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

Move test model folders (#17034)

Browse Source 
* move test model folders (TODO: fix imports and others)

* fix (potentially partially) imports (in model test modules)

* fix (potentially partially) imports (in tokenization test modules)

* fix (potentially partially) imports (in feature extraction test modules)

* fix import utils.test_modeling_tf_core

* fix path ../fixtures/

* fix imports about generation.test_generation_flax_utils

* fix more imports

* fix fixture path

* fix get_test_dir

* update module_to_test_file

* fix get_tests_dir from wrong transformers.utils

* update config.yml (CircleCI)

* fix style

* remove missing imports

* update new model script

* update check_repo

* update SPECIAL_MODULE_TO_TEST_MAP

* fix style

* add __init__

* update self-scheduled

* fix add_new_model scripts

* check one way to get location back

* python setup.py build install

* fix import in test auto

* update self-scheduled.yml

* update slack notification script

* Add comments about artifact names

* fix for yolos

Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>
This commit is contained in:
Yih-Dar
2022-05-03 14:42:02 +02:00
committed by GitHub
parent cd9274d010
commit 19420fd99e
408 changed files with 616 additions and 607 deletions
Download Patch File Download Diff File Expand all files Collapse all files

0
tests/models/mbart/__init__.py Normal file
View File

464
tests/models/mbart/test_modeling_flax_mbart.py Normal file
View File

@@ -0,0 +1,464 @@
# Copyright 2021 The HuggingFace 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
import numpy as np
import timeout_decorator # noqa
from transformers import MBartConfig, is_flax_available
from transformers.testing_utils import require_flax, require_sentencepiece, require_tokenizers, slow
from transformers.utils import cached_property
from ...generation.test_generation_flax_utils import FlaxGenerationTesterMixin
from ...test_modeling_flax_common import FlaxModelTesterMixin, ids_tensor
if is_flax_available():
import os
# The slow tests are often failing with OOM error on GPU
# This makes JAX allocate exactly what is needed on demand, and deallocate memory that is no longer needed
# but will be slower as stated here https://jax.readthedocs.io/en/latest/gpu_memory_allocation.html
os.environ["XLA_PYTHON_CLIENT_ALLOCATOR"] = "platform"
import jax
import jax.numpy as jnp
from transformers import AutoTokenizer
from transformers.models.mbart.modeling_flax_mbart import (
FlaxMBartForConditionalGeneration,
FlaxMBartForQuestionAnswering,
FlaxMBartForSequenceClassification,
FlaxMBartModel,
shift_tokens_right,
)
def prepare_mbart_inputs_dict(
config,
input_ids,
decoder_input_ids=None,
attention_mask=None,
decoder_attention_mask=None,
head_mask=None,
decoder_head_mask=None,
cross_attn_head_mask=None,
):
if attention_mask is None:
attention_mask = np.where(input_ids != config.pad_token_id, 1, 0)
if decoder_attention_mask is None:
decoder_attention_mask = np.where(decoder_input_ids != config.pad_token_id, 1, 0)
if head_mask is None:
head_mask = np.ones((config.encoder_layers, config.encoder_attention_heads))
if decoder_head_mask is None:
decoder_head_mask = np.ones((config.decoder_layers, config.decoder_attention_heads))
if cross_attn_head_mask is None:
cross_attn_head_mask = np.ones((config.decoder_layers, config.decoder_attention_heads))
return {
"input_ids": input_ids,
"decoder_input_ids": decoder_input_ids,
"attention_mask": attention_mask,
"decoder_attention_mask": decoder_attention_mask,
}
class FlaxMBartModelTester(unittest.TestCase):
def __init__(
self,
parent,
batch_size=13,
seq_length=7,
is_training=True,
use_labels=False,
vocab_size=99,
hidden_size=16,
num_hidden_layers=2,
num_attention_heads=4,
intermediate_size=4,
hidden_act="gelu",
hidden_dropout_prob=0.1,
attention_probs_dropout_prob=0.1,
max_position_embeddings=32,
eos_token_id=2,
pad_token_id=1,
bos_token_id=0,
decoder_start_token_id=2,
initializer_range=0.02,
):
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_act = hidden_act
self.hidden_dropout_prob = hidden_dropout_prob
self.attention_probs_dropout_prob = attention_probs_dropout_prob
self.max_position_embeddings = max_position_embeddings
self.eos_token_id = eos_token_id
self.pad_token_id = pad_token_id
self.bos_token_id = bos_token_id
self.decoder_start_token_id = decoder_start_token_id
self.initializer_range = initializer_range
def prepare_config_and_inputs(self):
input_ids = np.clip(ids_tensor([self.batch_size, self.seq_length - 1], self.vocab_size), 3, self.vocab_size)
input_ids = np.concatenate((input_ids, 2 * np.ones((self.batch_size, 1), dtype=np.int64)), -1)
decoder_input_ids = shift_tokens_right(input_ids, 1)
config = MBartConfig(
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_id=self.eos_token_id,
bos_token_id=self.bos_token_id,
pad_token_id=self.pad_token_id,
decoder_start_token_id=self.decoder_start_token_id,
initializer_range=self.initializer_range,
use_cache=False,
)
inputs_dict = prepare_mbart_inputs_dict(config, input_ids, decoder_input_ids)
return config, inputs_dict
def prepare_config_and_inputs_for_common(self):
config, inputs_dict = self.prepare_config_and_inputs()
return config, inputs_dict
def check_use_cache_forward(self, model_class_name, config, inputs_dict):
max_decoder_length = 20
model = model_class_name(config)
encoder_outputs = model.encode(inputs_dict["input_ids"])
decoder_input_ids, decoder_attention_mask = (
inputs_dict["decoder_input_ids"],
inputs_dict["decoder_attention_mask"],
)
past_key_values = model.init_cache(decoder_input_ids.shape[0], max_decoder_length, encoder_outputs)
decoder_attention_mask = jnp.ones((decoder_input_ids.shape[0], max_decoder_length), dtype="i4")
decoder_position_ids = jnp.broadcast_to(
jnp.arange(decoder_input_ids.shape[-1] - 1)[None, :],
(decoder_input_ids.shape[0], decoder_input_ids.shape[-1] - 1),
)
outputs_cache = model.decode(
decoder_input_ids[:, :-1],
encoder_outputs,
decoder_attention_mask=decoder_attention_mask,
past_key_values=past_key_values,
decoder_position_ids=decoder_position_ids,
)
decoder_position_ids = jnp.array(decoder_input_ids.shape[0] * [[decoder_input_ids.shape[-1] - 1]], dtype="i4")
outputs_cache_next = model.decode(
decoder_input_ids[:, -1:],
encoder_outputs,
decoder_attention_mask=decoder_attention_mask,
past_key_values=outputs_cache.past_key_values,
decoder_position_ids=decoder_position_ids,
)
outputs = model.decode(decoder_input_ids, encoder_outputs)
diff = np.max(np.abs((outputs_cache_next[0][:, -1, :5] - outputs[0][:, -1, :5])))
self.parent.assertTrue(diff < 1e-3, msg=f"Max diff is {diff}")
def check_use_cache_forward_with_attn_mask(self, model_class_name, config, inputs_dict):
max_decoder_length = 20
model = model_class_name(config)
encoder_outputs = model.encode(inputs_dict["input_ids"])
decoder_input_ids, decoder_attention_mask = (
inputs_dict["decoder_input_ids"],
inputs_dict["decoder_attention_mask"],
)
decoder_attention_mask_cache = jnp.concatenate(
[
decoder_attention_mask,
jnp.zeros((decoder_attention_mask.shape[0], max_decoder_length - decoder_attention_mask.shape[1])),
],
axis=-1,
)
past_key_values = model.init_cache(decoder_input_ids.shape[0], max_decoder_length, encoder_outputs)
decoder_position_ids = jnp.broadcast_to(
jnp.arange(decoder_input_ids.shape[-1] - 1)[None, :],
(decoder_input_ids.shape[0], decoder_input_ids.shape[-1] - 1),
)
outputs_cache = model.decode(
decoder_input_ids[:, :-1],
encoder_outputs,
decoder_attention_mask=decoder_attention_mask_cache,
past_key_values=past_key_values,
decoder_position_ids=decoder_position_ids,
)
decoder_position_ids = jnp.array(decoder_input_ids.shape[0] * [[decoder_input_ids.shape[-1] - 1]], dtype="i4")
outputs_cache_next = model.decode(
