SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

Add Ernie-M Model to huggingface (#21349)

Browse Source 
* config and tokenization(fast too) changed and ErnieEncoder added

* Slow Tokenization Added

* Tokenizer(slow) is now working and Fast Tokenizer removed

* Added Config code

* Added Base Model and utils

* ErnieMModel is now working

* All added except tests

* All tests passed except ErnieUIEM

* All tests passed

* all fixes done

* all fixes done

* fixed MAP

* fixed check_code_quality

* fixed Build PR Documentation issue

* Added changes(comments) and also updated to the latest upstream/main

* Added fixup

* Added # Copied comments

* Added fixup

* Added more comments and some nits

* Added fixup

* Fixed README_hd.md

* Added more fixes

* ErnieMTokenizer (being sentencepiece) protected and other docs edited

* Added code_quality fix

* Fixed for

* Added more fix

* modified AZ

* ernie-m tokenization test added!

* attention mask part fixed(with 0->self.config.pad_token_id)

* applied make fixup
This commit is contained in:
Susnato Dhar
2023-02-15 19:54:56 +05:30
committed by GitHub
parent 40ca13367e
commit 0c9c8472e6
30 changed files with 2343 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

0
tests/models/ernie_m/__init__.py Normal file
View File

304
tests/models/ernie_m/test_modeling_ernie_m.py Normal file
View File

@@ -0,0 +1,304 @@
# coding=utf-8
# Copyright 2023 The HuggingFace Inc. and Baidu 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.
""" Testing suite for the PyTorch ErnieM model. """
import unittest
from transformers import ErnieMConfig, is_torch_available
from transformers.testing_utils import require_torch, slow, torch_device
from ...test_configuration_common import ConfigTester
from ...test_modeling_common import ModelTesterMixin, ids_tensor, random_attention_mask
if is_torch_available():
import torch
from transformers import (
ErnieMForInformationExtraction,
ErnieMForMultipleChoice,
ErnieMForQuestionAnswering,
ErnieMForSequenceClassification,
ErnieMForTokenClassification,
ErnieMModel,
)
from transformers.models.ernie_m.modeling_ernie_m import ERNIE_M_PRETRAINED_MODEL_ARCHIVE_LIST
class ErnieMModelTester:
def __init__(
self,
parent,
batch_size=13,
seq_length=7,
is_training=True,
use_input_mask=True,
use_labels=True,
vocab_size=99,
hidden_size=32,
num_hidden_layers=5,
num_attention_heads=4,
intermediate_size=37,
hidden_act="gelu",
hidden_dropout_prob=0.1,
attention_probs_dropout_prob=0.1,
max_position_embeddings=512,
type_vocab_size=16,
type_sequence_label_size=2,
initializer_range=0.02,
num_labels=3,
num_choices=4,
scope=None,
):
self.parent = parent
self.batch_size = batch_size
self.seq_length = seq_length
self.is_training = is_training
self.use_input_mask = use_input_mask
self.use_labels = use_labels
self.vocab_size = vocab_size
self.hidden_size = hidden_size
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.intermediate_size = intermediate_size
self.hidden_act = hidden_act
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.type_sequence_label_size = type_sequence_label_size
self.initializer_range = initializer_range
self.num_labels = num_labels
self.num_choices = num_choices
self.scope = scope
def prepare_config_and_inputs(self):
input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
input_mask = None
if self.use_input_mask:
input_mask = random_attention_mask([self.batch_size, self.seq_length])
sequence_labels = None
token_labels = None
choice_labels = None
if self.use_labels:
sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
choice_labels = ids_tensor([self.batch_size], self.num_choices)
config = self.get_config()
return config, input_ids, input_mask, sequence_labels, token_labels, choice_labels
def prepare_config_and_inputs_for_uiem(self):
input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
input_mask = None
if self.use_input_mask:
input_mask = random_attention_mask([self.batch_size, self.seq_length])
config = self.get_config()
return config, input_ids, input_mask
def get_config(self):
return ErnieMConfig(
vocab_size=self.vocab_size,
hidden_size=self.hidden_size,
num_hidden_layers=self.num_hidden_layers,
num_attention_heads=self.num_attention_heads,
intermediate_size=self.intermediate_size,
hidden_act=self.hidden_act,
hidden_dropout_prob=self.hidden_dropout_prob,
attention_probs_dropout_prob=self.attention_probs_dropout_prob,
max_position_embeddings=self.max_position_embeddings,
type_vocab_size=self.type_vocab_size,
initializer_range=self.initializer_range,
)
def create_and_check_model(self, config, input_ids, input_mask, sequence_labels, token_labels, choice_labels):
model = ErnieMModel(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, return_dict=True)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_for_question_answering(
self, config, input_ids, input_mask, sequence_labels, token_labels, choice_labels
):
model = ErnieMForQuestionAnswering(config=config)
