Skip to content

GitLab

Unverified Commit 206f06f2 authored by yujun's avatar yujun Committed by GitHub
Browse files

Add new model RoFormer (use rotary position embedding ) (#11684) 



* add roformer

* Update docs/source/model_doc/roformer.rst
Co-authored-by: default avatarSuraj Patil <surajp815@gmail.com>

* Update docs/source/model_doc/roformer.rst
Co-authored-by: default avatarSuraj Patil <surajp815@gmail.com>

* update

* add TFRoFormerSinusoidalPositionalEmbedding and fix TFMarianSinusoidalPositionalEmbedding

* update docs

* make style and make quality

* roback

* unchanged

* rm copies from , this is a error in TFMarianSinusoidalPositionalEmbedding

* update Copyright year

* move # Add modeling imports here to the correct position

* max_position_embeddings can be set to 1536

* # Copied from transformers.models.bert.modeling_bert.BertOutput with Bert->RoFormer

* # Copied from transformers.models.bert.modeling_bert.BertLayer.__init__ with Bert->RoFormer

* update tokenization_roformer

* make style

* add staticmethod apply_rotary_position_embeddings

* add TF staticmethod apply_rotary_position_embeddings

* update torch apply_rotary_position_embeddings

* fix tf apply_rotary_position_embeddings error

* make style

* add pytorch RoFormerSelfAttentionRotaryPositionEmbeddingTest

* add TF rotary_position_embeddings test

* update test_modeling_rofomer

* Update docs/source/model_doc/roformer.rst
Co-authored-by: default avatarSylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/__init__.py
Co-authored-by: default avatarSylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/__init__.py
Co-authored-by: default avatarSylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/__init__.py
Co-authored-by: default avatarSylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/__init__.py
Co-authored-by: default avatarSylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/roformer/convert_roformer_original_tf_checkpoint_to_pytorch.py
Co-authored-by: default avatarSylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/roformer/modeling_roformer.py
Co-authored-by: default avatarSylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/roformer/modeling_roformer.py
Co-authored-by: default avatarSylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/roformer/modeling_tf_roformer.py
Co-authored-by: default avatarSylvain Gugger <35901082+sgugger@users.noreply.github.com>

* refact roformer tokenizer

* add RoFormerTokenizerFast

* add RoFormerTokenizationTest

* add require_jieba

* update Copyright

* update tokenizer & add copy from

* add option rotary_value

* use rust jieba

* use rjieba

* use rust jieba

* fix test_alignement_methods

* slice normalized_string is too slow

* add config.embedding_size when embedding_size!=hidden_size

* fix pickle tokenizer

* Update docs/source/model_doc/roformer.rst
Co-authored-by: default avatarPatrick von Platen <patrick.v.platen@gmail.com>