decoder_input_ids[:, -1:],
encoder_outputs,
past_key_values=outputs_cache.past_key_values,
decoder_attention_mask=decoder_attention_mask_cache,
decoder_position_ids=decoder_position_ids,
)
outputs = model.decode(decoder_input_ids, encoder_outputs, decoder_attention_mask=decoder_attention_mask)
diff = np.max(np.abs((outputs_cache_next[0][:, -1, :5] - outputs[0][:, -1, :5])))
self.parent.assertTrue(diff < 1e-3, msg=f"Max diff is {diff}")
@require_flax
class MBartHeadTests(unittest.TestCase):
vocab_size = 99
def _get_config_and_data(self):
input_ids = np.array(
[
[71, 82, 18, 33, 46, 91, 2],
[68, 34, 26, 58, 30, 82, 2],
[5, 97, 17, 39, 94, 40, 2],
[76, 83, 94, 25, 70, 78, 2],
[87, 59, 41, 35, 48, 66, 2],
[55, 13, 16, 58, 5, 2, 1], # note padding
[64, 27, 31, 51, 12, 75, 2],
[52, 64, 86, 17, 83, 39, 2],
[48, 61, 9, 24, 71, 82, 2],
[26, 1, 60, 48, 22, 13, 2],
[21, 5, 62, 28, 14, 76, 2],
[45, 98, 37, 86, 59, 48, 2],
[70, 70, 50, 9, 28, 0, 2],
],
dtype=np.int64,
)
batch_size = input_ids.shape[0]
config = MBartConfig(
vocab_size=self.vocab_size,
d_model=24,
encoder_layers=2,
decoder_layers=2,
encoder_attention_heads=2,
decoder_attention_heads=2,
encoder_ffn_dim=32,
decoder_ffn_dim=32,
max_position_embeddings=48,
eos_token_id=2,
pad_token_id=1,
bos_token_id=0,
)
return config, input_ids, batch_size
def test_sequence_classification_forward(self):
config, input_ids, batch_size = self._get_config_and_data()
model = FlaxMBartForSequenceClassification(config)
outputs = model(input_ids=input_ids, decoder_input_ids=input_ids)
expected_shape = (batch_size, config.num_labels)
self.assertEqual(outputs["logits"].shape, expected_shape)
def test_question_answering_forward(self):
config, input_ids, batch_size = self._get_config_and_data()
model = FlaxMBartForQuestionAnswering(config)
outputs = model(input_ids=input_ids)
self.assertEqual(outputs["start_logits"].shape, input_ids.shape)
self.assertEqual(outputs["end_logits"].shape, input_ids.shape)
# @timeout_decorator.timeout(1) # not working with the decorator so far
def test_lm_forward(self):
config, input_ids, batch_size = self._get_config_and_data()
lm_model = FlaxMBartForConditionalGeneration(config)
outputs = lm_model(input_ids=input_ids)
expected_shape = (batch_size, input_ids.shape[1], config.vocab_size)
self.assertEqual(outputs["logits"].shape, expected_shape)
def test_lm_uneven_forward(self):
config = MBartConfig(
vocab_size=self.vocab_size,
d_model=14,
encoder_layers=2,
decoder_layers=2,
encoder_attention_heads=2,
decoder_attention_heads=2,
encoder_ffn_dim=8,
decoder_ffn_dim=8,
max_position_embeddings=48,
)
lm_model = FlaxMBartForConditionalGeneration(config)
context = np.array([[71, 82, 18, 33, 46, 91, 2], [68, 34, 26, 58, 30, 2, 1]], dtype=np.int64)
summary = np.array([[82, 71, 82, 18, 2], [58, 68, 2, 1, 1]], dtype=np.int64)
outputs = lm_model(input_ids=context, decoder_input_ids=summary)
expected_shape = (*summary.shape, config.vocab_size)
self.assertEqual(outputs["logits"].shape, expected_shape)
def test_shift_tokens_right(self):
input_ids = np.array([[71, 82, 18, 33, 2, 1, 1], [68, 34, 26, 58, 30, 82, 2]], dtype=np.int64)
shifted = shift_tokens_right(input_ids, 1)
n_pad_before = np.equal(input_ids, 1).astype(np.float32).sum()
n_pad_after = np.equal(shifted, 1).astype(np.float32).sum()
self.assertEqual(shifted.shape, input_ids.shape)
self.assertEqual(n_pad_after, n_pad_before - 1)
self.assertTrue(np.equal(shifted[:, 0], 2).all())
@require_flax
class FlaxMBartModelTest(FlaxModelTesterMixin, unittest.TestCase, FlaxGenerationTesterMixin):
is_encoder_decoder = True
all_model_classes = (
(
FlaxMBartModel,
FlaxMBartForConditionalGeneration,
FlaxMBartForSequenceClassification,
FlaxMBartForQuestionAnswering,
)
if is_flax_available()
else ()
)
all_generative_model_classes = (FlaxMBartForConditionalGeneration,) if is_flax_available() else ()
def setUp(self):
self.model_tester = FlaxMBartModelTester(self)
def test_use_cache_forward(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs()
for model_class in self.all_model_classes:
self.model_tester.check_use_cache_forward(model_class, config, inputs_dict)
def test_use_cache_forward_with_attn_mask(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs()
for model_class in self.all_model_classes:
self.model_tester.check_use_cache_forward_with_attn_mask(model_class, config, inputs_dict)
def test_encode(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
with self.subTest(model_class.__name__):
prepared_inputs_dict = self._prepare_for_class(inputs_dict, model_class)
model = model_class(config)
@jax.jit
def encode_jitted(input_ids, attention_mask=None, **kwargs):
return model.encode(input_ids=input_ids, attention_mask=attention_mask)
with self.subTest("JIT Enabled"):
jitted_outputs = encode_jitted(**prepared_inputs_dict).to_tuple()
with self.subTest("JIT Disabled"):
with jax.disable_jit():
outputs = encode_jitted(**prepared_inputs_dict).to_tuple()
self.assertEqual(len(outputs), len(jitted_outputs))
for jitted_output, output in zip(jitted_outputs, outputs):
self.assertEqual(jitted_output.shape, output.shape)
def test_decode(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
with self.subTest(model_class.__name__):
model = model_class(config)
encoder_outputs = model.encode(inputs_dict["input_ids"], inputs_dict["attention_mask"])
prepared_inputs_dict = {
"decoder_input_ids": inputs_dict["decoder_input_ids"],
"decoder_attention_mask": inputs_dict["decoder_attention_mask"],
"encoder_outputs": encoder_outputs,
}
@jax.jit
def decode_jitted(decoder_input_ids, decoder_attention_mask, encoder_outputs):
return model.decode(
decoder_input_ids=decoder_input_ids,
decoder_attention_mask=decoder_attention_mask,
encoder_outputs=encoder_outputs,
)
with self.subTest("JIT Enabled"):
jitted_outputs = decode_jitted(**prepared_inputs_dict).to_tuple()
with self.subTest("JIT Disabled"):
with jax.disable_jit():
outputs = decode_jitted(**prepared_inputs_dict).to_tuple()
self.assertEqual(len(outputs), len(jitted_outputs))
for jitted_output, output in zip(jitted_outputs, outputs):
self.assertEqual(jitted_output.shape, output.shape)
@slow
def test_model_from_pretrained(self):
for model_class_name in self.all_model_classes:
model = model_class_name.from_pretrained("facebook/mbart-large-cc25", from_pt=True)
# FlaxMBartForSequenceClassification expects eos token in input_ids
input_ids = np.ones((1, 1)) * model.config.eos_token_id
outputs = model(input_ids)
self.assertIsNotNone(outputs)
@require_flax
@require_sentencepiece
@require_tokenizers
class FlaxMBartModelIntegrationTest(unittest.TestCase):
src_text = [
" UN Chief Says There Is No Military Solution in Syria",
]
expected_text = [
"Şeful ONU declară că nu există o soluţie militară în Siria",
]
model_name = "facebook/mbart-large-en-ro"
@cached_property
def tokenizer(self):
return AutoTokenizer.from_pretrained(self.model_name)
@cached_property
def model(self):
model = FlaxMBartForConditionalGeneration.from_pretrained(self.model_name, from_pt=True)
return model
def _assert_generated_batch_equal_expected(self, **tokenizer_kwargs):
generated_words = self.translate_src_text(**tokenizer_kwargs)
self.assertListEqual(self.expected_text, generated_words)
def translate_src_text(self, **tokenizer_kwargs):
model_inputs = self.tokenizer(self.src_text, **tokenizer_kwargs, return_tensors="np")
generated_ids = self.model.generate(
model_inputs.input_ids,
attention_mask=model_inputs.attention_mask,
decoder_start_token_id=self.tokenizer.lang_code_to_id["ro_RO"],
early_stopping=True,
num_beams=2,
).sequences
generated_words = self.tokenizer.batch_decode(generated_ids, skip_special_tokens=True)
return generated_words
@slow
def test_batch_generation_en_ro(self):
self._assert_generated_batch_equal_expected()

