EngineX-Cambricon/enginex-mlu370-any2any
10
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

init

Browse Source
This commit is contained in:
luopingyi
2025-10-09 16:47:16 +08:00
parent c8feb4deb5
commit e27e3f16bb
5248 changed files with 1778505 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

0
transformers/tests/models/clip/__init__.py Normal file
View File

128
transformers/tests/models/clip/test_image_processing_clip.py Normal file
View File

@@ -0,0 +1,128 @@
# Copyright 2021 HuggingFace Inc.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a 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.testing_utils import require_torch, require_vision
from transformers.utils import is_torchvision_available, is_vision_available
from ...test_image_processing_common import ImageProcessingTestMixin, prepare_image_inputs
if is_vision_available():
from transformers import CLIPImageProcessor
if is_torchvision_available():
from transformers import CLIPImageProcessorFast
class CLIPImageProcessingTester:
def __init__(
self,
parent,
batch_size=7,
num_channels=3,
image_size=18,
min_resolution=30,
max_resolution=400,
do_resize=True,
size=None,
do_center_crop=True,
crop_size=None,
do_normalize=True,
image_mean=[0.48145466, 0.4578275, 0.40821073],
image_std=[0.26862954, 0.26130258, 0.27577711],
do_convert_rgb=True,
):
super().__init__()
size = size if size is not None else {"shortest_edge": 20}
crop_size = crop_size if crop_size is not None else {"height": 18, "width": 18}
self.parent = parent
self.batch_size = batch_size
self.num_channels = num_channels
self.image_size = image_size
self.min_resolution = min_resolution
self.max_resolution = max_resolution
self.do_resize = do_resize
self.size = size
self.do_center_crop = do_center_crop
self.crop_size = crop_size
self.do_normalize = do_normalize
self.image_mean = image_mean
self.image_std = image_std
self.do_convert_rgb = do_convert_rgb
def prepare_image_processor_dict(self):
return {
"do_resize": self.do_resize,
"size": self.size,
"do_center_crop": self.do_center_crop,
"crop_size": self.crop_size,
"do_normalize": self.do_normalize,
"image_mean": self.image_mean,
"image_std": self.image_std,
"do_convert_rgb": self.do_convert_rgb,
}
def expected_output_image_shape(self, images):
return self.num_channels, self.crop_size["height"], self.crop_size["width"]
def prepare_image_inputs(self, equal_resolution=False, numpify=False, torchify=False):
return prepare_image_inputs(
batch_size=self.batch_size,
num_channels=self.num_channels,
min_resolution=self.min_resolution,
max_resolution=self.max_resolution,
equal_resolution=equal_resolution,
numpify=numpify,
torchify=torchify,
)
@require_torch
@require_vision
class CLIPImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
image_processing_class = CLIPImageProcessor if is_vision_available() else None
fast_image_processing_class = CLIPImageProcessorFast if is_torchvision_available() else None
def setUp(self):
super().setUp()
self.image_processor_tester = CLIPImageProcessingTester(self)
@property
def image_processor_dict(self):
return self.image_processor_tester.prepare_image_processor_dict()
def test_image_processor_properties(self):
for image_processing_class in self.image_processor_list:
image_processing = image_processing_class(**self.image_processor_dict)
self.assertTrue(hasattr(image_processing, "do_resize"))
self.assertTrue(hasattr(image_processing, "size"))
self.assertTrue(hasattr(image_processing, "do_center_crop"))
self.assertTrue(hasattr(image_processing, "center_crop"))
self.assertTrue(hasattr(image_processing, "do_normalize"))
self.assertTrue(hasattr(image_processing, "image_mean"))
self.assertTrue(hasattr(image_processing, "image_std"))
self.assertTrue(hasattr(image_processing, "do_convert_rgb"))
def test_image_processor_from_dict_with_kwargs(self):
for image_processing_class in self.image_processor_list:
image_processor = image_processing_class.from_dict(self.image_processor_dict)
self.assertEqual(image_processor.size, {"shortest_edge": 20})
self.assertEqual(image_processor.crop_size, {"height": 18, "width": 18})
image_processor = image_processing_class.from_dict(self.image_processor_dict, size=42, crop_size=84)
self.assertEqual(image_processor.size, {"shortest_edge": 42})
self.assertEqual(image_processor.crop_size, {"height": 84, "width": 84})

