[TFBart] Split TF-Bart (#9497)
* make templates ready * make add_new_model_command_ready * finish tf bart * prepare tf mbart * finish tf bart * add tf mbart * add marian * prep pegasus * add tf pegasus * push blenderbot tf * add blenderbot * add blenderbot small * clean-up * make fix copy * define blend bot tok * fix * up * make style * add to docs * add copy statements * overwrite changes * improve * fix docs * finish * fix last slow test * fix missing git conflict line * fix blenderbot * up * fix blenderbot small * load changes * finish copied from * upload fix
This commit is contained in:
Patrick von Platen
committed by
GitHub
GitHub
parent
0ecbb69806
commit
7f28613213
File diff suppressed because one or more lines are too long
@@ -1,5 +1,5 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2020 HuggingFace Inc. team.
|
||||
# Copyright 2021 The HuggingFace Inc. 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.
|
||||
@@ -13,35 +13,158 @@
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
|
||||
|
||||
import unittest
|
||||
|
||||
from tests.test_configuration_common import ConfigTester
|
||||
from tests.test_modeling_tf_bart import TFBartModelTester
|
||||
from tests.test_modeling_tf_common import TFModelTesterMixin
|
||||
from transformers import BlenderbotConfig, BlenderbotSmallTokenizer, is_tf_available
|
||||
from transformers import BlenderbotConfig, BlenderbotTokenizer, is_tf_available
|
||||
from transformers.file_utils import cached_property
|
||||
from transformers.testing_utils import is_pt_tf_cross_test, require_tf, require_tokenizers, slow
|
||||
from transformers.testing_utils import require_tf, require_tokenizers, slow
|
||||
|
||||
from .test_configuration_common import ConfigTester
|
||||
from .test_modeling_tf_common import TFModelTesterMixin, ids_tensor
|
||||
|
||||
|
||||
if is_tf_available():
|
||||
import tensorflow as tf
|
||||
|
||||
from transformers import TFAutoModelForSeq2SeqLM, TFBlenderbotForConditionalGeneration
|
||||
from transformers import TFAutoModelForSeq2SeqLM, TFBlenderbotForConditionalGeneration, TFBlenderbotModel
|
||||
|
||||
|
||||
class TFBlenderbotModelTester(TFBartModelTester):
|
||||
config_updates = dict(
|
||||
normalize_before=True,
|
||||
static_position_embeddings=True,
|
||||
do_blenderbot_90_layernorm=True,
|
||||
normalize_embeddings=True,
|
||||
)
|
||||
@require_tf
|
||||
class TFBlenderbotModelTester:
|
||||
config_cls = BlenderbotConfig
|
||||
config_updates = {}
|
||||
hidden_act = "gelu"
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
parent,
|
||||
batch_size=13,
|
||||
seq_length=7,
|
||||
is_training=True,
|
||||
use_labels=False,
|
||||
vocab_size=99,
|
||||
hidden_size=32,
|
||||
num_hidden_layers=5,
|
||||
num_attention_heads=4,
|
||||
intermediate_size=37,
|
||||
hidden_dropout_prob=0.1,
|
||||
attention_probs_dropout_prob=0.1,
|
||||
max_position_embeddings=20,
|
||||
eos_token_id=2,
|
||||
pad_token_id=1,
|
||||
bos_token_id=0,
|
||||
):
|
||||
self.parent = parent
|
||||
self.batch_size = batch_size
|
||||
self.seq_length = seq_length
|
||||
self.is_training = is_training
|
||||
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_dropout_prob = hidden_dropout_prob
|
||||
self.attention_probs_dropout_prob = attention_probs_dropout_prob
|
||||
self.max_position_embeddings = max_position_embeddings
|
||||
self.eos_token_id = eos_token_id
|
||||
self.pad_token_id = pad_token_id
|
||||
self.bos_token_id = bos_token_id
|
||||
|
||||
def prepare_config_and_inputs_for_common(self):
|
||||
input_ids = ids_tensor([self.batch_size, self.seq_length - 1], self.vocab_size)
|
||||
eos_tensor = tf.expand_dims(tf.constant([self.eos_token_id] * self.batch_size), 1)
|
||||
input_ids = tf.concat([input_ids, eos_tensor], axis=1)
|
||||
|
||||
decoder_input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
|
||||
|
||||
config = self.config_cls(
|
||||
vocab_size=self.vocab_size,
|
||||
d_model=self.hidden_size,
|
||||
encoder_layers=self.num_hidden_layers,
|
||||
decoder_layers=self.num_hidden_layers,
|
||||
encoder_attention_heads=self.num_attention_heads,
|
||||
decoder_attention_heads=self.num_attention_heads,
|
||||
encoder_ffn_dim=self.intermediate_size,
|
||||
decoder_ffn_dim=self.intermediate_size,
|
||||
dropout=self.hidden_dropout_prob,
|
||||
attention_dropout=self.attention_probs_dropout_prob,
|
||||
max_position_embeddings=self.max_position_embeddings,
|
||||
eos_token_ids=[2],
|
||||
bos_token_id=self.bos_token_id,
|
||||
pad_token_id=self.pad_token_id,
|
||||
decoder_start_token_id=self.pad_token_id,
|
||||
**self.config_updates,
|
||||
)
|
||||
inputs_dict = prepare_blenderbot_inputs_dict(config, input_ids, decoder_input_ids)
|
||||
return config, inputs_dict
|
||||
|
||||
def check_decoder_model_past_large_inputs(self, config, inputs_dict):
|
||||
model = TFBlenderbotModel(config=config).get_decoder()
|
||||
input_ids = inputs_dict["input_ids"]
|
||||
|
||||
input_ids = input_ids[:1, :]
|
||||
attention_mask = inputs_dict["attention_mask"][:1, :]
|
||||
self.batch_size = 1
|
||||
|
||||
# first forward pass
|
||||
outputs = model(input_ids, attention_mask=attention_mask, use_cache=True)
|
||||
|
||||
output, past_key_values = outputs.to_tuple()
|
||||
past_key_values = past_key_values[1]
|
||||
|
||||
# create hypothetical next token and extent to next_input_ids
|
||||
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
|
||||
next_attn_mask = tf.cast(ids_tensor((self.batch_size, 3), 2), tf.int8)
|
||||
|
||||
# append to next input_ids and
|
||||
next_input_ids = tf.concat([input_ids, next_tokens], axis=-1)
|
||||
next_attention_mask = tf.concat([attention_mask, next_attn_mask], axis=-1)
|
||||
|
||||
output_from_no_past = model(next_input_ids, attention_mask=next_attention_mask)[0]
|
||||
output_from_past = model(next_tokens, attention_mask=next_attention_mask, past_key_values=past_key_values)[0]
|
||||
|
||||
