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

T5 & mT5 (#8552)

Browse Source 
* add mt5 and t5v1_1 model

* fix tests

* correct some imports

* add tf model

* finish tf t5

* improve examples

* fix copies

* clean doc
This commit is contained in:
Patrick von Platen
2020-11-17 12:23:09 +01:00
committed by GitHub
parent 9e01f988dd
commit 86822a358b
21 changed files with 680 additions and 25 deletions
Download Patch File Download Diff File Expand all files Collapse all files

39
tests/test_modeling_mt5.py Normal file
View File

@@ -0,0 +1,39 @@
import unittest
from transformers import is_torch_available
from transformers.testing_utils import require_sentencepiece, require_tokenizers, require_torch, slow, torch_device
if is_torch_available():
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
@require_torch
@require_sentencepiece
@require_tokenizers
class MT5IntegrationTest(unittest.TestCase):
@slow
def test_small_integration_test(self):
"""
For comparision run:
>>> import t5 # pip install t5==0.7.1
>>> from t5.data.sentencepiece_vocabulary import SentencePieceVocabulary
>>> path_to_mtf_small_mt5_checkpoint = '<fill_in>'
>>> path_to_mtf_small_mt5_spm_model_path = '<fill_in>'
>>> t5_model = t5.models.MtfModel(model_dir=path_to_mtf_small_mt5_checkpoint, batch_size=1, tpu=None)
>>> vocab = SentencePieceVocabulary(path_to_mtf_small_mt5_spm_model_path)
>>> score = t5_model.score(inputs=["Hello there"], targets=["Hi I am"], vocabulary=vocab)
"""
model = AutoModelForSeq2SeqLM.from_pretrained("google/mt5-small", return_dict=True).to(torch_device)
tokenizer = AutoTokenizer.from_pretrained("google/mt5-small")
input_ids = tokenizer("Hello there", return_tensors="pt").input_ids
labels = tokenizer("Hi I am", return_tensors="pt").input_ids
loss = model(input_ids.to(torch_device), labels=labels.to(torch_device)).loss
mtf_score = -(labels.shape[-1] * loss.item())
EXPECTED_SCORE = -84.9127
self.assertTrue(abs(mtf_score - EXPECTED_SCORE) < 1e-4)

36
tests/test_modeling_t5.py
View File

@@ -490,6 +490,14 @@ class T5ModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_model(*config_and_inputs)
def test_model_v1_1(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
# check that gated gelu feed forward and different word embeddings work
config = config_and_inputs[0]
config.tie_word_embeddings = False
config.feed_forward_proj = "gated-gelu"
self.model_tester.create_and_check_model(config, *config_and_inputs[1:])
def test_with_lm_head(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_with_lm_head(*config_and_inputs)
@@ -569,7 +577,7 @@ class T5ModelIntegrationTests(unittest.TestCase):
>>> score = t5_model.score(inputs=["Hello there"], targets=["Hi I am"], vocabulary=vocab)
"""
model = T5ForConditionalGeneration.from_pretrained("t5-small", return_dict=True).to(torch_device)
model = T5ForConditionalGeneration.from_pretrained("t5-small").to(torch_device)
tokenizer = T5Tokenizer.from_pretrained("t5-small")
input_ids = tokenizer("Hello there", return_tensors="pt").input_ids
@@ -581,6 +589,32 @@ class T5ModelIntegrationTests(unittest.TestCase):
EXPECTED_SCORE = -19.0845
self.assertTrue(abs(mtf_score - EXPECTED_SCORE) < 1e-4)
@slow
def test_small_v1_1_integration_test(self):
"""
For comparision run:
>>> import t5 # pip install t5==0.7.1
>>> from t5.data.sentencepiece_vocabulary import SentencePieceVocabulary
>>> path_to_mtf_small_t5_v1_1_checkpoint = '<fill_in>'
>>> path_to_mtf_small_spm_model_path = '<fill_in>'
>>> t5_model = t5.models.MtfModel(model_dir=path_to_mtf_small_t5_v1_1_checkpoint, batch_size=1, tpu=None)
>>> vocab = SentencePieceVocabulary(path_to_mtf_small_spm_model_path, extra_ids=100)
>>> score = t5_model.score(inputs=["Hello there"], targets=["Hi I am"], vocabulary=vocab)
"""
model = T5ForConditionalGeneration.from_pretrained("google/t5-v1_1-small").to(torch_device)
tokenizer = T5Tokenizer.from_pretrained("google/t5-v1_1-small")
input_ids = tokenizer("Hello there", return_tensors="pt").input_ids
labels = tokenizer("Hi I am", return_tensors="pt").input_ids
loss = model(input_ids.to(torch_device), labels=labels.to(torch_device)).loss
mtf_score = -(labels.shape[-1] * loss.item())
EXPECTED_SCORE = -59.0293
self.assertTrue(abs(mtf_score - EXPECTED_SCORE) < 1e-4)
@slow
def test_summarization(self):
model = self.model

