Spaces:
Running
Running
File size: 22,225 Bytes
455a40f |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 |
# coding=utf-8
# Copyright 2023 The Intel Labs Team Authors, The Microsoft Research Team Authors and 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 BridgeTower model. """
import tempfile
import unittest
import numpy as np
from transformers import BridgeTowerConfig, is_torch_available, is_vision_available
from transformers.testing_utils import require_torch, require_vision, slow, torch_device
from transformers.utils import cached_property
from ...test_configuration_common import ConfigTester
from ...test_modeling_common import (
ModelTesterMixin,
_config_zero_init,
floats_tensor,
ids_tensor,
random_attention_mask,
)
from ...test_pipeline_mixin import PipelineTesterMixin
if is_torch_available():
import torch
from transformers import (
BridgeTowerForContrastiveLearning,
BridgeTowerForImageAndTextRetrieval,
BridgeTowerForMaskedLM,
BridgeTowerModel,
)
from transformers.models.bridgetower.modeling_bridgetower import BRIDGETOWER_PRETRAINED_MODEL_ARCHIVE_LIST
from transformers.pytorch_utils import is_torch_greater_or_equal_than_1_10
else:
is_torch_greater_or_equal_than_1_10 = False
if is_vision_available():
from PIL import Image
from transformers import BridgeTowerProcessor
class BridgeTowerModelTester:
def __init__(
self,
parent,
share_cross_modal_transformer_layers=True,
drop_rate=0.1,
head_hidden_scale=2,
hidden_act="gelu",
hidden_size=768,
initializer_factor=1,
is_encoder_decoder=False,
layer_norm_eps=1e-05,
share_link_tower_layers=False,
link_tower_type="add",
num_attention_heads=12,
num_hidden_layers=6,
tie_word_embeddings=False,
init_layernorm_from_vision_encoder=False,
output_hidden_states=False,
text_config=None,
vision_config=None,
image_size=288,
contrastive_hidden_size=512,
logit_scale_init_value=2.6592,
):
self.parent = parent
self.share_cross_modal_transformer_layers = share_cross_modal_transformer_layers
self.drop_rate = drop_rate
self.head_hidden_scale = head_hidden_scale
self.hidden_act = hidden_act
self.hidden_size = hidden_size
self.initializer_factor = initializer_factor
self.is_encoder_decoder = is_encoder_decoder
self.layer_norm_eps = layer_norm_eps
self.share_link_tower_layers = share_link_tower_layers
self.link_tower_type = link_tower_type
self.num_attention_heads = num_attention_heads
self.num_hidden_layers = num_hidden_layers
self.tie_word_embeddings = tie_word_embeddings
self.init_layernorm_from_vision_encoder = init_layernorm_from_vision_encoder
self.vocab_size = 99
self.num_channels = 3
self.seq_length = 4
self.num_image_features = 325
self.batch_size = 1
self.image_size = image_size
self.is_training = False
self.expected_num_hidden_layers = 32
self.output_hidden_states = output_hidden_states
self.contrastive_hidden_size = contrastive_hidden_size
self.logit_scale_init_value = logit_scale_init_value
def prepare_config_and_inputs(self):
input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
attention_mask = random_attention_mask([self.batch_size, self.seq_length])
pixel_values = floats_tensor([self.batch_size, self.num_channels, self.image_size, self.image_size])
pixel_mask = random_attention_mask([self.batch_size, self.image_size, self.image_size])
config = self.get_config()
return (config, input_ids, attention_mask, pixel_values, pixel_mask)
def get_config(self):
return BridgeTowerConfig(
share_cross_modal_transformer_layers=self.share_cross_modal_transformer_layers,
drop_rate=self.drop_rate,
head_hidden_scale=self.head_hidden_scale,
hidden_act=self.hidden_act,
hidden_size=self.hidden_size,
initializer_factor=self.initializer_factor,
image_size=self.image_size,
is_encoder_decoder=self.is_encoder_decoder,
layer_norm_eps=self.layer_norm_eps,
share_link_tower_layers=self.share_link_tower_layers,
link_tower_type=self.link_tower_type,
num_attention_heads=self.num_attention_heads,
num_hidden_layers=self.num_hidden_layers,
tie_word_embeddings=self.tie_word_embeddings,
init_layernorm_from_vision_encoder=self.init_layernorm_from_vision_encoder,
num_channels=self.num_channels,
output_hidden_states=self.output_hidden_states,
contrastive_hidden_size=self.contrastive_hidden_size,
logit_scale_init_value=self.logit_scale_init_value,
)
def create_and_check_model(
self,
config,
input_ids,
attention_mask,
pixel_values,
