TF generate refactor - Greedy Search (#15562)
* TF generate start refactor * Add tf tests for sample generate * re-organize * boom boom * Apply suggestions from code review * re-add * add all code * make random greedy pass * make encoder-decoder random work * further improvements * delete bogus file * make gpt2 and t5 tests work * finish logits tests * correct logits processors * correct past / encoder_outputs drama * refactor some methods * another fix * refactor shape_list * fix more shape list * import shape _list * finish docs * fix imports * make style * correct tf utils * Fix TFRag as well * Apply Lysandre's and Sylvais suggestions * Update tests/test_generation_tf_logits_process.py Co-authored-by: Matt <Rocketknight1@users.noreply.github.com> * Update src/transformers/tf_utils.py Co-authored-by: Matt <Rocketknight1@users.noreply.github.com> * remove cpu according to gante * correct logit processor Co-authored-by: Matt <Rocketknight1@users.noreply.github.com>
This commit is contained in:
Patrick von Platen
committed by
GitHub
GitHub
parent
a3dbbc3467
commit
2e12b907ae
172
tests/test_generation_tf_logits_process.py
Normal file
172
tests/test_generation_tf_logits_process.py
Normal file
@@ -0,0 +1,172 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2020 The HuggingFace Team 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 clone 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_tf
|
||||
|
||||
|
||||
if is_tf_available():
|
||||
import tensorflow as tf
|
||||
|
||||
from transformers.generation_tf_logits_process import (
|
||||
TFLogitsProcessorList,
|
||||
TFMinLengthLogitsProcessor,
|
||||
TFNoBadWordsLogitsProcessor,
|
||||
TFNoRepeatNGramLogitsProcessor,
|
||||
TFRepetitionPenaltyLogitsProcessor,
|
||||
)
|
||||
from transformers.tf_utils import set_tensor_by_indices_to_value
|
||||
|
||||
from .test_modeling_tf_common import ids_tensor
|
||||
|
||||
|
||||
@require_tf
|
||||
class TFLogitsProcessorTest(unittest.TestCase):
|
||||
def _get_uniform_logits(self, batch_size: int, length: int):
|
||||
scores = tf.ones((batch_size, length), dtype=tf.float32) / length
|
||||
return scores
|
||||
|
||||
def test_min_length_dist_processor(self):
|
||||
vocab_size = 20
|
||||
batch_size = 4
|
||||
eos_token_id = 0
|
||||
|
||||
min_dist_processor = TFMinLengthLogitsProcessor(min_length=10, eos_token_id=eos_token_id)
|
||||
|
||||
# check that min length is applied at length 5
|
||||
input_ids = ids_tensor((batch_size, 5), vocab_size=20)
|
||||
scores = self._get_uniform_logits(batch_size, vocab_size)
|
||||
scores_before_min_length = min_dist_processor(input_ids, scores)
|
||||
self.assertListEqual(scores_before_min_length[:, eos_token_id].numpy().tolist(), 4 * [-float("inf")])
|
||||
|
||||
# check that min length is not applied anymore at length 15
|
||||
input_ids = ids_tensor((batch_size, 15), vocab_size=20)
|
||||
scores = self._get_uniform_logits(batch_size, vocab_size)
|
||||
scores_before_min_length = min_dist_processor(input_ids, scores)
|
||||
self.assertFalse(tf.math.reduce_any(tf.math.is_inf(scores_before_min_length)).numpy())
|
||||
|
||||
def test_repetition_penalty_dist_process(self):
|
||||
input_ids = tf.constant([[0, 1], [5, 0]], dtype=tf.int32)
|
||||
vocab_size = 10
|
||||
|
||||
scores = self._get_uniform_logits(batch_size=2, length=vocab_size)
|
||||
|
||||
mask = tf.cast(tf.constant([[1] + 9 * [0], 10 * [0]]), tf.bool)
|
||||
scores = set_tensor_by_indices_to_value(scores, mask, -1 / vocab_size)
|
||||
mask = tf.cast(tf.constant([10 * [0], 5 * [0] + [1] + 4 * [0]]), tf.bool)
|
||||
scores = set_tensor_by_indices_to_value(scores, mask, 4 / vocab_size)
|
||||
|
||||
rep_penalty_proc = TFRepetitionPenaltyLogitsProcessor(penalty=2.0)
|
||||
|