* make style and make quality
Co-authored-by: default avatarSuraj Patil <surajp815@gmail.com>
Co-authored-by: default avatarSylvain Gugger <35901082+sgugger@users.noreply.github.com>
Co-authored-by: default avatarPatrick von Platen <patrick.v.platen@gmail.com>
parent 075fdab4
Hide whitespace changes
Inline Side-by-side
Showing with 4461 additions and 12 deletions
README.md
View file @ 206f06f2
......@@ -243,6 +243,7 @@ Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih.
1. **[ProphetNet](https://huggingface.co/transformers/model_doc/prophetnet.html)** (from Microsoft Research) released with the paper [ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training](https://arxiv.org/abs/2001.04063) by Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou.
1. **[Reformer](https://huggingface.co/transformers/model_doc/reformer.html)** (from Google Research) released with the paper [Reformer: The Efficient Transformer](https://arxiv.org/abs/2001.04451) by Nikita Kitaev, Łukasz Kaiser, Anselm Levskaya.
1. **[RoBERTa](https://huggingface.co/transformers/model_doc/roberta.html)** (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.
1. **[RoFormer](https://huggingface.co/transformers/model_doc/roformer.html)** (from ZhuiyiTechnology), released together with the paper a [RoFormer: Enhanced Transformer with Rotary Position Embedding](https://arxiv.org/pdf/2104.09864v1.pdf) by Jianlin Su and Yu Lu and Shengfeng Pan and Bo Wen and Yunfeng Liu.
1. **[SpeechToTextTransformer](https://huggingface.co/transformers/model_doc/speech_to_text.html)** (from Facebook), released together with the paper [fairseq S2T: Fast Speech-to-Text Modeling with fairseq](https://arxiv.org/abs/2010.05171) by Changhan Wang, Yun Tang, Xutai Ma, Anne Wu, Dmytro Okhonko, Juan Pino.
1. **[SqueezeBert](https://huggingface.co/transformers/model_doc/squeezebert.html)** released with the paper [SqueezeBERT: What can computer vision teach NLP about efficient neural networks?](https://arxiv.org/abs/2006.11316) by Forrest N. Iandola, Albert E. Shaw, Ravi Krishna, and Kurt W. Keutzer.
1. **[T5](https://huggingface.co/transformers/model_doc/t5.html)** (from Google AI) released with the paper [Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer](https://arxiv.org/abs/1910.10683) by Colin Raffel and Noam Shazeer and Adam Roberts and Katherine Lee and Sharan Narang and Michael Matena and Yanqi Zhou and Wei Li and Peter J. Liu.
......
docs/source/index.rst
View file @ 206f06f2
......@@ -231,41 +231,44 @@ Supported models
45. :doc:`RoBERTa <model_doc/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.
46. :doc:`SpeechToTextTransformer <model_doc/speech_to_text>` (from Facebook), released together with the paper
46. :doc:`RoFormer <model_doc/roformer>` (from ZhuiyiTechnology), released together with the paper a `RoFormer:
Enhanced Transformer with Rotary Position Embedding <https://arxiv.org/pdf/2104.09864v1.pdf>`__ by Jianlin Su and
Yu Lu and Shengfeng Pan and Bo Wen and Yunfeng Liu.
47. :doc:`SpeechToTextTransformer <model_doc/speech_to_text>` (from Facebook), released together with the paper
`fairseq S2T: Fast Speech-to-Text Modeling with fairseq <https://arxiv.org/abs/2010.05171>`__ by Changhan Wang, Yun
Tang, Xutai Ma, Anne Wu, Dmytro Okhonko, Juan Pino.
47. :doc:`SqueezeBert <model_doc/squeezebert>` released with the paper `SqueezeBERT: What can computer vision teach NLP
48. :doc:`SqueezeBert <model_doc/squeezebert>` released with the paper `SqueezeBERT: What can computer vision teach NLP
about efficient neural networks? <https://arxiv.org/abs/2006.11316>`__ by Forrest N. Iandola, Albert E. Shaw, Ravi
Krishna, and Kurt W. Keutzer.
48. :doc:`T5 <model_doc/t5>` (from Google AI) released with the paper `Exploring the Limits of Transfer Learning with a
49. :doc:`T5 <model_doc/t5>` (from Google AI) released with the paper `Exploring the Limits of Transfer Learning with a
Unified Text-to-Text Transformer <https://arxiv.org/abs/1910.10683>`__ by Colin Raffel and Noam Shazeer and Adam
Roberts and Katherine Lee and Sharan Narang and Michael Matena and Yanqi Zhou and Wei Li and Peter J. Liu.
49. :doc:`TAPAS <model_doc/tapas>` (from Google AI) released with the paper `TAPAS: Weakly Supervised Table Parsing via
50. :doc:`TAPAS <model_doc/tapas>` (from Google AI) released with the paper `TAPAS: Weakly Supervised Table Parsing via
Pre-training <https://arxiv.org/abs/2004.02349>`__ by Jonathan Herzig, Paweł Krzysztof Nowak, Thomas Müller,
Francesco Piccinno and Julian Martin Eisenschlos.
50. :doc:`Transformer-XL <model_doc/transformerxl>` (from Google/CMU) released with the paper `Transformer-XL:
51. :doc:`Transformer-XL <model_doc/transformerxl>` (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.
51. :doc:`Vision Transformer (ViT) <model_doc/vit>` (from Google AI) released with the paper `An Image is Worth 16x16
52. :doc:`Vision Transformer (ViT) <model_doc/vit>` (from Google AI) released with the paper `An Image is Worth 16x16
Words: Transformers for Image Recognition at Scale <https://arxiv.org/abs/2010.11929>`__ by Alexey Dosovitskiy,
Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias
Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby.
52. :doc:`Wav2Vec2 <model_doc/wav2vec2>` (from Facebook AI) released with the paper `wav2vec 2.0: A Framework for
53. :doc:`Wav2Vec2 <model_doc/wav2vec2>` (from Facebook AI) released with the paper `wav2vec 2.0: A Framework for
Self-Supervised Learning of Speech Representations <https://arxiv.org/abs/2006.11477>`__ by Alexei Baevski, Henry
Zhou, Abdelrahman Mohamed, Michael Auli.
53. :doc:`XLM <model_doc/xlm>` (from Facebook) released together with the paper `Cross-lingual Language Model
54. :doc:`XLM <model_doc/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.
54. :doc:`XLM-ProphetNet <model_doc/xlmprophetnet>` (from Microsoft Research) released with the paper `ProphetNet:
55. :doc:`XLM-ProphetNet <model_doc/xlmprophetnet>` (from Microsoft Research) released with the paper `ProphetNet:
Predicting Future N-gram for Sequence-to-Sequence Pre-training <https://arxiv.org/abs/2001.04063>`__ by Yu Yan,
Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou.
55. :doc:`XLM-RoBERTa <model_doc/xlmroberta>` (from Facebook AI), released together with the paper `Unsupervised
56. :doc:`XLM-RoBERTa <model_doc/xlmroberta>` (from Facebook AI), released together with the paper `Unsupervised
Cross-lingual Representation Learning at Scale <https://arxiv.org/abs/1911.02116>`__ by Alexis Conneau*, Kartikay
Khandelwal*, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke
Zettlemoyer and Veselin Stoyanov.
56. :doc:`XLNet <model_doc/xlnet>` (from Google/CMU) released with the paper `​XLNet: Generalized Autoregressive
57. :doc:`XLNet <model_doc/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.
57. :doc:`XLSR-Wav2Vec2 <model_doc/xlsr_wav2vec2>` (from Facebook AI) released with the paper `Unsupervised
58. :doc:`XLSR-Wav2Vec2 <model_doc/xlsr_wav2vec2>` (from Facebook AI) released with the paper `Unsupervised
Cross-Lingual Representation Learning For Speech Recognition <https://arxiv.org/abs/2006.13979>`__ by Alexis
Conneau, Alexei Baevski, Ronan Collobert, Abdelrahman Mohamed, Michael Auli.
......@@ -369,6 +372,8 @@ Flax), PyTorch, and/or TensorFlow.
+-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
| RoBERTa | ✅ | ✅ | ✅ | ✅ | ✅ |
+-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
| RoFormer | ✅ | ✅ | ✅ | ✅ | ❌ |
+-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
| Speech2Text | ✅ | ❌ | ✅ | ❌ | ❌ |
+-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
| SqueezeBERT | ✅ | ✅ | ✅ | ❌ | ❌ |
......@@ -520,6 +525,7 @@ Flax), PyTorch, and/or TensorFlow.
model_doc/reformer
model_doc/retribert
model_doc/roberta
model_doc/roformer
model_doc/speech_to_text
model_doc/squeezebert
model_doc/t5
......
docs/source/model_doc/roformer.rst 0 → 100644
View file @ 206f06f2
..
Copyright 2021 The HuggingFace Team. 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.
RoFormer
-----------------------------------------------------------------------------------------------------------------------
Overview
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
The RoFormer model was proposed in `RoFormer: Enhanced Transformer with Rotary Position Embedding
<https://arxiv.org/pdf/2104.09864v1.pdf>`__ by Jianlin Su and Yu Lu and Shengfeng Pan and Bo Wen and Yunfeng Liu.
The abstract from the paper is the following:
*Position encoding in transformer architecture provides supervision for dependency modeling between elements at
different positions in the sequence. We investigate various methods to encode positional information in
transformer-based language models and propose a novel implementation named Rotary Position Embedding(RoPE). The
proposed RoPE encodes absolute positional information with rotation matrix and naturally incorporates explicit relative
position dependency in self-attention formulation. Notably, RoPE comes with valuable properties such as flexibility of
being expand to any sequence lengths, decaying inter-token dependency with increasing relative distances, and
capability of equipping the linear self-attention with relative position encoding. As a result, the enhanced
transformer with rotary position embedding, or RoFormer, achieves superior performance in tasks with long texts. We
release the theoretical analysis along with some preliminary experiment results on Chinese data. The undergoing
experiment for English benchmark will soon be updated.*
Tips:
- RoFormer is a BERT-like autoencoding model with rotary position embeddings. Rotary position embeddings have shown
improved performance on classification tasks with long texts.
This model was contributed by `junnyu <https://huggingface.co/junnyu>`__. The original code can be found `here
<https://github.com/ZhuiyiTechnology/roformer>`__.
RoFormerConfig
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
.. autoclass:: transformers.RoFormerConfig
:members:
RoFormerTokenizer
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
.. autoclass:: transformers.RoFormerTokenizer
:members: build_inputs_with_special_tokens, get_special_tokens_mask,
create_token_type_ids_from_sequences, save_vocabulary
RobertaTokenizerFast
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
.. autoclass:: transformers.RoFormerTokenizerFast
:members: build_inputs_with_special_tokens
RoFormerModel
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
.. autoclass:: transformers.RoFormerModel
:members: forward
RoFormerForCausalLM
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
.. autoclass:: transformers.RoFormerForCausalLM
:members: forward
RoFormerForMaskedLM
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
.. autoclass:: transformers.RoFormerForMaskedLM
:members: forward
RoFormerForSequenceClassification
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
.. autoclass:: transformers.RoFormerForSequenceClassification
:members: forward
RoFormerForMultipleChoice
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
.. autoclass:: transformers.RoFormerForMultipleChoice
:members: forward
RoFormerForTokenClassification
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
.. autoclass:: transformers.RoFormerForTokenClassification
:members: forward
RoFormerForQuestionAnswering
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
.. autoclass:: transformers.RoFormerForQuestionAnswering
:members: forward
TFRoFormerModel
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
.. autoclass:: transformers.TFRoFormerModel
:members: call
TFRoFormerForMaskedLM
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
.. autoclass:: transformers.TFRoFormerForMaskedLM
:members: call
TFRoFormerForCausalLM
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
.. autoclass:: transformers.TFRoFormerForCausalLM
:members: call
TFRoFormerForSequenceClassification
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
.. autoclass:: transformers.TFRoFormerForSequenceClassification
:members: call
TFRoFormerForMultipleChoice
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
.. autoclass:: transformers.TFRoFormerForMultipleChoice
:members: call
TFRoFormerForTokenClassification
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
.. autoclass:: transformers.TFRoFormerForTokenClassification
:members: call
TFRoFormerForQuestionAnswering
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
.. autoclass:: transformers.TFRoFormerForQuestionAnswering
:members: call
src/transformers/__init__.py
View file @ 206f06f2
......@@ -218,6 +218,7 @@ _import_structure = {
"models.reformer": ["REFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP", "ReformerConfig"],
"models.retribert": ["RETRIBERT_PRETRAINED_CONFIG_ARCHIVE_MAP", "RetriBertConfig", "RetriBertTokenizer"],
"models.roberta": ["ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP", "RobertaConfig", "RobertaTokenizer"],
"models.roformer": ["ROFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP", "RoFormerConfig", "RoFormerTokenizer"],
"models.speech_to_text": [
"SPEECH_TO_TEXT_PRETRAINED_CONFIG_ARCHIVE_MAP",
"Speech2TextConfig",
......@@ -322,6 +323,7 @@ else:
# tokenizers-backed objects
if is_tokenizers_available():
# Fast tokenizers
_import_structure["models.roformer"].append("RoFormerTokenizerFast")
_import_structure["models.clip"].append("CLIPTokenizerFast")
_import_structure["models.convbert"].append("ConvBertTokenizerFast")
_import_structure["models.albert"].append("AlbertTokenizerFast")
......@@ -927,6 +929,21 @@ if is_torch_available():
"RobertaModel",
]
)
_import_structure["models.roformer"].extend(
[
"ROFORMER_PRETRAINED_MODEL_ARCHIVE_LIST",
"RoFormerForCausalLM",
"RoFormerForMaskedLM",
"RoFormerForMultipleChoice",
"RoFormerForQuestionAnswering",
"RoFormerForSequenceClassification",
"RoFormerForTokenClassification",
"RoFormerLayer",
"RoFormerModel",
"RoFormerPreTrainedModel",
"load_tf_weights_in_roformer",
]
)
_import_structure["models.speech_to_text"].extend(
[
"SPEECH_TO_TEXT_PRETRAINED_MODEL_ARCHIVE_LIST",
......@@ -1339,6 +1356,20 @@ if is_tf_available():
"TFRobertaPreTrainedModel",
]
)
_import_structure["models.roformer"].extend(
[
"TF_ROFORMER_PRETRAINED_MODEL_ARCHIVE_LIST",
"TFRoFormerForCausalLM",
"TFRoFormerForMaskedLM",
"TFRoFormerForMultipleChoice",
"TFRoFormerForQuestionAnswering",
"TFRoFormerForSequenceClassification",
"TFRoFormerForTokenClassification",
"TFRoFormerLayer",
"TFRoFormerModel",
"TFRoFormerPreTrainedModel",
]
)
_import_structure["models.t5"].extend(
[
"TF_T5_PRETRAINED_MODEL_ARCHIVE_LIST",
......@@ -1641,6 +1672,7 @@ if TYPE_CHECKING:
from .models.reformer import REFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP, ReformerConfig
from .models.retribert import RETRIBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, RetriBertConfig, RetriBertTokenizer
from .models.roberta import ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP, RobertaConfig, RobertaTokenizer
from .models.roformer import ROFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP, RoFormerConfig, RoFormerTokenizer
from .models.speech_to_text import SPEECH_TO_TEXT_PRETRAINED_CONFIG_ARCHIVE_MAP, Speech2TextConfig
from .models.squeezebert import SQUEEZEBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, SqueezeBertConfig, SqueezeBertTokenizer
from .models.t5 import T5_PRETRAINED_CONFIG_ARCHIVE_MAP, T5Config
......@@ -1767,6 +1799,7 @@ if TYPE_CHECKING:
from .models.reformer import ReformerTokenizerFast
from .models.retribert import RetriBertTokenizerFast
from .models.roberta import RobertaTokenizerFast
from .models.roformer import RoFormerTokenizerFast
from .models.squeezebert import SqueezeBertTokenizerFast
from .models.t5 import T5TokenizerFast
from .models.xlm_roberta import XLMRobertaTokenizerFast
......@@ -2232,6 +2265,19 @@ if TYPE_CHECKING:
RobertaForTokenClassification,
RobertaModel,
)
from .models.roformer import (
ROFORMER_PRETRAINED_MODEL_ARCHIVE_LIST,
RoFormerForCausalLM,
RoFormerForMaskedLM,
RoFormerForMultipleChoice,
RoFormerForQuestionAnswering,
RoFormerForSequenceClassification,
RoFormerForTokenClassification,
RoFormerLayer,
RoFormerModel,
RoFormerPreTrainedModel,
load_tf_weights_in_roformer,
)
from .models.speech_to_text import (
SPEECH_TO_TEXT_PRETRAINED_MODEL_ARCHIVE_LIST,
Speech2TextForConditionalGeneration,
......@@ -2575,6 +2621,18 @@ if TYPE_CHECKING:
TFRobertaModel,
TFRobertaPreTrainedModel,
)
from .models.roformer import (
TF_ROFORMER_PRETRAINED_MODEL_ARCHIVE_LIST,
TFRoFormerForCausalLM,
TFRoFormerForMaskedLM,
TFRoFormerForMultipleChoice,
TFRoFormerForQuestionAnswering,
TFRoFormerForSequenceClassification,
TFRoFormerForTokenClassification,
TFRoFormerLayer,
TFRoFormerModel,
TFRoFormerPreTrainedModel,
)
from .models.t5 import (
TF_T5_PRETRAINED_MODEL_ARCHIVE_LIST,
TFT5EncoderModel,
......
src/transformers/convert_slow_tokenizer.py
View file @ 206f06f2
......@@ -25,6 +25,7 @@ from tokenizers import Regex, Tokenizer, decoders, normalizers, pre_tokenizers,
from tokenizers.models import BPE, Unigram, WordPiece
from .file_utils import requires_backends
from .models.roformer.tokenization_utils import JiebaPreTokenizer
class SentencePieceExtractor:
......@@ -296,6 +297,43 @@ class RobertaConverter(Converter):
return tokenizer
class RoFormerConverter(Converter):
def converted(self) -> Tokenizer:
vocab = self.original_tokenizer.vocab
tokenizer = Tokenizer(WordPiece(vocab, unk_token=str(self.original_tokenizer.unk_token)))
strip_accents = False
do_lower_case = False
if hasattr(self.original_tokenizer, "basic_tokenizer"):
strip_accents = self.original_tokenizer.basic_tokenizer.strip_accents
do_lower_case = self.original_tokenizer.basic_tokenizer.do_lower_case
tokenizer.normalizer = normalizers.BertNormalizer(
clean_text=True,
handle_chinese_chars=False,
strip_accents=strip_accents,
lowercase=do_lower_case,
)
tokenizer.pre_tokenizer = pre_tokenizers.PreTokenizer.custom(JiebaPreTokenizer(vocab))
cls = str(self.original_tokenizer.cls_token)
sep = str(self.original_tokenizer.sep_token)
cls_token_id = self.original_tokenizer.cls_token_id
sep_token_id = self.original_tokenizer.sep_token_id
tokenizer.post_processor = processors.TemplateProcessing(
single=f"{cls}:0 $A:0 {sep}:0",
pair=f"{cls}:0 $A:0 {sep}:0 $B:1 {sep}:1",
special_tokens=[
(cls, cls_token_id),
(sep, sep_token_id),
],
)
tokenizer.decoder = decoders.WordPiece(prefix="##")
return tokenizer
class DebertaConverter(Converter):
def converted(self) -> Tokenizer:
ot = self.original_tokenizer
......@@ -755,6 +793,7 @@ SLOW_TO_FAST_CONVERTERS = {
"ReformerTokenizer": ReformerConverter,
"RetriBertTokenizer": BertConverter,
"RobertaTokenizer": RobertaConverter,
"RoFormerTokenizer": RoFormerConverter,
"SqueezeBertTokenizer": BertConverter,
"T5Tokenizer": T5Converter,
"XLMRobertaTokenizer": XLMRobertaConverter,
......
src/transformers/models/__init__.py
View file @ 206f06f2
......@@ -68,6 +68,7 @@ from . import (
reformer,
retribert,
roberta,
roformer,
speech_to_text,
squeezebert,
t5,
......
src/transformers/models/auto/configuration_auto.py
View file @ 206f06f2
......@@ -68,6 +68,7 @@ from ..rag.configuration_rag import RagConfig
from ..reformer.configuration_reformer import ReformerConfig
from ..retribert.configuration_retribert import RETRIBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, RetriBertConfig
from ..roberta.configuration_roberta import ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP, RobertaConfig
from ..roformer.configuration_roformer import ROFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP, RoFormerConfig
from ..speech_to_text.configuration_speech_to_text import (
SPEECH_TO_TEXT_PRETRAINED_CONFIG_ARCHIVE_MAP,
Speech2TextConfig,
......@@ -91,6 +92,7 @@ ALL_PRETRAINED_CONFIG_ARCHIVE_MAP = dict(
(key, value)
for pretrained_map in [
# Add archive maps here
ROFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP,
CLIP_PRETRAINED_CONFIG_ARCHIVE_MAP,
BIGBIRD_PEGASUS_PRETRAINED_CONFIG_ARCHIVE_MAP,
DEIT_PRETRAINED_CONFIG_ARCHIVE_MAP,
......@@ -146,6 +148,7 @@ ALL_PRETRAINED_CONFIG_ARCHIVE_MAP = dict(
CONFIG_MAPPING = OrderedDict(
[
# Add configs here
("roformer", RoFormerConfig),
("clip", CLIPConfig),
("bigbird_pegasus", BigBirdPegasusConfig),
("deit", DeiTConfig),
......@@ -207,6 +210,7 @@ CONFIG_MAPPING = OrderedDict(
MODEL_NAMES_MAPPING = OrderedDict(
[
# Add full (and cased) model names here
("roformer", "RoFormer"),
("clip", "CLIP"),
("bigbird_pegasus", "BigBirdPegasus"),
("deit", "DeiT"),
......
src/transformers/models/auto/modeling_auto.py
View file @ 206f06f2
......@@ -240,6 +240,15 @@ from ..roberta.modeling_roberta import (
RobertaForTokenClassification,
RobertaModel,
)
from ..roformer.modeling_roformer import (
RoFormerForCausalLM,
RoFormerForMaskedLM,
RoFormerForMultipleChoice,
RoFormerForQuestionAnswering,
RoFormerForSequenceClassification,
RoFormerForTokenClassification,
RoFormerModel,
)
from ..speech_to_text.modeling_speech_to_text import Speech2TextForConditionalGeneration, Speech2TextModel
from ..squeezebert.modeling_squeezebert import (
SqueezeBertForMaskedLM,
......@@ -334,6 +343,7 @@ from .configuration_auto import (
ReformerConfig,
RetriBertConfig,
RobertaConfig,
RoFormerConfig,
Speech2TextConfig,
SqueezeBertConfig,
T5Config,
......@@ -354,6 +364,7 @@ logger = logging.get_logger(__name__)
MODEL_MAPPING = OrderedDict(
[
# Base model mapping
(RoFormerConfig, RoFormerModel),
(CLIPConfig, CLIPModel),
(BigBirdPegasusConfig, BigBirdPegasusModel),
(DeiTConfig, DeiTModel),
......@@ -451,6 +462,7 @@ MODEL_FOR_PRETRAINING_MAPPING = OrderedDict(
MODEL_WITH_LM_HEAD_MAPPING = OrderedDict(
[
# Model with LM heads mapping
(RoFormerConfig, RoFormerForMaskedLM),
(BigBirdPegasusConfig, BigBirdPegasusForConditionalGeneration),
(GPTNeoConfig, GPTNeoForCausalLM),
(BigBirdConfig, BigBirdForMaskedLM),
......@@ -498,6 +510,7 @@ MODEL_WITH_LM_HEAD_MAPPING = OrderedDict(
MODEL_FOR_CAUSAL_LM_MAPPING = OrderedDict(
[
# Model for Causal LM mapping
(RoFormerConfig, RoFormerForCausalLM),
(BigBirdPegasusConfig, BigBirdPegasusForCausalLM),
(GPTNeoConfig, GPTNeoForCausalLM),
(BigBirdConfig, BigBirdForCausalLM),
......@@ -539,6 +552,7 @@ MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING = OrderedDict(
MODEL_FOR_MASKED_LM_MAPPING = OrderedDict(
[
# Model for Masked LM mapping
(RoFormerConfig, RoFormerForMaskedLM),
(BigBirdConfig, BigBirdForMaskedLM),
(Wav2Vec2Config, Wav2Vec2ForMaskedLM),
(ConvBertConfig, ConvBertForMaskedLM),
......@@ -592,6 +606,7 @@ MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING = OrderedDict(
MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING = OrderedDict(
[
# Model for Sequence Classification mapping
(RoFormerConfig, RoFormerForSequenceClassification),
(BigBirdPegasusConfig, BigBirdPegasusForSequenceClassification),
(BigBirdConfig, BigBirdForSequenceClassification),
(ConvBertConfig, ConvBertForSequenceClassification),
......@@ -630,6 +645,7 @@ MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING = OrderedDict(
MODEL_FOR_QUESTION_ANSWERING_MAPPING = OrderedDict(
[
# Model for Question Answering mapping
(RoFormerConfig, RoFormerForQuestionAnswering),
(BigBirdPegasusConfig, BigBirdPegasusForQuestionAnswering),
(BigBirdConfig, BigBirdForQuestionAnswering),
(ConvBertConfig, ConvBertForQuestionAnswering),
......@@ -670,6 +686,7 @@ MODEL_FOR_TABLE_QUESTION_ANSWERING_MAPPING = OrderedDict(
MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING = OrderedDict(
[