657
tests/models/mbart/test_modeling_mbart.py Normal file
View File

@@ -0,0 +1,657 @@
# 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.
""" Testing suite for the PyTorch MBART model. """
import copy
import tempfile
import unittest
from transformers import MBartConfig, is_torch_available
from transformers.testing_utils import require_sentencepiece, require_tokenizers, require_torch, slow, torch_device
from transformers.utils import cached_property
from ...generation.test_generation_utils import GenerationTesterMixin
from ...test_configuration_common import ConfigTester
from ...test_modeling_common import ModelTesterMixin, ids_tensor
if is_torch_available():
import torch
from transformers import (
AutoTokenizer,
BatchEncoding,
MBartForCausalLM,
MBartForConditionalGeneration,
MBartForQuestionAnswering,
MBartForSequenceClassification,
MBartModel,
)
from transformers.models.mbart.modeling_mbart import MBartDecoder, MBartEncoder
def prepare_mbart_inputs_dict(
config,
input_ids,
decoder_input_ids,
attention_mask=None,
decoder_attention_mask=None,
head_mask=None,
decoder_head_mask=None,
cross_attn_head_mask=None,
):
if attention_mask is None:
attention_mask = input_ids.ne(config.pad_token_id)
if decoder_attention_mask is None:
decoder_attention_mask = decoder_input_ids.ne(config.pad_token_id)
if head_mask is None:
head_mask = torch.ones(config.encoder_layers, config.encoder_attention_heads, device=torch_device)
if decoder_head_mask is None:
decoder_head_mask = torch.ones(config.decoder_layers, config.decoder_attention_heads, device=torch_device)
if cross_attn_head_mask is None:
cross_attn_head_mask = torch.ones(config.decoder_layers, config.decoder_attention_heads, device=torch_device)
return {
"input_ids": input_ids,
"decoder_input_ids": decoder_input_ids,
"attention_mask": attention_mask,
"decoder_attention_mask": attention_mask,
"head_mask": head_mask,
"decoder_head_mask": decoder_head_mask,
"cross_attn_head_mask": cross_attn_head_mask,
}
class MBartModelTester:
def __init__(
self,
parent,
batch_size=13,
seq_length=7,
is_training=True,
use_labels=False,
vocab_size=99,
hidden_size=16,
num_hidden_layers=2,
num_attention_heads=4,
intermediate_size=4,
hidden_act="gelu",
hidden_dropout_prob=0.1,
attention_probs_dropout_prob=0.1,
max_position_embeddings=100,
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_act = hidden_act
self.hidden_dropout_prob = hidden_dropout_prob
self.attention_probs_dropout_prob = attention_probs_dropout_prob
self.max_position_embeddings = max_position_embeddings
self.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(self):
input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size).clamp(
3,
)
input_ids[:, -1] = self.eos_token_id # Eos Token
decoder_input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
config = self.get_config()
inputs_dict = prepare_mbart_inputs_dict(config, input_ids, decoder_input_ids)
return config, inputs_dict
def get_config(self):
return MBartConfig(
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_id=self.eos_token_id,
bos_token_id=self.bos_token_id,
pad_token_id=self.pad_token_id,
)
def prepare_config_and_inputs_for_common(self):
config, inputs_dict = self.prepare_config_and_inputs()
return config, inputs_dict
def create_and_check_decoder_model_past_large_inputs(self, config, inputs_dict):
model = MBartModel(config=config).get_decoder().to(torch_device).eval()
input_ids = inputs_dict["input_ids"]
attention_mask = inputs_dict["attention_mask"]
head_mask = inputs_dict["head_mask"]
# first forward pass
outputs = model(input_ids, attention_mask=attention_mask, head_mask=head_mask, use_cache=True)
output, past_key_values = outputs.to_tuple()
# create hypothetical multiple next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
next_attn_mask = ids_tensor((self.batch_size, 3), 2)
# append to next input_ids and
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
next_attention_mask = torch.cat([attention_mask, next_attn_mask], dim=-1)
output_from_no_past = model(next_input_ids, attention_mask=next_attention_mask)["last_hidden_state"]
output_from_past = model(next_tokens, attention_mask=next_attention_mask, past_key_values=past_key_values)[
"last_hidden_state"
]
# select random slice
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
# test that outputs are equal for slice
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
def check_encoder_decoder_model_standalone(self, config, inputs_dict):
model = MBartModel(config=config).to(torch_device).eval()
outputs = model(**inputs_dict)
encoder_last_hidden_state = outputs.encoder_last_hidden_state
last_hidden_state = outputs.last_hidden_state
with tempfile.TemporaryDirectory() as tmpdirname:
encoder = model.get_encoder()
encoder.save_pretrained(tmpdirname)
encoder = MBartEncoder.from_pretrained(tmpdirname).to(torch_device)
encoder_last_hidden_state_2 = encoder(inputs_dict["input_ids"], attention_mask=inputs_dict["attention_mask"])[
0
]
self.parent.assertTrue((encoder_last_hidden_state_2 - encoder_last_hidden_state).abs().max().item() < 1e-3)
with tempfile.TemporaryDirectory() as tmpdirname:
decoder = model.get_decoder()
decoder.save_pretrained(tmpdirname)
decoder = MBartDecoder.from_pretrained(tmpdirname).to(torch_device)
last_hidden_state_2 = decoder(
input_ids=inputs_dict["decoder_input_ids"],
attention_mask=inputs_dict["decoder_attention_mask"],
encoder_hidden_states=encoder_last_hidden_state,
encoder_attention_mask=inputs_dict["attention_mask"],
)[0]
self.parent.assertTrue((last_hidden_state_2 - last_hidden_state).abs().max().item() < 1e-3)
@require_torch
class MBartModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):
all_model_classes = (
(MBartModel, MBartForConditionalGeneration, MBartForSequenceClassification, MBartForQuestionAnswering)
if is_torch_available()
else ()
)
all_generative_model_classes = (MBartForConditionalGeneration,) if is_torch_available() else ()
is_encoder_decoder = True
test_pruning = False
test_missing_keys = False
def setUp(self):
self.model_tester = MBartModelTester(self)
self.config_tester = ConfigTester(self, config_class=MBartConfig)
def test_config(self):
self.config_tester.run_common_tests()
def test_save_load_strict(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs()
for model_class in self.all_model_classes:
model = model_class(config)
with tempfile.TemporaryDirectory() as tmpdirname:
model.save_pretrained(tmpdirname)
model2, info = model_class.from_pretrained(tmpdirname, output_loading_info=True)
self.assertEqual(info["missing_keys"], [])
def test_decoder_model_past_with_large_inputs(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_decoder_model_past_large_inputs(*config_and_inputs)