||||
scores = rep_penalty_proc(input_ids, tf.identity(scores))
|
||||
|
||||
# check that values were correctly changed
|
||||
self.assertAlmostEqual(scores[0, 0].numpy(), -(1 / vocab_size) * 2)
|
||||
self.assertAlmostEqual(scores[0, 1].numpy(), (1 / vocab_size) / 2)
|
||||
|
||||
self.assertAlmostEqual(scores[1, 0].numpy(), (1 / vocab_size) / 2)
|
||||
self.assertAlmostEqual(scores[1, 5].numpy(), (4 / vocab_size) / 2)
|
||||
|
||||
def test_no_repeat_ngram_dist_processor(self):
|
||||
vocab_size = 3
|
||||
batch_size = 2
|
||||
|
||||
input_ids = tf.constant([[1, 1, 2, 1], [0, 1, 0, 1]], dtype=tf.int32)
|
||||
scores = self._get_uniform_logits(batch_size, vocab_size)
|
||||
|
||||
no_repeat_proc_2_gram = TFNoRepeatNGramLogitsProcessor(2)
|
||||
no_repeat_proc_3_gram = TFNoRepeatNGramLogitsProcessor(3)
|
||||
|
||||
filtered_scores_2_gram = no_repeat_proc_2_gram(input_ids, tf.identity(scores))
|
||||
filtered_scores_3_gram = no_repeat_proc_3_gram(input_ids, tf.identity(scores))
|
||||
|
||||
# 2-gram would forbid 2nd and 3rd token (1,2) at 1st batch and 1st token (0) at 2nd batch
|
||||
self.assertListEqual(
|
||||
tf.math.is_inf(filtered_scores_2_gram).numpy().tolist(), [[False, True, True], [True, False, False]]
|
||||
)
|
||||
|
||||
# 3-gram would forbid no token at 1st batch and 1st token (0) at 2nd batch
|
||||
self.assertListEqual(
|
||||
tf.math.is_inf(filtered_scores_3_gram).numpy().tolist(), [[False, False, False], [True, False, False]]
|
||||
)
|
||||
|
||||
def test_no_bad_words_dist_processor(self):
|
||||
vocab_size = 5
|
||||
batch_size = 2
|
||||
eos_token_id = 4
|
||||
|
||||
input_ids = tf.constant([[0, 1, 3, 1], [0, 1, 0, 1]], dtype=tf.int32)
|
||||
bad_word_tokens = [[1], [4], [1, 0], [0, 1, 2], [1, 3, 1, 3]]
|
||||
scores = self._get_uniform_logits(batch_size, vocab_size)
|
||||
|
||||
no_bad_words_dist_proc = TFNoBadWordsLogitsProcessor(bad_words_ids=bad_word_tokens, eos_token_id=eos_token_id)
|
||||
|
||||
filtered_scores = no_bad_words_dist_proc(input_ids, tf.identity(scores))
|
||||
|
||||
# batch 1: 1st, 2nd, and 4th (0, 1, 3) token are forbidden
|
||||
# batch 2: 1st, 2nd, and 3rd (0, 1, 2) token are forbidden
|
||||
self.assertListEqual(
|
||||
tf.math.is_inf(filtered_scores).numpy().tolist(),
|
||||
[[True, True, False, True, True], [True, True, True, False, True]],
|
||||
)
|
||||
|
||||
def test_processor_list(self):
|
||||
batch_size = 4
|
||||
sequence_length = 10
|
||||
vocab_size = 15
|
||||
eos_token_id = 0
|
||||
|
||||
# dummy input_ids and scores
|
||||
input_ids = ids_tensor((batch_size, sequence_length), vocab_size)
|
||||
input_ids_comp = tf.identity(input_ids)
|
||||
|
||||
scores = self._get_uniform_logits(batch_size, vocab_size)
|
||||
scores_comp = tf.identity(scores)
|
||||
|
||||
# instantiate all dist processors
|
||||
min_dist_proc = TFMinLengthLogitsProcessor(min_length=10, eos_token_id=eos_token_id)
|
||||
rep_penalty_proc = TFRepetitionPenaltyLogitsProcessor(penalty=2.0)
|
||||
no_repeat_proc = TFNoRepeatNGramLogitsProcessor(2)
|
||||
no_bad_words_dist_proc = TFNoBadWordsLogitsProcessor(bad_words_ids=[[1]], eos_token_id=eos_token_id)
|
||||
|
||||
# no processor list
|
||||
scores = min_dist_proc(input_ids, scores)
|
||||
scores = rep_penalty_proc(input_ids, scores)
|
||||
scores = no_repeat_proc(input_ids, scores)
|
||||
scores = no_bad_words_dist_proc(input_ids, scores)
|
||||
|
||||
# with processor list
|
||||
processor = TFLogitsProcessorList(
|
||||
[
|
||||
min_dist_proc,
|
||||
rep_penalty_proc,
|
||||
no_repeat_proc,
|
||||
no_bad_words_dist_proc,
|
||||
]
|
||||
)
|
||||
scores_comp = processor(input_ids, scores_comp)
|
||||
|
||||
# remove inf
|
||||
scores = set_tensor_by_indices_to_value(scores, tf.math.is_inf(scores), -1e9)