self.parent.assertEqual(next_tokens.shape[1], output_from_past.shape[1])
|
||||
|
||||
# select random slice
|
||||
random_slice_idx = int(ids_tensor((1,), output_from_past.shape[-1]))
|
||||
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx]
|
||||
output_from_past_slice = output_from_past[:, :, random_slice_idx]
|
||||
|
||||
# test that outputs are equal for slice
|
||||
tf.debugging.assert_near(output_from_past_slice, output_from_no_past_slice, rtol=1e-3)
|
||||
|
||||
|
||||
def prepare_blenderbot_inputs_dict(
|
||||
config,
|
||||
input_ids,
|
||||
decoder_input_ids,
|
||||
attention_mask=None,
|
||||
decoder_attention_mask=None,
|
||||
):
|
||||
if attention_mask is None:
|
||||
attention_mask = tf.cast(tf.math.not_equal(input_ids, config.pad_token_id), tf.int8)
|
||||
if decoder_attention_mask is None:
|
||||
decoder_attention_mask = tf.concat(
|
||||
[
|
||||
tf.ones(decoder_input_ids[:, :1].shape, dtype=tf.int8),
|
||||
tf.cast(tf.math.not_equal(decoder_input_ids[:, 1:], config.pad_token_id), tf.int8),
|
||||
],
|
||||
axis=-1,
|
||||
)
|
||||
return {
|
||||
"input_ids": input_ids,
|
||||
"decoder_input_ids": decoder_input_ids,
|
||||
"attention_mask": attention_mask,
|
||||
"decoder_attention_mask": decoder_attention_mask,
|
||||
}
|
||||
|
||||
|
||||
@require_tf
|
||||
class TFBlenderbotModelTest(TFModelTesterMixin, unittest.TestCase):
|
||||
all_model_classes = (TFBlenderbotForConditionalGeneration,) if is_tf_available() else ()
|
||||
all_model_classes = (TFBlenderbotForConditionalGeneration, TFBlenderbotModel) if is_tf_available() else ()
|
||||
all_generative_model_classes = (TFBlenderbotForConditionalGeneration,) if is_tf_available() else ()
|
||||
is_encoder_decoder = True
|
||||
test_pruning = False
|
||||
@@ -53,9 +176,9 @@ class TFBlenderbotModelTest(TFModelTesterMixin, unittest.TestCase):
|
||||
def test_config(self):
|
||||
self.config_tester.run_common_tests()
|
||||
|
||||
def test_inputs_embeds(self):
|
||||
# inputs_embeds not supported
|
||||
pass
|
||||
def test_decoder_model_past_large_inputs(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
self.model_tester.check_decoder_model_past_large_inputs(*config_and_inputs)
|
||||
|
||||
def test_model_common_attributes(self):
|
||||
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
@@ -77,8 +200,22 @@ class TFBlenderbotModelTest(TFModelTesterMixin, unittest.TestCase):
|
||||
name = model.get_bias()
|
||||
assert name is None
|
||||
|
||||
@slow
|
||||
def test_saved_model_with_hidden_states_output(self):
|
||||
# TODO(JPLU, PVP) - fix this with s2s tf-serving PR
|
||||
pass
|
||||
|
||||
@slow
|
||||
def test_saved_model_with_attentions_output(self):
|
||||
# TODO(JPLU, PVP) - fix this with s2s tf-serving PR
|
||||
pass
|
||||
|
||||
def test_saved_model_creation(self):
|
||||
# This test is too long (>30sec) and makes fail the CI
|
||||
# TODO(JPLU, PVP) - fix this with s2s tf-serving PR
|
||||
pass
|
||||
|
||||
def test_saved_model_creation_extended(self):
|
||||
# TODO(JPLU, PVP) - fix this with s2s tf-serving PR
|
||||
pass
|
||||
|
||||
def test_resize_token_embeddings(self):
|
||||
@@ -145,17 +282,33 @@ class TFBlenderbotModelTest(TFModelTesterMixin, unittest.TestCase):
|
||||
self.assertTrue(models_equal)
|
||||
|
||||
|
||||
@is_pt_tf_cross_test
|
||||
def _assert_tensors_equal(a, b, atol=1e-12, prefix=""):
|
||||
"""If tensors not close, or a and b arent both tensors, raise a nice Assertion error."""
|
||||
if a is None and b is None:
|
||||
return True
|
||||
try:
|
||||
if tf.debugging.assert_near(a, b, atol=atol):
|
||||
return True
|
||||
raise
|
||||
except Exception:
|
||||
msg = "{} != {}".format(a, b)
|
||||
if prefix:
|
||||
msg = prefix + ": " + msg
|
||||
raise AssertionError(msg)
|
||||
|
||||
|
||||
def _long_tensor(tok_lst):
|
||||
return tf.constant(tok_lst, dtype=tf.int32)
|
||||
|
||||
|
||||
@require_tokenizers
|
||||
class TFBlenderbot90MIntegrationTests(unittest.TestCase):
|
||||
src_text = [
|
||||
"Social anxiety\nWow, I am never shy. Do you have anxiety?\nYes. I end up sweating and blushing and feel like i'm going to throw up.\nand why is that?"
|
||||
]
|
||||
model_name = "facebook/blenderbot-90M"
|
||||
class TFBlenderbot400MIntegrationTests(unittest.TestCase):
|
||||
src_text = ["My friends are cool but they eat too many carbs."]
|
||||
model_name = "facebook/blenderbot-400M-distill"
|
||||
|
||||
@cached_property
|
||||
def tokenizer(self):
|
||||
return BlenderbotSmallTokenizer.from_pretrained(self.model_name)
|
||||
return BlenderbotTokenizer.from_pretrained(self.model_name)
|
||||
|
||||
@cached_property
|
||||
def model(self):
|
||||
@@ -163,17 +316,13 @@ class TFBlenderbot90MIntegrationTests(unittest.TestCase):
|
||||
return model
|
||||
|
||||
@slow
|
||||
def test_90_generation_from_long_input(self):
|
||||
def test_generation_from_long_input(self):
|
||||
model_inputs = self.tokenizer(self.src_text, return_tensors="tf")
|
||||
generated_ids = self.model.generate(
|
||||
model_inputs.input_ids,
|
||||
attention_mask=model_inputs.attention_mask,
|
||||
num_beams=2,
|
||||
use_cache=True,
|
||||
)
|
||||
generated_words = self.tokenizer.batch_decode(generated_ids.numpy(), skip_special_tokens=True)[0]
|
||||
assert generated_words in (
|
||||
"i don't know. i just feel like i'm going to throw up. it's not fun.",
|
||||
"i'm not sure. i just feel like i've been feeling like i have to be in a certain place",
|
||||
"i'm not sure. i just feel like i've been in a bad situation.",
|
||||
assert (
|
||||
generated_words
|
||||
== " That's unfortunate. Are they trying to lose weight or are they just trying to be healthier?"
|
||||
)
|
||||
|
||||
328
tests/test_modeling_tf_blenderbot_small.py
Normal file
328
tests/test_modeling_tf_blenderbot_small.py
Normal file
@@ -0,0 +1,328 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2021 The HuggingFace Inc. 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.