# Model for Token Classification mapping
(RoFormerConfig, RoFormerForTokenClassification),
(BigBirdConfig, BigBirdForTokenClassification),
(ConvBertConfig, ConvBertForTokenClassification),
(LayoutLMConfig, LayoutLMForTokenClassification),
......@@ -699,6 +716,7 @@ MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING = OrderedDict(
MODEL_FOR_MULTIPLE_CHOICE_MAPPING = OrderedDict(
[
# Model for Multiple Choice mapping
(RoFormerConfig, RoFormerForMultipleChoice),
(BigBirdConfig, BigBirdForMultipleChoice),
(ConvBertConfig, ConvBertForMultipleChoice),
(CamembertConfig, CamembertForMultipleChoice),
......
src/transformers/models/auto/modeling_tf_auto.py
View file @ 206f06f2
......@@ -148,6 +148,15 @@ from ..roberta.modeling_tf_roberta import (
TFRobertaForTokenClassification,
TFRobertaModel,
)
from ..roformer.modeling_tf_roformer import (
TFRoFormerForCausalLM,
TFRoFormerForMaskedLM,
TFRoFormerForMultipleChoice,
TFRoFormerForQuestionAnswering,
TFRoFormerForSequenceClassification,
TFRoFormerForTokenClassification,
TFRoFormerModel,
)
from ..t5.modeling_tf_t5 import TFT5ForConditionalGeneration, TFT5Model
from ..transfo_xl.modeling_tf_transfo_xl import (
TFTransfoXLForSequenceClassification,
......@@ -206,6 +215,7 @@ from .configuration_auto import (
OpenAIGPTConfig,
PegasusConfig,
RobertaConfig,
RoFormerConfig,
T5Config,
TransfoXLConfig,
XLMConfig,
......@@ -220,6 +230,7 @@ logger = logging.get_logger(__name__)
TF_MODEL_MAPPING = OrderedDict(
[
# Base model mapping
(RoFormerConfig, TFRoFormerModel),
(ConvBertConfig, TFConvBertModel),
(LEDConfig, TFLEDModel),
(LxmertConfig, TFLxmertModel),
......@@ -285,6 +296,7 @@ TF_MODEL_FOR_PRETRAINING_MAPPING = OrderedDict(
TF_MODEL_WITH_LM_HEAD_MAPPING = OrderedDict(
[
# Model with LM heads mapping
(RoFormerConfig, TFRoFormerForMaskedLM),
(ConvBertConfig, TFConvBertForMaskedLM),
(LEDConfig, TFLEDForConditionalGeneration),
(T5Config, TFT5ForConditionalGeneration),
......@@ -315,6 +327,7 @@ TF_MODEL_WITH_LM_HEAD_MAPPING = OrderedDict(
TF_MODEL_FOR_CAUSAL_LM_MAPPING = OrderedDict(
[
# Model for Causal LM mapping
(RoFormerConfig, TFRoFormerForCausalLM),
(BertConfig, TFBertLMHeadModel),
(OpenAIGPTConfig, TFOpenAIGPTLMHeadModel),
(GPT2Config, TFGPT2LMHeadModel),
......@@ -331,6 +344,7 @@ TF_MODEL_FOR_CAUSAL_LM_MAPPING = OrderedDict(
TF_MODEL_FOR_MASKED_LM_MAPPING = OrderedDict(
[
# Model for Masked LM mapping
(RoFormerConfig, TFRoFormerForMaskedLM),
(ConvBertConfig, TFConvBertForMaskedLM),
(DistilBertConfig, TFDistilBertForMaskedLM),
(AlbertConfig, TFAlbertForMaskedLM),
......@@ -368,6 +382,7 @@ TF_MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING = OrderedDict(
TF_MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING = OrderedDict(
[
# Model for Sequence Classification mapping
(RoFormerConfig, TFRoFormerForSequenceClassification),
(ConvBertConfig, TFConvBertForSequenceClassification),
(DistilBertConfig, TFDistilBertForSequenceClassification),
(AlbertConfig, TFAlbertForSequenceClassification),
......@@ -394,6 +409,7 @@ TF_MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING = OrderedDict(
TF_MODEL_FOR_QUESTION_ANSWERING_MAPPING = OrderedDict(
[
# Model for Question Answering mapping
(RoFormerConfig, TFRoFormerForQuestionAnswering),
(ConvBertConfig, TFConvBertForQuestionAnswering),
(DistilBertConfig, TFDistilBertForQuestionAnswering),
(AlbertConfig, TFAlbertForQuestionAnswering),
......@@ -415,6 +431,7 @@ TF_MODEL_FOR_QUESTION_ANSWERING_MAPPING = OrderedDict(
TF_MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING = OrderedDict(
[
# Model for Token Classification mapping
(RoFormerConfig, TFRoFormerForTokenClassification),
(ConvBertConfig, TFConvBertForTokenClassification),
(DistilBertConfig, TFDistilBertForTokenClassification),
(AlbertConfig, TFAlbertForTokenClassification),
......@@ -437,6 +454,7 @@ TF_MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING = OrderedDict(
TF_MODEL_FOR_MULTIPLE_CHOICE_MAPPING = OrderedDict(
[
# Model for Multiple Choice mapping
(RoFormerConfig, TFRoFormerForMultipleChoice),
(ConvBertConfig, TFConvBertForMultipleChoice),
(CamembertConfig, TFCamembertForMultipleChoice),
(XLMConfig, TFXLMForMultipleChoice),
......
src/transformers/models/auto/tokenization_auto.py
View file @ 206f06f2
......@@ -51,6 +51,7 @@ from ..prophetnet.tokenization_prophetnet import ProphetNetTokenizer
from ..rag.tokenization_rag import RagTokenizer
from ..retribert.tokenization_retribert import RetriBertTokenizer
from ..roberta.tokenization_roberta import RobertaTokenizer
from ..roformer.tokenization_roformer import RoFormerTokenizer
from ..squeezebert.tokenization_squeezebert import SqueezeBertTokenizer
from ..tapas.tokenization_tapas import TapasTokenizer
from ..transfo_xl.tokenization_transfo_xl import TransfoXLTokenizer
......@@ -98,6 +99,7 @@ from .configuration_auto import (
ReformerConfig,
RetriBertConfig,
RobertaConfig,
RoFormerConfig,
Speech2TextConfig,
SqueezeBertConfig,
T5Config,
......@@ -228,6 +230,7 @@ logger = logging.get_logger(__name__)
TOKENIZER_MAPPING = OrderedDict(
[
(RetriBertConfig, (RetriBertTokenizer, RetriBertTokenizerFast)),
(RoFormerConfig, (RoFormerTokenizer, None)),
(T5Config, (T5Tokenizer, T5TokenizerFast)),
(MT5Config, (MT5Tokenizer, MT5TokenizerFast)),
(MobileBertConfig, (MobileBertTokenizer, MobileBertTokenizerFast)),
......
src/transformers/models/roformer/__init__.py 0 → 100644
View file @ 206f06f2
# flake8: noqa
# There's no way to ignore "F401 '...' imported but unused" warnings in this
# module, but to preserve other warnings. So, don't check this module at all.
# Copyright 2021 The HuggingFace Team. 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.
from typing import TYPE_CHECKING
from ...file_utils import _BaseLazyModule, is_tf_available, is_tokenizers_available, is_torch_available
_import_structure = {
"configuration_roformer": ["ROFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP", "RoFormerConfig"],
"tokenization_roformer": ["RoFormerTokenizer"],
}
if is_tokenizers_available():
_import_structure["tokenization_roformer_fast"] = ["RoFormerTokenizerFast"]
if is_torch_available():
_import_structure["modeling_roformer"] = [
"ROFORMER_PRETRAINED_MODEL_ARCHIVE_LIST",
"RoFormerForCausalLM",
"RoFormerForMaskedLM",
"RoFormerForMultipleChoice",
"RoFormerForQuestionAnswering",
"RoFormerForSequenceClassification",
"RoFormerForTokenClassification",
"RoFormerLayer",
"RoFormerModel",
"RoFormerPreTrainedModel",
"load_tf_weights_in_roformer",
]
if is_tf_available():
_import_structure["modeling_tf_roformer"] = [
"TF_ROFORMER_PRETRAINED_MODEL_ARCHIVE_LIST",
"TFRoFormerForCausalLM",
"TFRoFormerForMaskedLM",
"TFRoFormerForMultipleChoice",
"TFRoFormerForQuestionAnswering",
"TFRoFormerForSequenceClassification",
"TFRoFormerForTokenClassification",
"TFRoFormerLayer",
"TFRoFormerModel",
"TFRoFormerPreTrainedModel",
]
if TYPE_CHECKING:
from .configuration_roformer import ROFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP, RoFormerConfig
from .tokenization_roformer import RoFormerTokenizer
if is_tokenizers_available():
from .tokenization_roformer_fast import RoFormerTokenizerFast
if is_torch_available():
from .modeling_roformer import (
ROFORMER_PRETRAINED_MODEL_ARCHIVE_LIST,
RoFormerForCausalLM,
RoFormerForMaskedLM,
RoFormerForMultipleChoice,
RoFormerForQuestionAnswering,
RoFormerForSequenceClassification,
RoFormerForTokenClassification,
RoFormerLayer,
RoFormerModel,
RoFormerPreTrainedModel,
load_tf_weights_in_roformer,
)
if is_tf_available():
from .modeling_tf_roformer import (
TF_ROFORMER_PRETRAINED_MODEL_ARCHIVE_LIST,
TFRoFormerForCausalLM,
TFRoFormerForMaskedLM,
TFRoFormerForMultipleChoice,
TFRoFormerForQuestionAnswering,
TFRoFormerForSequenceClassification,
TFRoFormerForTokenClassification,
TFRoFormerLayer,
TFRoFormerModel,
TFRoFormerPreTrainedModel,
)
else:
import importlib
import os
import sys
class _LazyModule(_BaseLazyModule):
"""
Module class that surfaces all objects but only performs associated imports when the objects are requested.
"""
__file__ = globals()["__file__"]
__path__ = [os.path.dirname(__file__)]
def _get_module(self, module_name: str):
return importlib.import_module("." + module_name, self.__name__)
sys.modules[__name__] = _LazyModule(__name__, _import_structure)
src/transformers/models/roformer/configuration_roformer.py 0 → 100644
View file @ 206f06f2
# coding=utf-8
# Copyright 2021 The HuggingFace Inc. team. 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.
""" RoFormer model configuration """
from ...configuration_utils import PretrainedConfig
from ...utils import logging
logger = logging.get_logger(__name__)
ROFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP = {
"junnyu/roformer_chinese_small": "https://huggingface.co/junnyu/roformer_chinese_small/resolve/main/config.json",
"junnyu/roformer_chinese_base": "https://huggingface.co/junnyu/roformer_chinese_base/resolve/main/config.json"
# See all RoFormer models at https://huggingface.co/models?filter=roformer
}
class RoFormerConfig(PretrainedConfig):
r"""
This is the configuration class to store the configuration of a :class:`~transformers.RoFormerModel`. It is used to
instantiate an RoFormer model according to the specified arguments, defining the model architecture. Instantiating
a configuration with the defaults will yield a similar configuration to that of the RoFormer
`junnyu/roformer_chinese_base <https://huggingface.co/junnyu/roformer_chinese_base>`__ architecture.
Configuration objects inherit from :class:`~transformers.PretrainedConfig` and can be used to control the model
outputs. Read the documentation from :class:`~transformers.PretrainedConfig` for more information.
Args:
vocab_size (:obj:`int`, `optional`, defaults to 50000):
Vocabulary size of the RoFormer model. Defines the number of different tokens that can be represented by
the :obj:`inputs_ids` passed when calling :class:`~transformers.RoFormerModel` or
:class:`~transformers.TFRoFormerModel`.
embedding_size (:obj:`int`, `optional`, defaults to 768):
Dimensionality of the encoder layers and the pooler layer.
hidden_size (:obj:`int`, `optional`, defaults to 768):
Dimension of the encoder layers and the pooler layer.
num_hidden_layers (:obj:`int`, `optional`, defaults to 12):
Number of hidden layers in the Transformer encoder.
num_attention_heads (:obj:`int`, `optional`, defaults to 12):
Number of attention heads for each attention layer in the Transformer encoder.
intermediate_size (:obj:`int`, `optional`, defaults to 3072):
Dimension of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
hidden_act (:obj:`str` or :obj:`function`, `optional`, defaults to :obj:`"gelu"`):
The non-linear activation function (function or string) in the encoder and pooler. If string,
:obj:`"gelu"`, :obj:`"relu"`, :obj:`"selu"` and :obj:`"gelu_new"` are supported.
hidden_dropout_prob (:obj:`float`, `optional`, defaults to 0.1):
The dropout probabilitiy for all fully connected layers in the embeddings, encoder, and pooler.
attention_probs_dropout_prob (:obj:`float`, `optional`, defaults to 0.1):
The dropout ratio for the attention probabilities.
max_position_embeddings (:obj:`int`, `optional`, defaults to 1536):
The maximum sequence length that this model might ever be used with. Typically set this to something large
just in case (e.g., 512 or 1024 or 1536).
type_vocab_size (:obj:`int`, `optional`, defaults to 2):
The vocabulary size of the :obj:`token_type_ids` passed when calling :class:`~transformers.RoFormerModel`
or :class:`~transformers.TFRoFormerModel`.
initializer_range (:obj:`float`, `optional`, defaults to 0.02):
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
layer_norm_eps (:obj:`float`, `optional`, defaults to 1e-12):
The epsilon used by the layer normalization layers.
use_cache (:obj:`bool`, `optional`, defaults to :obj:`True`):
Whether or not the model should return the last key/values attentions (not used by all models). Only
relevant if ``config.is_decoder=True``.
rotary_value (:obj:`bool`, `optional`, defaults to :obj:`False`):
Whether or not apply rotary position embeddings on value layer.
gradient_checkpointing (:obj:`bool`, `optional`, defaults to :obj:`False`):
If :obj:`True`, use gradient checkpointing to save memory at the expense of slower backward pass.
Example::
>>> from transformers import RoFormerModel, RoFormerConfig
>>> # Initializing a RoFormer junnyu/roformer_chinese_base style configuration
>>> configuration = RoFormerConfig()
>>> # Initializing a model from the junnyu/roformer_chinese_base style configuration
>>> model = RoFormerModel(configuration)
>>> # Accessing the model configuration
>>> configuration = model.config
"""
model_type = "roformer"
def __init__(
self,
vocab_size=50000,
embedding_size=768,
hidden_size=768,
num_hidden_layers=12,
num_attention_heads=12,
intermediate_size=3072,
hidden_act="gelu",
hidden_dropout_prob=0.1,
attention_probs_dropout_prob=0.1,
max_position_embeddings=1536,
type_vocab_size=2,
initializer_range=0.02,
layer_norm_eps=1e-12,
pad_token_id=0,
gradient_checkpointing=False,
rotary_value=False,
use_cache=True,
**kwargs
):
super().__init__(pad_token_id=pad_token_id, **kwargs)
self.vocab_size = vocab_size
self.embedding_size = embedding_size
self.hidden_size = hidden_size
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.hidden_act = hidden_act
self.intermediate_size = intermediate_size
self.hidden_dropout_prob = hidden_dropout_prob
self.attention_probs_dropout_prob = attention_probs_dropout_prob
self.max_position_embeddings = max_position_embeddings
self.type_vocab_size = type_vocab_size
self.initializer_range = initializer_range
self.layer_norm_eps = layer_norm_eps
self.gradient_checkpointing = gradient_checkpointing
self.rotary_value = rotary_value
self.use_cache = use_cache
src/transformers/models/roformer/convert_roformer_original_tf_checkpoint_to_pytorch.py 0 → 100755
View file @ 206f06f2
# coding=utf-8
# Copyright 2021 The HuggingFace Inc. team. 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.
"""Convert RoFormer checkpoint."""
import argparse
import torch
from transformers import RoFormerConfig, RoFormerForMaskedLM, load_tf_weights_in_roformer
from transformers.utils import logging
logging.set_verbosity_info()
def convert_tf_checkpoint_to_pytorch(tf_checkpoint_path, bert_config_file, pytorch_dump_path):
# Initialise PyTorch model
config = RoFormerConfig.from_json_file(bert_config_file)
print(f"Building PyTorch model from configuration: {config}")
model = RoFormerForMaskedLM(config)
# Load weights from tf checkpoint
load_tf_weights_in_roformer(model, config, tf_checkpoint_path)
# Save pytorch-model
print(f"Save PyTorch model to {pytorch_dump_path}")
torch.save(model.state_dict(), pytorch_dump_path, _use_new_zipfile_serialization=False)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
# Required parameters
parser.add_argument(
"--tf_checkpoint_path", default=None, type=str, required=True, help="Path to the TensorFlow checkpoint path."
)
parser.add_argument(
"--bert_config_file",
default=None,
type=str,
required=True,
help="The config json file corresponding to the pre-trained BERT model. \n"
"This specifies the model architecture.",
)
parser.add_argument(
"--pytorch_dump_path", default=None, type=str, required=True, help="Path to the output PyTorch model."
)
args = parser.parse_args()
convert_tf_checkpoint_to_pytorch(args.tf_checkpoint_path, args.bert_config_file, args.pytorch_dump_path)
src/transformers/models/roformer/modeling_roformer.py 0 → 100644
View file @ 206f06f2
# coding=utf-8
# Copyright 2021 The HuggingFace Inc. team. 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.
""" PyTorch RoFormer model. """
import math
import os
from typing import Optional
import numpy as np
import torch
import torch.utils.checkpoint
from torch import nn
from torch.nn import CrossEntropyLoss, MSELoss
from ...activations import ACT2FN
from ...file_utils import (
add_code_sample_docstrings,
add_start_docstrings,
add_start_docstrings_to_model_forward,
replace_return_docstrings,
)
from ...modeling_outputs import (
BaseModelOutputWithPastAndCrossAttentions,
CausalLMOutputWithCrossAttentions,
MaskedLMOutput,
MultipleChoiceModelOutput,
QuestionAnsweringModelOutput,
SequenceClassifierOutput,
TokenClassifierOutput,
)
from ...modeling_utils import (
PreTrainedModel,
SequenceSummary,
apply_chunking_to_forward,
find_pruneable_heads_and_indices,
prune_linear_layer,
)
from ...utils import logging
from .configuration_roformer import RoFormerConfig
logger = logging.get_logger(__name__)
_CHECKPOINT_FOR_DOC = "junnyu/roformer_chinese_base"
_CONFIG_FOR_DOC = "RoFormerConfig"
_TOKENIZER_FOR_DOC = "RoFormerTokenizer"
ROFORMER_PRETRAINED_MODEL_ARCHIVE_LIST = [
"junnyu/roformer_chinese_small",
"junnyu/roformer_chinese_base"
# See all RoFormer models at https://huggingface.co/models?filter=roformer
]
# Copied from transformers.models.marian.modeling_marian.MarianSinusoidalPositionalEmbedding with Marian->RoFormer
class RoFormerSinusoidalPositionalEmbedding(nn.Embedding):
"""This module produces sinusoidal positional embeddings of any length."""
def __init__(self, num_positions: int, embedding_dim: int, padding_idx: Optional[int] = None):
super().__init__(num_positions, embedding_dim)
self.weight = self._init_weight(self.weight)
@staticmethod
def _init_weight(out: nn.Parameter):
"""
Identical to the XLM create_sinusoidal_embeddings except features are not interleaved. The cos features are in
the 2nd half of the vector. [dim // 2:]
"""
n_pos, dim = out.shape
position_enc = np.array(
[[pos / np.power(10000, 2 * (j // 2) / dim) for j in range(dim)] for pos in range(n_pos)]
)
out.requires_grad = False # set early to avoid an error in pytorch-1.8+
sentinel = dim // 2 if dim % 2 == 0 else (dim // 2) + 1
out[:, 0:sentinel] = torch.FloatTensor(np.sin(position_enc[:, 0::2]))
out[:, sentinel:] = torch.FloatTensor(np.cos(position_enc[:, 1::2]))
out.detach_()
return out
@torch.no_grad()
def forward(self, input_ids_shape: torch.Size, past_key_values_length: int = 0):
"""`input_ids_shape` is expected to be [bsz x seqlen]."""
bsz, seq_len = input_ids_shape[:2]
positions = torch.arange(
past_key_values_length, past_key_values_length + seq_len, dtype=torch.long, device=self.weight.device
)
return super().forward(positions)
def load_tf_weights_in_roformer(model, config, tf_checkpoint_path):
"""Load tf checkpoints in a pytorch model."""
try:
import re
import numpy as np
import tensorflow as tf
except ImportError:
logger.error(
"Loading a TensorFlow model in PyTorch, requires TensorFlow to be installed. Please see "
"https://www.tensorflow.org/install/ for installation instructions."
)
raise
tf_path = os.path.abspath(tf_checkpoint_path)
logger.info(f"Converting TensorFlow checkpoint from {tf_path}")
# Load weights from TF model
init_vars = tf.train.list_variables(tf_path)
names = []
arrays = []
for name, shape in init_vars:
logger.info(f"Loading TF weight {name} with shape {shape}")
array = tf.train.load_variable(tf_path, name)
names.append(name.replace("bert", "roformer"))
arrays.append(array)
for name, array in zip(names, arrays):
name = name.split("/")
# adam_v and adam_m are variables used in AdamWeightDecayOptimizer to calculated m and v
# which are not required for using pretrained model
if any(
n in ["adam_v", "adam_m", "AdamWeightDecayOptimizer", "AdamWeightDecayOptimizer_1", "global_step"]
for n in name
):
logger.info(f"Skipping {'/'.join(name)}")
continue
pointer = model
for m_name in name:
if re.fullmatch(r"[A-Za-z]+_\d+", m_name):
scope_names = re.split(r"_(\d+)", m_name)
else:
scope_names = [m_name]
if scope_names[0] == "kernel" or scope_names[0] == "gamma":
pointer = getattr(pointer, "weight")
elif scope_names[0] == "output_bias" or scope_names[0] == "beta":
pointer = getattr(pointer, "bias")
elif scope_names[0] == "output_weights":
pointer = getattr(pointer, "weight")
elif scope_names[0] == "squad":
pointer = getattr(pointer, "classifier")
else:
try:
pointer = getattr(pointer, scope_names[0])
except AttributeError:
logger.info(f"Skipping {'/'.join(name)}")
continue
if len(scope_names) >= 2:
num = int(scope_names[1])
pointer = pointer[num]
if m_name[-11:] == "_embeddings":
pointer = getattr(pointer, "weight")
elif m_name == "kernel":
array = np.transpose(array)
try:
assert (
pointer.shape == array.shape
), f"Pointer shape {pointer.shape} and array shape {array.shape} mismatched"
except AssertionError as e:
e.args += (pointer.shape, array.shape)
raise
logger.info(f"Initialize PyTorch weight {name}")
pointer.data = torch.from_numpy(array)
return model
class RoFormerEmbeddings(nn.Module):
"""Construct the embeddings from word and token_type embeddings."""
def __init__(self, config):
super().__init__()
self.word_embeddings = nn.Embedding(config.vocab_size, config.embedding_size, padding_idx=config.pad_token_id)
self.token_type_embeddings = nn.Embedding(config.type_vocab_size, config.embedding_size)
# self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load
# any TensorFlow checkpoint file
self.LayerNorm = nn.LayerNorm(config.embedding_size, eps=config.layer_norm_eps)
self.dropout = nn.Dropout(config.hidden_dropout_prob)
def forward(self, input_ids=None, token_type_ids=None, inputs_embeds=None):
if input_ids is not None:
input_shape = input_ids.size()
else:
input_shape = inputs_embeds.size()[:-1]
if inputs_embeds is None:
inputs_embeds = self.word_embeddings(input_ids)
if token_type_ids is None:
token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=inputs_embeds.device)
token_type_embeddings = self.token_type_embeddings(token_type_ids)
embeddings = inputs_embeds + token_type_embeddings
embeddings = self.LayerNorm(embeddings)
embeddings = self.dropout(embeddings)
return embeddings
class RoFormerSelfAttention(nn.Module):
def __init__(self, config):
super().__init__()
if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"):
raise ValueError(
f"The hidden size ({config.hidden_size}) is not a multiple of the number of attention "
f"heads ({config.num_attention_heads})"
)
self.num_attention_heads = config.num_attention_heads
self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
self.all_head_size = self.num_attention_heads * self.attention_head_size
self.query = nn.Linear(config.hidden_size, self.all_head_size)
self.key = nn.Linear(config.hidden_size, self.all_head_size)
self.value = nn.Linear(config.hidden_size, self.all_head_size)
self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
self.is_decoder = config.is_decoder
self.rotary_value = config.rotary_value
def transpose_for_scores(self, x):
new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
x = x.view(*new_x_shape)
return x.permute(0, 2, 1, 3)
def forward(
self,
hidden_states,
attention_mask=None,
sinusoidal_pos=None,
head_mask=None,
encoder_hidden_states=None,
encoder_attention_mask=None,
past_key_value=None,
output_attentions=False,
):
mixed_query_layer = self.query(hidden_states)
query_layer = self.transpose_for_scores(mixed_query_layer)
# If this is instantiated as a cross-attention module, the keys
# and values come from an encoder; the attention mask needs to be
# such that the encoder's padding tokens are not attended to.
is_cross_attention = encoder_hidden_states is not None
if is_cross_attention and past_key_value is not None:
# reuse k,v, cross_attentions
key_layer = past_key_value[0]
value_layer = past_key_value[1]
attention_mask = encoder_attention_mask
elif is_cross_attention:
key_layer = self.transpose_for_scores(self.key(encoder_hidden_states))
