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

Integrate DeBERTa v2(the 1.5B model surpassed human performance on Su… (#10018)

Browse Source 
* Integrate DeBERTa v2(the 1.5B model surpassed human performance on SuperGLUE); Add DeBERTa v2 900M,1.5B models;

* DeBERTa-v2

* Fix v2 model loading issue (#10129)

* Doc members

* Update src/transformers/models/deberta/modeling_deberta.py

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Address Sylvain's comments

* Address Patrick's comments

Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Style

Co-authored-by: Lysandre <lysandre.debut@reseau.eseo.fr>
Co-authored-by: Lysandre Debut <lysandre@huggingface.co>
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
This commit is contained in:
Pengcheng He
2021-02-19 15:34:44 -08:00
committed by GitHub
parent f6e53e3c2b
commit 9a7e63729f
19 changed files with 3012 additions and 73 deletions
Download Patch File Download Diff File Expand all files Collapse all files

290
tests/test_modeling_deberta_v2.py Normal file
View File

@@ -0,0 +1,290 @@
# coding=utf-8
# Copyright 2018 Microsoft Authors and the HuggingFace Inc. team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import random
import unittest
import numpy as np
from transformers import is_torch_available
from transformers.testing_utils import require_sentencepiece, require_tokenizers, require_torch, slow, torch_device
from .test_configuration_common import ConfigTester
from .test_modeling_common import ModelTesterMixin, ids_tensor
if is_torch_available():
import torch
from transformers import (
DebertaV2Config,
DebertaV2ForMaskedLM,
DebertaV2ForQuestionAnswering,
DebertaV2ForSequenceClassification,
DebertaV2ForTokenClassification,
DebertaV2Model,
)
from transformers.models.deberta_v2.modeling_deberta_v2 import DEBERTA_V2_PRETRAINED_MODEL_ARCHIVE_LIST
@require_torch
class DebertaV2ModelTest(ModelTesterMixin, unittest.TestCase):
all_model_classes = (
(
DebertaV2Model,
DebertaV2ForMaskedLM,
DebertaV2ForSequenceClassification,
DebertaV2ForTokenClassification,
DebertaV2ForQuestionAnswering,
)
if is_torch_available()
else ()
)
test_torchscript = False
test_pruning = False
test_head_masking = False
is_encoder_decoder = False
class DebertaV2ModelTester(object):
def __init__(
self,
parent,
batch_size=13,
seq_length=7,
is_training=True,
use_input_mask=True,
use_token_type_ids=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,
relative_attention=False,
position_biased_input=True,
pos_att_type="None",
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_token_type_ids = use_token_type_ids
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.relative_attention = relative_attention
self.position_biased_input = position_biased_input
self.pos_att_type = pos_att_type
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 = ids_tensor([self.batch_size, self.seq_length], vocab_size=2)
token_type_ids = None
if self.use_token_type_ids:
token_type_ids = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size)
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 = DebertaV2Config(
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,
relative_attention=self.relative_attention,
position_biased_input=self.position_biased_input,
pos_att_type=self.pos_att_type,
)
return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
def check_loss_output(self, result):
self.parent.assertListEqual(list(result.loss.size()), [])
def create_and_check_deberta_model(
self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
):
model = DebertaV2Model(config=config)
model.to(torch_device)
model.eval()
sequence_output = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids)[0]
sequence_output = model(input_ids, token_type_ids=token_type_ids)[0]
sequence_output = model(input_ids)[0]
self.parent.assertListEqual(
list(sequence_output.size()), [self.batch_size, self.seq_length, self.hidden_size]
)
def create_and_check_deberta_for_masked_lm(
self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
):
model = DebertaV2ForMaskedLM(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
def create_and_check_deberta_for_sequence_classification(
self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
):
config.num_labels = self.num_labels
model = DebertaV2ForSequenceClassification(config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=sequence_labels)
self.parent.assertListEqual(list(result.logits.size()), [self.batch_size, self.num_labels])
self.check_loss_output(result)
def create_and_check_deberta_for_token_classification(
self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
):
config.num_labels = self.num_labels
model = DebertaV2ForTokenClassification(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.num_labels))
def create_and_check_deberta_for_question_answering(
self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
):
model = DebertaV2ForQuestionAnswering(config=config)
model.to(torch_device)
model.eval()
result = model(
input_ids,
attention_mask=input_mask,
token_type_ids=token_type_ids,
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 prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
(
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
) = config_and_inputs
inputs_dict = {"input_ids": input_ids, "token_type_ids": token_type_ids, "attention_mask": input_mask}
return config, inputs_dict
def setUp(self):
self.model_tester = DebertaV2ModelTest.DebertaV2ModelTester(self)
self.config_tester = ConfigTester(self, config_class=DebertaV2Config, hidden_size=37)
def test_config(self):
self.config_tester.run_common_tests()
def test_deberta_model(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_deberta_model(*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_deberta_for_sequence_classification(*config_and_inputs)
def test_for_masked_lm(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_deberta_for_masked_lm(*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_deberta_for_question_answering(*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_deberta_for_token_classification(*config_and_inputs)
@slow
def test_model_from_pretrained(self):
for model_name in DEBERTA_V2_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
model = DebertaV2Model.from_pretrained(model_name)
self.assertIsNotNone(model)
@require_torch
@require_sentencepiece
@require_tokenizers
class DebertaV2ModelIntegrationTest(unittest.TestCase):
@unittest.skip(reason="Model not available yet")
def test_inference_masked_lm(self):
pass
@slow
def test_inference_no_head(self):
random.seed(0)
np.random.seed(0)
torch.manual_seed(0)
torch.cuda.manual_seed_all(0)
model = DebertaV2Model.from_pretrained("microsoft/deberta-v2-xlarge")
input_ids = torch.tensor([[0, 31414, 232, 328, 740, 1140, 12695, 69, 46078, 1588, 2]])
output = model(input_ids)[0]
# compare the actual values for a slice.
expected_slice = torch.tensor(
[[[-0.2913, 0.2647, 0.5627], [-0.4318, 0.1389, 0.3881], [-0.2929, -0.2489, 0.3452]]]
)
self.assertTrue(torch.allclose(output[:, :3, :3], expected_slice, atol=1e-4), f"{output[:, :3, :3]}")