56
tests/test_modeling_tf_mt5.py Normal file
View File

@@ -0,0 +1,56 @@
# coding=utf-8
# Copyright 2018 The Google AI Language Team Authors.
#
# 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 unittest
from transformers import is_tf_available
from transformers.testing_utils import require_sentencepiece, require_tf, require_tokenizers, slow
if is_tf_available():
import tensorflow as tf
from transformers import AutoTokenizer, TFAutoModelForSeq2SeqLM
@require_tf
@require_sentencepiece
@require_tokenizers
class TFMT5ModelIntegrationTest(unittest.TestCase):
@slow
def test_small_integration_test(self):
"""
For comparision run:
>>> import t5 # pip install t5==0.7.1
>>> from t5.data.sentencepiece_vocabulary import SentencePieceVocabulary
>>> path_to_mtf_small_mt5_checkpoint = '<fill_in>'
>>> path_to_mtf_small_mt5_spm_model_path = '<fill_in>'
>>> t5_model = t5.models.MtfModel(model_dir=path_to_mtf_small_mt5_checkpoint, batch_size=1, tpu=None)
>>> vocab = SentencePieceVocabulary(path_to_mtf_small_mt5_spm_model_path, extra_ids=100)
>>> score = t5_model.score(inputs=["Hello there"], targets=["Hi I am"], vocabulary=vocab)
"""
model = TFAutoModelForSeq2SeqLM.from_pretrained("google/mt5-small")
tokenizer = AutoTokenizer.from_pretrained("google/mt5-small")
input_ids = tokenizer("Hello there", return_tensors="tf").input_ids
labels = tokenizer("Hi I am", return_tensors="tf").input_ids
loss = model(input_ids, labels=labels).loss
mtf_score = -tf.math.reduce_sum(loss).numpy()
EXPECTED_SCORE = -84.9127
self.assertTrue(abs(mtf_score - EXPECTED_SCORE) < 1e-4)

59
tests/test_modeling_tf_t5.py
View File

@@ -258,6 +258,13 @@ class TFT5ModelTest(TFModelTesterMixin, unittest.TestCase):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_t5_model(*config_and_inputs)
def test_t5_model_v1_1(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
config = config_and_inputs[0]
config.tie_word_embeddings = False
config.feed_forward_proj = "gated-gelu"
self.model_tester.create_and_check_t5_model(config, *config_and_inputs[1:])
def test_with_lm_head(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_t5_with_lm_head(*config_and_inputs)
@@ -296,6 +303,58 @@ class TFT5ModelIntegrationTests(unittest.TestCase):
def model(self):
return TFT5ForConditionalGeneration.from_pretrained("t5-base")
@slow
def test_small_integration_test(self):
"""
For comparision run:
>>> import t5 # pip install t5==0.7.1
>>> from t5.data.sentencepiece_vocabulary import SentencePieceVocabulary
>>> path_to_mtf_small_t5_checkpoint = '<fill_in>'
>>> path_to_mtf_small_spm_model_path = '<fill_in>'
>>> t5_model = t5.models.MtfModel(model_dir=path_to_mtf_small_t5_checkpoint, batch_size=1, tpu=None)
>>> vocab = SentencePieceVocabulary(path_to_mtf_small_spm_model_path, extra_ids=100)
>>> score = t5_model.score(inputs=["Hello there"], targets=["Hi I am"], vocabulary=vocab)
"""
model = TFT5ForConditionalGeneration.from_pretrained("t5-small")
tokenizer = T5Tokenizer.from_pretrained("t5-small")
input_ids = tokenizer("Hello there", return_tensors="tf").input_ids
labels = tokenizer("Hi I am", return_tensors="tf").input_ids
loss = model(input_ids, labels=labels).loss
mtf_score = -tf.math.reduce_sum(loss).numpy()
EXPECTED_SCORE = -19.0845
self.assertTrue(abs(mtf_score - EXPECTED_SCORE) < 1e-4)
@slow
def test_small_v1_1_integration_test(self):
"""
For comparision run:
>>> import t5 # pip install t5==0.7.1
>>> from t5.data.sentencepiece_vocabulary import SentencePieceVocabulary
>>> path_to_mtf_small_t5_v1.1_checkpoint = '<fill_in>'
>>> path_to_mtf_small_spm_model_path = '<fill_in>'
>>> t5_model = t5.models.MtfModel(model_dir=path_to_mtf_small_t5_v1.1_checkpoint, batch_size=1, tpu=None)
>>> vocab = SentencePieceVocabulary(path_to_mtf_small_spm_model_path, extra_ids=100)
>>> score = t5_model.score(inputs=["Hello there"], targets=["Hi I am"], vocabulary=vocab)
"""
model = TFT5ForConditionalGeneration.from_pretrained("google/t5-v1_1-small")
tokenizer = T5Tokenizer.from_pretrained("google/t5-v1_1-small")
input_ids = tokenizer("Hello there", return_tensors="tf").input_ids
labels = tokenizer("Hi I am", return_tensors="tf").input_ids
loss = model(input_ids, labels=labels).loss
mtf_score = -tf.math.reduce_sum(loss).numpy()
EXPECTED_SCORE = -59.0293
self.assertTrue(abs(mtf_score - EXPECTED_SCORE) < 1e-4)
@slow
def test_summarization(self):
model = self.model