|
||||
|
||||
|
||||
import unittest
|
||||
|
||||
from transformers import BlenderbotSmallConfig, BlenderbotSmallTokenizer, is_tf_available
|
||||
from transformers.file_utils import cached_property
|
||||
from transformers.testing_utils import require_tf, require_tokenizers, slow
|
||||
|
||||
from .test_configuration_common import ConfigTester
|
||||
from .test_modeling_tf_common import TFModelTesterMixin, ids_tensor
|
||||
|
||||
|
||||
if is_tf_available():
|
||||
import tensorflow as tf
|
||||
|
||||
from transformers import TFAutoModelForSeq2SeqLM, TFBlenderbotSmallForConditionalGeneration, TFBlenderbotSmallModel
|
||||
|
||||
|
||||
@require_tf
|
||||
class TFBlenderbotSmallModelTester:
|
||||
config_cls = BlenderbotSmallConfig
|
||||
config_updates = {}
|
||||
hidden_act = "gelu"
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
parent,
|
||||
batch_size=13,
|
||||
seq_length=7,
|
||||
is_training=True,
|
||||
use_labels=False,
|
||||
vocab_size=99,
|
||||
hidden_size=32,
|
||||
num_hidden_layers=5,
|
||||
num_attention_heads=4,
|
||||
intermediate_size=37,
|
||||
hidden_dropout_prob=0.1,
|
||||
attention_probs_dropout_prob=0.1,
|
||||
max_position_embeddings=20,
|
||||
eos_token_id=2,
|
||||
pad_token_id=1,
|
||||
bos_token_id=0,
|
||||
):
|
||||
self.parent = parent
|
||||
self.batch_size = batch_size
|
||||
self.seq_length = seq_length
|
||||
self.is_training = is_training
|
||||
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_dropout_prob = hidden_dropout_prob
|
||||
self.attention_probs_dropout_prob = attention_probs_dropout_prob
|
||||
self.max_position_embeddings = max_position_embeddings
|
||||
self.eos_token_id = eos_token_id
|
||||
self.pad_token_id = pad_token_id
|
||||
self.bos_token_id = bos_token_id
|
||||
|
||||
def prepare_config_and_inputs_for_common(self):
|
||||
input_ids = ids_tensor([self.batch_size, self.seq_length - 1], self.vocab_size)
|
||||
eos_tensor = tf.expand_dims(tf.constant([self.eos_token_id] * self.batch_size), 1)
|
||||
input_ids = tf.concat([input_ids, eos_tensor], axis=1)
|
||||
|
||||
decoder_input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
|
||||
|
||||
config = self.config_cls(
|
||||
vocab_size=self.vocab_size,
|
||||
d_model=self.hidden_size,
|
||||
encoder_layers=self.num_hidden_layers,
|
||||
decoder_layers=self.num_hidden_layers,
|
||||
encoder_attention_heads=self.num_attention_heads,
|
||||
decoder_attention_heads=self.num_attention_heads,
|
||||
encoder_ffn_dim=self.intermediate_size,
|
||||
decoder_ffn_dim=self.intermediate_size,
|
||||
dropout=self.hidden_dropout_prob,
|
||||
attention_dropout=self.attention_probs_dropout_prob,
|
||||
max_position_embeddings=self.max_position_embeddings,
|
||||
eos_token_ids=[2],
|
||||
bos_token_id=self.bos_token_id,
|
||||
pad_token_id=self.pad_token_id,
|
||||
decoder_start_token_id=self.pad_token_id,
|
||||
**self.config_updates,
|
||||
)
|
||||
inputs_dict = prepare_blenderbot_small_inputs_dict(config, input_ids, decoder_input_ids)
|
||||
return config, inputs_dict
|
||||
|
||||
def check_decoder_model_past_large_inputs(self, config, inputs_dict):
|
||||
model = TFBlenderbotSmallModel(config=config).get_decoder()
|
||||
input_ids = inputs_dict["input_ids"]
|
||||
|
||||
input_ids = input_ids[:1, :]
|
||||
attention_mask = inputs_dict["attention_mask"][:1, :]
|
||||
self.batch_size = 1
|
||||
|
||||
# first forward pass
|
||||
outputs = model(input_ids, attention_mask=attention_mask, use_cache=True)
|
||||
|
||||
output, past_key_values = outputs.to_tuple()
|
||||
past_key_values = past_key_values[1]
|
||||
|
||||
# create hypothetical next token and extent to next_input_ids
|
||||
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
|
||||
next_attn_mask = tf.cast(ids_tensor((self.batch_size, 3), 2), tf.int8)
|
||||
|
||||
# append to next input_ids and
|
||||
next_input_ids = tf.concat([input_ids, next_tokens], axis=-1)
|
||||
next_attention_mask = tf.concat([attention_mask, next_attn_mask], axis=-1)
|
||||
|
||||
output_from_no_past = model(next_input_ids, attention_mask=next_attention_mask)[0]
|
||||
output_from_past = model(next_tokens, attention_mask=next_attention_mask, past_key_values=past_key_values)[0]
|
||||
|
||||
self.parent.assertEqual(next_tokens.shape[1], output_from_past.shape[1])
|
||||
|
||||
# select random slice
|
||||
random_slice_idx = int(ids_tensor((1,), output_from_past.shape[-1]))
|
||||
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx]
|
||||
output_from_past_slice = output_from_past[:, :, random_slice_idx]
|
||||
|
||||
# test that outputs are equal for slice
|
||||
tf.debugging.assert_near(output_from_past_slice, output_from_no_past_slice, rtol=1e-3)
|
||||
|
||||
|
||||
def prepare_blenderbot_small_inputs_dict(
|
||||
config,
|
||||
input_ids,
|
||||
decoder_input_ids,
|
||||
attention_mask=None,
|
||||
decoder_attention_mask=None,
|
||||
):
|
||||
if attention_mask is None:
|
||||
attention_mask = tf.cast(tf.math.not_equal(input_ids, config.pad_token_id), tf.int8)
|
||||
if decoder_attention_mask is None:
|
||||
decoder_attention_mask = tf.concat(
|
||||
[
|
||||
tf.ones(decoder_input_ids[:, :1].shape, dtype=tf.int8),
|
||||
tf.cast(tf.math.not_equal(decoder_input_ids[:, 1:], config.pad_token_id), tf.int8),
|
||||
],
|
||||
axis=-1,
|
||||
)
|
||||
return {
|
||||
"input_ids": input_ids,
|
||||
"decoder_input_ids": decoder_input_ids,
|
||||
"attention_mask": attention_mask,
|
||||
"decoder_attention_mask": decoder_attention_mask,
|
||||
}
|
||||
|
||||
|
||||
@require_tf
|
||||
class TFBlenderbotSmallModelTest(TFModelTesterMixin, unittest.TestCase):
|
||||
all_model_classes = (
|
||||
(TFBlenderbotSmallForConditionalGeneration, TFBlenderbotSmallModel) if is_tf_available() else ()
|
||||
)
|
||||
all_generative_model_classes = (TFBlenderbotSmallForConditionalGeneration,) if is_tf_available() else ()
|
||||
is_encoder_decoder = True
|
||||
test_pruning = False
|
||||
|
||||
def setUp(self):
|
||||
self.model_tester = TFBlenderbotSmallModelTester(self)
|
||||
self.config_tester = ConfigTester(self, config_class=BlenderbotSmallConfig)
|
||||
|
||||
def test_config(self):
|
||||
self.config_tester.run_common_tests()
|
||||
|
||||
def test_decoder_model_past_large_inputs(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
self.model_tester.check_decoder_model_past_large_inputs(*config_and_inputs)
|
||||
|
||||
def test_model_common_attributes(self):
|
||||
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
|
||||
for model_class in self.all_model_classes:
|
||||
model = model_class(config)
|
||||
assert isinstance(model.get_input_embeddings(), tf.keras.layers.Layer)
|
||||
x = model.get_output_layer_with_bias()
|
||||
assert x is None
|
||||
name = model.get_prefix_bias_name()
|
||||
assert name is None
|
||||
|
||||
@slow
|
||||
def test_saved_model_with_hidden_states_output(self):
|
||||
# TODO(JPLU, PVP) - fix this with s2s tf-serving PR
|
||||
pass
|
||||
|
||||
@slow
|
||||
def test_saved_model_with_attentions_output(self):
|
||||
# TODO(JPLU, PVP) - fix this with s2s tf-serving PR
|
||||
pass
|
||||
|
||||
def test_saved_model_creation(self):
|
||||
# TODO(JPLU, PVP) - fix this with s2s tf-serving PR
|
||||
pass
|
||||
|
||||
def test_saved_model_creation_extended(self):
|
||||
# TODO(JPLU, PVP) - fix this with s2s tf-serving PR
|
||||
pass
|
||||
|
||||
def test_resize_token_embeddings(self):
|
||||
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
|
||||
def _get_word_embedding_weight(model, embedding_layer):
|
||||
if hasattr(embedding_layer, "weight"):
|
||||
return embedding_layer.weight
|
||||
else:
|
||||
# Here we build the word embeddings weights if not exists.
|
||||
# And then we retry to get the attribute once built.
|
||||
model(model.dummy_inputs)
|
||||
if hasattr(embedding_layer, "weight"):
|
||||
return embedding_layer.weight
|
||||
else:
|
||||
return None
|
||||
|
||||
for model_class in self.all_model_classes:
|
||||
for size in [config.vocab_size - 10, config.vocab_size + 10, None]:
|
||||
# build the embeddings
|
||||
model = model_class(config=config)
|
||||
old_input_embeddings = _get_word_embedding_weight(model, model.get_input_embeddings())
|
||||
old_output_embeddings = _get_word_embedding_weight(model, model.get_output_embeddings())
|
||||
old_final_logits_bias = model.get_bias()
|
||||
|
||||
# reshape the embeddings
|
||||
model.resize_token_embeddings(size)
|
||||
new_input_embeddings = _get_word_embedding_weight(model, model.get_input_embeddings())
|
||||
new_output_embeddings = _get_word_embedding_weight(model, model.get_output_embeddings())
|
||||
new_final_logits_bias = model.get_bias()