|
||||
scores_comp = set_tensor_by_indices_to_value(scores_comp, tf.math.is_inf(scores_comp), -1e9)
|
||||
|
||||
# scores should be equal
|
||||
tf.debugging.assert_near(scores, scores_comp, atol=1e-3)
|
||||
|
||||
# input_ids should never be changed
|
||||
self.assertListEqual(input_ids.numpy().tolist(), input_ids_comp.numpy().tolist())
|
||||
@@ -955,7 +955,7 @@ class TFModelTesterMixin:
|
||||
# Models with non-text inputs won't work here; num_return_sequences = 1
|
||||
self._check_generated_ids(model.generate(do_sample=True, max_length=5))
|
||||
|
||||
with self.assertRaises(AssertionError):
|
||||
with self.assertRaises(ValueError):
|
||||
# generating multiple sequences when no beam search generation
|
||||
# is not allowed as it would always generate the same sequences
|
||||
model.generate(input_ids, do_sample=False, num_return_sequences=2)
|
||||
|
||||
@@ -26,14 +26,15 @@ from .test_modeling_tf_core import TFCoreModelTesterMixin
|
||||
if is_tf_available():
|
||||
import tensorflow as tf
|
||||
|
||||
from transformers import GPT2Tokenizer
|
||||
from transformers.models.gpt2.modeling_tf_gpt2 import (
|
||||
TF_GPT2_PRETRAINED_MODEL_ARCHIVE_LIST,
|
||||
TFGPT2DoubleHeadsModel,
|
||||
TFGPT2ForSequenceClassification,
|
||||
TFGPT2LMHeadModel,
|
||||
TFGPT2Model,
|
||||
shape_list,
|
||||
)
|
||||
from transformers.tf_utils import shape_list
|
||||
|
||||
|
||||
class TFGPT2ModelTester:
|
||||
@@ -428,60 +429,53 @@ class TFGPT2ModelTest(TFModelTesterMixin, TFCoreModelTesterMixin, unittest.TestC
|
||||
@require_tf
|
||||
class TFGPT2ModelLanguageGenerationTest(unittest.TestCase):
|
||||
@slow
|
||||
def test_lm_generate_gpt2(self):
|
||||
model = TFGPT2LMHeadModel.from_pretrained("gpt2")
|
||||
input_ids = tf.convert_to_tensor([[464, 3290]], dtype=tf.int32) # The dog
|
||||
expected_output_ids = [
|
||||
464,
|
||||
3290,
|
||||
373,
|
||||
1043,
|
||||
287,
|
||||
257,
|
||||
2214,
|
||||
1474,
|
||||
262,
|
||||
16246,
|
||||
286,
|
||||
2688,
|
||||
290,
|
||||
2688,
|
||||
27262,
|
||||
13,
|
||||
198,
|
||||
198,
|
||||
464,
|
||||
3290,
|
||||
] # The dog was found in a field near the intersection of West and West Streets.\n\nThe dog
|
||||
def test_lm_generate_distilgpt2(self):
|
||||
model = TFGPT2LMHeadModel.from_pretrained("distilgpt2")
|
||||
input_ids = tf.convert_to_tensor([[464, 1893]], dtype=tf.int32) # The president
|
||||
|
||||
# The president of the United States, and the president of the United Kingdom, have been in the White
|
||||
# fmt: off
|
||||
expected_output_ids = [464, 1893, 286, 262, 1578, 1829, 11, 290, 262, 1893, 286, 262, 1578, 7526, 11, 423, 587, 287, 262, 2635]
|
||||
# fmt: on
|
||||
|
||||
output_ids = model.generate(input_ids, do_sample=False)
|
||||
self.assertListEqual(output_ids[0].numpy().tolist(), expected_output_ids)
|
||||
|
||||
@slow
|
||||
def test_lm_generate_distilgpt2(self):
|
||||
def test_lm_generate_distilgpt2_batch_special(self):
|
||||
model = TFGPT2LMHeadModel.from_pretrained("distilgpt2")
|
||||
input_ids = tf.convert_to_tensor([[464, 1893]], dtype=tf.int32) # The president
|
||||
expected_output_ids = [
|
||||
464,
|
||||
1893,
|
||||
286,
|
||||
262,
|
||||
1578,
|
||||
1829,
|
||||
11,
|
||||
290,
|
||||
262,
|
||||
1893,
|
||||
286,
|
||||
262,
|
||||
1578,
|
||||
7526,
|
||||
11,
|
||||
423,
|
||||
587,
|
||||
287,
|
||||
262,
|
||||
2635,
|
||||
] # The president of the United States, and the president of the United Kingdom, have been in the White
|
||||
tokenizer = GPT2Tokenizer.from_pretrained("distilgpt2")
|
||||
|
||||
tokenizer.pad_token = tokenizer.eos_token
|
||||
tokenizer.padding_side = "left"
|
||||
|
||||
sentences = ["Today is a beautiful day and", "Yesterday was"]
|