pixel_mask,
):
model = BridgeTowerModel(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=attention_mask, pixel_values=pixel_values, pixel_mask=pixel_mask)
result = model(input_ids, attention_mask=attention_mask, pixel_values=pixel_values)
self.parent.assertEqual(result["text_features"].shape, (self.batch_size, self.seq_length, self.hidden_size))
self.parent.assertEqual(
result["image_features"].shape, (self.batch_size, self.num_image_features, self.hidden_size)
)
self.parent.assertEqual(result["pooler_output"].shape, (self.batch_size, 2 * self.hidden_size))
def create_and_check_for_image_and_text_retrieval(
self,
config,
input_ids,
attention_mask,
pixel_values,
pixel_mask,
):
bridgetower_itm_output_last_dimension = 2
model = BridgeTowerForImageAndTextRetrieval(config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=attention_mask, pixel_values=pixel_values, pixel_mask=pixel_mask)
result = model(input_ids, attention_mask=attention_mask, pixel_values=pixel_values)
self.parent.assertEqual(result.logits.shape, (self.batch_size, bridgetower_itm_output_last_dimension))
def create_and_check_for_masked_language_modeling(
self,
config,
input_ids,
attention_mask,
pixel_values,
pixel_mask,
):
model = BridgeTowerForMaskedLM(config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=attention_mask, pixel_values=pixel_values, pixel_mask=pixel_mask)
result = model(input_ids, attention_mask=attention_mask, pixel_values=pixel_values)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, 50265))
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
(config, input_ids, attention_mask, pixel_values, pixel_mask) = config_and_inputs
inputs_dict = {
"input_ids": input_ids,
"attention_mask": attention_mask,
"pixel_values": pixel_values,
"pixel_mask": pixel_mask,
}
return config, inputs_dict
@slow
@require_torch
@unittest.skipIf(not is_torch_greater_or_equal_than_1_10, "BridgeTower is only available in torch v1.10+")
class BridgeTowerModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
all_model_classes = (
(
BridgeTowerModel,
BridgeTowerForImageAndTextRetrieval,
BridgeTowerForMaskedLM,
BridgeTowerForContrastiveLearning,
)
if is_torch_available()
else ()
)
pipeline_model_mapping = {"feature-extraction": BridgeTowerModel} if is_torch_available() else {}
is_training = False
test_headmasking = False
test_pruning = False
test_torchscript = False
test_resize_embeddings = False
has_attentions = False
# function to extract meaningful tensor from output per different model_class
def extract_output(self, outputs, model_class):
return outputs["pooler_output"] if model_class == "BridgeTowerModel" else outputs["logits"]
def setUp(self):
self.model_tester = BridgeTowerModelTester(self)
self.config_tester = ConfigTester(self, config_class=BridgeTowerConfig, hidden_size=37, vocab_size=99)
def test_config(self):
self.config_tester.run_common_tests()
def test_model(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_model(*config_and_inputs)
def test_for_image_and_text_retrieval(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_for_image_and_text_retrieval(*config_and_inputs)
def test_for_masked_language_modeling(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_for_masked_language_modeling(*config_and_inputs)
@slow
def test_model_from_pretrained(self):
for model_name in BRIDGETOWER_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
model = BridgeTowerModel.from_pretrained(model_name)
self.assertIsNotNone(model)
@slow
def test_save_load_fast_init_from_base(self):
# Override as it is a slow test on this model
super().test_save_load_fast_init_from_base()
# Override as extracting meaningful tensor from output is different for BridgeTower
def test_save_load(self):
config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(**input_dict)
out_2 = self.extract_output(outputs, model_class.__name__)
out_2 = out_2.cpu().numpy()
out_2[np.isnan(out_2)] = 0
with tempfile.TemporaryDirectory() as tmpdirname:
model.save_pretrained(tmpdirname)
model = model_class.from_pretrained(tmpdirname)
model.to(torch_device)
with torch.no_grad():
after_outputs = model(**input_dict)
# Make sure we don't have nans
out_1 = self.extract_output(after_outputs, model_class.__name__)
out_1 = out_1.cpu().numpy()
out_1[np.isnan(out_1)] = 0
max_diff = np.amax(np.abs(out_1 - out_2))