model.to(torch_device)
model.eval()
result = model(
input_ids,
attention_mask=input_mask,
start_positions=sequence_labels,
end_positions=sequence_labels,
)
self.parent.assertEqual(result.start_logits.shape, (self.batch_size, self.seq_length))
self.parent.assertEqual(result.end_logits.shape, (self.batch_size, self.seq_length))
def create_and_check_for_information_extraction(
self, config, input_ids, input_mask, sequence_labels, token_labels, choice_labels
):
model = ErnieMForInformationExtraction(config=config)
model.to(torch_device)
model.eval()
sequence_labels = torch.ones_like(input_ids, dtype=torch.float32)
result = model(
input_ids,
attention_mask=input_mask,
start_positions=sequence_labels,
end_positions=sequence_labels,
)
self.parent.assertEqual(result.start_logits.shape, (self.batch_size, self.seq_length))
self.parent.assertEqual(result.end_logits.shape, (self.batch_size, self.seq_length))
def create_and_check_for_sequence_classification(
self, config, input_ids, input_mask, sequence_labels, token_labels, choice_labels
):
config.num_labels = self.num_labels
model = ErnieMForSequenceClassification(config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, labels=sequence_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_labels))
def create_and_check_for_token_classification(
self, config, input_ids, input_mask, sequence_labels, token_labels, choice_labels
):
config.num_labels = self.num_labels
model = ErnieMForTokenClassification(config=config)
model.to(torch_device)
model.eval()
input_ids.to(torch_device)
input_mask.to(torch_device)
token_labels.to(torch_device)
result = model(input_ids, attention_mask=input_mask, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.num_labels))
def create_and_check_for_multiple_choice(
self, config, input_ids, input_mask, sequence_labels, token_labels, choice_labels
):
config.num_choices = self.num_choices
model = ErnieMForMultipleChoice(config=config)
model.to(torch_device)
model.eval()
multiple_choice_inputs_ids = input_ids.unsqueeze(1).expand(-1, self.num_choices, -1).contiguous()
multiple_choice_input_mask = input_mask.unsqueeze(1).expand(-1, self.num_choices, -1).contiguous()
result = model(
multiple_choice_inputs_ids,
attention_mask=multiple_choice_input_mask,
labels=choice_labels,
)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_choices))
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
(
config,
input_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
) = config_and_inputs
inputs_dict = {"input_ids": input_ids, "attention_mask": input_mask}
return config, inputs_dict
@require_torch
class ErnieMModelTest(ModelTesterMixin, unittest.TestCase):
all_model_classes = (
(
ErnieMModel,
ErnieMForMultipleChoice,
ErnieMForQuestionAnswering,
ErnieMForSequenceClassification,
ErnieMForTokenClassification,
)
if is_torch_available()
else ()
)
all_generative_model_classes = ()
test_torchscript = False
def setUp(self):
self.model_tester = ErnieMModelTester(self)
self.config_tester = ConfigTester(self, config_class=ErnieMConfig, hidden_size=37)
def test_config(self):
self.config_tester.run_common_tests()
def test_model(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_model(*config_and_inputs)
def test_model_various_embeddings(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
for type in ["absolute", "relative_key", "relative_key_query"]:
config_and_inputs[0].position_embedding_type = type
self.model_tester.create_and_check_model(*config_and_inputs)
def test_for_multiple_choice(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_for_multiple_choice(*config_and_inputs)
def test_for_question_answering(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_for_question_answering(*config_and_inputs)
def test_for_sequence_classification(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_for_sequence_classification(*config_and_inputs)
def test_for_information_extraction(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_for_information_extraction(*config_and_inputs)
def test_for_token_classification(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_for_token_classification(*config_and_inputs)
@slow
def test_model_from_pretrained(self):
for model_name in ERNIE_M_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
model = ErnieMModel.from_pretrained(model_name)
self.assertIsNotNone(model)
@require_torch
class ErnieMModelIntegrationTest(unittest.TestCase):
@slow
def test_inference_model(self):
model = ErnieMModel.from_pretrained("susnato/ernie-m-base_pytorch")
model.eval()
input_ids = torch.tensor([[0, 1, 2, 3, 4, 5]])
output = model(input_ids)[0]
# TODO Replace vocab size
hidden_size = 768
expected_shape = torch.Size((1, 6, hidden_size))
self.assertEqual(output.shape, expected_shape)
expected_slice = torch.tensor(
[[[-0.0012, 0.1245, -0.0214], [-0.0742, 0.0244, -0.0771], [-0.0333, 0.1164, -0.1554]]]
)
self.assertTrue(torch.allclose(output[:, :3, :3], expected_slice, atol=1e-3))