|
||||
|
||||
# check that the resized embeddings size matches the desired size.
|
||||
assert_size = size if size is not None else config.vocab_size
|
||||
|
||||
self.assertEqual(new_input_embeddings.shape[0], assert_size)
|
||||
|
||||
# check that weights remain the same after resizing
|
||||
models_equal = True
|
||||
for p1, p2 in zip(old_input_embeddings.value(), new_input_embeddings.value()):
|
||||
if tf.math.reduce_sum(tf.math.abs(p1 - p2)) > 0:
|
||||
models_equal = False
|
||||
self.assertTrue(models_equal)
|
||||
|
||||
if old_output_embeddings is not None and new_output_embeddings is not None:
|
||||
self.assertEqual(new_output_embeddings.shape[0], assert_size)
|
||||
|
||||
models_equal = True
|
||||
for p1, p2 in zip(old_output_embeddings.value(), new_output_embeddings.value()):
|
||||
if tf.math.reduce_sum(tf.math.abs(p1 - p2)) > 0:
|
||||
models_equal = False
|
||||
self.assertTrue(models_equal)
|
||||
|
||||
if old_final_logits_bias is not None and new_final_logits_bias is not None:
|
||||
old_final_logits_bias = old_final_logits_bias["final_logits_bias"]
|
||||
new_final_logits_bias = new_final_logits_bias["final_logits_bias"]
|
||||
self.assertEqual(new_final_logits_bias.shape[0], 1)
|
||||
self.assertEqual(new_final_logits_bias.shape[1], assert_size)
|
||||
|
||||
models_equal = True
|
||||
for old, new in zip(old_final_logits_bias.value(), new_final_logits_bias.value()):
|
||||
for p1, p2 in zip(old, new):
|
||||
if tf.math.reduce_sum(tf.math.abs(p1 - p2)) > 0:
|
||||
models_equal = False
|
||||
self.assertTrue(models_equal)
|
||||
|
||||
|
||||
def _assert_tensors_equal(a, b, atol=1e-12, prefix=""):
|
||||
"""If tensors not close, or a and b arent both tensors, raise a nice Assertion error."""
|
||||
if a is None and b is None:
|
||||
return True
|
||||
try:
|
||||
if tf.debugging.assert_near(a, b, atol=atol):
|
||||
return True
|
||||
raise
|
||||
except Exception:
|
||||
msg = "{} != {}".format(a, b)
|
||||
if prefix:
|
||||
msg = prefix + ": " + msg
|
||||
raise AssertionError(msg)
|
||||
|
||||
|
||||
def _long_tensor(tok_lst):
|
||||
return tf.constant(tok_lst, dtype=tf.int32)
|
||||
|
||||
|
||||
@require_tokenizers
|
||||
class TFBlenderbot90MIntegrationTests(unittest.TestCase):
|
||||
src_text = [
|
||||
"Social anxiety\nWow, I am never shy. Do you have anxiety?\nYes. I end up sweating and blushing and feel like i'm going to throw up.\nand why is that?"
|
||||
]
|
||||
model_name = "facebook/blenderbot_small-90M"
|
||||
|
||||
@cached_property
|
||||
def tokenizer(self):
|
||||
# use "old" tokenizer here because of bug when downloading new tokenizer
|
||||
return BlenderbotSmallTokenizer.from_pretrained("facebook/blenderbot-90M")
|
||||
|
||||
@cached_property
|
||||
def model(self):
|
||||
model = TFAutoModelForSeq2SeqLM.from_pretrained(self.model_name)
|
||||
return model
|
||||
|
||||
@slow
|
||||
def test_90_generation_from_long_input(self):
|
||||
model_inputs = self.tokenizer(self.src_text, return_tensors="tf")
|
||||
generated_ids = self.model.generate(
|
||||
model_inputs.input_ids,
|
||||
attention_mask=model_inputs.attention_mask,
|
||||
num_beams=2,
|
||||
use_cache=True,
|
||||
)
|
||||
generated_words = self.tokenizer.batch_decode(generated_ids.numpy(), skip_special_tokens=True)[0]
|
||||
assert generated_words in (
|
||||
"i don't know. i just feel like i'm going to throw up. it's not fun.",
|
||||
"i'm not sure. i just feel like i've been feeling like i have to be in a certain place",
|
||||
"i'm not sure. i just feel like i've been in a bad situation.",
|
||||
)
|
||||
@@ -1,5 +1,5 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2020 HuggingFace Inc. team.
|
||||
# Copyright 2021 The HuggingFace Inc. 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.
|
||||
@@ -13,48 +13,174 @@
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
|
||||
|
||||
import tempfile
|
||||
import unittest
|
||||
import warnings
|
||||
|
||||
from transformers import AutoTokenizer, MarianConfig, MarianTokenizer, TranslationPipeline, is_tf_available
|
||||
from transformers.file_utils import cached_property
|
||||
from transformers.testing_utils import is_pt_tf_cross_test, require_sentencepiece, require_tf, require_tokenizers, slow
|
||||
from transformers.testing_utils import require_sentencepiece, require_tf, require_tokenizers, slow
|
||||
|
||||
from .test_configuration_common import ConfigTester
|
||||
from .test_modeling_tf_bart import TFBartModelTester
|
||||
from .test_modeling_tf_common import TFModelTesterMixin
|
||||
from .test_modeling_tf_common import TFModelTesterMixin, ids_tensor
|
||||
|
||||
|
||||
if is_tf_available():
|
||||
import tensorflow as tf
|
||||
|
||||
from transformers import TFAutoModelForSeq2SeqLM, TFMarianMTModel
|
||||
|
||||
|
||||
class ModelTester(TFBartModelTester):
|
||||
config_updates = dict(static_position_embeddings=True, add_bias_logits=True)
|
||||
config_cls = MarianConfig
|
||||
from transformers import TFAutoModelForSeq2SeqLM, TFMarianModel, TFMarianMTModel
|
||||
|
||||
|
||||
@require_tf
|
||||
class TFMarianMTModelTest(TFModelTesterMixin, unittest.TestCase):
|
||||
all_model_classes = (TFMarianMTModel,) if is_tf_available() else ()
|
||||
class TFMarianModelTester:
|
||||
config_cls = MarianConfig
|
||||
config_updates = {}
|
||||
hidden_act = "gelu"
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
parent,
|
||||
batch_size=13,
|
||||
seq_length=7,
|
||||
is_training=True,
|
||||
use_labels=False,
|
||||
vocab_size=99,
|
||||
hidden_size=32,
|
||||
num_hidden_layers=5,
|
||||
num_attention_heads=4,
|
||||
intermediate_size=37,
|
||||
hidden_dropout_prob=0.1,
|
||||
attention_probs_dropout_prob=0.1,
|
||||
max_position_embeddings=20,
|
||||
eos_token_id=2,
|
||||
pad_token_id=1,
|
||||
bos_token_id=0,
|
||||
):
|
||||
self.parent = parent
|
||||
self.batch_size = batch_size
|
||||
self.seq_length = seq_length
|
||||
self.is_training = is_training
|
||||
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_dropout_prob = hidden_dropout_prob
|
||||
self.attention_probs_dropout_prob = attention_probs_dropout_prob
|
||||
self.max_position_embeddings = max_position_embeddings
|
||||
self.eos_token_id = eos_token_id
|
||||
self.pad_token_id = pad_token_id
|
||||
self.bos_token_id = bos_token_id
|
||||
|
||||
def prepare_config_and_inputs_for_common(self):
|
||||
input_ids = ids_tensor([self.batch_size, self.seq_length - 1], self.vocab_size)
|
||||
eos_tensor = tf.expand_dims(tf.constant([self.eos_token_id] * self.batch_size), 1)
|
||||
input_ids = tf.concat([input_ids, eos_tensor], axis=1)
|
||||
|
||||
decoder_input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
|
||||
|
||||
config = self.config_cls(
|
||||
vocab_size=self.vocab_size,
|
||||
d_model=self.hidden_size,
|
||||
encoder_layers=self.num_hidden_layers,
|
||||
decoder_layers=self.num_hidden_layers,
|