self.assertLessEqual(max_diff, 1e-5)
# Override this as `hidden states output` is different for BridgeTower
def test_hidden_states_output(self):
def check_hidden_states_output(inputs_dict, config, model_class):
model = model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(**self._prepare_for_class(inputs_dict, model_class))
hidden_states_text, hidden_states_vision, hidden_states_cross = (
outputs.encoder_hidden_states if config.is_encoder_decoder else outputs.hidden_states
)
expected_num_layers = getattr(
self.model_tester, "expected_num_hidden_layers", self.model_tester.num_hidden_layers + 1
)
self.assertEqual(
sum((len(hidden_states_text), len(hidden_states_vision), len(hidden_states_cross))),
expected_num_layers,
)
seq_length = self.model_tester.seq_length
num_image_features = self.model_tester.num_image_features
self.assertListEqual(
list(hidden_states_text[0].shape[-2:]),
[seq_length, self.model_tester.hidden_size],
)
self.assertListEqual(
list(hidden_states_vision[0].shape),
[num_image_features, 1, self.model_tester.hidden_size],
)
self.assertListEqual(
list(hidden_states_cross[0][0].shape[-2:]),
[seq_length, self.model_tester.hidden_size],
)
self.assertListEqual(
list(hidden_states_cross[0][1].shape[-2:]),
[num_image_features, self.model_tester.hidden_size],
)
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
inputs_dict["output_hidden_states"] = True
check_hidden_states_output(inputs_dict, config, model_class)
# check that output_hidden_states also work using config
del inputs_dict["output_hidden_states"]
config.output_hidden_states = True
check_hidden_states_output(inputs_dict, config, model_class)
# Override as `hidden states output` is different for BridgeTower
def test_retain_grad_hidden_states_attentions(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.output_hidden_states = True
config.output_attentions = self.has_attentions
# no need to test all models as different heads yield the same functionality
model_class = self.all_model_classes[0]
model = model_class(config)
model.to(torch_device)
inputs = self._prepare_for_class(inputs_dict, model_class)
outputs = model(**inputs)
output = outputs[0]
# Encoder-/Decoder-only models
hidden_states = outputs.hidden_states[0][0]
hidden_states.retain_grad()
if self.has_attentions:
attentions = outputs.attentions[0][0]
attentions.retain_grad()
output.flatten()[0].backward(retain_graph=True)
self.assertIsNotNone(hidden_states.grad)
if self.has_attentions:
self.assertIsNotNone(attentions.grad)
# override as the `logit_scale` parameter initilization is different for BRIDGE TOWER
def test_initialization(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
configs_no_init = _config_zero_init(config)
for model_class in self.all_model_classes:
model = model_class(config=configs_no_init)
for name, param in model.named_parameters():
if param.requires_grad:
if name == "logit_scale":
self.assertAlmostEqual(
param.data.item(),
config.logit_scale_init_value,
delta=1e-3,
msg=f"Parameter {name} of model {model_class} seems not properly initialized",
)
else:
self.assertIn(
((param.data.mean() * 1e9).round() / 1e9).item(),
[0.0, 1.0],
msg=f"Parameter {name} of model {model_class} seems not properly initialized",
)
@unittest.skip(reason="""Bridge Tower does not have input/output embeddings. So this test is not applicable.""")
def test_model_common_attributes(self):
pass
@unittest.skip(reason="""Bridge Tower does not have input/output embeddings. Thus this test is not applicable.""")
def test_inputs_embeds(self):
pass
# We will verify our results on an image of cute cats
def prepare_img():
image = Image.open("./tests/fixtures/tests_samples/COCO/000000039769.png")
return image
@require_torch
@require_vision
@unittest.skipIf(not is_torch_greater_or_equal_than_1_10, "BridgeTower is only available in torch v1.10+")
class BridgeTowerModelIntegrationTest(unittest.TestCase):
@cached_property
def default_processor(self):
return (
BridgeTowerProcessor.from_pretrained("BridgeTower/bridgetower-base-itm-mlm")
if is_vision_available()
else None
)
@slow
def test_image_and_text_retrieval(self):
model = BridgeTowerForImageAndTextRetrieval.from_pretrained("BridgeTower/bridgetower-base-itm-mlm").to(
torch_device
)
model.eval()
processor = self.default_processor
image = prepare_img()
text = "a bunch of cats laying on a tower."