value_layer = self.transpose_for_scores(self.value(encoder_hidden_states))
attention_mask = encoder_attention_mask
elif past_key_value is not None:
key_layer = self.transpose_for_scores(self.key(hidden_states))
value_layer = self.transpose_for_scores(self.value(hidden_states))
key_layer = torch.cat([past_key_value[0], key_layer], dim=2)
value_layer = torch.cat([past_key_value[1], value_layer], dim=2)
else:
key_layer = self.transpose_for_scores(self.key(hidden_states))
value_layer = self.transpose_for_scores(self.value(hidden_states))
if sinusoidal_pos is not None:
if self.rotary_value:
query_layer, key_layer, value_layer = self.apply_rotary_position_embeddings(
sinusoidal_pos, query_layer, key_layer, value_layer
)
else:
query_layer, key_layer = self.apply_rotary_position_embeddings(
sinusoidal_pos, query_layer, key_layer
)
if self.is_decoder:
# if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
# Further calls to cross_attention layer can then reuse all cross-attention
# key/value_states (first "if" case)
# if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
# all previous decoder key/value_states. Further calls to uni-directional self-attention
# can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
# if encoder bi-directional self-attention `past_key_value` is always `None`
past_key_value = (key_layer, value_layer)
# Take the dot product between "query" and "key" to get the raw attention scores.
attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
attention_scores = attention_scores / math.sqrt(self.attention_head_size)
if attention_mask is not None:
# Apply the attention mask is (precomputed for all layers in RoFormerModel forward() function)
attention_scores = attention_scores + attention_mask
# Normalize the attention scores to probabilities.
attention_probs = nn.Softmax(dim=-1)(attention_scores)
# This is actually dropping out entire tokens to attend to, which might
# seem a bit unusual, but is taken from the original Transformer paper.
attention_probs = self.dropout(attention_probs)
# Mask heads if we want to
if head_mask is not None:
attention_probs = attention_probs * head_mask
context_layer = torch.matmul(attention_probs, value_layer)
context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
context_layer = context_layer.view(*new_context_layer_shape)
outputs = (context_layer, attention_probs) if output_attentions else (context_layer,)
if self.is_decoder:
outputs = outputs + (past_key_value,)
return outputs
@staticmethod
def apply_rotary_position_embeddings(sinusoidal_pos, query_layer, key_layer, value_layer=None):
# https://kexue.fm/archives/8265
# sin [batch_size, num_heads, sequence_length, embed_size_per_head//2]
# cos [batch_size, num_heads, sequence_length, embed_size_per_head//2]
sin, cos = sinusoidal_pos.chunk(2, dim=-1)
# sin [θ0,θ1,θ2......θd/2-1] -> sin_pos [θ0,θ0,θ1,θ1,θ2,θ2......θd/2-1,θd/2-1]
sin_pos = torch.repeat_interleave(sin, 2, dim=-1)
# cos [θ0,θ1,θ2......θd/2-1] -> cos_pos [θ0,θ0,θ1,θ1,θ2,θ2......θd/2-1,θd/2-1]
cos_pos = torch.repeat_interleave(cos, 2, dim=-1)
# rotate_half_query_layer [-q1,q0,-q3,q2......,-qd-1,qd-2]
rotate_half_query_layer = torch.stack([-query_layer[..., 1::2], query_layer[..., ::2]], dim=-1).reshape_as(
query_layer
)
query_layer = query_layer * cos_pos + rotate_half_query_layer * sin_pos
# rotate_half_key_layer [-k1,k0,-k3,k2......,-kd-1,kd-2]
rotate_half_key_layer = torch.stack([-key_layer[..., 1::2], key_layer[..., ::2]], dim=-1).reshape_as(key_layer)
key_layer = key_layer * cos_pos + rotate_half_key_layer * sin_pos
if value_layer is not None:
# rotate_half_value_layer [-v1,v0,-v3,v2......,-vd-1,vd-2]
rotate_half_value_layer = torch.stack([-value_layer[..., 1::2], value_layer[..., ::2]], dim=-1).reshape_as(
value_layer
)
value_layer = value_layer * cos_pos + rotate_half_value_layer * sin_pos
return query_layer, key_layer, value_layer
return query_layer, key_layer
# Copied from transformers.models.bert.modeling_bert.BertSelfOutput with Bert->RoFormer
class RoFormerSelfOutput(nn.Module):
def __init__(self, config):
super().__init__()
self.dense = nn.Linear(config.hidden_size, config.hidden_size)
self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
self.dropout = nn.Dropout(config.hidden_dropout_prob)
def forward(self, hidden_states, input_tensor):
hidden_states = self.dense(hidden_states)
hidden_states = self.dropout(hidden_states)
hidden_states = self.LayerNorm(hidden_states + input_tensor)
return hidden_states
class RoFormerAttention(nn.Module):
# Copied from transformers.models.bert.modeling_bert.BertAttention.__init__ with Bert->RoFormer
def __init__(self, config):
super().__init__()
self.self = RoFormerSelfAttention(config)
self.output = RoFormerSelfOutput(config)
self.pruned_heads = set()
# End Copy
# Copied from transformers.models.bert.modeling_bert.BertAttention.prune_heads
def prune_heads(self, heads):
if len(heads) == 0:
return
heads, index = find_pruneable_heads_and_indices(
heads, self.self.num_attention_heads, self.self.attention_head_size, self.pruned_heads
)
# Prune linear layers
self.self.query = prune_linear_layer(self.self.query, index)
self.self.key = prune_linear_layer(self.self.key, index)
self.self.value = prune_linear_layer(self.self.value, index)
self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)
# Update hyper params and store pruned heads
self.self.num_attention_heads = self.self.num_attention_heads - len(heads)
self.self.all_head_size = self.self.attention_head_size * self.self.num_attention_heads
self.pruned_heads = self.pruned_heads.union(heads)
# End Copy
def forward(
self,
hidden_states,
attention_mask=None,
sinusoidal_pos=None,
head_mask=None,
encoder_hidden_states=None,
encoder_attention_mask=None,
past_key_value=None,
output_attentions=False,
):
self_outputs = self.self(
hidden_states,
attention_mask,
sinusoidal_pos,
head_mask,
encoder_hidden_states,
encoder_attention_mask,
past_key_value,
output_attentions,
)
attention_output = self.output(self_outputs[0], hidden_states)
outputs = (attention_output,) + self_outputs[1:] # add attentions if we output them
return outputs
# Copied from transformers.models.bert.modeling_bert.BertIntermediate with Bert->RoFormer
class RoFormerIntermediate(nn.Module):
def __init__(self, config):
super().__init__()
self.dense = nn.Linear(config.hidden_size, config.intermediate_size)
if isinstance(config.hidden_act, str):
self.intermediate_act_fn = ACT2FN[config.hidden_act]
else:
self.intermediate_act_fn = config.hidden_act
def forward(self, hidden_states):
hidden_states = self.dense(hidden_states)
hidden_states = self.intermediate_act_fn(hidden_states)
return hidden_states
# Copied from transformers.models.bert.modeling_bert.BertOutput with Bert->RoFormer
class RoFormerOutput(nn.Module):
def __init__(self, config):
super().__init__()
self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
self.dropout = nn.Dropout(config.hidden_dropout_prob)
def forward(self, hidden_states, input_tensor):
hidden_states = self.dense(hidden_states)
hidden_states = self.dropout(hidden_states)
hidden_states = self.LayerNorm(hidden_states + input_tensor)
return hidden_states
class RoFormerLayer(nn.Module):
# Copied from transformers.models.bert.modeling_bert.BertLayer.__init__ with Bert->RoFormer
def __init__(self, config):
super().__init__()
self.chunk_size_feed_forward = config.chunk_size_feed_forward
self.seq_len_dim = 1
self.attention = RoFormerAttention(config)
self.is_decoder = config.is_decoder
self.add_cross_attention = config.add_cross_attention
if self.add_cross_attention:
assert self.is_decoder, f"{self} should be used as a decoder model if cross attention is added"
self.crossattention = RoFormerAttention(config)
self.intermediate = RoFormerIntermediate(config)
self.output = RoFormerOutput(config)
# End Copy
def forward(
self,
hidden_states,
attention_mask=None,
sinusoidal_pos=None,
head_mask=None,
encoder_hidden_states=None,
encoder_attention_mask=None,
past_key_value=None,
output_attentions=False,
):
# decoder uni-directional self-attention cached key/values tuple is at positions 1,2
self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
self_attention_outputs = self.attention(
hidden_states,
attention_mask,
sinusoidal_pos,
head_mask,
output_attentions=output_attentions,
past_key_value=self_attn_past_key_value,
)
attention_output = self_attention_outputs[0]
# if decoder, the last output is tuple of self-attn cache
if self.is_decoder:
outputs = self_attention_outputs[1:-1]
present_key_value = self_attention_outputs[-1]
else:
outputs = self_attention_outputs[1:] # add self attentions if we output attention weights
cross_attn_present_key_value = None
if self.is_decoder and encoder_hidden_states is not None:
if not hasattr(self, "crossattention"):
raise ValueError(
f"If `encoder_hidden_states` are passed, {self} has to be instantiated with cross-attention "
"layers by setting `config.add_cross_attention=True`"
)
# cross_attn cached key/values tuple is at positions 3,4 of past_key_value tuple
cross_attn_past_key_value = past_key_value[-2:] if past_key_value is not None else None
cross_attention_outputs = self.crossattention(
attention_output,
attention_mask,
sinusoidal_pos,
head_mask,
encoder_hidden_states,
encoder_attention_mask,
cross_attn_past_key_value,
output_attentions,
)
attention_output = cross_attention_outputs[0]
outputs = outputs + cross_attention_outputs[1:-1] # add cross attentions if we output attention weights
# add cross-attn cache to positions 3,4 of present_key_value tuple
cross_attn_present_key_value = cross_attention_outputs[-1]
present_key_value = present_key_value + cross_attn_present_key_value
layer_output = apply_chunking_to_forward(
self.feed_forward_chunk, self.chunk_size_feed_forward, self.seq_len_dim, attention_output
)
outputs = (layer_output,) + outputs
# if decoder, return the attn key/values as the last output
if self.is_decoder:
outputs = outputs + (present_key_value,)
return outputs
def feed_forward_chunk(self, attention_output):
intermediate_output = self.intermediate(attention_output)
layer_output = self.output(intermediate_output, attention_output)
return layer_output
class RoFormerEncoder(nn.Module):
def __init__(self, config):
super().__init__()
self.config = config
self.embed_positions = RoFormerSinusoidalPositionalEmbedding(
config.max_position_embeddings, config.hidden_size // config.num_attention_heads
)
self.layer = nn.ModuleList([RoFormerLayer(config) for _ in range(config.num_hidden_layers)])
def forward(
self,
hidden_states,
attention_mask=None,
head_mask=None,
encoder_hidden_states=None,
encoder_attention_mask=None,
past_key_values=None,
use_cache=None,
output_attentions=False,
output_hidden_states=False,
return_dict=True,
):
all_hidden_states = () if output_hidden_states else None
all_self_attentions = () if output_attentions else None
all_cross_attentions = () if output_attentions and self.config.add_cross_attention else None
# [sequence_length, embed_size_per_head] -> [batch_size, num_heads, sequence_length, embed_size_per_head]
sinusoidal_pos = self.embed_positions(hidden_states.shape[:-1])[None, None, :, :]
next_decoder_cache = () if use_cache else None
for i, layer_module in enumerate(self.layer):
if output_hidden_states:
all_hidden_states = all_hidden_states + (hidden_states,)
layer_head_mask = head_mask[i] if head_mask is not None else None
past_key_value = past_key_values[i] if past_key_values is not None else None
if getattr(self.config, "gradient_checkpointing", False) and self.training:
if use_cache:
logger.warning(
"`use_cache=True` is incompatible with `config.gradient_checkpointing=True`. Setting "
"`use_cache=False`..."
)
use_cache = False
def create_custom_forward(module):
def custom_forward(*inputs):
return module(*inputs, past_key_value, output_attentions)
return custom_forward
layer_outputs = torch.utils.checkpoint.checkpoint(
create_custom_forward(layer_module),
hidden_states,
attention_mask,
sinusoidal_pos,
layer_head_mask,
encoder_hidden_states,
encoder_attention_mask,
)
else:
layer_outputs = layer_module(
hidden_states,
attention_mask,
sinusoidal_pos,
layer_head_mask,
encoder_hidden_states,
encoder_attention_mask,
past_key_value,
output_attentions,
)
hidden_states = layer_outputs[0]
if use_cache:
next_decoder_cache += (layer_outputs[-1],)
if output_attentions:
all_self_attentions = all_self_attentions + (layer_outputs[1],)
if self.config.add_cross_attention:
all_cross_attentions = all_cross_attentions + (layer_outputs[2],)
if output_hidden_states:
all_hidden_states = all_hidden_states + (hidden_states,)
if not return_dict:
return tuple(
v
for v in [
hidden_states,
next_decoder_cache,
all_hidden_states,
all_self_attentions,
all_cross_attentions,
]
if v is not None
)
return BaseModelOutputWithPastAndCrossAttentions(
last_hidden_state=hidden_states,
past_key_values=next_decoder_cache,
hidden_states=all_hidden_states,
attentions=all_self_attentions,
cross_attentions=all_cross_attentions,
)
class RoFormerPredictionHeadTransform(nn.Module):
def __init__(self, config):
super().__init__()
self.dense = nn.Linear(config.hidden_size, config.embedding_size)
if isinstance(config.hidden_act, str):
self.transform_act_fn = ACT2FN[config.hidden_act]
else:
self.transform_act_fn = config.hidden_act
self.LayerNorm = nn.LayerNorm(config.embedding_size, eps=config.layer_norm_eps)
def forward(self, hidden_states):
hidden_states = self.dense(hidden_states)
hidden_states = self.transform_act_fn(hidden_states)
hidden_states = self.LayerNorm(hidden_states)
return hidden_states
class RoFormerLMPredictionHead(nn.Module):
def __init__(self, config):
super().__init__()
self.transform = RoFormerPredictionHeadTransform(config)
# The output weights are the same as the input embeddings, but there is
# an output-only bias for each token.
self.decoder = nn.Linear(config.embedding_size, config.vocab_size, bias=False)
self.bias = nn.Parameter(torch.zeros(config.vocab_size))
# Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`
self.decoder.bias = self.bias
def forward(self, hidden_states):
hidden_states = self.transform(hidden_states)
hidden_states = self.decoder(hidden_states)
return hidden_states
# Copied from transformers.models.bert.modeling_bert.BertOnlyMLMHead with Bert->RoFormer
class RoFormerOnlyMLMHead(nn.Module):
def __init__(self, config):
super().__init__()
self.predictions = RoFormerLMPredictionHead(config)
def forward(self, sequence_output):
prediction_scores = self.predictions(sequence_output)
return prediction_scores
class RoFormerPreTrainedModel(PreTrainedModel):
"""
An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
models.
"""
config_class = RoFormerConfig
load_tf_weights = load_tf_weights_in_roformer
base_model_prefix = "roformer"
_keys_to_ignore_on_load_missing = []
_keys_to_ignore_on_load_unexpected = [
r"roformer\.embeddings_project\.weight",
r"roformer\.embeddings_project\.bias",
]
def _init_weights(self, module):
"""Initialize the weights"""
if isinstance(module, nn.Linear):
# Slightly different from the TF version which uses truncated_normal for initialization
# cf https://github.com/pytorch/pytorch/pull/5617
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
if module.bias is not None:
module.bias.data.zero_()
elif isinstance(module, RoFormerSinusoidalPositionalEmbedding):
pass
elif isinstance(module, nn.Embedding):
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
if module.padding_idx is not None:
module.weight.data[module.padding_idx].zero_()
elif isinstance(module, nn.LayerNorm):
module.bias.data.zero_()
module.weight.data.fill_(1.0)
ROFORMER_START_DOCSTRING = r"""
This model is a PyTorch `torch.nn.Module <https://pytorch.org/docs/stable/nn.html#torch.nn.Module>`_ sub-class. Use
it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and
behavior.
Parameters:
config (:class:`~transformers.RoFormerConfig`): Model configuration class with all the parameters of the model.
Initializing with a config file does not load the weights associated with the model, only the
configuration. Check out the :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model
weights.
"""
ROFORMER_INPUTS_DOCSTRING = r"""
Args:
input_ids (:obj:`torch.LongTensor` of shape :obj:`{0}`):
Indices of input sequence tokens in the vocabulary.
Indices can be obtained using :class:`transformers.RoFormerTokenizer`. See
:func:`transformers.PreTrainedTokenizer.encode` and :func:`transformers.PreTrainedTokenizer.__call__` for
details.
`What are input IDs? <../glossary.html#input-ids>`__
attention_mask (:obj:`torch.FloatTensor` of shape :obj:`{0}`, `optional`):
Mask to avoid performing attention on padding token indices. Mask values selected in ``[0, 1]``:
- 1 for tokens that are **not masked**,
- 0 for tokens that are **masked**.
`What are attention masks? <../glossary.html#attention-mask>`__
token_type_ids (:obj:`torch.LongTensor` of shape :obj:`{0}`, `optional`):
Segment token indices to indicate first and second portions of the inputs. Indices are selected in ``[0,
1]``:
- 0 corresponds to a `sentence A` token,
- 1 corresponds to a `sentence B` token.
`What are token type IDs? <../glossary.html#token-type-ids>`_
head_mask (:obj:`torch.FloatTensor` of shape :obj:`(num_heads,)` or :obj:`(num_layers, num_heads)`, `optional`):
Mask to nullify selected heads of the self-attention modules. Mask values selected in ``[0, 1]``:
- 1 indicates the head is **not masked**,
- 0 indicates the head is **masked**.
inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded representation.
This is useful if you want more control over how to convert `input_ids` indices into associated vectors
than the model's internal embedding lookup matrix.
output_attentions (:obj:`bool`, `optional`):
Whether or not to return the attentions tensors of all attention layers. See ``attentions`` under returned
tensors for more detail.
output_hidden_states (:obj:`bool`, `optional`):
Whether or not to return the hidden states of all layers. See ``hidden_states`` under returned tensors for
more detail.
return_dict (:obj:`bool`, `optional`):
Whether or not to return a :class:`~transformers.file_utils.ModelOutput` instead of a plain tuple.
"""
@add_start_docstrings(
"The bare RoFormer Model transformer outputting raw hidden-states without any specific head on top.",
ROFORMER_START_DOCSTRING,
)
class RoFormerModel(RoFormerPreTrainedModel):
"""
The model can behave as an encoder (with only self-attention) as well as a decoder, in which case a layer of
cross-attention is added between the self-attention layers, following the architecture described in `Attention is
all you need <https://arxiv.org/abs/1706.03762>`__ by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit,
Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin.
To behave as an decoder the model needs to be initialized with the :obj:`is_decoder` argument of the configuration
set to :obj:`True`. To be used in a Seq2Seq model, the model needs to initialized with both :obj:`is_decoder`
argument and :obj:`add_cross_attention` set to :obj:`True`; an :obj:`encoder_hidden_states` is then expected as an
input to the forward pass.
"""
def __init__(self, config):
super().__init__(config)
self.config = config
self.embeddings = RoFormerEmbeddings(config)
if config.embedding_size != config.hidden_size:
self.embeddings_project = nn.Linear(config.embedding_size, config.hidden_size)
self.encoder = RoFormerEncoder(config)
self.init_weights()
def get_input_embeddings(self):
return self.embeddings.word_embeddings
def set_input_embeddings(self, value):
self.embeddings.word_embeddings = value
def _prune_heads(self, heads_to_prune):
"""
Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
class PreTrainedModel
"""
for layer, heads in heads_to_prune.items():
self.encoder.layer[layer].attention.prune_heads(heads)
@add_start_docstrings_to_model_forward(ROFORMER_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
@add_code_sample_docstrings(
tokenizer_class=_TOKENIZER_FOR_DOC,
checkpoint=_CHECKPOINT_FOR_DOC,
output_type=BaseModelOutputWithPastAndCrossAttentions,
config_class=_CONFIG_FOR_DOC,
)
def forward(
self,
input_ids=None,
attention_mask=None,
token_type_ids=None,
head_mask=None,
inputs_embeds=None,
encoder_hidden_states=None,
encoder_attention_mask=None,
past_key_values=None,
use_cache=None,
output_attentions=None,
output_hidden_states=None,
return_dict=None,
):
r"""
encoder_hidden_states (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
the model is configured as a decoder.
encoder_attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
the cross-attention if the model is configured as a decoder. Mask values selected in ``[0, 1]``:
- 1 for tokens that are **not masked**,
- 0 for tokens that are **masked**.
past_key_values (:obj:`tuple(tuple(torch.FloatTensor))` of length :obj:`config.n_layers` with each tuple having 4 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
If :obj:`past_key_values` are used, the user can optionally input only the last :obj:`decoder_input_ids`
(those that don't have their past key value states given to this model) of shape :obj:`(batch_size, 1)`
instead of all :obj:`decoder_input_ids` of shape :obj:`(batch_size, sequence_length)`.
use_cache (:obj:`bool`, `optional`):
If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
decoding (see :obj:`past_key_values`).
"""
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
output_hidden_states = (
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
)
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
if self.config.is_decoder:
use_cache = use_cache if use_cache is not None else self.config.use_cache
else:
use_cache = False
if input_ids is not None and inputs_embeds is not None:
raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
elif input_ids is not None:
input_shape = input_ids.size()
batch_size, seq_length = input_shape
elif inputs_embeds is not None:
input_shape = inputs_embeds.size()[:-1]
batch_size, seq_length = input_shape
else:
raise ValueError("You have to specify either input_ids or inputs_embeds")
device = input_ids.device if input_ids is not None else inputs_embeds.device
# past_key_values_length
past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0
if attention_mask is None:
attention_mask = torch.ones(((batch_size, seq_length + past_key_values_length)), device=device)
if token_type_ids is None:
token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
# We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
# ourselves in which case we just need to make it broadcastable to all heads.
extended_attention_mask: torch.Tensor = self.get_extended_attention_mask(attention_mask, input_shape, device)
# If a 2D or 3D attention mask is provided for the cross-attention
# we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
if self.config.is_decoder and encoder_hidden_states is not None:
encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()
encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
if encoder_attention_mask is None:
encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask)
else:
encoder_extended_attention_mask = None
# Prepare head mask if needed
# 1.0 in head_mask indicate we keep the head
# attention_probs has shape bsz x n_heads x N x N
# input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
# and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
embedding_output = self.embeddings(
input_ids=input_ids, token_type_ids=token_type_ids, inputs_embeds=inputs_embeds