def test_encoder_decoder_model_standalone(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs_for_common()
self.model_tester.check_encoder_decoder_model_standalone(*config_and_inputs)
# MBartForSequenceClassification does not support inputs_embeds
def test_inputs_embeds(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in (MBartModel, MBartForConditionalGeneration, MBartForQuestionAnswering):
model = model_class(config)
model.to(torch_device)
model.eval()
inputs = copy.deepcopy(self._prepare_for_class(inputs_dict, model_class))
if not self.is_encoder_decoder:
input_ids = inputs["input_ids"]
del inputs["input_ids"]
else:
encoder_input_ids = inputs["input_ids"]
decoder_input_ids = inputs.get("decoder_input_ids", encoder_input_ids)
del inputs["input_ids"]
inputs.pop("decoder_input_ids", None)
wte = model.get_input_embeddings()
if not self.is_encoder_decoder:
inputs["inputs_embeds"] = wte(input_ids)
else:
inputs["inputs_embeds"] = wte(encoder_input_ids)
inputs["decoder_inputs_embeds"] = wte(decoder_input_ids)
with torch.no_grad():
model(**inputs)[0]
def test_generate_fp16(self):
config, input_dict = self.model_tester.prepare_config_and_inputs()
input_ids = input_dict["input_ids"]
attention_mask = input_ids.ne(1).to(torch_device)
model = MBartForConditionalGeneration(config).eval().to(torch_device)
if torch_device == "cuda":
model.half()
model.generate(input_ids, attention_mask=attention_mask)
model.generate(num_beams=4, do_sample=True, early_stopping=False, num_return_sequences=3)
def assert_tensors_close(a, b, atol=1e-12, prefix=""):
"""If tensors have different shapes, different values or a and b are not both tensors, raise a nice Assertion error."""
if a is None and b is None:
return True
try:
if torch.allclose(a, b, atol=atol):
return True
raise
except Exception:
pct_different = (torch.gt((a - b).abs(), atol)).float().mean().item()
if a.numel() > 100:
msg = f"tensor values are {pct_different:.1%} percent different."
else:
msg = f"{a} != {b}"
if prefix:
msg = prefix + ": " + msg
raise AssertionError(msg)
def _long_tensor(tok_lst):
return torch.tensor(tok_lst, dtype=torch.long, device=torch_device)
@require_torch
@require_sentencepiece
@require_tokenizers
class AbstractSeq2SeqIntegrationTest(unittest.TestCase):
maxDiff = 1000 # longer string compare tracebacks
checkpoint_name = None
@classmethod
def setUpClass(cls):
cls.tokenizer = AutoTokenizer.from_pretrained(cls.checkpoint_name, use_fast=False)
return cls
@cached_property
def model(self):
"""Only load the model if needed."""
model = MBartForConditionalGeneration.from_pretrained(self.checkpoint_name).to(torch_device)
if "cuda" in torch_device:
model = model.half()
return model
@require_torch
@require_sentencepiece
@require_tokenizers
class MBartEnroIntegrationTest(AbstractSeq2SeqIntegrationTest):
checkpoint_name = "facebook/mbart-large-en-ro"
src_text = [
" UN Chief Says There Is No Military Solution in Syria",
""" Secretary-General Ban Ki-moon says his response to Russia's stepped up military support for Syria is that "there is no military solution" to the nearly five-year conflict and more weapons will only worsen the violence and misery for millions of people.""",
]
tgt_text = [
"Şeful ONU declară că nu există o soluţie militară în Siria",
'Secretarul General Ban Ki-moon declară că răspunsul său la intensificarea sprijinului militar al Rusiei pentru Siria este că "nu există o soluţie militară" la conflictul de aproape cinci ani şi că noi arme nu vor face decât să înrăutăţească violenţa şi mizeria pentru milioane de oameni.',
]
expected_src_tokens = [8274, 127873, 25916, 7, 8622, 2071, 438, 67485, 53, 187895, 23, 51712, 2, 250004]
@slow
def test_enro_generate_one(self):
batch: BatchEncoding = self.tokenizer(
["UN Chief Says There Is No Military Solution in Syria"], return_tensors="pt"
).to(torch_device)
translated_tokens = self.model.generate(**batch)
decoded = self.tokenizer.batch_decode(translated_tokens, skip_special_tokens=True)
self.assertEqual(self.tgt_text[0], decoded[0])
# self.assertEqual(self.tgt_text[1], decoded[1])
@slow
def test_enro_generate_batch(self):
batch: BatchEncoding = self.tokenizer(self.src_text, return_tensors="pt", padding=True, truncation=True).to(
torch_device
)
translated_tokens = self.model.generate(**batch)
decoded = self.tokenizer.batch_decode(translated_tokens, skip_special_tokens=True)
assert self.tgt_text == decoded
def test_mbart_enro_config(self):
mbart_models = ["facebook/mbart-large-en-ro"]
expected = {"scale_embedding": True, "output_past": True}
for name in mbart_models:
config = MBartConfig.from_pretrained(name)
for k, v in expected.items():
try:
self.assertEqual(v, getattr(config, k))
except AssertionError as e:
e.args += (name, k)
raise
def test_mbart_fast_forward(self):
config = MBartConfig(
vocab_size=99,
d_model=24,
encoder_layers=2,
decoder_layers=2,
encoder_attention_heads=2,
decoder_attention_heads=2,
encoder_ffn_dim=32,
decoder_ffn_dim=32,
max_position_embeddings=48,
add_final_layer_norm=True,
)
lm_model = MBartForConditionalGeneration(config).to(torch_device)
context = torch.tensor(
[[71, 82, 18, 33, 46, 91, 2], [68, 34, 26, 58, 30, 2, 1]], device=torch_device, dtype=torch.long
)
summary = torch.tensor([[82, 71, 82, 18, 2], [58, 68, 2, 1, 1]], device=torch_device, dtype=torch.long)
result = lm_model(input_ids=context, decoder_input_ids=summary, labels=summary)
expected_shape = (*summary.shape, config.vocab_size)
self.assertEqual(result.logits.shape, expected_shape)
@require_torch
@require_sentencepiece
@require_tokenizers
class MBartCC25IntegrationTest(AbstractSeq2SeqIntegrationTest):
checkpoint_name = "facebook/mbart-large-cc25"
src_text = [
" UN Chief Says There Is No Military Solution in Syria",
" I ate lunch twice yesterday",
]
tgt_text = ["Şeful ONU declară că nu există o soluţie militară în Siria", "to be padded"]
@unittest.skip("This test is broken, still generates english")
def test_cc25_generate(self):
inputs = self.tokenizer([self.src_text[0]], return_tensors="pt").to(torch_device)
translated_tokens = self.model.generate(
input_ids=inputs["input_ids"].to(torch_device),
decoder_start_token_id=self.tokenizer.lang_code_to_id["ro_RO"],
)
decoded = self.tokenizer.batch_decode(translated_tokens, skip_special_tokens=True)
self.assertEqual(self.tgt_text[0], decoded[0])
@slow
def test_fill_mask(self):
inputs = self.tokenizer(["One of the best <mask> I ever read!"], return_tensors="pt").to(torch_device)
outputs = self.model.generate(
inputs["input_ids"], decoder_start_token_id=self.tokenizer.lang_code_to_id["en_XX"], num_beams=1