846
transformers/tests/models/clip/test_modeling_clip.py Normal file
View File

@@ -0,0 +1,846 @@
# 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 CLIP model."""
import inspect
import os
import tempfile
import unittest
import numpy as np
import pytest
import requests
from parameterized import parameterized
from transformers import CLIPConfig, CLIPTextConfig, CLIPVisionConfig
from transformers.testing_utils import (
require_torch,
require_vision,
slow,
torch_device,
)
from transformers.utils import (
is_torch_available,
is_vision_available,
)
from ...test_configuration_common import ConfigTester
from ...test_modeling_common import (
TEST_EAGER_MATCHES_SDPA_INFERENCE_PARAMETERIZATION,
ModelTesterMixin,
_config_zero_init,
floats_tensor,
ids_tensor,
is_flaky,
random_attention_mask,
)
from ...test_pipeline_mixin import PipelineTesterMixin
if is_torch_available():
import torch
from torch import nn
from transformers import (
CLIPForImageClassification,
CLIPModel,
CLIPTextModel,
CLIPTextModelWithProjection,
CLIPVisionModel,
CLIPVisionModelWithProjection,
)
if is_vision_available():
from PIL import Image
from transformers import CLIPProcessor
class CLIPVisionModelTester:
def __init__(
self,
parent,
batch_size=12,
image_size=30,
patch_size=2,
num_channels=3,
is_training=True,
hidden_size=32,
projection_dim=32,
num_hidden_layers=2,
num_attention_heads=4,
intermediate_size=37,
dropout=0.1,
attention_dropout=0.1,
initializer_range=0.02,
scope=None,
):
self.parent = parent
self.batch_size = batch_size
self.image_size = image_size
self.patch_size = patch_size
self.num_channels = num_channels
self.is_training = is_training
self.hidden_size = hidden_size
self.projection_dim = projection_dim
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.intermediate_size = intermediate_size
self.dropout = dropout
self.attention_dropout = attention_dropout
self.initializer_range = initializer_range
self.scope = scope
# in ViT, the seq length equals the number of patches + 1 (we add 1 for the [CLS] token)
num_patches = (image_size // patch_size) ** 2
self.seq_length = num_patches + 1
def prepare_config_and_inputs(self):
pixel_values = floats_tensor([self.batch_size, self.num_channels, self.image_size, self.image_size])
config = self.get_config()
return config, pixel_values
def get_config(self):
return CLIPVisionConfig(
image_size=self.image_size,
patch_size=self.patch_size,
num_channels=self.num_channels,
hidden_size=self.hidden_size,
projection_dim=self.projection_dim,
num_hidden_layers=self.num_hidden_layers,
num_attention_heads=self.num_attention_heads,
intermediate_size=self.intermediate_size,
dropout=self.dropout,
attention_dropout=self.attention_dropout,
initializer_range=self.initializer_range,
)
def create_and_check_model(self, config, pixel_values):
model = CLIPVisionModel(config=config)
model.to(torch_device)
model.eval()
with torch.no_grad():
result = model(pixel_values)
# expected sequence length = num_patches + 1 (we add 1 for the [CLS] token)
image_size = (self.image_size, self.image_size)
patch_size = (self.patch_size, self.patch_size)
num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, num_patches + 1, self.hidden_size))
self.parent.assertEqual(result.pooler_output.shape, (self.batch_size, self.hidden_size))
def create_and_check_model_with_projection(self, config, pixel_values):
model = CLIPVisionModelWithProjection(config=config)
model.to(torch_device)
model.eval()
with torch.no_grad():
result = model(pixel_values)
# expected sequence length = num_patches + 1 (we add 1 for the [CLS] token)
image_size = (self.image_size, self.image_size)
patch_size = (self.patch_size, self.patch_size)
num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, num_patches + 1, self.hidden_size))
self.parent.assertEqual(result.image_embeds.shape, (self.batch_size, self.projection_dim))
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
config, pixel_values = config_and_inputs
inputs_dict = {"pixel_values": pixel_values}
return config, inputs_dict
@parameterized.expand(TEST_EAGER_MATCHES_SDPA_INFERENCE_PARAMETERIZATION)
def test_eager_matches_sdpa_inference(self, *args):
return getattr(ModelTesterMixin, self._testMethodName)(self)