)
if hasattr(self, "embeddings_project"):
embedding_output = self.embeddings_project(embedding_output)
encoder_outputs = self.encoder(
embedding_output,
attention_mask=extended_attention_mask,
head_mask=head_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_extended_attention_mask,
past_key_values=past_key_values,
use_cache=use_cache,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
sequence_output = encoder_outputs[0]
if not return_dict:
return (sequence_output,) + encoder_outputs[1:]
return BaseModelOutputWithPastAndCrossAttentions(
last_hidden_state=sequence_output,
past_key_values=encoder_outputs.past_key_values,
hidden_states=encoder_outputs.hidden_states,
attentions=encoder_outputs.attentions,
cross_attentions=encoder_outputs.cross_attentions,
)
@add_start_docstrings("""RoFormer Model with a `language modeling` head on top. """, ROFORMER_START_DOCSTRING)
class RoFormerForMaskedLM(RoFormerPreTrainedModel):
def __init__(self, config):
super().__init__(config)
if config.is_decoder:
logger.warning(
"If you want to use `RoFormerForMaskedLM` make sure `config.is_decoder=False` for "
"bi-directional self-attention."
)
self.roformer = RoFormerModel(config)
self.cls = RoFormerOnlyMLMHead(config)
self.init_weights()
def get_output_embeddings(self):
return self.cls.predictions.decoder
def set_output_embeddings(self, new_embeddings):
self.cls.predictions.decoder = new_embeddings
@add_start_docstrings_to_model_forward(ROFORMER_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
@add_code_sample_docstrings(
tokenizer_class=_TOKENIZER_FOR_DOC,
checkpoint=_CHECKPOINT_FOR_DOC,
output_type=MaskedLMOutput,
config_class=_CONFIG_FOR_DOC,
)
def forward(
self,
input_ids=None,
attention_mask=None,
token_type_ids=None,
head_mask=None,
inputs_embeds=None,
encoder_hidden_states=None,
encoder_attention_mask=None,
labels=None,
output_attentions=None,
output_hidden_states=None,
return_dict=None,
):
r"""
labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
Labels for computing the masked language modeling loss. Indices should be in ``[-100, 0, ...,
config.vocab_size]`` (see ``input_ids`` docstring) Tokens with indices set to ``-100`` are ignored
(masked), the loss is only computed for the tokens with labels in ``[0, ..., config.vocab_size]``.
"""
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
outputs = self.roformer(
input_ids,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
head_mask=head_mask,
inputs_embeds=inputs_embeds,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
sequence_output = outputs[0]
prediction_scores = self.cls(sequence_output)
masked_lm_loss = None
if labels is not None:
loss_fct = CrossEntropyLoss() # -100 index = padding token
masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), labels.view(-1))
if not return_dict:
output = (prediction_scores,) + outputs[1:]
return ((masked_lm_loss,) + output) if masked_lm_loss is not None else output
return MaskedLMOutput(
loss=masked_lm_loss,
logits=prediction_scores,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
def prepare_inputs_for_generation(self, input_ids, attention_mask=None, **model_kwargs):
input_shape = input_ids.shape
effective_batch_size = input_shape[0]
# add a dummy token
assert self.config.pad_token_id is not None, "The PAD token should be defined for generation"
attention_mask = torch.cat([attention_mask, attention_mask.new_zeros((attention_mask.shape[0], 1))], dim=-1)
dummy_token = torch.full(
(effective_batch_size, 1), self.config.pad_token_id, dtype=torch.long, device=input_ids.device
)
input_ids = torch.cat([input_ids, dummy_token], dim=1)
return {"input_ids": input_ids, "attention_mask": attention_mask}
@add_start_docstrings(
"""RoFormer Model with a `language modeling` head on top for CLM fine-tuning. """, ROFORMER_START_DOCSTRING
)
class RoFormerForCausalLM(RoFormerPreTrainedModel):
_keys_to_ignore_on_load_missing = [r"predictions.decoder.bias"]
def __init__(self, config):
super().__init__(config)
if not config.is_decoder:
logger.warning("If you want to use `RoFormerForCausalLM` as a standalone, add `is_decoder=True.`")
self.roformer = RoFormerModel(config)
self.cls = RoFormerOnlyMLMHead(config)
self.init_weights()
def get_output_embeddings(self):
return self.cls.predictions.decoder
def set_output_embeddings(self, new_embeddings):
self.cls.predictions.decoder = new_embeddings
@add_start_docstrings_to_model_forward(ROFORMER_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
@replace_return_docstrings(output_type=CausalLMOutputWithCrossAttentions, config_class=_CONFIG_FOR_DOC)
def forward(
self,
input_ids=None,
attention_mask=None,
token_type_ids=None,
inputs_embeds=None,
encoder_hidden_states=None,
encoder_attention_mask=None,
head_mask=None,
cross_attn_head_mask=None,
past_key_values=None,
labels=None,
use_cache=None,
output_attentions=None,
output_hidden_states=None,
return_dict=None,
):
r"""
encoder_hidden_states (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
the model is configured as a decoder.
encoder_attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
the cross-attention if the model is configured as a decoder. Mask values selected in ``[0, 1]``:
- 1 for tokens that are **not masked**,
- 0 for tokens that are **masked**.
past_key_values (:obj:`tuple(tuple(torch.FloatTensor))` of length :obj:`config.n_layers` with each tuple having 4 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
If :obj:`past_key_values` are used, the user can optionally input only the last :obj:`decoder_input_ids`
(those that don't have their past key value states given to this model) of shape :obj:`(batch_size, 1)`
instead of all :obj:`decoder_input_ids` of shape :obj:`(batch_size, sequence_length)`.
labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
Labels for computing the left-to-right language modeling loss (next word prediction). Indices should be in
``[-100, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring) Tokens with indices set to ``-100`` are
ignored (masked), the loss is only computed for the tokens with labels n ``[0, ..., config.vocab_size]``.
use_cache (:obj:`bool`, `optional`):
If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
decoding (see :obj:`past_key_values`).
Returns:
Example::
>>> from transformers import RoFormerTokenizer, RoFormerForCausalLM, RoFormerConfig
>>> import torch
>>> tokenizer = RoFormerTokenizer.from_pretrained('junnyu/roformer_chinese_base')
>>> config = RoFormerConfig.from_pretrained("junnyu/roformer_chinese_base")
>>> config.is_decoder = True
>>> model = RoFormerForCausalLM.from_pretrained('junnyu/roformer_chinese_base', config=config)
>>> inputs = tokenizer("今天天气非常好。", return_tensors="pt")
>>> outputs = model(**inputs)
>>> prediction_logits = outputs.logits
"""
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
outputs = self.roformer(
input_ids,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
head_mask=head_mask,
inputs_embeds=inputs_embeds,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
past_key_values=past_key_values,
use_cache=use_cache,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
sequence_output = outputs[0]
prediction_scores = self.cls(sequence_output)
lm_loss = None
if labels is not None:
# we are doing next-token prediction; shift prediction scores and input ids by one
shifted_prediction_scores = prediction_scores[:, :-1, :].contiguous()
labels = labels[:, 1:].contiguous()
loss_fct = CrossEntropyLoss()
lm_loss = loss_fct(shifted_prediction_scores.view(-1, self.config.vocab_size), labels.view(-1))
if not return_dict:
output = (prediction_scores,) + outputs[1:]
return ((lm_loss,) + output) if lm_loss is not None else output
return CausalLMOutputWithCrossAttentions(
loss=lm_loss,
logits=prediction_scores,
past_key_values=outputs.past_key_values,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
cross_attentions=outputs.cross_attentions,
)
def prepare_inputs_for_generation(self, input_ids, past=None, attention_mask=None, **model_kwargs):
input_shape = input_ids.shape
# if model is used as a decoder in encoder-decoder model, the decoder attention mask is created on the fly
if attention_mask is None:
attention_mask = input_ids.new_ones(input_shape)
# cut decoder_input_ids if past is used
if past is not None:
input_ids = input_ids[:, -1:]
return {"input_ids": input_ids, "attention_mask": attention_mask, "past_key_values": past}
def _reorder_cache(self, past, beam_idx):
reordered_past = ()
for layer_past in past:
reordered_past += (
tuple(past_state.index_select(0, beam_idx) for past_state in layer_past[:2]) + layer_past[2:],
)
return reordered_past
class RoFormerClassificationHead(nn.Module):
"""Head for sentence-level classification tasks."""
def __init__(self, config):
super().__init__()
self.dense = nn.Linear(config.hidden_size, config.hidden_size)
self.dropout = nn.Dropout(config.hidden_dropout_prob)
self.out_proj = nn.Linear(config.hidden_size, config.num_labels)
self.config = config
def forward(self, features, **kwargs):
x = features[:, 0, :] # take <s> token (equiv. to [CLS])
x = self.dropout(x)
x = self.dense(x)
x = ACT2FN[self.config.hidden_act](x)
x = self.dropout(x)
x = self.out_proj(x)
return x
@add_start_docstrings(
"""
RoFormer Model transformer with a sequence classification/regression head on top (a linear layer on top of the
pooled output) e.g. for GLUE tasks.
""",
ROFORMER_START_DOCSTRING,
)
class RoFormerForSequenceClassification(RoFormerPreTrainedModel):
def __init__(self, config):
super().__init__(config)
self.num_labels = config.num_labels
self.roformer = RoFormerModel(config)
self.classifier = RoFormerClassificationHead(config)
self.init_weights()
@add_start_docstrings_to_model_forward(ROFORMER_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
@add_code_sample_docstrings(
tokenizer_class=_TOKENIZER_FOR_DOC,
checkpoint=_CHECKPOINT_FOR_DOC,
output_type=SequenceClassifierOutput,
config_class=_CONFIG_FOR_DOC,
)
def forward(
self,
input_ids=None,
attention_mask=None,
token_type_ids=None,
head_mask=None,
inputs_embeds=None,
labels=None,
output_attentions=None,
output_hidden_states=None,
return_dict=None,
):
r"""
labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
Labels for computing the sequence classification/regression loss. Indices should be in :obj:`[0, ...,
config.num_labels - 1]`. If :obj:`config.num_labels == 1` a regression loss is computed (Mean-Square loss),
If :obj:`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
"""
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
outputs = self.roformer(
input_ids,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
head_mask=head_mask,
inputs_embeds=inputs_embeds,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
sequence_output = outputs[0]
logits = self.classifier(sequence_output)
loss = None
if labels is not None:
if self.num_labels == 1:
# We are doing regression
loss_fct = MSELoss()
loss = loss_fct(logits.view(-1), labels.view(-1))
else:
loss_fct = CrossEntropyLoss()
loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
if not return_dict:
output = (logits,) + outputs[1:]
return ((loss,) + output) if loss is not None else output
return SequenceClassifierOutput(
loss=loss,
logits=logits,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
@add_start_docstrings(
"""
RoFormer Model with a multiple choice classification head on top (a linear layer on top of the pooled output and a
softmax) e.g. for RocStories/SWAG tasks.
""",
ROFORMER_START_DOCSTRING,
)
class RoFormerForMultipleChoice(RoFormerPreTrainedModel):
def __init__(self, config):
super().__init__(config)
self.roformer = RoFormerModel(config)
self.sequence_summary = SequenceSummary(config)
self.classifier = nn.Linear(config.hidden_size, 1)
self.init_weights()
@add_start_docstrings_to_model_forward(
ROFORMER_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length")
)
@add_code_sample_docstrings(
tokenizer_class=_TOKENIZER_FOR_DOC,
checkpoint=_CHECKPOINT_FOR_DOC,
output_type=MultipleChoiceModelOutput,
config_class=_CONFIG_FOR_DOC,
)
def forward(
self,
input_ids=None,
attention_mask=None,
token_type_ids=None,
head_mask=None,
inputs_embeds=None,
labels=None,
output_attentions=None,
output_hidden_states=None,
return_dict=None,
):
r"""
labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
Labels for computing the multiple choice classification loss. Indices should be in ``[0, ...,
num_choices-1]`` where :obj:`num_choices` is the size of the second dimension of the input tensors. (See
:obj:`input_ids` above)
"""
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
num_choices = input_ids.shape[1] if input_ids is not None else inputs_embeds.shape[1]
input_ids = input_ids.view(-1, input_ids.size(-1)) if input_ids is not None else None
attention_mask = attention_mask.view(-1, attention_mask.size(-1)) if attention_mask is not None else None
token_type_ids = token_type_ids.view(-1, token_type_ids.size(-1)) if token_type_ids is not None else None
inputs_embeds = (
inputs_embeds.view(-1, inputs_embeds.size(-2), inputs_embeds.size(-1))
if inputs_embeds is not None
else None
)
outputs = self.roformer(
input_ids,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
head_mask=head_mask,
inputs_embeds=inputs_embeds,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
sequence_output = outputs[0]
pooled_output = self.sequence_summary(sequence_output)
logits = self.classifier(pooled_output)
reshaped_logits = logits.view(-1, num_choices)
loss = None
if labels is not None:
loss_fct = CrossEntropyLoss()
loss = loss_fct(reshaped_logits, labels)
if not return_dict:
output = (reshaped_logits,) + outputs[1:]
return ((loss,) + output) if loss is not None else output
return MultipleChoiceModelOutput(
loss=loss,
logits=reshaped_logits,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
@add_start_docstrings(
"""
RoFormer Model with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for
Named-Entity-Recognition (NER) tasks.
""",
ROFORMER_START_DOCSTRING,
)
class RoFormerForTokenClassification(RoFormerPreTrainedModel):
def __init__(self, config):
super().__init__(config)
self.num_labels = config.num_labels
self.roformer = RoFormerModel(config)
self.dropout = nn.Dropout(config.hidden_dropout_prob)
self.classifier = nn.Linear(config.hidden_size, config.num_labels)
self.init_weights()
@add_start_docstrings_to_model_forward(ROFORMER_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
@add_code_sample_docstrings(
tokenizer_class=_TOKENIZER_FOR_DOC,
checkpoint=_CHECKPOINT_FOR_DOC,
output_type=TokenClassifierOutput,
config_class=_CONFIG_FOR_DOC,
)
def forward(
self,
input_ids=None,
attention_mask=None,
token_type_ids=None,
head_mask=None,
inputs_embeds=None,
labels=None,
output_attentions=None,
output_hidden_states=None,
return_dict=None,
):
r"""
labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
Labels for computing the token classification loss. Indices should be in ``[0, ..., config.num_labels -
1]``.
"""
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
outputs = self.roformer(
input_ids,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
head_mask=head_mask,
inputs_embeds=inputs_embeds,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
sequence_output = outputs[0]
sequence_output = self.dropout(sequence_output)
logits = self.classifier(sequence_output)
loss = None
if labels is not None:
loss_fct = CrossEntropyLoss()
# Only keep active parts of the loss
if attention_mask is not None:
active_loss = attention_mask.view(-1) == 1
active_logits = logits.view(-1, self.num_labels)
active_labels = torch.where(
active_loss, labels.view(-1), torch.tensor(loss_fct.ignore_index).type_as(labels)
)
loss = loss_fct(active_logits, active_labels)
else:
loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
if not return_dict:
output = (logits,) + outputs[1:]
return ((loss,) + output) if loss is not None else output
return TokenClassifierOutput(
loss=loss,
logits=logits,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
@add_start_docstrings(
"""
RoFormer Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear
layers on top of the hidden-states output to compute `span start logits` and `span end logits`).
""",
ROFORMER_START_DOCSTRING,
)
class RoFormerForQuestionAnswering(RoFormerPreTrainedModel):
def __init__(self, config):
super().__init__(config)
config.num_labels = 2
self.num_labels = config.num_labels
self.roformer = RoFormerModel(config)
self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)
self.init_weights()
@add_start_docstrings_to_model_forward(ROFORMER_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
@add_code_sample_docstrings(
tokenizer_class=_TOKENIZER_FOR_DOC,
checkpoint=_CHECKPOINT_FOR_DOC,
output_type=QuestionAnsweringModelOutput,
config_class=_CONFIG_FOR_DOC,
)
def forward(
self,
input_ids=None,
attention_mask=None,
token_type_ids=None,
head_mask=None,
inputs_embeds=None,
start_positions=None,
end_positions=None,
output_attentions=None,
output_hidden_states=None,
return_dict=None,
):
r"""
start_positions (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
Labels for position (index) of the start of the labelled span for computing the token classification loss.
Positions are clamped to the length of the sequence (:obj:`sequence_length`). Position outside of the
sequence are not taken into account for computing the loss.
end_positions (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
Labels for position (index) of the end of the labelled span for computing the token classification loss.
Positions are clamped to the length of the sequence (:obj:`sequence_length`). Position outside of the
sequence are not taken into account for computing the loss.
"""
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
outputs = self.roformer(
input_ids,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
head_mask=head_mask,
inputs_embeds=inputs_embeds,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
sequence_output = outputs[0]
logits = self.qa_outputs(sequence_output)
start_logits, end_logits = logits.split(1, dim=-1)
start_logits = start_logits.squeeze(-1)
end_logits = end_logits.squeeze(-1)
total_loss = None
if start_positions is not None and end_positions is not None:
# If we are on multi-GPU, split add a dimension
if len(start_positions.size()) > 1:
start_positions = start_positions.squeeze(-1)
if len(end_positions.size()) > 1:
end_positions = end_positions.squeeze(-1)
# sometimes the start/end positions are outside our model inputs, we ignore these terms
ignored_index = start_logits.size(1)
start_positions.clamp_(0, ignored_index)
end_positions.clamp_(0, ignored_index)
loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
start_loss = loss_fct(start_logits, start_positions)
end_loss = loss_fct(end_logits, end_positions)
total_loss = (start_loss + end_loss) / 2
if not return_dict:
output = (start_logits, end_logits) + outputs[1:]
return ((total_loss,) + output) if total_loss is not None else output
return QuestionAnsweringModelOutput(
loss=total_loss,
start_logits=start_logits,
end_logits=end_logits,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
src/transformers/models/roformer/modeling_tf_roformer.py 0 → 100644
View file @ 206f06f2
# coding=utf-8
# Copyright 2021 The HuggingFace Inc. team. 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.
""" TF 2.0 RoFormer model. """
import math
from typing import Dict, Optional, Tuple, Union
import numpy as np
import tensorflow as tf
from ...activations_tf import get_tf_activation
from ...file_utils import (
MULTIPLE_CHOICE_DUMMY_INPUTS,
add_code_sample_docstrings,
add_start_docstrings,
add_start_docstrings_to_model_forward,
)
from ...modeling_tf_outputs import (
TFBaseModelOutput,
TFBaseModelOutputWithPooling,
TFCausalLMOutput,
TFMaskedLMOutput,
TFMultipleChoiceModelOutput,
TFQuestionAnsweringModelOutput,
TFSequenceClassifierOutput,
TFTokenClassifierOutput,
)
from ...modeling_tf_utils import (
TFCausalLanguageModelingLoss,
TFMaskedLanguageModelingLoss,
TFModelInputType,
TFMultipleChoiceLoss,
TFPreTrainedModel,
TFQuestionAnsweringLoss,
TFSequenceClassificationLoss,
TFSequenceSummary,
TFTokenClassificationLoss,
get_initializer,
input_processing,
keras_serializable,
shape_list,
)
from ...utils import logging
from .configuration_roformer import RoFormerConfig
logger = logging.get_logger(__name__)
_CHECKPOINT_FOR_DOC = "junnyu/roformer_chinese_base"
_CONFIG_FOR_DOC = "RoFormerConfig"
_TOKENIZER_FOR_DOC = "RoFormerTokenizer"
TF_ROFORMER_PRETRAINED_MODEL_ARCHIVE_LIST = [
"junnyu/roformer_chinese_small",
"junnyu/roformer_chinese_base"
# See all RoFormer models at https://huggingface.co/models?filter=roformer
]
class TFRoFormerSinusoidalPositionalEmbedding(tf.keras.layers.Layer):
"""This module produces sinusoidal positional embeddings of any length."""
def __init__(self, num_positions: int, embedding_dim: int, **kwargs):
super().__init__(**kwargs)
if embedding_dim % 2 != 0:
raise NotImplementedError(f"odd embedding_dim {embedding_dim} not supported")
self.embedding_dim = embedding_dim
self.num_positions = num_positions
def build(self, input_shape: tf.TensorShape):
"""
Build shared token embedding layer Shared weights logic adapted from
https://github.com/tensorflow/models/blob/a009f4fb9d2fc4949e32192a944688925ef78659/official/transformer/v2/embedding_layer.py#L24
"""
weight = self._init_weight(self.num_positions, self.embedding_dim)
self.weight = self.add_weight(
name="embeddings",
shape=[self.num_positions, self.embedding_dim],
)
weight = tf.cast(weight, dtype=self.weight.dtype)
self.weight.assign(weight)
super().build(input_shape)
@staticmethod
def _init_weight(n_pos: int, dim: int):
"""
Identical to the XLM create_sinusoidal_embeddings except features are not interleaved. The cos features are in
the 2nd half of the vector. [dim // 2:]
"""
position_enc = np.array(
[[pos / np.power(10000, 2 * (j // 2) / dim) for j in range(dim)] for pos in range(n_pos)]
)
table = np.zeros_like(position_enc)
# index 0 is all zero
table[:, 0 : dim // 2] = np.sin(position_enc[:, 0::2])
table[:, dim // 2 :] = np.cos(position_enc[:, 1::2])
# convert to tensor
table = tf.convert_to_tensor(table)
tf.stop_gradient(table)
return table
def call(self, input_shape: tf.TensorShape, past_key_values_length: int = 0):
"""Input is expected to be of size [bsz x seqlen]."""
bsz, seq_len = input_shape[:2]
positions = tf.range(past_key_values_length, seq_len + past_key_values_length, delta=1, name="range")
return tf.gather(self.weight, positions)
class TFRoFormerEmbeddings(tf.keras.layers.Layer):
"""Construct the embeddings from word, position and token_type embeddings."""
def __init__(self, config: RoFormerConfig, **kwargs):
super().__init__(**kwargs)
self.vocab_size = config.vocab_size
self.type_vocab_size = config.type_vocab_size
self.embedding_size = config.embedding_size
self.initializer_range = config.initializer_range
self.embeddings_sum = tf.keras.layers.Add()
self.LayerNorm = tf.keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name="LayerNorm")
self.dropout = tf.keras.layers.Dropout(rate=config.hidden_dropout_prob)
def build(self, input_shape: tf.TensorShape):
with tf.name_scope("word_embeddings"):
self.weight = self.add_weight(
name="weight",
shape=[self.vocab_size, self.embedding_size],
initializer=get_initializer(self.initializer_range),
)
with tf.name_scope("token_type_embeddings"):
self.token_type_embeddings = self.add_weight(
name="embeddings",
shape=[self.type_vocab_size, self.embedding_size],