)
prediction: str = self.tokenizer.batch_decode(
outputs, clean_up_tokenization_spaces=True, skip_special_tokens=True
)[0]
self.assertEqual(prediction, "of the best books I ever read!")
class MBartStandaloneDecoderModelTester:
def __init__(
self,
parent,
vocab_size=99,
batch_size=13,
d_model=16,
decoder_seq_length=7,
is_training=True,
is_decoder=True,
use_attention_mask=True,
use_cache=False,
use_labels=True,
decoder_start_token_id=2,
decoder_ffn_dim=32,
decoder_layers=4,
encoder_attention_heads=4,
decoder_attention_heads=4,
max_position_embeddings=30,
is_encoder_decoder=False,
pad_token_id=0,
bos_token_id=1,
eos_token_id=2,
scope=None,
):
self.parent = parent
self.batch_size = batch_size
self.decoder_seq_length = decoder_seq_length
# For common tests
self.seq_length = self.decoder_seq_length
self.is_training = is_training
self.use_attention_mask = use_attention_mask
self.use_labels = use_labels
self.vocab_size = vocab_size
self.d_model = d_model
self.hidden_size = d_model
self.num_hidden_layers = decoder_layers
self.decoder_layers = decoder_layers
self.decoder_ffn_dim = decoder_ffn_dim
self.encoder_attention_heads = encoder_attention_heads
self.decoder_attention_heads = decoder_attention_heads
self.num_attention_heads = decoder_attention_heads
self.eos_token_id = eos_token_id
self.bos_token_id = bos_token_id
self.pad_token_id = pad_token_id
self.decoder_start_token_id = decoder_start_token_id
self.use_cache = use_cache
self.max_position_embeddings = max_position_embeddings
self.is_encoder_decoder = is_encoder_decoder
self.scope = None
self.decoder_key_length = decoder_seq_length
self.base_model_out_len = 2
self.decoder_attention_idx = 1
def prepare_config_and_inputs(self):
input_ids = ids_tensor([self.batch_size, self.decoder_seq_length], self.vocab_size)
attention_mask = None
if self.use_attention_mask:
attention_mask = ids_tensor([self.batch_size, self.decoder_seq_length], vocab_size=2)
lm_labels = None
if self.use_labels:
lm_labels = ids_tensor([self.batch_size, self.decoder_seq_length], self.vocab_size)
config = MBartConfig(
vocab_size=self.vocab_size,
d_model=self.d_model,
decoder_layers=self.decoder_layers,
decoder_ffn_dim=self.decoder_ffn_dim,
encoder_attention_heads=self.encoder_attention_heads,
decoder_attention_heads=self.decoder_attention_heads,
eos_token_id=self.eos_token_id,
bos_token_id=self.bos_token_id,
use_cache=self.use_cache,
pad_token_id=self.pad_token_id,
decoder_start_token_id=self.decoder_start_token_id,
max_position_embeddings=self.max_position_embeddings,
is_encoder_decoder=self.is_encoder_decoder,
)
return (
config,
input_ids,
attention_mask,
lm_labels,
)
def create_and_check_decoder_model_past(
self,
config,
input_ids,
attention_mask,
lm_labels,
):
config.use_cache = True
model = MBartDecoder(config=config).to(torch_device).eval()
# first forward pass
outputs = model(input_ids, use_cache=True)
outputs_use_cache_conf = model(input_ids)
outputs_no_past = model(input_ids, use_cache=False)
self.parent.assertTrue(len(outputs) == len(outputs_use_cache_conf))
self.parent.assertTrue(len(outputs) == len(outputs_no_past) + 1)
past_key_values = outputs["past_key_values"]
# create hypothetical next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size)
# append to next input_ids and
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
output_from_no_past = model(next_input_ids)["last_hidden_state"]
output_from_past = model(next_tokens, past_key_values=past_key_values)["last_hidden_state"]
# select random slice
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
output_from_no_past_slice = output_from_no_past[:, next_input_ids.shape[-1] - 1, random_slice_idx].detach()
output_from_past_slice = output_from_past[:, 0, random_slice_idx].detach()
# test that outputs are equal for slice
assert torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3)
def create_and_check_decoder_model_attention_mask_past(
self,
config,
input_ids,
attention_mask,
lm_labels,
):
model = MBartDecoder(config=config).to(torch_device).eval()
# create attention mask
attn_mask = torch.ones(input_ids.shape, dtype=torch.long, device=torch_device)
half_seq_length = input_ids.shape[-1] // 2
attn_mask[:, half_seq_length:] = 0
# first forward pass
past_key_values = model(input_ids, attention_mask=attn_mask, use_cache=True)["past_key_values"]
# create hypothetical next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size)
# change a random masked slice from input_ids
random_seq_idx_to_change = ids_tensor((1,), half_seq_length).item() + 1
random_other_next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size).squeeze(-1)
input_ids[:, -random_seq_idx_to_change] = random_other_next_tokens
# append to next input_ids and attn_mask
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
attn_mask = torch.cat(
[attn_mask, torch.ones((attn_mask.shape[0], 1), dtype=torch.long, device=torch_device)],
dim=1,
)
# get two different outputs
output_from_no_past = model(next_input_ids, attention_mask=attn_mask)["last_hidden_state"]
output_from_past = model(next_tokens, attention_mask=attn_mask, past_key_values=past_key_values)[
"last_hidden_state"
]
# select random slice
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
output_from_no_past_slice = output_from_no_past[:, next_input_ids.shape[-1] - 1, random_slice_idx].detach()
output_from_past_slice = output_from_past[:, 0, random_slice_idx].detach()
# test that outputs are equal for slice
assert torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3)
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
(
config,
input_ids,
attention_mask,
lm_labels,
) = config_and_inputs
inputs_dict = {
"input_ids": input_ids,
"attention_mask": attention_mask,
}
return config, inputs_dict
@require_torch
class MBartStandaloneDecoderModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):
all_model_classes = (MBartDecoder, MBartForCausalLM) if is_torch_available() else ()
all_generative_model_classes = (MBartForCausalLM,) if is_torch_available() else ()
test_pruning = False
is_encoder_decoder = False
def setUp(
self,
):
self.model_tester = MBartStandaloneDecoderModelTester(self, is_training=False)
self.config_tester = ConfigTester(self, config_class=MBartConfig)
def test_config(self):
self.config_tester.run_common_tests()
def test_decoder_model_past(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_decoder_model_past(*config_and_inputs)
def test_decoder_model_attn_mask_past(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_decoder_model_attention_mask_past(*config_and_inputs)
def test_retain_grad_hidden_states_attentions(self):
# decoder cannot keep gradients
return