||||
input_ids = tokenizer(sentences, return_tensors="tf", padding=True).input_ids
|
||||
|
||||
generation_kwargs = {
|
||||
"bad_words_ids": [tokenizer("is").input_ids, tokenizer("angry about").input_ids],
|
||||
"no_repeat_ngram_size": 2,
|
||||
"do_sample": False,
|
||||
"repetition_penalty": 1.3,
|
||||
}
|
||||
|
||||
output_ids = model.generate(input_ids, **generation_kwargs)
|
||||
|
||||
output_strings = tokenizer.batch_decode(output_ids, skip_special_tokens=True)
|
||||
expected_output_string = [
|
||||
"Today is a beautiful day and I am so happy to be able take part in this amazing event.",
|
||||
"Yesterday was a very busy day for the first time since I started writing this post",
|
||||
]
|
||||
self.assertListEqual(output_strings, expected_output_string)
|
||||
|
||||
@slow
|
||||
def test_lm_generate_gpt2(self):
|
||||
model = TFGPT2LMHeadModel.from_pretrained("gpt2")
|
||||
input_ids = tf.convert_to_tensor([[464, 3290]], dtype=tf.int32) # The dog
|
||||
|
||||
# The dog was found in a field near the intersection of West and West Streets.\n\nThe dog
|
||||
# fmt: off
|
||||
expected_output_ids = [464, 3290, 373, 1043, 287, 257, 2214, 1474, 262, 16246, 286, 2688, 290, 2688, 27262, 13, 198, 198, 464, 3290]
|
||||
# fmt: on
|
||||
output_ids = model.generate(input_ids, do_sample=False)
|
||||
self.assertListEqual(output_ids[0].numpy().tolist(), expected_output_ids)
|
||||
|
||||
@@ -36,14 +36,7 @@ if is_tf_available():
|
||||
TFLongformerModel,
|
||||
TFLongformerSelfAttention,
|
||||
)
|
||||
|
||||
def shape_list(x):
|
||||
"""
|
||||
copied from transformers.modeling_tf_utils
|
||||
"""
|
||||
static = x.shape.as_list()
|
||||
dynamic = tf.shape(x)
|
||||
return [dynamic[i] if s is None else s for i, s in enumerate(static)]
|
||||
from transformers.tf_utils import shape_list
|
||||
|
||||
|
||||
class TFLongformerModelTester:
|
||||
|
||||
@@ -474,7 +474,7 @@ class TFSpeech2TextModelTest(TFModelTesterMixin, unittest.TestCase):
|
||||
# num_return_sequences = 1
|
||||
self._check_generated_ids(model.generate(input_features, do_sample=True))
|
||||
|
||||
with self.assertRaises(AssertionError):
|
||||
with self.assertRaises(ValueError):
|
||||
# generating multiple sequences when no beam search generation
|
||||
# is not allowed as it would always generate the same sequences
|
||||
model.generate(input_features, do_sample=False, num_return_sequences=2)
|
||||
|
||||
@@ -453,6 +453,34 @@ class TFT5EncoderOnlyModelTest(TFModelTesterMixin, unittest.TestCase):
|
||||
pass
|
||||
|
||||
|
||||
@require_tf
|
||||
@require_sentencepiece
|
||||
@require_tokenizers
|
||||
class TFT5GenerationIntegrationTests(unittest.TestCase):
|
||||
@slow
|
||||
def test_greedy_generate(self):
|
||||
model = TFT5ForConditionalGeneration.from_pretrained("t5-small")
|
||||
tokenizer = T5Tokenizer.from_pretrained("t5-small")
|
||||
|
||||
sentences = ["Yesterday, my name was", "Today is a beautiful day and"]
|
||||
input_ids = tokenizer(sentences, return_tensors="tf", padding=True).input_ids
|
||||
|
||||
generation_kwargs = {
|
||||
"bad_words_ids": [tokenizer("my").input_ids, tokenizer("ein schöner").input_ids],
|
||||
"no_repeat_ngram_size": 3,
|
||||
"do_sample": False,
|
||||
"repetition_penalty": 2.2,
|
||||
}
|
||||
|
||||
output_ids = model.generate(input_ids, **generation_kwargs)
|
||||
|
||||
output_strings = tokenizer.batch_decode(output_ids, skip_special_tokens=True)
|
||||
|
||||
expected_output_string = ["Yesterday, my name was", "Heute ist ein schöne Tag und"]
|
||||
|
||||
self.assertListEqual(expected_output_string, output_strings)
|
||||
|
||||
|
||||
@require_tf
|
||||
@require_sentencepiece
|
||||
@require_tokenizers
|
||||
|
||||
Reference in New Issue
Block a user