initializer=get_initializer(self.initializer_range),
)
super().build(input_shape)
def call(
self,
input_ids: tf.Tensor = None,
token_type_ids: tf.Tensor = None,
inputs_embeds: tf.Tensor = None,
training: bool = False,
) -> tf.Tensor:
"""
Applies embedding based on inputs tensor.
Returns:
final_embeddings (:obj:`tf.Tensor`): output embedding tensor.
"""
assert not (input_ids is None and inputs_embeds is None)
if input_ids is not None:
inputs_embeds = tf.gather(params=self.weight, indices=input_ids)
input_shape = shape_list(inputs_embeds)[:-1]
if token_type_ids is None:
token_type_ids = tf.fill(dims=input_shape, value=0)
token_type_embeds = tf.gather(params=self.token_type_embeddings, indices=token_type_ids)
final_embeddings = self.embeddings_sum(inputs=[inputs_embeds, token_type_embeds])
final_embeddings = self.LayerNorm(inputs=final_embeddings)
final_embeddings = self.dropout(inputs=final_embeddings, training=training)
return final_embeddings
class TFRoFormerSelfAttention(tf.keras.layers.Layer):
def __init__(self, config: RoFormerConfig, **kwargs):
super().__init__(**kwargs)
if config.hidden_size % config.num_attention_heads != 0:
raise ValueError(
f"The hidden size ({config.hidden_size}) is not a multiple of the number "
f"of attention heads ({config.num_attention_heads})"
)
self.num_attention_heads = config.num_attention_heads
self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
self.all_head_size = self.num_attention_heads * self.attention_head_size
self.sqrt_att_head_size = math.sqrt(self.attention_head_size)
self.query = tf.keras.layers.Dense(
units=self.all_head_size, kernel_initializer=get_initializer(config.initializer_range), name="query"
)
self.key = tf.keras.layers.Dense(
units=self.all_head_size, kernel_initializer=get_initializer(config.initializer_range), name="key"
)
self.value = tf.keras.layers.Dense(
units=self.all_head_size, kernel_initializer=get_initializer(config.initializer_range), name="value"
)
self.dropout = tf.keras.layers.Dropout(rate=config.attention_probs_dropout_prob)
self.rotary_value = config.rotary_value
def transpose_for_scores(self, tensor: tf.Tensor, batch_size: int) -> tf.Tensor:
# Reshape from [batch_size, seq_length, all_head_size] to [batch_size, seq_length, num_attention_heads, attention_head_size]
tensor = tf.reshape(tensor=tensor, shape=(batch_size, -1, self.num_attention_heads, self.attention_head_size))
# Transpose the tensor from [batch_size, seq_length, num_attention_heads, attention_head_size] to [batch_size, num_attention_heads, seq_length, attention_head_size]
return tf.transpose(tensor, perm=[0, 2, 1, 3])
def call(
self,
hidden_states: tf.Tensor,
attention_mask: tf.Tensor,
sinusoidal_pos: tf.Tensor,
head_mask: tf.Tensor,
output_attentions: bool,
training: bool = False,
) -> Tuple[tf.Tensor]:
batch_size = shape_list(hidden_states)[0]
mixed_query_layer = self.query(inputs=hidden_states)
mixed_key_layer = self.key(inputs=hidden_states)
mixed_value_layer = self.value(inputs=hidden_states)
query_layer = self.transpose_for_scores(mixed_query_layer, batch_size)
key_layer = self.transpose_for_scores(mixed_key_layer, batch_size)
value_layer = self.transpose_for_scores(mixed_value_layer, batch_size)
if sinusoidal_pos is not None:
if self.rotary_value:
query_layer, key_layer, value_layer = self.apply_rotary_position_embeddings(
sinusoidal_pos, query_layer, key_layer, value_layer
)
else:
query_layer, key_layer = self.apply_rotary_position_embeddings(sinusoidal_pos, query_layer, key_layer)
# Take the dot product between "query" and "key" to get the raw attention scores.
# (batch size, num_heads, seq_len_q, seq_len_k)
attention_scores = tf.matmul(query_layer, key_layer, transpose_b=True)
dk = tf.cast(self.sqrt_att_head_size, dtype=attention_scores.dtype)
attention_scores = tf.divide(attention_scores, dk)
if attention_mask is not None:
# Apply the attention mask is (precomputed for all layers in TFRoFormerModel call() function)
attention_scores = tf.add(attention_scores, attention_mask)
# Normalize the attention scores to probabilities.
attention_probs = tf.nn.softmax(logits=attention_scores, axis=-1)
# This is actually dropping out entire tokens to attend to, which might
# seem a bit unusual, but is taken from the original Transformer paper.
attention_probs = self.dropout(inputs=attention_probs, training=training)
# Mask heads if we want to
if head_mask is not None:
attention_probs = tf.multiply(attention_probs, head_mask)
attention_output = tf.matmul(attention_probs, value_layer)
attention_output = tf.transpose(attention_output, perm=[0, 2, 1, 3])
# (batch_size, seq_len_q, all_head_size)
attention_output = tf.reshape(tensor=attention_output, shape=(batch_size, -1, self.all_head_size))
outputs = (attention_output, attention_probs) if output_attentions else (attention_output,)
return outputs
@staticmethod
def apply_rotary_position_embeddings(sinusoidal_pos, query_layer, key_layer, value_layer=None):
# https://kexue.fm/archives/8265
# sin [batch_size, num_heads, sequence_length, embed_size_per_head//2]
# cos [batch_size, num_heads, sequence_length, embed_size_per_head//2]
sin, cos = tf.split(sinusoidal_pos, num_or_size_splits=2, axis=-1)
# sin [θ0,θ1,θ2......θd/2-1]-> sin_pos [θ0,θ0,θ1,θ1,θ2,θ2......θd/2-1,θd/2-1]
# cos [θ0,θ1,θ2......θd/2-1]-> cos_pos [θ0,θ0,θ1,θ1,θ2,θ2......θd/2-1,θd/2-1]
sin_pos = tf.repeat(sin, 2, axis=-1)
cos_pos = tf.repeat(cos, 2, axis=-1)
# rotate_half_query_layer [-q1,q0,-q3,q2......,-qd-1,qd-2]
rotate_half_query_layer = tf.stack([-query_layer[..., 1::2], query_layer[..., ::2]], axis=-1)
rotate_half_query_layer = tf.reshape(rotate_half_query_layer, shape_list(query_layer))
query_layer = query_layer * cos_pos + rotate_half_query_layer * sin_pos
# rotate_half_key_layer [-k1,k0,-k3,k2......,-kd-1,kd-2]
rotate_half_key_layer = tf.stack([-key_layer[..., 1::2], key_layer[..., ::2]], axis=-1)
rotate_half_key_layer = tf.reshape(rotate_half_key_layer, shape_list(key_layer))
key_layer = key_layer * cos_pos + rotate_half_key_layer * sin_pos
if value_layer is not None:
# rotate_half_value_layer [-v1,v0,-v3,v2......,-vd-1,vd-2]
rotate_half_value_layer = tf.stack([-value_layer[..., 1::2], value_layer[..., ::2]], axis=-1)
rotate_half_value_layer = tf.reshape(rotate_half_value_layer, shape_list(value_layer))
value_layer = value_layer * cos_pos + rotate_half_value_layer * sin_pos
return query_layer, key_layer, value_layer
return query_layer, key_layer
# Copied from transformers.models.bert.modeling_tf_bert.TFBertSelfOutput with Bert->RoFormer
class TFRoFormerSelfOutput(tf.keras.layers.Layer):
def __init__(self, config: RoFormerConfig, **kwargs):
super().__init__(**kwargs)
self.dense = tf.keras.layers.Dense(
units=config.hidden_size, kernel_initializer=get_initializer(config.initializer_range), name="dense"
)
self.LayerNorm = tf.keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name="LayerNorm")
self.dropout = tf.keras.layers.Dropout(rate=config.hidden_dropout_prob)
def call(self, hidden_states: tf.Tensor, input_tensor: tf.Tensor, training: bool = False) -> tf.Tensor:
hidden_states = self.dense(inputs=hidden_states)
hidden_states = self.dropout(inputs=hidden_states, training=training)
hidden_states = self.LayerNorm(inputs=hidden_states + input_tensor)
return hidden_states
class TFRoFormerAttention(tf.keras.layers.Layer):
def __init__(self, config: RoFormerConfig, **kwargs):
super().__init__(**kwargs)
self.self_attention = TFRoFormerSelfAttention(config, name="self")
self.dense_output = TFRoFormerSelfOutput(config, name="output")
def prune_heads(self, heads):
raise NotImplementedError
def call(
self,
input_tensor: tf.Tensor,
attention_mask: tf.Tensor,
sinusoidal_pos: tf.Tensor,
head_mask: tf.Tensor,
output_attentions: bool,
training: bool = False,
) -> Tuple[tf.Tensor]:
self_outputs = self.self_attention(
hidden_states=input_tensor,
attention_mask=attention_mask,
sinusoidal_pos=sinusoidal_pos,
head_mask=head_mask,
output_attentions=output_attentions,
training=training,
)
attention_output = self.dense_output(
hidden_states=self_outputs[0], input_tensor=input_tensor, training=training
)
outputs = (attention_output,) + self_outputs[1:] # add attentions if we output them
return outputs
# Copied from transformers.models.bert.modeling_tf_bert.TFBertIntermediate with Bert->RoFormer
class TFRoFormerIntermediate(tf.keras.layers.Layer):
def __init__(self, config: RoFormerConfig, **kwargs):
super().__init__(**kwargs)
self.dense = tf.keras.layers.Dense(
units=config.intermediate_size, kernel_initializer=get_initializer(config.initializer_range), name="dense"
)
if isinstance(config.hidden_act, str):
self.intermediate_act_fn = get_tf_activation(config.hidden_act)
else:
self.intermediate_act_fn = config.hidden_act
def call(self, hidden_states: tf.Tensor) -> tf.Tensor:
hidden_states = self.dense(inputs=hidden_states)
hidden_states = self.intermediate_act_fn(hidden_states)
return hidden_states
# Copied from transformers.models.bert.modeling_tf_bert.TFBertOutput with Bert->RoFormer
class TFRoFormerOutput(tf.keras.layers.Layer):
def __init__(self, config: RoFormerConfig, **kwargs):
super().__init__(**kwargs)
self.dense = tf.keras.layers.Dense(
units=config.hidden_size, kernel_initializer=get_initializer(config.initializer_range), name="dense"
)
self.LayerNorm = tf.keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name="LayerNorm")
self.dropout = tf.keras.layers.Dropout(rate=config.hidden_dropout_prob)
def call(self, hidden_states: tf.Tensor, input_tensor: tf.Tensor, training: bool = False) -> tf.Tensor:
hidden_states = self.dense(inputs=hidden_states)
hidden_states = self.dropout(inputs=hidden_states, training=training)
hidden_states = self.LayerNorm(inputs=hidden_states + input_tensor)
return hidden_states
class TFRoFormerLayer(tf.keras.layers.Layer):
def __init__(self, config: RoFormerConfig, **kwargs):
super().__init__(**kwargs)
self.attention = TFRoFormerAttention(config, name="attention")
self.intermediate = TFRoFormerIntermediate(config, name="intermediate")
self.roformer_output = TFRoFormerOutput(config, name="output")
def call(
self,
hidden_states: tf.Tensor,
attention_mask: tf.Tensor,
sinusoidal_pos: tf.Tensor,
head_mask: tf.Tensor,
output_attentions: bool,
training: bool = False,
) -> Tuple[tf.Tensor]:
attention_outputs = self.attention(
input_tensor=hidden_states,
attention_mask=attention_mask,
sinusoidal_pos=sinusoidal_pos,
head_mask=head_mask,
output_attentions=output_attentions,
training=training,
)
attention_output = attention_outputs[0]
intermediate_output = self.intermediate(hidden_states=attention_output)
layer_output = self.roformer_output(
hidden_states=intermediate_output, input_tensor=attention_output, training=training
)
outputs = (layer_output,) + attention_outputs[1:] # add attentions if we output them
return outputs
class TFRoFormerEncoder(tf.keras.layers.Layer):
def __init__(self, config: RoFormerConfig, **kwargs):
super().__init__(**kwargs)
self.embed_positions = TFRoFormerSinusoidalPositionalEmbedding(
config.max_position_embeddings,
config.hidden_size // config.num_attention_heads,
name="embed_positions",
)
self.layer = [TFRoFormerLayer(config, name=f"layer_._{i}") for i in range(config.num_hidden_layers)]
def call(
self,
hidden_states: tf.Tensor,
attention_mask: tf.Tensor,
head_mask: tf.Tensor,
output_attentions: bool,
output_hidden_states: bool,
return_dict: bool,
training: bool = False,
) -> Union[TFBaseModelOutput, Tuple[tf.Tensor]]:
all_hidden_states = () if output_hidden_states else None
all_attentions = () if output_attentions else None
# [sequence_length, embed_size_per_head] -> [batch_size, num_heads, sequence_length, embed_size_per_head]
sinusoidal_pos = self.embed_positions(shape_list(hidden_states)[:-1])[None, None, :, :]
for i, layer_module in enumerate(self.layer):
if output_hidden_states:
all_hidden_states = all_hidden_states + (hidden_states,)
layer_outputs = layer_module(
hidden_states=hidden_states,
attention_mask=attention_mask,
sinusoidal_pos=sinusoidal_pos,
head_mask=head_mask[i],
output_attentions=output_attentions,
training=training,
)
hidden_states = layer_outputs[0]
if output_attentions:
all_attentions = all_attentions + (layer_outputs[1],)
# Add last layer
if output_hidden_states:
all_hidden_states = all_hidden_states + (hidden_states,)
if not return_dict:
return tuple(v for v in [hidden_states, all_hidden_states, all_attentions] if v is not None)
return TFBaseModelOutput(
last_hidden_state=hidden_states, hidden_states=all_hidden_states, attentions=all_attentions
)
class TFRoFormerPredictionHeadTransform(tf.keras.layers.Layer):
def __init__(self, config: RoFormerConfig, **kwargs):
super().__init__(**kwargs)
self.dense = tf.keras.layers.Dense(
units=config.embedding_size,
kernel_initializer=get_initializer(config.initializer_range),
name="dense",
)
if isinstance(config.hidden_act, str):
self.transform_act_fn = get_tf_activation(config.hidden_act)
else:
self.transform_act_fn = config.hidden_act
self.LayerNorm = tf.keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name="LayerNorm")
def call(self, hidden_states: tf.Tensor) -> tf.Tensor:
hidden_states = self.dense(inputs=hidden_states)
hidden_states = self.transform_act_fn(hidden_states)
hidden_states = self.LayerNorm(inputs=hidden_states)
return hidden_states
class TFRoFormerLMPredictionHead(tf.keras.layers.Layer):
def __init__(self, config: RoFormerConfig, input_embeddings: tf.keras.layers.Layer, **kwargs):
super().__init__(**kwargs)
self.vocab_size = config.vocab_size
self.embedding_size = config.embedding_size
self.transform = TFRoFormerPredictionHeadTransform(config, name="transform")
# The output weights are the same as the input embeddings, but there is
# an output-only bias for each token.
self.input_embeddings = input_embeddings
def build(self, input_shape: tf.TensorShape):
self.bias = self.add_weight(shape=(self.vocab_size,), initializer="zeros", trainable=True, name="bias")
super().build(input_shape)
def get_output_embeddings(self) -> tf.keras.layers.Layer:
return self.input_embeddings
def set_output_embeddings(self, value: tf.Variable):
self.input_embeddings.weight = value
self.input_embeddings.vocab_size = shape_list(value)[0]
def get_bias(self) -> Dict[str, tf.Variable]:
return {"bias": self.bias}
def set_bias(self, value: tf.Variable):
self.bias = value["bias"]
self.vocab_size = shape_list(value["bias"])[0]
def call(self, hidden_states: tf.Tensor) -> tf.Tensor:
hidden_states = self.transform(hidden_states=hidden_states)
seq_length = shape_list(hidden_states)[1]
hidden_states = tf.reshape(tensor=hidden_states, shape=[-1, self.embedding_size])
hidden_states = tf.matmul(a=hidden_states, b=self.input_embeddings.weight, transpose_b=True)
hidden_states = tf.reshape(tensor=hidden_states, shape=[-1, seq_length, self.vocab_size])
hidden_states = tf.nn.bias_add(value=hidden_states, bias=self.bias)
return hidden_states
# Copied from transformers.models.bert.modeling_tf_bert.TFBertMLMHead with Bert->RoFormer
class TFRoFormerMLMHead(tf.keras.layers.Layer):
def __init__(self, config: RoFormerConfig, input_embeddings: tf.keras.layers.Layer, **kwargs):
super().__init__(**kwargs)
self.predictions = TFRoFormerLMPredictionHead(config, input_embeddings, name="predictions")
def call(self, sequence_output: tf.Tensor) -> tf.Tensor:
prediction_scores = self.predictions(hidden_states=sequence_output)
return prediction_scores
@keras_serializable
class TFRoFormerMainLayer(tf.keras.layers.Layer):
config_class = RoFormerConfig
def __init__(self, config: RoFormerConfig, add_pooling_layer: bool = True, **kwargs):
super().__init__(**kwargs)
self.config = config
self.embeddings = TFRoFormerEmbeddings(config, name="embeddings")
if config.embedding_size != config.hidden_size:
self.embeddings_project = tf.keras.layers.Dense(config.hidden_size, name="embeddings_project")
self.encoder = TFRoFormerEncoder(config, name="encoder")
def get_input_embeddings(self) -> tf.keras.layers.Layer:
return self.embeddings
def set_input_embeddings(self, value: tf.Variable):
self.embeddings.weight = value
self.embeddings.vocab_size = shape_list(value)[0]
def _prune_heads(self, heads_to_prune):
"""
Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
class PreTrainedModel
"""
raise NotImplementedError
def call(
self,
input_ids: Optional[TFModelInputType] = None,
attention_mask: Optional[Union[np.ndarray, tf.Tensor]] = None,
token_type_ids: Optional[Union[np.ndarray, tf.Tensor]] = None,
head_mask: Optional[Union[np.ndarray, tf.Tensor]] = None,
inputs_embeds: Optional[Union[np.ndarray, tf.Tensor]] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
training: bool = False,
**kwargs,
) -> Union[TFBaseModelOutput, Tuple[tf.Tensor]]:
inputs = input_processing(
func=self.call,
config=self.config,
input_ids=input_ids,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
head_mask=head_mask,
inputs_embeds=inputs_embeds,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
training=training,
kwargs_call=kwargs,
)
if inputs["input_ids"] is not None and inputs["inputs_embeds"] is not None:
raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
elif inputs["input_ids"] is not None:
input_shape = shape_list(inputs["input_ids"])
elif inputs["inputs_embeds"] is not None:
input_shape = shape_list(inputs["inputs_embeds"])[:-1]
else:
raise ValueError("You have to specify either input_ids or inputs_embeds")
if inputs["attention_mask"] is None:
inputs["attention_mask"] = tf.fill(dims=input_shape, value=1)
if inputs["token_type_ids"] is None:
inputs["token_type_ids"] = tf.fill(dims=input_shape, value=0)
embedding_output = self.embeddings(
input_ids=inputs["input_ids"],
token_type_ids=inputs["token_type_ids"],
inputs_embeds=inputs["inputs_embeds"],
training=inputs["training"],
)
if hasattr(self, "embeddings_project"):
embedding_output = self.embeddings_project(embedding_output, training=inputs["training"])
# We create a 3D attention mask from a 2D tensor mask.
# Sizes are [batch_size, 1, 1, to_seq_length]
# So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
# this attention mask is more simple than the triangular masking of causal attention
# used in OpenAI GPT, we just need to prepare the broadcast dimension here.
extended_attention_mask = tf.reshape(inputs["attention_mask"], (input_shape[0], 1, 1, input_shape[1]))
# Since attention_mask is 1.0 for positions we want to attend and 0.0 for
# masked positions, this operation will create a tensor which is 0.0 for
# positions we want to attend and -10000.0 for masked positions.
# Since we are adding it to the raw scores before the softmax, this is
# effectively the same as removing these entirely.
extended_attention_mask = tf.cast(extended_attention_mask, dtype=embedding_output.dtype)
one_cst = tf.constant(1.0, dtype=embedding_output.dtype)
ten_thousand_cst = tf.constant(-10000.0, dtype=embedding_output.dtype)
extended_attention_mask = tf.multiply(tf.subtract(one_cst, extended_attention_mask), ten_thousand_cst)
# Prepare head mask if needed
# 1.0 in head_mask indicate we keep the head
# attention_probs has shape bsz x n_heads x N x N
# input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
# and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
if inputs["head_mask"] is not None:
raise NotImplementedError
else:
inputs["head_mask"] = [None] * self.config.num_hidden_layers
encoder_outputs = self.encoder(
hidden_states=embedding_output,
attention_mask=extended_attention_mask,
head_mask=inputs["head_mask"],
output_attentions=inputs["output_attentions"],
output_hidden_states=inputs["output_hidden_states"],
return_dict=inputs["return_dict"],
training=inputs["training"],
)
sequence_output = encoder_outputs[0]
if not inputs["return_dict"]:
return (sequence_output,) + encoder_outputs[1:]
return TFBaseModelOutput(
last_hidden_state=sequence_output,
hidden_states=encoder_outputs.hidden_states,
attentions=encoder_outputs.attentions,
)
class TFRoFormerPreTrainedModel(TFPreTrainedModel):
"""
An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
models.
"""
config_class = RoFormerConfig
base_model_prefix = "roformer"
ROFORMER_START_DOCSTRING = r"""
This model inherits from :class:`~transformers.TFPreTrainedModel`. Check the superclass documentation for the
generic methods the library implements for all its model (such as downloading or saving, resizing the input
embeddings, pruning heads etc.)
This model is also a `tf.keras.Model <https://www.tensorflow.org/api_docs/python/tf/keras/Model>`__ subclass. Use
it as a regular TF 2.0 Keras Model and refer to the TF 2.0 documentation for all matter related to general usage
and behavior.
.. note::
TF 2.0 models accepts two formats as inputs:
- having all inputs as keyword arguments (like PyTorch models), or
- having all inputs as a list, tuple or dict in the first positional arguments.
This second option is useful when using :meth:`tf.keras.Model.fit` method which currently requires having all
the tensors in the first argument of the model call function: :obj:`model(inputs)`.
If you choose this second option, there are three possibilities you can use to gather all the input Tensors in
the first positional argument :
- a single Tensor with :obj:`input_ids` only and nothing else: :obj:`model(inputs_ids)`
- a list of varying length with one or several input Tensors IN THE ORDER given in the docstring:
:obj:`model([input_ids, attention_mask])` or :obj:`model([input_ids, attention_mask, token_type_ids])`
- a dictionary with one or several input Tensors associated to the input names given in the docstring:
:obj:`model({"input_ids": input_ids, "token_type_ids": token_type_ids})`
Args:
config (:class:`~transformers.RoFormerConfig`): Model configuration class with all the parameters of the model.
Initializing with a config file does not load the weights associated with the model, only the
configuration. Check out the :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model
weights.
"""
ROFORMER_INPUTS_DOCSTRING = r"""
Args:
input_ids (:obj:`np.ndarray`, :obj:`tf.Tensor`, :obj:`List[tf.Tensor]` :obj:`Dict[str, tf.Tensor]` or :obj:`Dict[str, np.ndarray]` and each example must have the shape :obj:`({0})`):
Indices of input sequence tokens in the vocabulary.
Indices can be obtained using :class:`~transformers.RoFormerTokenizer`. See
:func:`transformers.PreTrainedTokenizer.__call__` and :func:`transformers.PreTrainedTokenizer.encode` for
details.
`What are input IDs? <../glossary.html#input-ids>`__
attention_mask (:obj:`np.ndarray` or :obj:`tf.Tensor` of shape :obj:`({0})`, `optional`):
Mask to avoid performing attention on padding token indices. Mask values selected in ``[0, 1]``:
- 1 for tokens that are **not masked**,
- 0 for tokens that are **masked**.
`What are attention masks? <../glossary.html#attention-mask>`__
token_type_ids (:obj:`np.ndarray` or :obj:`tf.Tensor` of shape :obj:`({0})`, `optional`):
Segment token indices to indicate first and second portions of the inputs. Indices are selected in ``[0,
1]``:
- 0 corresponds to a `sentence A` token,
- 1 corresponds to a `sentence B` token.
`What are token type IDs? <../glossary.html#token-type-ids>`__
head_mask (:obj:`np.ndarray` or :obj:`tf.Tensor` of shape :obj:`(num_heads,)` or :obj:`(num_layers, num_heads)`, `optional`):
Mask to nullify selected heads of the self-attention modules. Mask values selected in ``[0, 1]``:
- 1 indicates the head is **not masked**,
- 0 indicates the head is **masked**.