||||
encoder_attention_heads=self.num_attention_heads,
|
||||
decoder_attention_heads=self.num_attention_heads,
|
||||
encoder_ffn_dim=self.intermediate_size,
|
||||
decoder_ffn_dim=self.intermediate_size,
|
||||
dropout=self.hidden_dropout_prob,
|
||||
attention_dropout=self.attention_probs_dropout_prob,
|
||||
max_position_embeddings=self.max_position_embeddings,
|
||||
eos_token_ids=[2],
|
||||
bos_token_id=self.bos_token_id,
|
||||
pad_token_id=self.pad_token_id,
|
||||
decoder_start_token_id=self.pad_token_id,
|
||||
**self.config_updates,
|
||||
)
|
||||
inputs_dict = prepare_marian_inputs_dict(config, input_ids, decoder_input_ids)
|
||||
return config, inputs_dict
|
||||
|
||||
def check_decoder_model_past_large_inputs(self, config, inputs_dict):
|
||||
model = TFMarianModel(config=config).get_decoder()
|
||||
input_ids = inputs_dict["input_ids"]
|
||||
|
||||
input_ids = input_ids[:1, :]
|
||||
attention_mask = inputs_dict["attention_mask"][:1, :]
|
||||
self.batch_size = 1
|
||||
|
||||
# first forward pass
|
||||
outputs = model(input_ids, attention_mask=attention_mask, use_cache=True)
|
||||
|
||||
output, past_key_values = outputs.to_tuple()
|
||||
past_key_values = past_key_values[1]
|
||||
|
||||
# create hypothetical next token and extent to next_input_ids
|
||||
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
|
||||
next_attn_mask = tf.cast(ids_tensor((self.batch_size, 3), 2), tf.int8)
|
||||
|
||||
# append to next input_ids and
|
||||
next_input_ids = tf.concat([input_ids, next_tokens], axis=-1)
|
||||
next_attention_mask = tf.concat([attention_mask, next_attn_mask], axis=-1)
|
||||
|
||||
output_from_no_past = model(next_input_ids, attention_mask=next_attention_mask)[0]
|
||||
output_from_past = model(next_tokens, attention_mask=next_attention_mask, past_key_values=past_key_values)[0]
|
||||
|
||||
self.parent.assertEqual(next_tokens.shape[1], output_from_past.shape[1])
|
||||
|
||||
# select random slice
|
||||
random_slice_idx = int(ids_tensor((1,), output_from_past.shape[-1]))
|
||||
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx]
|
||||
output_from_past_slice = output_from_past[:, :, random_slice_idx]
|
||||
|
||||
# test that outputs are equal for slice
|
||||
tf.debugging.assert_near(output_from_past_slice, output_from_no_past_slice, rtol=1e-3)
|
||||
|
||||
|
||||
def prepare_marian_inputs_dict(
|
||||
config,
|
||||
input_ids,
|
||||
decoder_input_ids,
|
||||
attention_mask=None,
|
||||
decoder_attention_mask=None,
|
||||
):
|
||||
if attention_mask is None:
|
||||
attention_mask = tf.cast(tf.math.not_equal(input_ids, config.pad_token_id), tf.int8)
|
||||
if decoder_attention_mask is None:
|
||||
decoder_attention_mask = tf.concat(
|
||||
[
|
||||
tf.ones(decoder_input_ids[:, :1].shape, dtype=tf.int8),
|
||||
tf.cast(tf.math.not_equal(decoder_input_ids[:, 1:], config.pad_token_id), tf.int8),
|
||||
],
|
||||
axis=-1,
|
||||
)
|
||||
return {
|
||||
"input_ids": input_ids,
|
||||
"decoder_input_ids": decoder_input_ids,
|
||||
"attention_mask": attention_mask,
|
||||
"decoder_attention_mask": decoder_attention_mask,
|
||||
}
|
||||
|
||||
|
||||
@require_tf
|
||||
class TFMarianModelTest(TFModelTesterMixin, unittest.TestCase):
|
||||
all_model_classes = (TFMarianMTModel, TFMarianModel) if is_tf_available() else ()
|
||||
all_generative_model_classes = (TFMarianMTModel,) if is_tf_available() else ()
|
||||
model_tester_cls = ModelTester
|
||||
is_encoder_decoder = True
|
||||
test_pruning = False
|
||||
|
||||
def setUp(self):
|
||||
self.model_tester = self.model_tester_cls(self)
|
||||
self.model_tester = TFMarianModelTester(self)
|
||||
self.config_tester = ConfigTester(self, config_class=MarianConfig)
|
||||
|
||||
def test_config(self):
|
||||
self.config_tester.run_common_tests()
|
||||
|
||||
def test_inputs_embeds(self):
|
||||
# inputs_embeds not supported
|
||||
pass
|
||||
def test_decoder_model_past_large_inputs(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
self.model_tester.check_decoder_model_past_large_inputs(*config_and_inputs)
|
||||
|
||||
def test_compile_tf_model(self):
|
||||
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
@@ -107,8 +233,22 @@ class TFMarianMTModelTest(TFModelTesterMixin, unittest.TestCase):
|
||||
name = model.get_bias()
|
||||
assert name is None
|
||||
|
||||
@slow
|
||||
def test_saved_model_with_hidden_states_output(self):
|
||||
# TODO(JPLU, PVP) - fix this with s2s tf-serving PR
|
||||
pass
|
||||
|
||||
@slow
|
||||
def test_saved_model_with_attentions_output(self):
|
||||
# TODO(JPLU, PVP) - fix this with s2s tf-serving PR
|
||||
pass
|
||||
|
||||
def test_saved_model_creation(self):
|
||||
# This test is too long (>30sec) and makes fail the CI
|
||||
# TODO(JPLU, PVP) - fix this with s2s tf-serving PR
|
||||
pass
|
||||
|
||||
def test_saved_model_creation_extended(self):
|
||||
# TODO(JPLU, PVP) - fix this with s2s tf-serving PR
|
||||
pass
|
||||
|
||||
def test_resize_token_embeddings(self):
|
||||
@@ -175,6 +315,25 @@ class TFMarianMTModelTest(TFModelTesterMixin, unittest.TestCase):
|
||||
self.assertTrue(models_equal)
|
||||
|
||||
|
||||
def _assert_tensors_equal(a, b, atol=1e-12, prefix=""):
|
||||
"""If tensors not close, or a and b arent both tensors, raise a nice Assertion error."""
|
||||
if a is None and b is None:
|
||||
return True
|
||||
try:
|
||||
if tf.debugging.assert_near(a, b, atol=atol):
|
||||
return True
|
||||
raise
|
||||
except Exception:
|
||||
msg = "{} != {}".format(a, b)
|
||||
if prefix:
|
||||
msg = prefix + ": " + msg
|
||||
raise AssertionError(msg)
|
||||
|
||||
|
||||
def _long_tensor(tok_lst):
|
||||
return tf.constant(tok_lst, dtype=tf.int32)
|
||||
|
||||
|
||||
class AbstractMarianIntegrationTest(unittest.TestCase):
|
||||
maxDiff = 1000 # show more chars for failing integration tests
|
||||
|
||||
@@ -219,7 +378,6 @@ class AbstractMarianIntegrationTest(unittest.TestCase):
|
||||
|
||||
@require_sentencepiece
|
||||
@require_tokenizers
|
||||
@is_pt_tf_cross_test
|
||||
class TestMarian_MT_EN(AbstractMarianIntegrationTest):
|
||||
"""Cover low resource/high perplexity setting. This breaks if pad_token_id logits not set to LARGE_NEGATIVE."""
|
||||
|
||||
@@ -233,7 +391,6 @@ class TestMarian_MT_EN(AbstractMarianIntegrationTest):
|
||||
self._assert_generated_batch_equal_expected()
|
||||
|
||||
|
||||
@is_pt_tf_cross_test
|
||||
@require_sentencepiece
|
||||
@require_tokenizers
|
||||
class TestMarian_en_zh(AbstractMarianIntegrationTest):
|
||||
@@ -247,7 +404,6 @@ class TestMarian_en_zh(AbstractMarianIntegrationTest):
|
||||
self._assert_generated_batch_equal_expected()
|
||||
|
||||
|
||||
@is_pt_tf_cross_test
|
||||
@require_sentencepiece
|
||||
@require_tokenizers
|
||||
class TestMarian_en_ROMANCE(AbstractMarianIntegrationTest):
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2020 HuggingFace Inc. team.