144
tests/models/ernie_m/test_tokenization_ernie_m.py Normal file
View File

@@ -0,0 +1,144 @@
# coding=utf-8
# Copyright 2023 The HuggingFace Inc. and Baidu 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.
""" Testing suite for the PyTorch ErnieM model. """
import unittest
from transformers import ErnieMTokenizer
from transformers.testing_utils import get_tests_dir, require_sentencepiece, require_tokenizers, slow
from ...test_tokenization_common import TokenizerTesterMixin
SAMPLE_VOCAB = get_tests_dir("fixtures/spiece.model")
@require_sentencepiece
@require_tokenizers
class ErnieMTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
tokenizer_class = ErnieMTokenizer
test_seq2seq = False
test_sentencepiece = True
test_rust_tokenizer = False
test_sentencepiece_ignore_case = False
def setUp(self):
super().setUp()
# We have a SentencePiece fixture for testing
tokenizer = ErnieMTokenizer(SAMPLE_VOCAB, unk_token="<unk>", pad_token="<pad>")
tokenizer.save_pretrained(self.tmpdirname)
def get_input_output_texts(self, tokenizer):
input_text = "this is a test"
output_text = "this is a test"
return input_text, output_text
def test_convert_token_and_id(self):
"""Test ``_convert_token_to_id`` and ``_convert_id_to_token``."""
token = "<pad>"
token_id = 0
self.assertEqual(self.get_tokenizer()._convert_token_to_id(token), token_id)
self.assertEqual(self.get_tokenizer()._convert_id_to_token(token_id), token)
def test_get_vocab(self):
vocab_keys = list(self.get_tokenizer().get_vocab().keys())
self.assertEqual(vocab_keys[0], "<pad>")
self.assertEqual(vocab_keys[1], "<unk>")
self.assertEqual(vocab_keys[-1], "▁eloquent")
self.assertEqual(len(vocab_keys), 30_000)
def test_vocab_size(self):
self.assertEqual(self.get_tokenizer().vocab_size, 30_000)
def test_rust_and_python_full_tokenizers(self):
if not self.test_rust_tokenizer:
return
tokenizer = self.get_tokenizer()
rust_tokenizer = self.get_rust_tokenizer()
sequence = "I was born in 92000, and this is falsé."
tokens = tokenizer.tokenize(sequence)
rust_tokens = rust_tokenizer.tokenize(sequence)
self.assertListEqual(tokens, rust_tokens)
ids = tokenizer.encode(sequence, add_special_tokens=False)
rust_ids = rust_tokenizer.encode(sequence, add_special_tokens=False)
self.assertListEqual(ids, rust_ids)
rust_tokenizer = self.get_rust_tokenizer()
ids = tokenizer.encode(sequence)
rust_ids = rust_tokenizer.encode(sequence)
self.assertListEqual(ids, rust_ids)
def test_full_tokenizer(self):
tokenizer = ErnieMTokenizer(SAMPLE_VOCAB, do_lower_case=True, unk_token="<unk>", pad_token="<pad>")
tokens = tokenizer.tokenize("This is a test")
self.assertListEqual(tokens, ["▁this", "▁is", "▁a", "▁test"])
self.assertListEqual(tokenizer.convert_tokens_to_ids(tokens), [48, 25, 21, 1289])
tokens = tokenizer.tokenize("I was born in 92000, and this is falsé.")