162
tests/test_tokenization_deberta_v2.py Normal file
View File

@@ -0,0 +1,162 @@
# coding=utf-8
# Copyright 2019 Hugging Face inc.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import unittest
from transformers import DebertaV2Tokenizer
from transformers.testing_utils import require_sentencepiece, require_tokenizers
from .test_tokenization_common import TokenizerTesterMixin
SAMPLE_VOCAB = os.path.join(os.path.dirname(os.path.abspath(__file__)), "fixtures/spiece.model")
@require_sentencepiece
@require_tokenizers
class DebertaV2TokenizationTest(TokenizerTesterMixin, unittest.TestCase):
tokenizer_class = DebertaV2Tokenizer
rust_tokenizer_class = None
test_rust_tokenizer = False
def setUp(self):
super().setUp()
# We have a SentencePiece fixture for testing
tokenizer = DebertaV2Tokenizer(SAMPLE_VOCAB)
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_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 = DebertaV2Tokenizer(SAMPLE_VOCAB, keep_accents=True)
tokens = tokenizer.tokenize("This is a test")
self.assertListEqual(tokens, ["▁", "[UNK]", "his", "▁is", "▁a", "▁test"])
self.assertListEqual(tokenizer.convert_tokens_to_ids(tokens), [13, 1, 4398, 25, 21, 1289])
tokens = tokenizer.tokenize("I was born in 92000, and this is falsé.")
# fmt: off
self.assertListEqual(
tokens,
["▁", "[UNK]", "▁was", "▁born", "▁in", "▁9", "2000", ",", "▁and", "▁this", "▁is", "▁fal", "s", "[UNK]", "."],
)
ids = tokenizer.convert_tokens_to_ids(tokens)
self.assertListEqual(ids, [13, 1, 23, 386, 19, 561, 3050, 15, 17, 48, 25, 8256, 18, 1, 9])
back_tokens = tokenizer.convert_ids_to_tokens(ids)
self.assertListEqual(
back_tokens,
["▁", "<unk>", "▁was", "▁born", "▁in", "▁9", "2000", ",", "▁and", "▁this", "▁is", "▁fal", "s", "<unk>", "."],
)
# fmt: on
def test_sequence_builders(self):
tokenizer = DebertaV2Tokenizer(SAMPLE_VOCAB)
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] + text_2 + [
tokenizer.sep_token_id
]
def test_tokenizer_integration(self):
tokenizer_classes = [self.tokenizer_class]
if self.test_rust_tokenizer:
tokenizer_classes.append(self.rust_tokenizer_class)
for tokenizer_class in tokenizer_classes:
tokenizer = tokenizer_class.from_pretrained("microsoft/deberta-xlarge-v2")
sequences = [
[
"DeBERTa: Decoding-enhanced BERT with Disentangled Attention",
"DeBERTa: Decoding-enhanced BERT with Disentangled Attention",
],
[
"Recent progress in pre-trained neural language models has significantly improved the performance of many natural language processing (NLP) tasks.",
"DeBERTa: Decoding-enhanced BERT with Disentangled Attention",
],
[
"In this paper we propose a new model architecture DeBERTa",
"DeBERTa: Decoding-enhanced BERT with Disentangled Attention",
],
]
encoding = tokenizer(sequences, padding=True)
decoded_sequences = [tokenizer.decode(seq, skip_special_tokens=True) for seq in encoding["input_ids"]]
# fmt: off
expected_encoding = {
'input_ids': [
[1, 1804, 69418, 191, 43, 117056, 18, 44596, 448, 37132, 19, 8655, 10625, 69860, 21149, 2, 1804, 69418, 191, 43, 117056, 18, 44596, 448, 37132, 19, 8655, 10625, 69860, 21149, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
[1, 9755, 1944, 11, 1053, 18, 16899, 12730, 1072, 1506, 45, 2497, 2510, 5, 610, 9, 127, 699, 1072, 2101, 36, 99388, 53, 2930, 4, 2, 1804, 69418, 191, 43, 117056, 18, 44596, 448, 37132, 19, 8655, 10625, 69860, 21149, 2],
[1, 84, 32, 778, 42, 9441, 10, 94, 735, 3372, 1804, 69418, 191, 2, 1804, 69418, 191, 43, 117056, 18, 44596, 448, 37132, 19, 8655, 10625, 69860, 21149, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]],
'token_type_ids': [
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]],
'attention_mask': [
[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, 0, 0, 0, 0, 0, 0, 0, 0, 0],
[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, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
]
}
expected_decoded_sequences = [
'DeBERTa: Decoding-enhanced BERT with Disentangled Attention DeBERTa: Decoding-enhanced BERT with Disentangled Attention',
'Recent progress in pre-trained neural language models has significantly improved the performance of many natural language processing (NLP) tasks. DeBERTa: Decoding-enhanced BERT with Disentangled Attention',
'In this paper we propose a new model architecture DeBERTa DeBERTa: Decoding-enhanced BERT with Disentangled Attention'
]
# fmt: on
self.assertDictEqual(encoding.data, expected_encoding)
for expected, decoded in zip(expected_decoded_sequences, decoded_sequences):
self.assertEqual(expected, decoded)