|
||||
# Copyright 2021 The HuggingFace Inc. 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.
|
||||
@@ -12,47 +12,107 @@
|
||||
# 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 tempfile
|
||||
import unittest
|
||||
|
||||
from tests.test_configuration_common import ConfigTester
|
||||
from tests.test_modeling_tf_bart import TFBartModelTester
|
||||
from tests.test_modeling_tf_common import TFModelTesterMixin
|
||||
from transformers import AutoTokenizer, MBartConfig, is_tf_available
|
||||
from transformers.file_utils import cached_property
|
||||
from transformers.testing_utils import is_pt_tf_cross_test, require_sentencepiece, require_tf, require_tokenizers, slow
|
||||
from transformers.testing_utils import require_sentencepiece, require_tf, require_tokenizers, slow
|
||||
|
||||
from .test_configuration_common import ConfigTester
|
||||
from .test_modeling_tf_common import TFModelTesterMixin, ids_tensor
|
||||
|
||||
|
||||
if is_tf_available():
|
||||
|
||||
import tensorflow as tf
|
||||
|
||||
from transformers import TFAutoModelForSeq2SeqLM, TFMBartForConditionalGeneration
|
||||
|
||||
|
||||
class ModelTester(TFBartModelTester):
|
||||
config_updates = dict(normalize_before=True, add_final_layer_norm=True)
|
||||
config_cls = MBartConfig
|
||||
from transformers import TFAutoModelForSeq2SeqLM, TFMBartForConditionalGeneration, TFMBartModel
|
||||
|
||||
|
||||
@require_tf
|
||||
class TFMBartModelTest(TFModelTesterMixin, unittest.TestCase):
|
||||
all_model_classes = (TFMBartForConditionalGeneration,) if is_tf_available() else ()
|
||||
all_generative_model_classes = (TFMBartForConditionalGeneration,) if is_tf_available() else ()
|
||||
model_tester_cls = ModelTester
|
||||
is_encoder_decoder = True
|
||||
test_pruning = False
|
||||
class TFMBartModelTester:
|
||||
config_cls = MBartConfig
|
||||
config_updates = {}
|
||||
hidden_act = "gelu"
|
||||
|
||||
def setUp(self):
|
||||
self.model_tester = self.model_tester_cls(self)
|
||||
self.config_tester = ConfigTester(self, config_class=MBartConfig)
|
||||
def __init__(
|
||||
self,
|
||||
parent,
|
||||
batch_size=13,
|
||||
seq_length=7,
|
||||
is_training=True,
|
||||
use_labels=False,
|
||||
vocab_size=99,
|
||||
hidden_size=32,
|
||||
num_hidden_layers=5,
|
||||
num_attention_heads=4,
|
||||
intermediate_size=37,
|
||||
hidden_dropout_prob=0.1,
|
||||
attention_probs_dropout_prob=0.1,
|
||||
max_position_embeddings=20,
|
||||
eos_token_id=2,
|
||||
pad_token_id=1,
|
||||
bos_token_id=0,
|
||||
):
|
||||
self.parent = parent
|
||||
self.batch_size = batch_size
|
||||
self.seq_length = seq_length
|
||||
self.is_training = is_training
|
||||
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_dropout_prob = hidden_dropout_prob
|
||||
self.attention_probs_dropout_prob = attention_probs_dropout_prob
|
||||
self.max_position_embeddings = max_position_embeddings
|
||||
self.eos_token_id = eos_token_id
|
||||
self.pad_token_id = pad_token_id
|
||||
self.bos_token_id = bos_token_id
|
||||
|
||||
def test_config(self):
|
||||
self.config_tester.run_common_tests()
|
||||
def prepare_config_and_inputs_for_common(self):
|
||||
input_ids = ids_tensor([self.batch_size, self.seq_length - 1], self.vocab_size)
|
||||
eos_tensor = tf.expand_dims(tf.constant([self.eos_token_id] * self.batch_size), 1)
|
||||
input_ids = tf.concat([input_ids, eos_tensor], axis=1)
|
||||
|
||||
def test_inputs_embeds(self):
|
||||
# inputs_embeds not supported
|
||||
pass
|
||||
decoder_input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
|
||||
|
||||
config = self.config_cls(
|
||||
vocab_size=self.vocab_size,
|
||||
d_model=self.hidden_size,
|
||||
encoder_layers=self.num_hidden_layers,
|
||||
decoder_layers=self.num_hidden_layers,
|
||||
encoder_attention_heads=self.num_attention_heads,
|
||||
decoder_attention_heads=self.num_attention_heads,
|
||||
encoder_ffn_dim=self.intermediate_size,
|
||||
decoder_ffn_dim=self.intermediate_size,
|
||||
dropout=self.hidden_dropout_prob,
|
||||
attention_dropout=self.attention_probs_dropout_prob,
|
||||
max_position_embeddings=self.max_position_embeddings,
|
||||
eos_token_ids=[2],
|
||||
bos_token_id=self.bos_token_id,
|
||||
pad_token_id=self.pad_token_id,
|
||||
decoder_start_token_id=self.pad_token_id,
|
||||
**self.config_updates,
|
||||
)
|
||||
inputs_dict = prepare_mbart_inputs_dict(config, input_ids, decoder_input_ids)
|
||||
return config, inputs_dict
|
||||
|
||||
def check_decoder_model_past_large_inputs(self, config, inputs_dict):
|
||||
model = TFMBartModel(config=config).get_decoder()
|
||||
input_ids = inputs_dict["input_ids"]
|
||||
|
||||
input_ids = input_ids[:1, :]
|
||||
attention_mask = inputs_dict["attention_mask"][:1, :]
|
||||
self.batch_size = 1
|
||||
|
||||
# first forward pass
|
||||
outputs = model(input_ids, attention_mask=attention_mask, use_cache=True)
|
||||
|
||||
output, past_key_values = outputs.to_tuple()
|
||||
past_key_values = past_key_values[1]
|
||||
|
||||
def test_compile_tf_model(self):
|
||||
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
@@ -60,13 +120,11 @@ class TFMBartModelTest(TFModelTesterMixin, unittest.TestCase):
|
||||
optimizer = tf.keras.optimizers.Adam(learning_rate=3e-5, epsilon=1e-08, clipnorm=1.0)
|
||||
loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
|
||||
metric = tf.keras.metrics.SparseCategoricalAccuracy("accuracy")
|
||||
|
||||
model_class = self.all_generative_model_classes[0]
|
||||
input_ids = {
|
||||
"decoder_input_ids": tf.keras.Input(batch_shape=(2, 2000), name="decoder_input_ids", dtype="int32"),
|
||||
"input_ids": tf.keras.Input(batch_shape=(2, 2000), name="input_ids", dtype="int32"),
|
||||
}
|
||||
|
||||
# Prepare our model
|
||||
model = model_class(config)
|
||||
model(self._prepare_for_class(inputs_dict, model_class)) # Model must be called before saving.