inputs_embeds (:obj:`np.ndarray` or :obj:`tf.Tensor` of shape :obj:`({0}, hidden_size)`, `optional`):
Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded representation.
This is useful if you want more control over how to convert :obj:`input_ids` indices into associated
vectors than the model's internal embedding lookup matrix.
output_attentions (:obj:`bool`, `optional`):
Whether or not to return the attentions tensors of all attention layers. See ``attentions`` under returned
tensors for more detail. This argument can be used only in eager mode, in graph mode the value in the
config will be used instead.
output_hidden_states (:obj:`bool`, `optional`):
Whether or not to return the hidden states of all layers. See ``hidden_states`` under returned tensors for
more detail. This argument can be used only in eager mode, in graph mode the value in the config will be
used instead.
return_dict (:obj:`bool`, `optional`):
Whether or not to return a :class:`~transformers.file_utils.ModelOutput` instead of a plain tuple. This
argument can be used in eager mode, in graph mode the value will always be set to True.
training (:obj:`bool`, `optional`, defaults to :obj:`False`):
Whether or not to use the model in training mode (some modules like dropout modules have different
behaviors between training and evaluation).
"""
@add_start_docstrings(
"The bare RoFormer Model transformer outputing raw hidden-states without any specific head on top.",
ROFORMER_START_DOCSTRING,
)
class TFRoFormerModel(TFRoFormerPreTrainedModel):
def __init__(self, config: RoFormerConfig, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
self.roformer = TFRoFormerMainLayer(config, name="roformer")
@add_start_docstrings_to_model_forward(ROFORMER_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
@add_code_sample_docstrings(
tokenizer_class=_TOKENIZER_FOR_DOC,
checkpoint=_CHECKPOINT_FOR_DOC,
output_type=TFBaseModelOutputWithPooling,
config_class=_CONFIG_FOR_DOC,
)
def call(
self,
input_ids: Optional[TFModelInputType] = None,
attention_mask: Optional[Union[np.ndarray, tf.Tensor]] = None,
token_type_ids: Optional[Union[np.ndarray, tf.Tensor]] = None,
head_mask: Optional[Union[np.ndarray, tf.Tensor]] = None,
inputs_embeds: Optional[Union[np.ndarray, tf.Tensor]] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
training: Optional[bool] = False,
**kwargs,
) -> Union[TFBaseModelOutputWithPooling, Tuple[tf.Tensor]]:
inputs = input_processing(
func=self.call,
config=self.config,
input_ids=input_ids,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
head_mask=head_mask,
inputs_embeds=inputs_embeds,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
training=training,
kwargs_call=kwargs,
)
outputs = self.roformer(
input_ids=inputs["input_ids"],
attention_mask=inputs["attention_mask"],
token_type_ids=inputs["token_type_ids"],
head_mask=inputs["head_mask"],
inputs_embeds=inputs["inputs_embeds"],
output_attentions=inputs["output_attentions"],
output_hidden_states=inputs["output_hidden_states"],
return_dict=inputs["return_dict"],
training=inputs["training"],
)
return outputs
def serving_output(self, output: TFBaseModelOutput) -> TFBaseModelOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFBaseModelOutput(last_hidden_state=output.last_hidden_state, hidden_states=hs, attentions=attns)
@add_start_docstrings("""RoFormer Model with a `language modeling` head on top. """, ROFORMER_START_DOCSTRING)
class TFRoFormerForMaskedLM(TFRoFormerPreTrainedModel, TFMaskedLanguageModelingLoss):
def __init__(self, config: RoFormerConfig, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
if config.is_decoder:
logger.warning(
"If you want to use `TFRoFormerForMaskedLM` make sure `config.is_decoder=False` for "
"bi-directional self-attention."
)
self.roformer = TFRoFormerMainLayer(config, name="roformer")
self.mlm = TFRoFormerMLMHead(config, input_embeddings=self.roformer.embeddings, name="mlm___cls")
def get_lm_head(self) -> tf.keras.layers.Layer:
return self.mlm.predictions
@add_start_docstrings_to_model_forward(ROFORMER_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
@add_code_sample_docstrings(
tokenizer_class=_TOKENIZER_FOR_DOC,
checkpoint=_CHECKPOINT_FOR_DOC,
output_type=TFMaskedLMOutput,
config_class=_CONFIG_FOR_DOC,
)
def call(
self,
input_ids: Optional[TFModelInputType] = None,
attention_mask: Optional[Union[np.ndarray, tf.Tensor]] = None,
token_type_ids: Optional[Union[np.ndarray, tf.Tensor]] = None,
head_mask: Optional[Union[np.ndarray, tf.Tensor]] = None,
inputs_embeds: Optional[Union[np.ndarray, tf.Tensor]] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
training: Optional[bool] = False,
**kwargs,
) -> Union[TFMaskedLMOutput, Tuple[tf.Tensor]]:
r"""
labels (:obj:`tf.Tensor` or :obj:`np.ndarray` of shape :obj:`(batch_size, sequence_length)`, `optional`):
Labels for computing the masked language modeling loss. Indices should be in ``[-100, 0, ...,
config.vocab_size]`` (see ``input_ids`` docstring) Tokens with indices set to ``-100`` are ignored
(masked), the loss is only computed for the tokens with labels in ``[0, ..., config.vocab_size]``
"""
inputs = input_processing(
func=self.call,
config=self.config,
input_ids=input_ids,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
head_mask=head_mask,
inputs_embeds=inputs_embeds,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
labels=labels,
training=training,
kwargs_call=kwargs,
)
outputs = self.roformer(
input_ids=inputs["input_ids"],
attention_mask=inputs["attention_mask"],
token_type_ids=inputs["token_type_ids"],
head_mask=inputs["head_mask"],
inputs_embeds=inputs["inputs_embeds"],
output_attentions=inputs["output_attentions"],
output_hidden_states=inputs["output_hidden_states"],
return_dict=inputs["return_dict"],
training=inputs["training"],
)
sequence_output = outputs[0]
prediction_scores = self.mlm(sequence_output=sequence_output, training=inputs["training"])
loss = (
None if inputs["labels"] is None else self.compute_loss(labels=inputs["labels"], logits=prediction_scores)
)
if not inputs["return_dict"]:
output = (prediction_scores,) + outputs[2:]
return ((loss,) + output) if loss is not None else output
return TFMaskedLMOutput(
loss=loss,
logits=prediction_scores,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
def serving_output(self, output: TFMaskedLMOutput) -> TFMaskedLMOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFMaskedLMOutput(logits=output.logits, hidden_states=hs, attentions=attns)
@add_start_docstrings(
"""RoFormer Model with a `language modeling` head on top for CLM fine-tuning. """, ROFORMER_START_DOCSTRING
)
class TFRoFormerForCausalLM(TFRoFormerPreTrainedModel, TFCausalLanguageModelingLoss):
def __init__(self, config: RoFormerConfig, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
if not config.is_decoder:
logger.warning("If you want to use `TFRoFormerForCausalLM` as a standalone, add `is_decoder=True.`")
self.roformer = TFRoFormerMainLayer(config, name="roformer")
self.mlm = TFRoFormerMLMHead(config, input_embeddings=self.roformer.embeddings, name="mlm___cls")
def get_lm_head(self) -> tf.keras.layers.Layer:
return self.mlm.predictions
@add_code_sample_docstrings(
tokenizer_class=_TOKENIZER_FOR_DOC,
checkpoint=_CHECKPOINT_FOR_DOC,
output_type=TFCausalLMOutput,
config_class=_CONFIG_FOR_DOC,
)
def call(
self,
input_ids: Optional[TFModelInputType] = None,
attention_mask: Optional[Union[np.ndarray, tf.Tensor]] = None,
token_type_ids: Optional[Union[np.ndarray, tf.Tensor]] = None,
head_mask: Optional[Union[np.ndarray, tf.Tensor]] = None,
inputs_embeds: Optional[Union[np.ndarray, tf.Tensor]] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
training: Optional[bool] = False,
**kwargs,
) -> Union[TFCausalLMOutput, Tuple[tf.Tensor]]:
r"""
labels (:obj:`tf.Tensor` or :obj:`np.ndarray` of shape :obj:`(batch_size, sequence_length)`, `optional`):
Labels for computing the cross entropy classification loss. Indices should be in ``[0, ...,
config.vocab_size - 1]``.
"""
inputs = input_processing(
func=self.call,
config=self.config,
input_ids=input_ids,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
head_mask=head_mask,
inputs_embeds=inputs_embeds,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
labels=labels,
training=training,
kwargs_call=kwargs,
)
outputs = self.roformer(
input_ids=inputs["input_ids"],
attention_mask=inputs["attention_mask"],
token_type_ids=inputs["token_type_ids"],
head_mask=inputs["head_mask"],
inputs_embeds=inputs["inputs_embeds"],
output_attentions=inputs["output_attentions"],
output_hidden_states=inputs["output_hidden_states"],
return_dict=inputs["return_dict"],
training=inputs["training"],
)
sequence_output = outputs[0]
logits = self.mlm(sequence_output=sequence_output, training=inputs["training"])
loss = None
if inputs["labels"] is not None:
# shift labels to the left and cut last logit token
logits = logits[:, :-1]
labels = inputs["labels"][:, 1:]
loss = self.compute_loss(labels=labels, logits=logits)
if not inputs["return_dict"]:
output = (logits,) + outputs[2:]
return ((loss,) + output) if loss is not None else output
return TFCausalLMOutput(
loss=loss,
logits=logits,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
def serving_output(self, output: TFCausalLMOutput) -> TFCausalLMOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFCausalLMOutput(logits=output.logits, hidden_states=hs, attentions=attns)
class TFRoFormerClassificationHead(tf.keras.layers.Layer):
"""Head for sentence-level classification tasks."""
def __init__(self, config: RoFormerConfig, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
self.dense = tf.keras.layers.Dense(
units=config.hidden_size, kernel_initializer=get_initializer(config.initializer_range), name="dense"
)
self.dropout = tf.keras.layers.Dropout(rate=config.hidden_dropout_prob)
self.out_proj = tf.keras.layers.Dense(
units=config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="out_proj"
)
if isinstance(config.hidden_act, str):
self.classifier_act_fn = get_tf_activation(config.hidden_act)
else:
self.classifier_act_fn = config.hidden_act
def call(self, hidden_states: tf.Tensor, training: bool = False) -> tf.Tensor:
hidden_states = hidden_states[:, 0, :] # take <s> token (equiv. to [CLS])
hidden_states = self.dropout(inputs=hidden_states, training=training)
hidden_states = self.dense(inputs=hidden_states)
hidden_states = self.classifier_act_fn(hidden_states)
hidden_states = self.dropout(inputs=hidden_states, training=training)
hidden_states = self.out_proj(hidden_states)
return hidden_states
@add_start_docstrings(
"""
RoFormer Model transformer with a sequence classification/regression head on top e.g., for GLUE tasks.
""",
ROFORMER_START_DOCSTRING,
)
class TFRoFormerForSequenceClassification(TFRoFormerPreTrainedModel, TFSequenceClassificationLoss):
def __init__(self, config: RoFormerConfig, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
self.num_labels = config.num_labels
self.roformer = TFRoFormerMainLayer(config, name="roformer")
self.classifier = TFRoFormerClassificationHead(config, name="classifier")
@add_start_docstrings_to_model_forward(ROFORMER_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
@add_code_sample_docstrings(
tokenizer_class=_TOKENIZER_FOR_DOC,
checkpoint=_CHECKPOINT_FOR_DOC,
output_type=TFSequenceClassifierOutput,
config_class=_CONFIG_FOR_DOC,
)
def call(
self,
input_ids: Optional[TFModelInputType] = None,
attention_mask: Optional[Union[np.ndarray, tf.Tensor]] = None,
token_type_ids: Optional[Union[np.ndarray, tf.Tensor]] = None,
head_mask: Optional[Union[np.ndarray, tf.Tensor]] = None,
inputs_embeds: Optional[Union[np.ndarray, tf.Tensor]] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
training: Optional[bool] = False,
**kwargs,
) -> Union[TFSequenceClassifierOutput, Tuple[tf.Tensor]]:
r"""
labels (:obj:`tf.Tensor` or :obj:`np.ndarray` of shape :obj:`(batch_size,)`, `optional`):
Labels for computing the sequence classification/regression loss. Indices should be in :obj:`[0, ...,
config.num_labels - 1]`. If :obj:`config.num_labels == 1` a regression loss is computed (Mean-Square loss),
If :obj:`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
"""
inputs = input_processing(
func=self.call,
config=self.config,
input_ids=input_ids,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
head_mask=head_mask,
inputs_embeds=inputs_embeds,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
labels=labels,
training=training,
kwargs_call=kwargs,
)
outputs = self.roformer(
input_ids=inputs["input_ids"],
attention_mask=inputs["attention_mask"],
token_type_ids=inputs["token_type_ids"],
head_mask=inputs["head_mask"],
inputs_embeds=inputs["inputs_embeds"],
output_attentions=inputs["output_attentions"],
output_hidden_states=inputs["output_hidden_states"],
return_dict=inputs["return_dict"],
training=inputs["training"],
)
logits = self.classifier(hidden_states=outputs[0], training=inputs["training"])
loss = None if inputs["labels"] is None else self.compute_loss(labels=inputs["labels"], logits=logits)
if not inputs["return_dict"]:
output = (logits,) + outputs[1:]
return ((loss,) + output) if loss is not None else output
return TFSequenceClassifierOutput(
loss=loss,
logits=logits,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
def serving_output(self, output: TFSequenceClassifierOutput) -> TFSequenceClassifierOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFSequenceClassifierOutput(logits=output.logits, hidden_states=hs, attentions=attns)
@add_start_docstrings(
"""
RoFormer Model with a multiple choice classification head on top (a linear layer on top of the pooled output and a
softmax) e.g. for RocStories/SWAG tasks.
""",
ROFORMER_START_DOCSTRING,
)
class TFRoFormerForMultipleChoice(TFRoFormerPreTrainedModel, TFMultipleChoiceLoss):
def __init__(self, config: RoFormerConfig, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
self.roformer = TFRoFormerMainLayer(config, name="roformer")
self.sequence_summary = TFSequenceSummary(config, config.initializer_range, name="sequence_summary")
self.classifier = tf.keras.layers.Dense(
units=1, kernel_initializer=get_initializer(config.initializer_range), name="classifier"
)
@property
def dummy_inputs(self) -> Dict[str, tf.Tensor]:
"""
Dummy inputs to build the network.
Returns:
tf.Tensor with dummy inputs
"""
return {"input_ids": tf.constant(MULTIPLE_CHOICE_DUMMY_INPUTS)}
@add_start_docstrings_to_model_forward(
ROFORMER_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length")
)
@add_code_sample_docstrings(
tokenizer_class=_TOKENIZER_FOR_DOC,
checkpoint=_CHECKPOINT_FOR_DOC,
output_type=TFMultipleChoiceModelOutput,
config_class=_CONFIG_FOR_DOC,
)
def call(
self,
input_ids: Optional[TFModelInputType] = None,
attention_mask: Optional[Union[np.ndarray, tf.Tensor]] = None,
token_type_ids: Optional[Union[np.ndarray, tf.Tensor]] = None,
head_mask: Optional[Union[np.ndarray, tf.Tensor]] = None,
inputs_embeds: Optional[Union[np.ndarray, tf.Tensor]] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
training: Optional[bool] = False,
**kwargs,
) -> Union[TFMultipleChoiceModelOutput, Tuple[tf.Tensor]]:
r"""
labels (:obj:`tf.Tensor` or :obj:`np.ndarray` of shape :obj:`(batch_size,)`, `optional`):
Labels for computing the multiple choice classification loss. Indices should be in ``[0, ...,
num_choices]`` where :obj:`num_choices` is the size of the second dimension of the input tensors. (See
:obj:`input_ids` above)
"""
inputs = input_processing(
func=self.call,
config=self.config,
input_ids=input_ids,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
head_mask=head_mask,
inputs_embeds=inputs_embeds,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
labels=labels,
training=training,
kwargs_call=kwargs,
)
if inputs["input_ids"] is not None:
num_choices = shape_list(inputs["input_ids"])[1]
seq_length = shape_list(inputs["input_ids"])[2]
else:
num_choices = shape_list(inputs["inputs_embeds"])[1]
seq_length = shape_list(inputs["inputs_embeds"])[2]
flat_input_ids = (
tf.reshape(tensor=inputs["input_ids"], shape=(-1, seq_length)) if inputs["input_ids"] is not None else None
)
flat_attention_mask = (
tf.reshape(tensor=inputs["attention_mask"], shape=(-1, seq_length))
if inputs["attention_mask"] is not None
else None
)
flat_token_type_ids = (
tf.reshape(tensor=inputs["token_type_ids"], shape=(-1, seq_length))
if inputs["token_type_ids"] is not None
else None
)
flat_inputs_embeds = (
tf.reshape(tensor=inputs["inputs_embeds"], shape=(-1, seq_length, shape_list(inputs["inputs_embeds"])[3]))
if inputs["inputs_embeds"] is not None
else None
)
outputs = self.roformer(
input_ids=flat_input_ids,
attention_mask=flat_attention_mask,
token_type_ids=flat_token_type_ids,
head_mask=inputs["head_mask"],
inputs_embeds=flat_inputs_embeds,
output_attentions=inputs["output_attentions"],
output_hidden_states=inputs["output_hidden_states"],
return_dict=inputs["return_dict"],
training=inputs["training"],
)
logits = self.sequence_summary(inputs=outputs[0], training=inputs["training"])
logits = self.classifier(inputs=logits)
reshaped_logits = tf.reshape(tensor=logits, shape=(-1, num_choices))
loss = None if inputs["labels"] is None else self.compute_loss(labels=inputs["labels"], logits=reshaped_logits)
if not inputs["return_dict"]:
output = (reshaped_logits,) + outputs[1:]
return ((loss,) + output) if loss is not None else output
return TFMultipleChoiceModelOutput(
loss=loss,
logits=reshaped_logits,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
@tf.function(
input_signature=[
{
"input_ids": tf.TensorSpec((None, None, None), tf.int32, name="input_ids"),
"attention_mask": tf.TensorSpec((None, None, None), tf.int32, name="attention_mask"),
"token_type_ids": tf.TensorSpec((None, None, None), tf.int32, name="token_type_ids"),
}
]
)
def serving(self, inputs: Dict[str, tf.Tensor]) -> TFMultipleChoiceModelOutput:
output = self.call(input_ids=inputs)
return self.serving_output(output)
def serving_output(self, output: TFMultipleChoiceModelOutput) -> TFMultipleChoiceModelOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFMultipleChoiceModelOutput(logits=output.logits, hidden_states=hs, attentions=attns)
@add_start_docstrings(
"""
RoFormer Model with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for
Named-Entity-Recognition (NER) tasks.
""",
ROFORMER_START_DOCSTRING,
)
class TFRoFormerForTokenClassification(TFRoFormerPreTrainedModel, TFTokenClassificationLoss):
def __init__(self, config: RoFormerConfig, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
self.num_labels = config.num_labels
self.roformer = TFRoFormerMainLayer(config, name="roformer")
self.dropout = tf.keras.layers.Dropout(rate=config.hidden_dropout_prob)
self.classifier = tf.keras.layers.Dense(
units=config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="classifier"
)
@add_start_docstrings_to_model_forward(ROFORMER_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
@add_code_sample_docstrings(
tokenizer_class=_TOKENIZER_FOR_DOC,
checkpoint=_CHECKPOINT_FOR_DOC,
output_type=TFTokenClassifierOutput,
config_class=_CONFIG_FOR_DOC,
)
def call(
self,
input_ids: Optional[TFModelInputType] = None,
attention_mask: Optional[Union[np.ndarray, tf.Tensor]] = None,
token_type_ids: Optional[Union[np.ndarray, tf.Tensor]] = None,
head_mask: Optional[Union[np.ndarray, tf.Tensor]] = None,
inputs_embeds: Optional[Union[np.ndarray, tf.Tensor]] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
training: Optional[bool] = False,
**kwargs,
) -> Union[TFTokenClassifierOutput, Tuple[tf.Tensor]]:
r"""
labels (:obj:`tf.Tensor` or :obj:`np.ndarray` of shape :obj:`(batch_size, sequence_length)`, `optional`):
Labels for computing the token classification loss. Indices should be in ``[0, ..., config.num_labels -
1]``.
"""
inputs = input_processing(
func=self.call,
config=self.config,
input_ids=input_ids,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
head_mask=head_mask,
inputs_embeds=inputs_embeds,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
labels=labels,
training=training,
kwargs_call=kwargs,
)
outputs = self.roformer(
input_ids=inputs["input_ids"],
attention_mask=inputs["attention_mask"],
token_type_ids=inputs["token_type_ids"],
head_mask=inputs["head_mask"],
inputs_embeds=inputs["inputs_embeds"],
output_attentions=inputs["output_attentions"],
output_hidden_states=inputs["output_hidden_states"],
return_dict=inputs["return_dict"],
training=inputs["training"],
)
sequence_output = outputs[0]
sequence_output = self.dropout(inputs=sequence_output, training=inputs["training"])
logits = self.classifier(inputs=sequence_output)
loss = None if inputs["labels"] is None else self.compute_loss(labels=inputs["labels"], logits=logits)
if not inputs["return_dict"]:
output = (logits,) + outputs[1:]
return ((loss,) + output) if loss is not None else output
return TFTokenClassifierOutput(
loss=loss,
logits=logits,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
def serving_output(self, output: TFTokenClassifierOutput) -> TFTokenClassifierOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFTokenClassifierOutput(logits=output.logits, hidden_states=hs, attentions=attns)
@add_start_docstrings(
"""
RoFormer Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear
layer on top of the hidden-states output to compute `span start logits` and `span end logits`).
""",
ROFORMER_START_DOCSTRING,
)
class TFRoFormerForQuestionAnswering(TFRoFormerPreTrainedModel, TFQuestionAnsweringLoss):
def __init__(self, config: RoFormerConfig, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
self.num_labels = config.num_labels
self.roformer = TFRoFormerMainLayer(config, name="roformer")
self.qa_outputs = tf.keras.layers.Dense(
units=config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="qa_outputs"
)
@add_start_docstrings_to_model_forward(ROFORMER_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
@add_code_sample_docstrings(
tokenizer_class=_TOKENIZER_FOR_DOC,
checkpoint=_CHECKPOINT_FOR_DOC,
output_type=TFQuestionAnsweringModelOutput,
config_class=_CONFIG_FOR_DOC,
)
def call(
self,
input_ids: Optional[TFModelInputType] = None,
attention_mask: Optional[Union[np.ndarray, tf.Tensor]] = None,
token_type_ids: Optional[Union[np.ndarray, tf.Tensor]] = None,
head_mask: Optional[Union[np.ndarray, tf.Tensor]] = None,
inputs_embeds: Optional[Union[np.ndarray, tf.Tensor]] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
start_positions: Optional[Union[np.ndarray, tf.Tensor]] = None,
end_positions: Optional[Union[np.ndarray, tf.Tensor]] = None,
training: Optional[bool] = False,
**kwargs,
) -> Union[TFQuestionAnsweringModelOutput, Tuple[tf.Tensor]]:
r"""
start_positions (:obj:`tf.Tensor` or :obj:`np.ndarray` of shape :obj:`(batch_size,)`, `optional`):
Labels for position (index) of the start of the labelled span for computing the token classification loss.
Positions are clamped to the length of the sequence (:obj:`sequence_length`). Position outside of the
sequence are not taken into account for computing the loss.
end_positions (:obj:`tf.Tensor` or :obj:`np.ndarray` of shape :obj:`(batch_size,)`, `optional`):
Labels for position (index) of the end of the labelled span for computing the token classification loss.
Positions are clamped to the length of the sequence (:obj:`sequence_length`). Position outside of the
sequence are not taken into account for computing the loss.
"""
inputs = input_processing(
func=self.call,
config=self.config,
input_ids=input_ids,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