inputs = processor(image, text, return_tensors="pt").to(torch_device)
# forward pass
with torch.no_grad():
outputs = model(**inputs)
# verify the logits
expected_shape = torch.Size([1, 2])
self.assertEqual(outputs.logits.shape, expected_shape)
self.assertTrue(outputs.logits[0, 1].item() > outputs.logits[0, 0].item())
# verify loss
inputs["labels"] = torch.ones(1, dtype=torch.long, device=torch_device)
inputs = inputs.to(torch_device)
with torch.no_grad():
outputs = model(**inputs)
self.assertAlmostEqual(outputs.loss.item(), 0.5108, places=4)
@slow
def test_masked_language_modeling(self):
model = BridgeTowerForMaskedLM.from_pretrained("BridgeTower/bridgetower-base-itm-mlm").to(torch_device)
model.eval()
processor = self.default_processor
image = prepare_img()
text = "a bunch of <mask> laying on a tower."
inputs = processor(image, text, return_tensors="pt").to(torch_device)
# forward pass
with torch.no_grad():
outputs = model(**inputs)
# verify the logits
expected_shape = torch.Size([1, 11, 50265])
self.assertEqual(outputs.logits.shape, expected_shape)
# verify predicted word
predicted_id = outputs.logits.argmax(dim=-1).squeeze(0).tolist()[4]
self.assertTrue(processor.decode([predicted_id]) == " cats")
# verify loss
inputs["labels"] = inputs["input_ids"].clone()
inputs = inputs.to(torch_device)
with torch.no_grad():
outputs = model(**inputs)
self.assertAlmostEqual(outputs.loss.item(), 5.7373, places=4)
@slow
def test_constrastive_learning(self):
model = BridgeTowerForContrastiveLearning.from_pretrained("BridgeTower/bridgetower-large-itm-mlm-itc").to(
torch_device
)
model.eval()
processor = BridgeTowerProcessor.from_pretrained("BridgeTower/bridgetower-large-itm-mlm-itc")
image = prepare_img()
text = "a bunch of cats laying on a tower."
inputs = processor(image, text, padding=True, return_tensors="pt").to(torch_device)
with torch.no_grad():
outputs = model(**inputs, output_hidden_states=True, return_loss=True)
# verify the logits
expected_shape = torch.Size([1, 3, 512])
self.assertEqual(outputs.logits.shape, expected_shape)
@slow
@require_torch
@unittest.skipIf(not is_torch_greater_or_equal_than_1_10, "BridgeTower is only available in torch v1.10+")
class BridgeTowerModelTrainingTest(unittest.TestCase):
all_training_supported_model_classes = (
(BridgeTowerForImageAndTextRetrieval, BridgeTowerForMaskedLM, BridgeTowerForContrastiveLearning)
if is_torch_available()
else ()
)
def setUp(self):
self.model_tester = BridgeTowerModelTester(self)
self.config_tester = ConfigTester(self, config_class=BridgeTowerConfig, hidden_size=37, vocab_size=99)
def _prepare_inputs_for_training(self, model_class):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
if model_class == BridgeTowerForMaskedLM:
inputs_dict["labels"] = inputs_dict["input_ids"]
elif model_class == BridgeTowerForImageAndTextRetrieval:
inputs_dict["labels"] = ids_tensor([1], 2)
elif model_class == BridgeTowerForContrastiveLearning:
inputs_dict["return_loss"] = True
return config, inputs_dict
def _get_non_used_layer_names(self, model_class):
non_used_layer_names = ["text_model.pooler"]
if model_class == BridgeTowerForMaskedLM:
non_used_layer_names = non_used_layer_names + [
"cross_modal_image_layers.5",
"cross_modal_image_pooler",
"cross_modal_text_pooler",
]
return non_used_layer_names
def _is_layer_used(self, model_class, layer_name):
non_used_layer_names = self._get_non_used_layer_names(model_class)
for non_used_layer_name in non_used_layer_names:
if non_used_layer_name in layer_name:
return False
return True
def test_training(self):
for model_class in self.all_training_supported_model_classes:
config, inputs_dict = self._prepare_inputs_for_training(model_class)
model = model_class(config)
model.to(torch_device)
model.train()
loss = model(**inputs_dict).loss
loss.backward()
# verify the gradients of used layers' weight are not None
for name, param in model.named_parameters():
if self._is_layer_used(model_class, name):
self.assertIsNotNone(param.grad, f"Gradients should not be None - got {param.grad} for {name}")
|