|
||||
@@ -74,17 +132,58 @@ class TFMBartModelTest(TFModelTesterMixin, unittest.TestCase):
|
||||
with tempfile.TemporaryDirectory() as tmpdirname:
|
||||
model.save_pretrained(tmpdirname)
|
||||
model = model_class.from_pretrained(tmpdirname)
|
||||
|
||||
outputs_dict = model(input_ids)
|
||||
hidden_states = outputs_dict[0]
|
||||
|
||||
# Add a dense layer on top to test integration with other keras modules
|
||||
outputs = tf.keras.layers.Dense(2, activation="softmax", name="outputs")(hidden_states)
|
||||
|
||||
# Compile extended model
|
||||
extended_model = tf.keras.Model(inputs=[input_ids], outputs=[outputs])
|
||||
extended_model.compile(optimizer=optimizer, loss=loss, metrics=[metric])
|
||||
|
||||
|
||||
def prepare_mbart_inputs_dict(
|
||||
config,
|
||||
input_ids,
|
||||
decoder_input_ids,
|
||||
attention_mask=None,
|
||||
decoder_attention_mask=None,
|
||||
):
|
||||
if attention_mask is None:
|
||||
attention_mask = tf.cast(tf.math.not_equal(input_ids, config.pad_token_id), tf.int8)
|
||||
if decoder_attention_mask is None:
|
||||
decoder_attention_mask = tf.concat(
|
||||
[
|
||||
tf.ones(decoder_input_ids[:, :1].shape, dtype=tf.int8),
|
||||
tf.cast(tf.math.not_equal(decoder_input_ids[:, 1:], config.pad_token_id), tf.int8),
|
||||
],
|
||||
axis=-1,
|
||||
)
|
||||
return {
|
||||
"input_ids": input_ids,
|
||||
"decoder_input_ids": decoder_input_ids,
|
||||
"attention_mask": attention_mask,
|
||||
"decoder_attention_mask": decoder_attention_mask,
|
||||
}
|
||||
|
||||
|
||||
@require_tf
|
||||
class TFMBartModelTest(TFModelTesterMixin, unittest.TestCase):
|
||||
all_model_classes = (TFMBartForConditionalGeneration, TFMBartModel) if is_tf_available() else ()
|
||||
all_generative_model_classes = (TFMBartForConditionalGeneration,) if is_tf_available() else ()
|
||||
is_encoder_decoder = True
|
||||
test_pruning = False
|
||||
|
||||
def setUp(self):
|
||||
self.model_tester = TFMBartModelTester(self)
|
||||
self.config_tester = ConfigTester(self, config_class=MBartConfig)
|
||||
|
||||
def test_config(self):
|
||||
self.config_tester.run_common_tests()
|
||||
|
||||
def test_decoder_model_past_large_inputs(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
self.model_tester.check_decoder_model_past_large_inputs(*config_and_inputs)
|
||||
|
||||
def test_model_common_attributes(self):
|
||||
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
|
||||
@@ -105,8 +204,22 @@ class TFMBartModelTest(TFModelTesterMixin, unittest.TestCase):
|
||||
name = model.get_bias()
|
||||
assert name is None
|
||||
|
||||
@slow
|
||||
def test_saved_model_with_hidden_states_output(self):
|
||||
# TODO(JPLU, PVP) - fix this with s2s tf-serving PR
|
||||
pass
|
||||
|
||||
@slow
|
||||
def test_saved_model_with_attentions_output(self):
|
||||
# TODO(JPLU, PVP) - fix this with s2s tf-serving PR
|
||||
pass
|
||||
|
||||
def test_saved_model_creation(self):
|
||||
# This test is too long (>30sec) and makes fail the CI
|
||||
# TODO(JPLU, PVP) - fix this with s2s tf-serving PR
|
||||
pass
|
||||
|
||||
def test_saved_model_creation_extended(self):
|
||||
# TODO(JPLU, PVP) - fix this with s2s tf-serving PR
|
||||
pass
|
||||
|
||||
def test_resize_token_embeddings(self):
|
||||
@@ -173,10 +286,31 @@ class TFMBartModelTest(TFModelTesterMixin, unittest.TestCase):
|
||||
self.assertTrue(models_equal)
|
||||
|
||||
|
||||
@is_pt_tf_cross_test
|
||||
def _assert_tensors_equal(a, b, atol=1e-12, prefix=""):
|
||||
"""If tensors not close, or a and b arent both tensors, raise a nice Assertion error."""
|
||||
if a is None and b is None:
|
||||
return True
|
||||
try:
|
||||
if tf.debugging.assert_near(a, b, atol=atol):
|
||||
return True
|
||||
raise
|
||||
except Exception:
|
||||
msg = "{} != {}".format(a, b)
|
||||
if prefix:
|
||||
msg = prefix + ": " + msg
|
||||
raise AssertionError(msg)
|
||||
|
||||
|
||||
def _long_tensor(tok_lst):
|
||||
return tf.constant(tok_lst, dtype=tf.int32)
|
||||
|
||||
|
||||
TOLERANCE = 1e-4
|
||||
|
||||
|
||||
@require_sentencepiece
|
||||
@require_tokenizers
|
||||
class TestMBartEnRO(unittest.TestCase):
|
||||
class TFMBartModelIntegrationTest(unittest.TestCase):
|
||||
src_text = [
|
||||
" UN Chief Says There Is No Military Solution in Syria",
|
||||
]
|
||||
@@ -191,7 +325,7 @@ class TestMBartEnRO(unittest.TestCase):
|
||||
|
||||
@cached_property
|
||||
def model(self):
|
||||
model = TFAutoModelForSeq2SeqLM.from_pretrained(self.model_name, from_pt=True)
|
||||
model = TFAutoModelForSeq2SeqLM.from_pretrained(self.model_name)
|
||||
return model
|
||||
|
||||
def _assert_generated_batch_equal_expected(self, **tokenizer_kwargs):
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2020 HuggingFace Inc. team.
|
||||
# Copyright 2021 The HuggingFace Inc. 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.
|
||||
@@ -18,46 +18,167 @@ import unittest
|
||||
|
||||
from transformers import AutoTokenizer, PegasusConfig, is_tf_available
|
||||
from transformers.file_utils import cached_property
|
||||
from transformers.testing_utils import is_pt_tf_cross_test, require_sentencepiece, require_tf, require_tokenizers, slow
|
||||
from transformers.testing_utils import require_sentencepiece, require_tf, require_tokenizers, slow
|
||||
|
||||
from .test_configuration_common import ConfigTester
|
||||
from .test_modeling_tf_bart import TFBartModelTester
|
||||
from .test_modeling_tf_common import TFModelTesterMixin
|
||||
from .test_modeling_tf_common import TFModelTesterMixin, ids_tensor
|
||||
|
||||
|
||||
if is_tf_available():
|
||||
import tensorflow as tf
|
||||
|
||||
from transformers import TFAutoModelForSeq2SeqLM, TFPegasusForConditionalGeneration
|
||||
from transformers import TFAutoModelForSeq2SeqLM, TFPegasusForConditionalGeneration, TFPegasusModel
|
||||
|
||||
|
||||
class ModelTester(TFBartModelTester):
|
||||
config_updates = dict(
|
||||
normalize_before=True,
|
||||
static_position_embeddings=True,
|
||||
)
|
||||
hidden_act = "relu"
|
||||
@require_tf
|
||||
class TFPegasusModelTester:
|
||||
config_cls = PegasusConfig
|
||||
config_updates = {}
|
||||
hidden_act = "gelu"
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
parent,
|
||||
batch_size=13,
|
||||
seq_length=7,
|
||||
is_training=True,
|
||||
use_labels=False,
|
||||
vocab_size=99,
|
||||
hidden_size=32,
|
||||
num_hidden_layers=5,
|
||||
num_attention_heads=4,
|
||||
intermediate_size=37,
|
||||
hidden_dropout_prob=0.1,
|
||||
attention_probs_dropout_prob=0.1,
|
||||
max_position_embeddings=20,
|
||||
eos_token_id=2,
|
||||
pad_token_id=1,
|
||||
bos_token_id=0,
|
||||
):
|
||||
self.parent = parent
|
||||
self.batch_size = batch_size
|
||||
self.seq_length = seq_length
|
||||
self.is_training = is_training
|
||||
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_dropout_prob = hidden_dropout_prob
|
||||
self.attention_probs_dropout_prob = attention_probs_dropout_prob
|
||||
self.max_position_embeddings = max_position_embeddings
|
||||
self.eos_token_id = eos_token_id
|
||||
self.pad_token_id = pad_token_id
