer, Conv1D)
 

pytorch_transformers/modeling_auto.py

@@ -30,6 +30,7 @@ from .modeling_transfo_xl import TransfoXLConfig, TransfoXLModel
 from .modeling_xlnet import XLNetConfig, XLNetModel
 from .modeling_xlm import XLMConfig, XLMModel
 from .modeling_roberta import RobertaConfig, RobertaModel
+from .modeling_distilbert import DistilBertConfig, DistilBertModel
 
 from .modeling_utils import PreTrainedModel, SequenceSummary
 
@@ -110,7 +111,9 @@ class AutoConfig(object):
             assert unused_kwargs == {'foo': False}
 
         """
-        if 'roberta' in pretrained_model_name_or_path:
+        if 'distilbert' in pretrained_model_name_or_path:
+            return DistilBertConfig.from_pretrained(pretrained_model_name_or_path, **kwargs)
+        elif 'roberta' in pretrained_model_name_or_path:
             return RobertaConfig.from_pretrained(pretrained_model_name_or_path, **kwargs)
         elif 'bert' in pretrained_model_name_or_path:
             return BertConfig.from_pretrained(pretrained_model_name_or_path, **kwargs)
@@ -225,7 +228,9 @@ class AutoModel(object):
             model = AutoModel.from_pretrained('./tf_model/bert_tf_checkpoint.ckpt.index', from_tf=True, config=config)
 
         """
-        if 'roberta' in pretrained_model_name_or_path:
+        if 'distilbert' in pretrained_model_name_or_path:
+            return DistilBertModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'roberta' in pretrained_model_name_or_path:
             return RobertaModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'bert' in pretrained_model_name_or_path:
             return BertModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)

pytorch_transformers/modeling_bert.py

@@ -216,7 +216,7 @@ class BertConfig(PretrainedConfig):
             self.layer_norm_eps = layer_norm_eps
         else:
             raise ValueError("First argument must be either a vocabulary size (int)"
-                             "or the path to a pretrained model config file (str)")
+                             " or the path to a pretrained model config file (str)")
 
 
 
@@ -449,7 +449,7 @@ class BertEncoder(nn.Module):
             outputs = outputs + (all_hidden_states,)
         if self.output_attentions:
             outputs = outputs + (all_attentions,)
-        return outputs  # outputs, (hidden states), (attentions)
+        return outputs  # last-layer hidden state, (all hidden states), (all attentions)
 
 
 class BertPooler(nn.Module):

pytorch_transformers/modeling_transfo_xl.py

@@ -285,7 +285,7 @@ class TransfoXLConfig(PretrainedConfig):
             self.init_std = init_std
         else:
             raise ValueError("First argument must be either a vocabulary size (int)"
-                             "or the path to a pretrained model config file (str)")
+                             " or the path to a pretrained model config file (str)")
 
     @property
     def max_position_embeddings(self):

pytorch_transformers/modeling_xlm.py

@@ -178,7 +178,7 @@ class XLMConfig(PretrainedConfig):
             self.end_n_top = end_n_top
         else:
             raise ValueError("First argument must be either a vocabulary size (int)"
-                             "or the path to a pretrained model config file (str)")
+                             " or the path to a pretrained model config file (str)")
 
     @property
     def vocab_size(self):

pytorch_transformers/modeling_xlnet.py

@@ -306,7 +306,7 @@ class XLNetConfig(PretrainedConfig):
             self.end_n_top = end_n_top
         else:
             raise ValueError("First argument must be either a vocabulary size (int)"
-                             "or the path to a pretrained model config file (str)")
+                             " or the path to a pretrained model config file (str)")
 
     @property
     def max_position_embeddings(self):