345
tests/models/mbart/test_modeling_tf_mbart.py Normal file
View File

@@ -0,0 +1,345 @@
# 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 tempfile
import unittest
from transformers import AutoTokenizer, MBartConfig, is_tf_available
from transformers.testing_utils import require_sentencepiece, require_tf, require_tokenizers, slow
from transformers.utils import cached_property
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, TFMBartModel
@require_tf
class TFMBartModelTester:
config_cls = MBartConfig
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_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, :]
head_mask = inputs_dict["head_mask"]
self.batch_size = 1
# first forward pass
outputs = model(input_ids, attention_mask=attention_mask, head_mask=head_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()
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.
# Let's load it from the disk to be sure we can use pretrained weights
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,
head_mask=None,
decoder_head_mask=None,
cross_attn_head_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,
)
if head_mask is None:
head_mask = tf.ones((config.encoder_layers, config.encoder_attention_heads))
if decoder_head_mask is None:
decoder_head_mask = tf.ones((config.decoder_layers, config.decoder_attention_heads))
if cross_attn_head_mask is None:
cross_attn_head_mask = tf.ones((config.decoder_layers, config.decoder_attention_heads))
return {
"input_ids": input_ids,
"decoder_input_ids": decoder_input_ids,
"attention_mask": attention_mask,
"decoder_attention_mask": decoder_attention_mask,
"head_mask": head_mask,
"decoder_head_mask": decoder_head_mask,
"cross_attn_head_mask": cross_attn_head_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
test_onnx = 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()
for model_class in self.all_model_classes:
model = model_class(config)
assert isinstance(model.get_input_embeddings(), tf.keras.layers.Layer)
if model_class in self.all_generative_model_classes:
x = model.get_output_embeddings()
assert isinstance(x, tf.keras.layers.Layer)
name = model.get_bias()
assert isinstance(name, dict)
for k, v in name.items():
assert isinstance(v, tf.Variable)
else:
x = model.get_output_embeddings()
assert x is None
name = model.get_bias()
assert name is None
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 test_saved_model_creation(self):
# This test is too long (>30sec) and makes fail the CI
pass
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:
if len(prefix) > 0:
prefix = f"{prefix}: "
raise AssertionError(f"{prefix}{a} != {b}")
def _long_tensor(tok_lst):
return tf.constant(tok_lst, dtype=tf.int32)
TOLERANCE = 1e-4
@require_sentencepiece
@require_tokenizers
@require_tf
class TFMBartModelIntegrationTest(unittest.TestCase):
src_text = [
" UN Chief Says There Is No Military Solution in Syria",
]
expected_text = [
"Şeful ONU declară că nu există o soluţie militară în Siria",
]
model_name = "facebook/mbart-large-en-ro"
@cached_property
def tokenizer(self):
return AutoTokenizer.from_pretrained(self.model_name)
@cached_property
def model(self):
model = TFAutoModelForSeq2SeqLM.from_pretrained(self.model_name)
return model
def _assert_generated_batch_equal_expected(self, **tokenizer_kwargs):
generated_words = self.translate_src_text(**tokenizer_kwargs)
self.assertListEqual(self.expected_text, generated_words)
def translate_src_text(self, **tokenizer_kwargs):
model_inputs = self.tokenizer(self.src_text, **tokenizer_kwargs, return_tensors="tf")
generated_ids = self.model.generate(
model_inputs.input_ids, attention_mask=model_inputs.attention_mask, num_beams=2
)
generated_words = self.tokenizer.batch_decode(generated_ids, skip_special_tokens=True)
return generated_words
@slow
def test_batch_generation_en_ro(self):
self._assert_generated_batch_equal_expected()