|
||||
self.bos_token_id = bos_token_id
|
||||
|
||||
def prepare_config_and_inputs_for_common(self):
|
||||
input_ids = ids_tensor([self.batch_size, self.seq_length - 1], self.vocab_size)
|
||||
eos_tensor = tf.expand_dims(tf.constant([self.eos_token_id] * self.batch_size), 1)
|
||||
input_ids = tf.concat([input_ids, eos_tensor], axis=1)
|
||||
|
||||
decoder_input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
|
||||
|
||||
config = self.config_cls(
|
||||
vocab_size=self.vocab_size,
|
||||
d_model=self.hidden_size,
|
||||
encoder_layers=self.num_hidden_layers,
|
||||
decoder_layers=self.num_hidden_layers,
|
||||
encoder_attention_heads=self.num_attention_heads,
|
||||
decoder_attention_heads=self.num_attention_heads,
|
||||
encoder_ffn_dim=self.intermediate_size,
|
||||
decoder_ffn_dim=self.intermediate_size,
|
||||
dropout=self.hidden_dropout_prob,
|
||||
attention_dropout=self.attention_probs_dropout_prob,
|
||||
max_position_embeddings=self.max_position_embeddings,
|
||||
eos_token_ids=[2],
|
||||
bos_token_id=self.bos_token_id,
|
||||
pad_token_id=self.pad_token_id,
|
||||
decoder_start_token_id=self.pad_token_id,
|
||||
**self.config_updates,
|
||||
)
|
||||
inputs_dict = prepare_pegasus_inputs_dict(config, input_ids, decoder_input_ids)
|
||||
return config, inputs_dict
|
||||
|
||||
def check_decoder_model_past_large_inputs(self, config, inputs_dict):
|
||||
model = TFPegasusModel(config=config).get_decoder()
|
||||
input_ids = inputs_dict["input_ids"]
|
||||
|
||||
input_ids = input_ids[:1, :]
|
||||
attention_mask = inputs_dict["attention_mask"][:1, :]
|
||||
self.batch_size = 1
|
||||
|
||||
# first forward pass
|
||||
outputs = model(input_ids, attention_mask=attention_mask, use_cache=True)
|
||||
|
||||
output, past_key_values = outputs.to_tuple()
|
||||
past_key_values = past_key_values[1]
|
||||
|
||||
# create hypothetical next token and extent to next_input_ids
|
||||
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
|
||||
next_attn_mask = tf.cast(ids_tensor((self.batch_size, 3), 2), tf.int8)
|
||||
|
||||
# append to next input_ids and
|
||||
next_input_ids = tf.concat([input_ids, next_tokens], axis=-1)
|
||||
next_attention_mask = tf.concat([attention_mask, next_attn_mask], axis=-1)
|
||||
|
||||
output_from_no_past = model(next_input_ids, attention_mask=next_attention_mask)[0]
|
||||
output_from_past = model(next_tokens, attention_mask=next_attention_mask, past_key_values=past_key_values)[0]
|
||||
|
||||
self.parent.assertEqual(next_tokens.shape[1], output_from_past.shape[1])
|
||||
|
||||
# select random slice
|
||||
random_slice_idx = int(ids_tensor((1,), output_from_past.shape[-1]))
|
||||
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx]
|
||||
output_from_past_slice = output_from_past[:, :, random_slice_idx]
|
||||
|
||||
# test that outputs are equal for slice
|
||||
tf.debugging.assert_near(output_from_past_slice, output_from_no_past_slice, rtol=1e-3)
|
||||
|
||||
|
||||
def prepare_pegasus_inputs_dict(
|
||||
config,
|
||||
input_ids,
|
||||
decoder_input_ids,
|
||||
attention_mask=None,
|
||||
decoder_attention_mask=None,
|
||||
):
|
||||
if attention_mask is None:
|
||||
attention_mask = tf.cast(tf.math.not_equal(input_ids, config.pad_token_id), tf.int8)
|
||||
if decoder_attention_mask is None:
|
||||
decoder_attention_mask = tf.concat(
|
||||
[
|
||||
tf.ones(decoder_input_ids[:, :1].shape, dtype=tf.int8),
|
||||
tf.cast(tf.math.not_equal(decoder_input_ids[:, 1:], config.pad_token_id), tf.int8),
|
||||
],
|
||||
axis=-1,
|
||||
)
|
||||
return {
|
||||
"input_ids": input_ids,
|
||||
"decoder_input_ids": decoder_input_ids,
|
||||
"attention_mask": attention_mask,
|
||||
"decoder_attention_mask": decoder_attention_mask,
|
||||
}
|
||||
|
||||
|
||||
@require_tf
|
||||
class TFPegasusModelTest(TFModelTesterMixin, unittest.TestCase):
|
||||
all_model_classes = (TFPegasusForConditionalGeneration,) if is_tf_available() else ()
|
||||
all_model_classes = (TFPegasusForConditionalGeneration, TFPegasusModel) if is_tf_available() else ()
|
||||
all_generative_model_classes = (TFPegasusForConditionalGeneration,) if is_tf_available() else ()
|
||||
model_tester_cls = ModelTester
|
||||
is_encoder_decoder = True
|
||||
test_pruning = False
|
||||
|
||||
def setUp(self):
|
||||
self.model_tester = self.model_tester_cls(self)
|
||||
self.model_tester = TFPegasusModelTester(self)
|
||||
self.config_tester = ConfigTester(self, config_class=PegasusConfig)
|
||||
|
||||
def test_config(self):
|
||||
self.config_tester.run_common_tests()
|
||||
|
||||
def test_inputs_embeds(self):
|
||||
# inputs_embeds not supported
|
||||
pass
|
||||
def test_decoder_model_past_large_inputs(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
self.model_tester.check_decoder_model_past_large_inputs(*config_and_inputs)
|
||||
|
||||
def test_compile_tf_model(self):
|
||||
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
@@ -110,8 +231,22 @@ class TFPegasusModelTest(TFModelTesterMixin, unittest.TestCase):
|
||||
name = model.get_bias()
|
||||
assert name is None
|
||||
|
||||
@slow
|
||||
def test_saved_model_with_hidden_states_output(self):
|
||||
# TODO(JPLU, PVP) - fix this with s2s tf-serving PR
|
||||
pass
|
||||
|
||||
@slow
|
||||
def test_saved_model_with_attentions_output(self):
|
||||
# TODO(JPLU, PVP) - fix this with s2s tf-serving PR
|
||||
pass
|
||||
|
||||
def test_saved_model_creation(self):
|
||||
# This test is too long (>30sec) and makes fail the CI
|
||||
# TODO(JPLU, PVP) - fix this with s2s tf-serving PR
|
||||
pass
|
||||
|
||||
def test_saved_model_creation_extended(self):
|
||||
# TODO(JPLU, PVP) - fix this with s2s tf-serving PR
|
||||
pass
|
||||
|
||||
def test_resize_token_embeddings(self):
|
||||
@@ -178,7 +313,25 @@ class TFPegasusModelTest(TFModelTesterMixin, unittest.TestCase):
|
||||
self.assertTrue(models_equal)
|
||||
|
||||
|
||||
@is_pt_tf_cross_test
|
||||
def _assert_tensors_equal(a, b, atol=1e-12, prefix=""):
|
||||
"""If tensors not close, or a and b arent both tensors, raise a nice Assertion error."""
|
||||
if a is None and b is None:
|
||||
return True
|
||||
try:
|
||||
if tf.debugging.assert_near(a, b, atol=atol):
|
||||
return True
|
||||
raise
|
||||
except Exception:
|
||||
msg = "{} != {}".format(a, b)
|
||||
if prefix:
|
||||
msg = prefix + ": " + msg
|
||||
raise AssertionError(msg)
|
||||
|
||||
|
||||
def _long_tensor(tok_lst):
|
||||
return tf.constant(tok_lst, dtype=tf.int32)
|
||||
|
||||
|
||||
@require_sentencepiece
|
||||
@require_tokenizers
|
||||
class TFPegasusIntegrationTests(unittest.TestCase):
|
||||
@@ -198,7 +351,7 @@ class TFPegasusIntegrationTests(unittest.TestCase):
|
||||
|
||||
@cached_property
|
||||
def model(self):
|
||||
model = TFAutoModelForSeq2SeqLM.from_pretrained(self.model_name, from_pt=True)
|
||||
model = TFAutoModelForSeq2SeqLM.from_pretrained(self.model_name)
|
||||
return model
|
||||
|
||||
def _assert_generated_batch_equal_expected(self, **tokenizer_kwargs):
|
||||
|
||||
Reference in New Issue
Block a user