# ErnieMTokenizer(paddlenlp implementation) outputs '9' instead of '_9' so to mimic that '_9' is changed to '9'
self.assertListEqual(
tokens, ["▁i", "▁was", "▁born", "▁in", "9", "2000", ",", "▁and", "▁this", "▁is", "▁fal", "s", "é", "."]
)
ids = tokenizer.convert_tokens_to_ids(tokens)
self.assertListEqual(ids, [31, 23, 386, 19, 518, 3050, 15, 17, 48, 25, 8256, 18, 1, 9])
back_tokens = tokenizer.convert_ids_to_tokens(ids)
self.assertListEqual(
back_tokens,
["▁i", "▁was", "▁born", "▁in", "9", "2000", ",", "▁and", "▁this", "▁is", "▁fal", "s", "<unk>", "."],
)
def test_sequence_builders(self):
tokenizer = ErnieMTokenizer(SAMPLE_VOCAB, unk_token="<unk>", pad_token="<pad>")
text = tokenizer.encode("sequence builders")
text_2 = tokenizer.encode("multi-sequence build")
encoded_sentence = tokenizer.build_inputs_with_special_tokens(text)
encoded_pair = tokenizer.build_inputs_with_special_tokens(text, text_2)
assert encoded_sentence == [tokenizer.cls_token_id] + text + [tokenizer.sep_token_id]
assert encoded_pair == [tokenizer.cls_token_id] + text + [tokenizer.sep_token_id] + [
tokenizer.sep_token_id
] + text_2 + [tokenizer.sep_token_id]
@slow
def test_tokenizer_integration(self):
# fmt: off
expected_encoding = {'input_ids': [[0, 11062, 82772, 7, 15, 82772, 538, 51529, 237, 17198, 1290, 206, 9, 215175, 1314, 136, 17198, 1290, 206, 9, 56359, 42, 122009, 9, 16466, 16, 87344, 4537, 9, 4717, 78381, 6, 159958, 7, 15, 24480, 618, 4, 527, 22693, 9, 304, 4, 2777, 24480, 9874, 4, 43523, 594, 4, 803, 18392, 33189, 18, 4, 43523, 24447, 5, 5, 5, 16, 100, 24955, 83658, 9626, 144057, 15, 839, 22335, 16, 136, 24955, 83658, 83479, 15, 39102, 724, 16, 678, 645, 6460, 1328, 4589, 42, 122009, 115774, 23, 3559, 1328, 46876, 7, 136, 53894, 1940, 42227, 41159, 17721, 823, 425, 4, 27512, 98722, 206, 136, 5531, 4970, 919, 17336, 5, 2], [0, 20080, 618, 83, 82775, 47, 479, 9, 1517, 73, 53894, 333, 80581, 110117, 18811, 5256, 1295, 51, 152526, 297, 7986, 390, 124416, 538, 35431, 214, 98, 15044, 25737, 136, 7108, 43701, 23, 756, 135355, 7, 5, 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [0, 581, 63773, 119455, 6, 147797, 88203, 7, 645, 70, 21, 3285, 10269, 5, 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]]}
# fmt: on
self.tokenizer_integration_test_util(
expected_encoding=expected_encoding,
model_name="susnato/ernie-m-base_pytorch",
sequences=[
"Transformers (formerly known as pytorch-transformers and pytorch-pretrained-bert) provides "
"general-purpose architectures (BERT, GPT-2, RoBERTa, XLM, DistilBert, XLNet...) for Natural "
"Language Understanding (NLU) and Natural Language Generation (NLG) with over32+ pretrained "
"models in100+ languages and deep interoperability between Jax, PyTorch and TensorFlow.",
"BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly "
"conditioning on both left and right context in all layers.",
"The quick brown fox jumps over the lazy dog.",
],
)