330
tests/models/mbart/test_tokenization_mbart.py Normal file
View File

@@ -0,0 +1,330 @@
# Copyright 2020 The HuggingFace 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 shutil
import tempfile
import unittest
from transformers import SPIECE_UNDERLINE, BatchEncoding, MBartTokenizer, MBartTokenizerFast, is_torch_available
from transformers.testing_utils import (
get_tests_dir,
nested_simplify,
require_sentencepiece,
require_tokenizers,
require_torch,
)
from ...test_tokenization_common import TokenizerTesterMixin
SAMPLE_VOCAB = get_tests_dir("fixtures/test_sentencepiece.model")
if is_torch_available():
from transformers.models.mbart.modeling_mbart import shift_tokens_right
EN_CODE = 250004
RO_CODE = 250020
@require_sentencepiece
@require_tokenizers
class MBartTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
tokenizer_class = MBartTokenizer
rust_tokenizer_class = MBartTokenizerFast
test_rust_tokenizer = True
test_sentencepiece = True
def setUp(self):
super().setUp()
# We have a SentencePiece fixture for testing
tokenizer = MBartTokenizer(SAMPLE_VOCAB, keep_accents=True)
tokenizer.save_pretrained(self.tmpdirname)
def test_full_tokenizer(self):
tokenizer = MBartTokenizer(SAMPLE_VOCAB, keep_accents=True)
tokens = tokenizer.tokenize("This is a test")
self.assertListEqual(tokens, ["▁This", "▁is", "▁a", "▁t", "est"])
self.assertListEqual(
tokenizer.convert_tokens_to_ids(tokens),
[value + tokenizer.fairseq_offset for value in [285, 46, 10, 170, 382]],
)
tokens = tokenizer.tokenize("I was born in 92000, and this is falsé.")
self.assertListEqual(
tokens,
[
SPIECE_UNDERLINE + "I",
SPIECE_UNDERLINE + "was",
SPIECE_UNDERLINE + "b",
"or",
"n",
SPIECE_UNDERLINE + "in",
SPIECE_UNDERLINE + "",
"9",
"2",
"0",
"0",
"0",
",",
SPIECE_UNDERLINE + "and",
SPIECE_UNDERLINE + "this",
SPIECE_UNDERLINE + "is",
SPIECE_UNDERLINE + "f",
"al",
"s",
"é",
".",
],
)
ids = tokenizer.convert_tokens_to_ids(tokens)
self.assertListEqual(
ids,
[
value + tokenizer.fairseq_offset
for value in [8, 21, 84, 55, 24, 19, 7, 2, 602, 347, 347, 347, 3, 12, 66, 46, 72, 80, 6, 2, 4]
# ^ unk: 2 + 1 = 3 unk: 2 + 1 = 3 ^
],
)
back_tokens = tokenizer.convert_ids_to_tokens(ids)
self.assertListEqual(
back_tokens,
[
SPIECE_UNDERLINE + "I",
SPIECE_UNDERLINE + "was",
SPIECE_UNDERLINE + "b",
"or",
"n",
SPIECE_UNDERLINE + "in",
SPIECE_UNDERLINE + "",
"<unk>",
"2",
"0",
"0",
"0",
",",
SPIECE_UNDERLINE + "and",
SPIECE_UNDERLINE + "this",
SPIECE_UNDERLINE + "is",
SPIECE_UNDERLINE + "f",
"al",
"s",
"<unk>",
".",
],
)
# overwrite from test_tokenization_common to speed up test
def test_save_pretrained(self):
if not self.test_slow_tokenizer:
# as we don't have a slow version, we can't compare the outputs between slow and fast versions
return
self.tokenizers_list[0] = (self.rust_tokenizer_class, "hf-internal-testing/tiny-random-mbart", {})
for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"):
tokenizer_r = self.rust_tokenizer_class.from_pretrained(pretrained_name, **kwargs)
tokenizer_p = self.tokenizer_class.from_pretrained(pretrained_name, **kwargs)
tmpdirname2 = tempfile.mkdtemp()
tokenizer_r_files = tokenizer_r.save_pretrained(tmpdirname2)
tokenizer_p_files = tokenizer_p.save_pretrained(tmpdirname2)
# Checks it save with the same files + the tokenizer.json file for the fast one
self.assertTrue(any("tokenizer.json" in f for f in tokenizer_r_files))
tokenizer_r_files = tuple(f for f in tokenizer_r_files if "tokenizer.json" not in f)
self.assertSequenceEqual(tokenizer_r_files, tokenizer_p_files)
# Checks everything loads correctly in the same way
tokenizer_rp = tokenizer_r.from_pretrained(tmpdirname2)
tokenizer_pp = tokenizer_p.from_pretrained(tmpdirname2)
# Check special tokens are set accordingly on Rust and Python
for key in tokenizer_pp.special_tokens_map:
self.assertTrue(hasattr(tokenizer_rp, key))
# self.assertEqual(getattr(tokenizer_rp, key), getattr(tokenizer_pp, key))
# self.assertEqual(getattr(tokenizer_rp, key + "_id"), getattr(tokenizer_pp, key + "_id"))
shutil.rmtree(tmpdirname2)
# Save tokenizer rust, legacy_format=True
tmpdirname2 = tempfile.mkdtemp()
tokenizer_r_files = tokenizer_r.save_pretrained(tmpdirname2, legacy_format=True)
tokenizer_p_files = tokenizer_p.save_pretrained(tmpdirname2)
# Checks it save with the same files
self.assertSequenceEqual(tokenizer_r_files, tokenizer_p_files)
# Checks everything loads correctly in the same way
tokenizer_rp = tokenizer_r.from_pretrained(tmpdirname2)
tokenizer_pp = tokenizer_p.from_pretrained(tmpdirname2)
# Check special tokens are set accordingly on Rust and Python
for key in tokenizer_pp.special_tokens_map:
self.assertTrue(hasattr(tokenizer_rp, key))
shutil.rmtree(tmpdirname2)
# Save tokenizer rust, legacy_format=False
tmpdirname2 = tempfile.mkdtemp()
tokenizer_r_files = tokenizer_r.save_pretrained(tmpdirname2, legacy_format=False)
tokenizer_p_files = tokenizer_p.save_pretrained(tmpdirname2)
# Checks it saved the tokenizer.json file
self.assertTrue(any("tokenizer.json" in f for f in tokenizer_r_files))
# Checks everything loads correctly in the same way
tokenizer_rp = tokenizer_r.from_pretrained(tmpdirname2)
tokenizer_pp = tokenizer_p.from_pretrained(tmpdirname2)
# Check special tokens are set accordingly on Rust and Python
for key in tokenizer_pp.special_tokens_map:
self.assertTrue(hasattr(tokenizer_rp, key))
