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

[TFBart] Split TF-Bart (#9497)

Browse Source 
* 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
2021-01-12 02:06:32 +01:00
committed by GitHub
parent 0ecbb69806
commit 7f28613213
39 changed files with 7883 additions and 605 deletions
Download Patch File Download Diff File Expand all files Collapse all files

328
tests/test_modeling_tf_blenderbot_small.py Normal file
View 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.",
)