Integrate fast tokenizers library inside transformers (#2674)
* Implemented fast version of tokenizers Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Bumped tokenizers version requirements to latest 0.2.1 Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Added matching tests Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Matching OpenAI GPT tokenization ! Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Matching GPT2 on tokenizers Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Expose add_prefix_space as constructor parameter for GPT2 Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Matching Roberta tokenization ! Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Removed fast implementation of CTRL. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Binding TransformerXL tokenizers to Rust. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Updating tests accordingly. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Added tokenizers as top-level modules. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Black & isort. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Rename LookupTable to WordLevel to match Rust side. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Black. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Use "fast" suffix instead of "ru" for rust tokenizers implementations. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Introduce tokenize() method on fast tokenizers. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * encode_plus dispatchs to batch_encode_plus Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * batch_encode_plus now dispatchs to encode if there is only one input element. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Bind all the encode_plus parameter to the forwarded batch_encode_plus call. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Bump tokenizers dependency to 0.3.0 Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Formatting. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Fix tokenization_auto with support for new (python, fast) mapping schema. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Give correct fixtures path in test_tokenization_fast.py for the CLI. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Expose max_len_ properties on BertTokenizerFast Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Move max_len_ properties to PreTrainedTokenizerFast and override in specific subclasses. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * _convert_encoding should keep the batch axis tensor if only one sample in the batch. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Add warning message for RobertaTokenizerFast if used for MLM. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Added use_fast (bool) parameter on AutoTokenizer.from_pretrained(). This allows to easily enable/disable Rust-based tokenizer instantiation. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Let's tokenizers handle all the truncation and padding stuff. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Allow to provide tokenizer arguments during pipeline creation. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Update test_fill_mask pipeline to not use fast tokenizers. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Fix too much parameters for convert_encoding. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * When enabling padding, max_length should be set to None. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Avoid returning nested tensors of length 1 when calling encode_plus Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Ensure output is padded when return_tensor is not None. Tensor creation requires the inital list input to be of the exact same size. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Disable transfoxl unittest if pytorch is not available (required to load the model) Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * encode_plus should not remove the leading batch axis if return_tensor is set Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Temporary disable fast tokenizers on QA pipelines. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Fix formatting issues. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Update tokenizers to 0.4.0 * Update style * Enable truncation + stride unit test on fast tokenizers. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Add unittest ensuring special_tokens set match between Python and Rust. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Ensure special_tokens are correctly set during construction. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Give more warning feedback to the user in case of padding without pad_token. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * quality & format. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Added possibility to add a single token as str Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Added unittest for add_tokens and add_special_tokens on fast tokenizers. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Fix rebase mismatch on pipelines qa default model. QA requires cased input while the tokenizers would be uncased. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Addressing review comment: Using offset mapping relative to the original string + unittest. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Addressing review comment: save_vocabulary requires folder and file name Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Addressing review comment: Simplify import for Bert. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Addressing review comment: truncate_and_pad disables padding according to the same heuristic than the one enabling padding. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Addressing review comment: Remove private member access in tokenize() Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Addressing review comment: Bump tokenizers dependency to 0.4.2 Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * format & quality. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Addressing review comment: Use named arguments when applicable. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Addressing review comment: Add Github link to Roberta/GPT2 space issue on masked input. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Addressing review comment: Move max_len_single_sentence / max_len_sentences_pair to PreTrainedTokenizerFast + tests. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Addressing review comment: Relax type checking to include tuple and list object. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Addressing review comment: Document the truncate_and_pad manager behavior. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Raise an exception if return_offsets_mapping is not available with the current tokenizer. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Ensure padding is set on the tokenizers before setting any padding strategy + unittest. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * On pytorch we need to stack tensor to get proper new axis. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Generalize tests to different framework removing hard written return_tensors="..." Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Bump tokenizer dependency for num_special_tokens_to_add Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Overflowing tokens in batch_encode_plus are now stacked over the batch axis. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Improved error message for padding strategy without pad token. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Bumping tokenizers dependency to 0.5.0 for release. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Optimizing convert_encoding around 4x improvement. 🚀 Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * expose pad_to_max_length in encode_plus to avoid duplicating the parameters in kwargs Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Generate a proper overflow_to_sampling_mapping when return_overflowing_tokens is True. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Fix unittests for overflow_to_sampling_mapping not being returned as tensor. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Format & quality. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Remove perfect alignment constraint for Roberta (allowing 1% difference max) Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Triggering final CI Co-authored-by: MOI Anthony <xn1t0x@gmail.com>
This commit is contained in:
Funtowicz Morgan
@@ -17,7 +17,9 @@
|
||||
|
||||
import logging
|
||||
|
||||
from .tokenization_gpt2 import GPT2Tokenizer
|
||||
from tokenizers.processors import RobertaProcessing
|
||||
|
||||
from .tokenization_gpt2 import GPT2Tokenizer, GPT2TokenizerFast
|
||||
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
@@ -163,3 +165,48 @@ class RobertaTokenizer(GPT2Tokenizer):
|
||||
if add_prefix_space and not text[0].isspace():
|
||||
text = " " + text
|
||||
return text
|
||||
|
||||
|
||||
class RobertaTokenizerFast(GPT2TokenizerFast):
|
||||
vocab_files_names = VOCAB_FILES_NAMES
|
||||
pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
|
||||
max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
vocab_file,
|
||||
merges_file,
|
||||
errors="replace",
|
||||
bos_token="<s>",
|
||||
eos_token="</s>",
|
||||
sep_token="</s>",
|
||||
cls_token="<s>",
|
||||
unk_token="<unk>",
|
||||
pad_token="<pad>",
|
||||
mask_token="<mask>",
|
||||
add_prefix_space=False,
|
||||
**kwargs
|
||||
):
|
||||
kwargs.setdefault("pad_token", pad_token)
|
||||
kwargs.setdefault("sep_token", sep_token)
|
||||
kwargs.setdefault("cls_token", cls_token)
|
||||
kwargs.setdefault("mask_token", mask_token)
|
||||
|
||||
super().__init__(
|
||||
vocab_file=vocab_file,
|
||||
merges_file=merges_file,
|
||||
unk_token=unk_token,
|
||||
bos_token=bos_token,
|
||||
eos_token=eos_token,
|
||||
add_prefix_space=add_prefix_space,
|
||||
**kwargs,
|
||||
)
|
||||
|
||||
self.tokenizer._tokenizer.post_processor = RobertaProcessing(
|
||||
(sep_token, self.sep_token_id), (cls_token, self.cls_token_id)
|
||||
)
|
||||
|
||||
# As we override the post_processor post super.__init__ the computed num_added_tokens is wrong in super().
|
||||
# We need to recompute max_len according to the newly register post_processor to get real values.
|
||||
self.max_len_single_sentence = self.max_len - self.num_added_tokens(False) # take into account special tokens
|
||||
self.max_len_sentences_pair = self.max_len - self.num_added_tokens(True) # take into account special tokens
|
||||
|
||||
Reference in New Issue
Block a user