shutil.rmtree(tmpdirname2)
@require_torch
@require_sentencepiece
@require_tokenizers
class MBartEnroIntegrationTest(unittest.TestCase):
checkpoint_name = "facebook/mbart-large-en-ro"
src_text = [
" UN Chief Says There Is No Military Solution in Syria",
""" Secretary-General Ban Ki-moon says his response to Russia's stepped up military support for Syria is that "there is no military solution" to the nearly five-year conflict and more weapons will only worsen the violence and misery for millions of people.""",
]
tgt_text = [
"Şeful ONU declară că nu există o soluţie militară în Siria",
'Secretarul General Ban Ki-moon declară că răspunsul său la intensificarea sprijinului militar al Rusiei pentru Siria este că "nu există o soluţie militară" la conflictul de aproape cinci ani şi că noi arme nu vor face decât să înrăutăţească violenţele şi mizeria pentru milioane de oameni.',
]
expected_src_tokens = [8274, 127873, 25916, 7, 8622, 2071, 438, 67485, 53, 187895, 23, 51712, 2, EN_CODE]
@classmethod
def setUpClass(cls):
cls.tokenizer: MBartTokenizer = MBartTokenizer.from_pretrained(
cls.checkpoint_name, src_lang="en_XX", tgt_lang="ro_RO"
)
cls.pad_token_id = 1
return cls
def check_language_codes(self):
self.assertEqual(self.tokenizer.fairseq_tokens_to_ids["ar_AR"], 250001)
self.assertEqual(self.tokenizer.fairseq_tokens_to_ids["en_EN"], 250004)
self.assertEqual(self.tokenizer.fairseq_tokens_to_ids["ro_RO"], 250020)
def test_enro_tokenizer_batch_encode_plus(self):
ids = self.tokenizer.batch_encode_plus(self.src_text).input_ids[0]
self.assertListEqual(self.expected_src_tokens, ids)
def test_enro_tokenizer_decode_ignores_language_codes(self):
self.assertIn(RO_CODE, self.tokenizer.all_special_ids)
generated_ids = [RO_CODE, 884, 9019, 96, 9, 916, 86792, 36, 18743, 15596, 5, 2]
result = self.tokenizer.decode(generated_ids, skip_special_tokens=True)
expected_romanian = self.tokenizer.decode(generated_ids[1:], skip_special_tokens=True)
self.assertEqual(result, expected_romanian)
self.assertNotIn(self.tokenizer.eos_token, result)
def test_enro_tokenizer_truncation(self):
src_text = ["this is gunna be a long sentence " * 20]
assert isinstance(src_text[0], str)
desired_max_length = 10
ids = self.tokenizer(src_text, max_length=desired_max_length, truncation=True).input_ids[0]
self.assertEqual(ids[-2], 2)
self.assertEqual(ids[-1], EN_CODE)
self.assertEqual(len(ids), desired_max_length)
def test_mask_token(self):
self.assertListEqual(self.tokenizer.convert_tokens_to_ids(["<mask>", "ar_AR"]), [250026, 250001])
def test_special_tokens_unaffacted_by_save_load(self):
tmpdirname = tempfile.mkdtemp()
original_special_tokens = self.tokenizer.fairseq_tokens_to_ids
self.tokenizer.save_pretrained(tmpdirname)
new_tok = MBartTokenizer.from_pretrained(tmpdirname)
self.assertDictEqual(new_tok.fairseq_tokens_to_ids, original_special_tokens)
@require_torch
def test_batch_fairseq_parity(self):
batch = self.tokenizer(self.src_text, padding=True)
with self.tokenizer.as_target_tokenizer():
targets = self.tokenizer(self.tgt_text, padding=True, return_tensors="pt")
labels = targets["input_ids"]
batch["decoder_input_ids"] = shift_tokens_right(labels, self.tokenizer.pad_token_id).tolist()
# fairseq batch: https://gist.github.com/sshleifer/cba08bc2109361a74ac3760a7e30e4f4
assert batch.input_ids[1][-2:] == [2, EN_CODE]
assert batch.decoder_input_ids[1][0] == RO_CODE
assert batch.decoder_input_ids[1][-1] == 2
assert labels[1][-2:].tolist() == [2, RO_CODE]
@require_torch
def test_enro_tokenizer_prepare_batch(self):
batch = self.tokenizer(
self.src_text, padding=True, truncation=True, max_length=len(self.expected_src_tokens), return_tensors="pt"
)
with self.tokenizer.as_target_tokenizer():
targets = self.tokenizer(
self.tgt_text,
padding=True,
truncation=True,
max_length=len(self.expected_src_tokens),
return_tensors="pt",
)
labels = targets["input_ids"]
batch["decoder_input_ids"] = shift_tokens_right(labels, self.tokenizer.pad_token_id)
self.assertIsInstance(batch, BatchEncoding)
self.assertEqual((2, 14), batch.input_ids.shape)
self.assertEqual((2, 14), batch.attention_mask.shape)
result = batch.input_ids.tolist()[0]
self.assertListEqual(self.expected_src_tokens, result)
self.assertEqual(2, batch.decoder_input_ids[0, -1]) # EOS
# Test that special tokens are reset
self.assertEqual(self.tokenizer.prefix_tokens, [])
self.assertEqual(self.tokenizer.suffix_tokens, [self.tokenizer.eos_token_id, EN_CODE])
def test_seq2seq_max_length(self):
batch = self.tokenizer(self.src_text, padding=True, truncation=True, max_length=3, return_tensors="pt")
with self.tokenizer.as_target_tokenizer():
targets = self.tokenizer(self.tgt_text, padding=True, truncation=True, max_length=10, return_tensors="pt")
labels = targets["input_ids"]
batch["decoder_input_ids"] = shift_tokens_right(labels, self.tokenizer.pad_token_id)
self.assertEqual(batch.input_ids.shape[1], 3)
self.assertEqual(batch.decoder_input_ids.shape[1], 10)
@require_torch
def test_tokenizer_translation(self):
inputs = self.tokenizer._build_translation_inputs(
"A test", return_tensors="pt", src_lang="en_XX", tgt_lang="ar_AR"
)
self.assertEqual(
nested_simplify(inputs),
{
# A, test, EOS, en_XX
"input_ids": [[62, 3034, 2, 250004]],
"attention_mask": [[1, 1, 1, 1]],
# ar_AR
"forced_bos_token_id": 250001,
},
)