head_mask=head_mask,
inputs_embeds=inputs_embeds,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
start_positions=start_positions,
end_positions=end_positions,
training=training,
kwargs_call=kwargs,
)
outputs = self.roformer(
input_ids=inputs["input_ids"],
attention_mask=inputs["attention_mask"],
token_type_ids=inputs["token_type_ids"],
head_mask=inputs["head_mask"],
inputs_embeds=inputs["inputs_embeds"],
output_attentions=inputs["output_attentions"],
output_hidden_states=inputs["output_hidden_states"],
return_dict=inputs["return_dict"],
training=inputs["training"],
)
sequence_output = outputs[0]
logits = self.qa_outputs(inputs=sequence_output)
start_logits, end_logits = tf.split(value=logits, num_or_size_splits=2, axis=-1)
start_logits = tf.squeeze(input=start_logits, axis=-1)
end_logits = tf.squeeze(input=end_logits, axis=-1)
loss = None
if inputs["start_positions"] is not None and inputs["end_positions"] is not None:
labels = {"start_position": inputs["start_positions"]}
labels["end_position"] = inputs["end_positions"]
loss = self.compute_loss(labels=labels, logits=(start_logits, end_logits))
if not inputs["return_dict"]:
output = (start_logits, end_logits) + outputs[2:]
return ((loss,) + output) if loss is not None else output
return TFQuestionAnsweringModelOutput(
loss=loss,
start_logits=start_logits,
end_logits=end_logits,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
def serving_output(self, output: TFQuestionAnsweringModelOutput) -> TFQuestionAnsweringModelOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFQuestionAnsweringModelOutput(
start_logits=output.start_logits, end_logits=output.end_logits, hidden_states=hs, attentions=attns
)
src/transformers/models/roformer/tokenization_roformer.py 0 → 100644
View file @ 206f06f2
# coding=utf-8
# Copyright 2021 The HuggingFace Inc. team. 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.
"""Tokenization classes for RoFormer."""
import collections
import os
from typing import List, Optional, Tuple
from ...tokenization_utils import PreTrainedTokenizer
from ...utils import logging
from ..bert.tokenization_bert import BasicTokenizer, WordpieceTokenizer, load_vocab
logger = logging.get_logger(__name__)
VOCAB_FILES_NAMES = {"vocab_file": "vocab.txt"}
PRETRAINED_VOCAB_FILES_MAP = {
"vocab_file": {
"junnyu/roformer_chinese_small": "https://huggingface.co/junnyu/roformer_chinese_small/resolve/main/vocab.txt",
"junnyu/roformer_chinese_base": "https://huggingface.co/junnyu/roformer_chinese_base/resolve/main/vocab.txt",
}
}
PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {"junnyu/roformer_chinese_small": 1536, "junnyu/roformer_chinese_base": 1536}
PRETRAINED_INIT_CONFIGURATION = {
"junnyu/roformer_chinese_small": {"do_lower_case": True},
"junnyu/roformer_chinese_base": {"do_lower_case": True},
}
class RoFormerTokenizer(PreTrainedTokenizer):
r"""
Construct a RoFormer tokenizer. Based on `Rust Jieba <https://pypi.org/project/rjieba/>`.
This tokenizer inherits from :class:`~transformers.PreTrainedTokenizer` which contains most of the main methods.
Users should refer to this superclass for more information regarding those methods.
Args:
vocab_file (:obj:`str`):
File containing the vocabulary.
do_lower_case (:obj:`bool`, `optional`, defaults to :obj:`True`):
Whether or not to lowercase the input when tokenizing.
do_basic_tokenize (:obj:`bool`, `optional`, defaults to :obj:`True`):
Whether or not to do basic tokenization before WordPiece.
never_split (:obj:`Iterable`, `optional`):
Collection of tokens which will never be split during tokenization. Only has an effect when
:obj:`do_basic_tokenize=True`
unk_token (:obj:`str`, `optional`, defaults to :obj:`"[UNK]"`):
The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
token instead.
sep_token (:obj:`str`, `optional`, defaults to :obj:`"[SEP]"`):
The separator token, which is used when building a sequence from multiple sequences, e.g. two sequences for
sequence classification or for a text and a question for question answering. It is also used as the last
token of a sequence built with special tokens.
pad_token (:obj:`str`, `optional`, defaults to :obj:`"[PAD]"`):
The token used for padding, for example when batching sequences of different lengths.
cls_token (:obj:`str`, `optional`, defaults to :obj:`"[CLS]"`):
The classifier token which is used when doing sequence classification (classification of the whole sequence
instead of per-token classification). It is the first token of the sequence when built with special tokens.
mask_token (:obj:`str`, `optional`, defaults to :obj:`"[MASK]"`):
The token used for masking values. This is the token used when training this model with masked language
modeling. This is the token which the model will try to predict.
tokenize_chinese_chars (:obj:`bool`, `optional`, defaults to :obj:`True`):
Whether or not to tokenize Chinese characters.
This should likely be deactivated for Japanese (see this `issue
<https://github.com/huggingface/transformers/issues/328>`__).
strip_accents: (:obj:`bool`, `optional`):
Whether or not to strip all accents. If this option is not specified, then it will be determined by the
value for :obj:`lowercase` (as in the original BERT).
Example::
>>> from transformers import RoFormerTokenizer
>>> tokenizer = RoFormerTokenizer.from_pretrained('junnyu/roformer_chinese_base')
>>> tokenizer.tokenize("今天天气非常好。")
# ['今', '天', '天', '气', '非常', '好', '。']
"""
vocab_files_names = VOCAB_FILES_NAMES
pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
pretrained_init_configuration = PRETRAINED_INIT_CONFIGURATION
def __init__(
self,
vocab_file,
do_lower_case=True,
do_basic_tokenize=True,
never_split=None,
unk_token="[UNK]",
sep_token="[SEP]",
pad_token="[PAD]",
cls_token="[CLS]",
mask_token="[MASK]",
tokenize_chinese_chars=True,
strip_accents=None,
**kwargs
):
super().__init__(
do_lower_case=do_lower_case,
do_basic_tokenize=do_basic_tokenize,
never_split=never_split,
unk_token=unk_token,
sep_token=sep_token,
pad_token=pad_token,
cls_token=cls_token,
mask_token=mask_token,
tokenize_chinese_chars=tokenize_chinese_chars,
strip_accents=strip_accents,
**kwargs,
)
if not os.path.isfile(vocab_file):
raise ValueError(
f"Can't find a vocabulary file at path '{vocab_file}'. To load the vocabulary from a Google pretrained "
"model use `tokenizer = RoFormerTokenizer.from_pretrained(PRETRAINED_MODEL_NAME)`"
)
self.vocab = load_vocab(vocab_file)
self.ids_to_tokens = collections.OrderedDict([(ids, tok) for tok, ids in self.vocab.items()])
self.do_basic_tokenize = do_basic_tokenize
if do_basic_tokenize:
self.basic_tokenizer = BasicTokenizer(
do_lower_case=do_lower_case,
never_split=never_split,
tokenize_chinese_chars=tokenize_chinese_chars,
strip_accents=strip_accents,
)
self.wordpiece_tokenizer = WordpieceTokenizer(vocab=self.vocab, unk_token=self.unk_token)
try:
import rjieba
except ImportError:
raise ImportError(
"You need to install rjieba to use RoFormerTokenizer."
"See https://pypi.org/project/rjieba/ for installation."
)
self.jieba = rjieba
@property
def do_lower_case(self):
return self.basic_tokenizer.do_lower_case
@property
def vocab_size(self):
return len(self.vocab)
def __getstate__(self):
state = self.__dict__.copy()
state["jieba"] = None
return state
def __setstate__(self, d):
self.__dict__ = d
try:
import rjieba
except ImportError:
raise ImportError(
"You need to install rjieba to use RoFormerTokenizer."
"See https://pypi.org/project/rjieba/ for installation."
)
self.jieba = rjieba
def get_vocab(self):
return dict(self.vocab, **self.added_tokens_encoder)
def _tokenize(self, text, use_jieba=True):
split_tokens = []
if use_jieba:
for wholword in self.jieba.cut(text, False):
if wholword in self.vocab:
split_tokens.append(wholword)
else:
# use bert tokenizer to _tokenize
char_list = self._tokenize(wholword, use_jieba=False)
split_tokens.extend(char_list)
else:
if self.do_basic_tokenize:
for token in self.basic_tokenizer.tokenize(text, never_split=self.all_special_tokens):
# If the token is part of the never_split set
if token in self.basic_tokenizer.never_split:
split_tokens.append(token)
else:
split_tokens += self.wordpiece_tokenizer.tokenize(token)
else:
split_tokens = self.wordpiece_tokenizer.tokenize(text)
return split_tokens
def _convert_token_to_id(self, token):
"""Converts a token (str) in an id using the vocab."""
return self.vocab.get(token, self.vocab.get(self.unk_token))
def _convert_id_to_token(self, index):
"""Converts an index (integer) in a token (str) using the vocab."""
return self.ids_to_tokens.get(index, self.unk_token)
def convert_tokens_to_string(self, tokens):
"""Converts a sequence of tokens (string) in a single string."""
out_string = " ".join(tokens).replace(" ##", "").strip()
return out_string
def build_inputs_with_special_tokens(
self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
) -> List[int]:
"""
Build model inputs from a sequence or a pair of sequence for sequence classification tasks by concatenating and
adding special tokens. A RoFormer sequence has the following format:
- single sequence: ``[CLS] X [SEP]``
- pair of sequences: ``[CLS] A [SEP] B [SEP]``
Args:
token_ids_0 (:obj:`List[int]`):
List of IDs to which the special tokens will be added.
token_ids_1 (:obj:`List[int]`, `optional`):
Optional second list of IDs for sequence pairs.
Returns:
:obj:`List[int]`: List of `input IDs <../glossary.html#input-ids>`__ with the appropriate special tokens.
"""
if token_ids_1 is None:
return [self.cls_token_id] + token_ids_0 + [self.sep_token_id]
cls = [self.cls_token_id]
sep = [self.sep_token_id]
return cls + token_ids_0 + sep + token_ids_1 + sep
def get_special_tokens_mask(
self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None, already_has_special_tokens: bool = False
) -> List[int]:
"""
Retrieve sequence ids from a token list that has no special tokens added. This method is called when adding
special tokens using the tokenizer ``prepare_for_model`` method.
Args:
token_ids_0 (:obj:`List[int]`):
List of IDs.
token_ids_1 (:obj:`List[int]`, `optional`):
Optional second list of IDs for sequence pairs.
already_has_special_tokens (:obj:`bool`, `optional`, defaults to :obj:`False`):
Whether or not the token list is already formatted with special tokens for the model.
Returns:
:obj:`List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
"""
if already_has_special_tokens:
return super().get_special_tokens_mask(
token_ids_0=token_ids_0, token_ids_1=token_ids_1, already_has_special_tokens=True
)
if token_ids_1 is not None:
return [1] + ([0] * len(token_ids_0)) + [1] + ([0] * len(token_ids_1)) + [1]
return [1] + ([0] * len(token_ids_0)) + [1]
def create_token_type_ids_from_sequences(
self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
) -> List[int]:
"""
Create a mask from the two sequences passed to be used in a sequence-pair classification task. A RoFormer
sequence pair mask has the following format:
::
0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1
| first sequence | second sequence |
If :obj:`token_ids_1` is :obj:`None`, this method only returns the first portion of the mask (0s).
Args:
token_ids_0 (:obj:`List[int]`):
List of IDs.
token_ids_1 (:obj:`List[int]`, `optional`):
Optional second list of IDs for sequence pairs.
Returns:
:obj:`List[int]`: List of `token type IDs <../glossary.html#token-type-ids>`_ according to the given
sequence(s).
"""
sep = [self.sep_token_id]
cls = [self.cls_token_id]
if token_ids_1 is None:
return len(cls + token_ids_0 + sep) * [0]
return len(cls + token_ids_0 + sep) * [0] + len(token_ids_1 + sep) * [1]
def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
index = 0
if os.path.isdir(save_directory):
vocab_file = os.path.join(
save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
)
else:
vocab_file = (filename_prefix + "-" if filename_prefix else "") + save_directory
with open(vocab_file, "w", encoding="utf-8") as writer:
for token, token_index in sorted(self.vocab.items(), key=lambda kv: kv[1]):
if index != token_index:
logger.warning(
f"Saving vocabulary to {vocab_file}: vocabulary indices are not consecutive."
" Please check that the vocabulary is not corrupted!"
)
index = token_index
writer.write(token + "\n")
index += 1
return (vocab_file,)
src/transformers/models/roformer/tokenization_roformer_fast.py 0 → 100644
View file @ 206f06f2
# coding=utf-8
# Copyright 2021 The HuggingFace Inc. team. 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.
"""Tokenization classes for RoFormer."""
import json
from typing import List, Optional, Tuple
from tokenizers import normalizers
from tokenizers.pre_tokenizers import BertPreTokenizer, PreTokenizer
from ...tokenization_utils_fast import PreTrainedTokenizerFast
from ...utils import logging
from .tokenization_roformer import RoFormerTokenizer
from .tokenization_utils import JiebaPreTokenizer
logger = logging.get_logger(__name__)
VOCAB_FILES_NAMES = {"vocab_file": "vocab.txt"}
PRETRAINED_VOCAB_FILES_MAP = {
"vocab_file": {
"junnyu/roformer_chinese_small": "https://huggingface.co/junnyu/roformer_chinese_small/resolve/main/vocab.txt",
"junnyu/roformer_chinese_base": "https://huggingface.co/junnyu/roformer_chinese_base/resolve/main/vocab.txt",
}
}
PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {"junnyu/roformer_chinese_small": 1536, "junnyu/roformer_chinese_base": 1536}
PRETRAINED_INIT_CONFIGURATION = {
"junnyu/roformer_chinese_small": {"do_lower_case": True},
"junnyu/roformer_chinese_base": {"do_lower_case": True},
}
class RoFormerTokenizerFast(PreTrainedTokenizerFast):
r"""
Construct a "fast" RoFormer tokenizer (backed by HuggingFace's `tokenizers` library).
:class:`~transformers.RoFormerTokenizerFast` is almost identical to :class:`~transformers.BertTokenizerFast` and
runs end-to-end tokenization: punctuation splitting and wordpiece. There are some difference between them when
tokenizing Chinese.
This tokenizer inherits from :class:`~transformers.PreTrainedTokenizerFast` which contains most of the main
methods. Users should refer to this superclass for more information regarding those methods.
Example::
>>> from transformers import RoFormerTokenizerFast
>>> tokenizer = RoFormerTokenizerFast.from_pretrained('junnyu/roformer_chinese_base')
>>> tokenizer.tokenize("今天天气非常好。")
# ['今', '天', '天', '气', '非常', '好', '。']
"""
vocab_files_names = VOCAB_FILES_NAMES
pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
pretrained_init_configuration = PRETRAINED_INIT_CONFIGURATION
slow_tokenizer_class = RoFormerTokenizer
def __init__(
self,
vocab_file,
tokenizer_file=None,
do_lower_case=True,
unk_token="[UNK]",
sep_token="[SEP]",
pad_token="[PAD]",
cls_token="[CLS]",
mask_token="[MASK]",
tokenize_chinese_chars=True,
strip_accents=None,
**kwargs
):
super().__init__(
vocab_file,
tokenizer_file=tokenizer_file,
do_lower_case=do_lower_case,
unk_token=unk_token,
sep_token=sep_token,
pad_token=pad_token,
cls_token=cls_token,
mask_token=mask_token,
tokenize_chinese_chars=tokenize_chinese_chars,
strip_accents=strip_accents,
**kwargs,
)
pre_tok_state = json.loads(self.backend_tokenizer.normalizer.__getstate__())
if (
pre_tok_state.get("lowercase", do_lower_case) != do_lower_case
or pre_tok_state.get("strip_accents", strip_accents) != strip_accents
):
pre_tok_class = getattr(normalizers, pre_tok_state.pop("type"))
pre_tok_state["lowercase"] = do_lower_case
pre_tok_state["strip_accents"] = strip_accents
self.backend_tokenizer.normalizer = pre_tok_class(**pre_tok_state)
self.do_lower_case = do_lower_case
def __getstate__(self):
state = self.__dict__.copy()
state["_tokenizer"].pre_tokenizer = BertPreTokenizer()
return state
def __setstate__(self, d):
self.__dict__ = d
vocab = self.__dict__["_tokenizer"].get_vocab()
self.__dict__["_tokenizer"].pre_tokenizer = PreTokenizer.custom(JiebaPreTokenizer(vocab))
def build_inputs_with_special_tokens(self, token_ids_0, token_ids_1=None):
"""
Build model inputs from a sequence or a pair of sequence for sequence classification tasks by concatenating and
adding special tokens. A RoFormer sequence has the following format:
- single sequence: ``[CLS] X [SEP]``
- pair of sequences: ``[CLS] A [SEP] B [SEP]``
Args:
token_ids_0 (:obj:`List[int]`):
List of IDs to which the special tokens will be added.
token_ids_1 (:obj:`List[int]`, `optional`):
Optional second list of IDs for sequence pairs.
Returns:
:obj:`List[int]`: List of `input IDs <../glossary.html#input-ids>`__ with the appropriate special tokens.
"""
output = [self.cls_token_id] + token_ids_0 + [self.sep_token_id]
if token_ids_1:
output += token_ids_1 + [self.sep_token_id]
return output
def create_token_type_ids_from_sequences(
self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
) -> List[int]:
"""
Create a mask from the two sequences passed to be used in a sequence-pair classification task. A RoFormer
sequence pair mask has the following format:
::
0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1
| first sequence | second sequence |
If :obj:`token_ids_1` is :obj:`None`, this method only returns the first portion of the mask (0s).
Args:
token_ids_0 (:obj:`List[int]`):
List of IDs.
token_ids_1 (:obj:`List[int]`, `optional`):
Optional second list of IDs for sequence pairs.
Returns:
:obj:`List[int]`: List of `token type IDs <../glossary.html#token-type-ids>`_ according to the given
sequence(s).
"""
sep = [self.sep_token_id]
cls = [self.cls_token_id]
if token_ids_1 is None:
return len(cls + token_ids_0 + sep) * [0]
return len(cls + token_ids_0 + sep) * [0] + len(token_ids_1 + sep) * [1]
def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
files = self._tokenizer.model.save(save_directory, name=filename_prefix)
return tuple(files)
def save_pretrained(
self,
save_directory,
legacy_format=None,
filename_prefix=None,
push_to_hub=False,
**kwargs,
):
self.backend_tokenizer.pre_tokenizer = BertPreTokenizer()
return super().save_pretrained(save_directory, legacy_format, filename_prefix, push_to_hub, **kwargs)
src/transformers/models/roformer/tokenization_utils.py 0 → 100644
View file @ 206f06f2
# coding=utf-8
# Copyright 2021 The HuggingFace Inc. team. 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.
"""Tokenization utils for RoFormer."""
from typing import List
from tokenizers import NormalizedString, PreTokenizedString, normalizers
class JiebaPreTokenizer:
def __init__(self, vocab) -> None:
self.vocab = vocab
self.normalizers = normalizers.BertNormalizer(
clean_text=False,
handle_chinese_chars=True,
strip_accents=False,
lowercase=False,
)
try:
import rjieba
except ImportError:
raise ImportError(
"You need to install rjieba to use RoFormerTokenizer."
"See https://pypi.org/project/rjieba/ for installation."
)
self.jieba = rjieba
def jieba_split(self, i: int, normalized_string: NormalizedString) -> List[NormalizedString]:
splits = []
# this code slice normalized_string is too slow (6s) but test_alignement_methods can pass
# for token, start, end in self.jieba.tokenize(str(normalized_string), hmm=False):
# if token in self.vocab:
# splits.append(normalized_string.slice((start, end)))
# else:
# token_list = self.normalizers.normalize_str(token).split()
# for token in token_list:
# if token:
# end = start + len(token)
# splits.append(normalized_string.slice((start, end)))
# start = end
# this code test_alignement_methods can't pass but fast (300ms)
for token in self.jieba.cut(str(normalized_string), False):
if token in self.vocab:
splits.append(NormalizedString(token))
else:
token_list = self.normalizers.normalize_str(token).split()
for token in token_list:
if token:
splits.append(NormalizedString(token))
return splits
def pre_tokenize(self, pretok: PreTokenizedString):
pretok.split(self.jieba_split)
src/transformers/utils/dummy_pt_objects.py
View file @ 206f06f2
......@@ -2553,6 +2553,86 @@ class RobertaModel:
requires_backends(self, ["torch"])
ROFORMER_PRETRAINED_MODEL_ARCHIVE_LIST = None
class RoFormerForCausalLM:
def __init__(self, *args, **kwargs):
requires_backends(self, ["torch"])
class RoFormerForMaskedLM:
def __init__(self, *args, **kwargs):
requires_backends(self, ["torch"])
@classmethod
def from_pretrained(self, *args, **kwargs):
requires_backends(self, ["torch"])
class RoFormerForMultipleChoice:
def __init__(self, *args, **kwargs):
requires_backends(self, ["torch"])
@classmethod
def from_pretrained(self, *args, **kwargs):
requires_backends(self, ["torch"])
class RoFormerForQuestionAnswering:
def __init__(self, *args, **kwargs):
requires_backends(self, ["torch"])
@classmethod
def from_pretrained(self, *args, **kwargs):
requires_backends(self, ["torch"])
class RoFormerForSequenceClassification:
def __init__(self, *args, **kwargs):
requires_backends(self, ["torch"])
@classmethod
def from_pretrained(self, *args, **kwargs):
requires_backends(self, ["torch"])
class RoFormerForTokenClassification:
def __init__(self, *args, **kwargs):
requires_backends(self, ["torch"])
@classmethod
def from_pretrained(self, *args, **kwargs):
requires_backends(self, ["torch"])
class RoFormerLayer:
def __init__(self, *args, **kwargs):
requires_backends(self, ["torch"])
class RoFormerModel:
def __init__(self, *args, **kwargs):
requires_backends(self, ["torch"])
@classmethod
def from_pretrained(self, *args, **kwargs):
requires_backends(self, ["torch"])
class RoFormerPreTrainedModel:
def __init__(self, *args, **kwargs):
requires_backends(self, ["torch"])
@classmethod
def from_pretrained(self, *args, **kwargs):
requires_backends(self, ["torch"])
def load_tf_weights_in_roformer(*args, **kwargs):
requires_backends(load_tf_weights_in_roformer, ["torch"])
SPEECH_TO_TEXT_PRETRAINED_MODEL_ARCHIVE_LIST = None
......
src/transformers/utils/dummy_tf_objects.py
View file @ 206f06f2
......@@ -1479,6 +1479,82 @@ class TFRobertaPreTrainedModel:
requires_backends(self, ["tf"])
TF_ROFORMER_PRETRAINED_MODEL_ARCHIVE_LIST = None
class TFRoFormerForCausalLM:
def __init__(self, *args, **kwargs):
requires_backends(self, ["tf"])
class TFRoFormerForMaskedLM:
def __init__(self, *args, **kwargs):
requires_backends(self, ["tf"])
@classmethod
def from_pretrained(self, *args, **kwargs):
requires_backends(self, ["tf"])
class TFRoFormerForMultipleChoice:
def __init__(self, *args, **kwargs):
requires_backends(self, ["tf"])
@classmethod
def from_pretrained(self, *args, **kwargs):
requires_backends(self, ["tf"])
class TFRoFormerForQuestionAnswering:
def __init__(self, *args, **kwargs):
requires_backends(self, ["tf"])
@classmethod
def from_pretrained(self, *args, **kwargs):
requires_backends(self, ["tf"])
class TFRoFormerForSequenceClassification:
def __init__(self, *args, **kwargs):
requires_backends(self, ["tf"])
@classmethod
def from_pretrained(self, *args, **kwargs):
requires_backends(self, ["tf"])
class TFRoFormerForTokenClassification:
def __init__(self, *args, **kwargs):
requires_backends(self, ["tf"])
@classmethod
def from_pretrained(self, *args, **kwargs):
requires_backends(self, ["tf"])
class TFRoFormerLayer:
def __init__(self, *args, **kwargs):
requires_backends(self, ["tf"])
class TFRoFormerModel:
def __init__(self, *args, **kwargs):
requires_backends(self, ["tf"])
@classmethod
def from_pretrained(self, *args, **kwargs):
requires_backends(self, ["tf"])
class TFRoFormerPreTrainedModel:
def __init__(self, *args, **kwargs):
requires_backends(self, ["tf"])
@classmethod
def from_pretrained(self, *args, **kwargs):
requires_backends(self, ["tf"])
TF_T5_PRETRAINED_MODEL_ARCHIVE_LIST = None
......
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