class CLIPModelTesterMixin(ModelTesterMixin):
"""
Subclass of ModelTesterMixin with methods specific to testing CLIP models.
The SDPA equivalence test is overridden here because CLIP models may have test/vision/text+vision inputs,
different output logits, and are not supposed to be used or tested with padding_side="left".
"""
def test_sdpa_can_dispatch_composite_models(self):
for model_class in self.all_model_classes:
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
model = model_class(config)
with tempfile.TemporaryDirectory() as tmpdirname:
model.save_pretrained(tmpdirname)
# Load the model with SDPA (it is the default, but we explicit it for clarity)
model_sdpa = model_class.from_pretrained(tmpdirname, attn_implementation="sdpa")
model_sdpa = model_sdpa.eval().to(torch_device)
# Load model with eager attention
model_eager = model_class.from_pretrained(
tmpdirname,
attn_implementation="eager",
)
model_eager = model_eager.eval().to(torch_device)
if hasattr(model_sdpa, "vision_model"):
self.assertTrue(model_sdpa.vision_model.config._attn_implementation == "sdpa")
self.assertTrue(model_eager.vision_model.config._attn_implementation == "eager")
if hasattr(model_sdpa, "text_model"):
self.assertTrue(model_sdpa.text_model.config._attn_implementation == "sdpa")
self.assertTrue(model_eager.text_model.config._attn_implementation == "eager")
self.assertTrue(model_sdpa.config._attn_implementation == "sdpa")
self.assertTrue(model_eager.config._attn_implementation == "eager")
@require_torch
class CLIPVisionModelTest(CLIPModelTesterMixin, unittest.TestCase):
"""
Here we also overwrite some of the tests of test_modeling_common.py, as CLIP does not use input_ids, inputs_embeds,
attention_mask and seq_length.
"""
all_model_classes = (CLIPVisionModel, CLIPVisionModelWithProjection) if is_torch_available() else ()
fx_compatible = True
test_pruning = False
test_resize_embeddings = False
test_head_masking = False
def setUp(self):
self.model_tester = CLIPVisionModelTester(self)
self.config_tester = ConfigTester(self, config_class=CLIPVisionConfig, has_text_modality=False, hidden_size=37)
def test_config(self):
self.config_tester.run_common_tests()
@unittest.skip(reason="CLIP does not use inputs_embeds")
def test_inputs_embeds(self):
pass
def test_model_get_set_embeddings(self):
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
self.assertIsInstance(model.get_input_embeddings(), (nn.Module))
x = model.get_output_embeddings()
self.assertTrue(x is None or isinstance(x, nn.Linear))
def test_forward_signature(self):
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
signature = inspect.signature(model.forward)
# signature.parameters is an OrderedDict => so arg_names order is deterministic
arg_names = [*signature.parameters.keys()]
expected_arg_names = ["pixel_values"]
self.assertListEqual(arg_names[:1], expected_arg_names)
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_model_with_projection(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_model_with_projection(*config_and_inputs)
@unittest.skip
def test_training(self):
pass
@unittest.skip
def test_training_gradient_checkpointing(self):
pass
@unittest.skip(
reason="This architecture seem to not compute gradients properly when using GC, check: https://github.com/huggingface/transformers/pull/27124"
)
def test_training_gradient_checkpointing_use_reentrant(self):
pass
@unittest.skip(
reason="This architecture seem to not compute gradients properly when using GC, check: https://github.com/huggingface/transformers/pull/27124"
)
def test_training_gradient_checkpointing_use_reentrant_false(self):
pass
@slow
def test_model_from_pretrained(self):
model_name = "openai/clip-vit-base-patch32"
model = CLIPVisionModel.from_pretrained(model_name)
self.assertIsNotNone(model)
@slow
def test_model_with_projection_from_pretrained(self):
model_name = "openai/clip-vit-base-patch32"
model = CLIPVisionModelWithProjection.from_pretrained(model_name)
self.assertIsNotNone(model)
self.assertTrue(hasattr(model, "visual_projection"))
@parameterized.expand(TEST_EAGER_MATCHES_SDPA_INFERENCE_PARAMETERIZATION)
@is_flaky()
def test_eager_matches_sdpa_inference(self, *args):
# adding only flaky decorator here and call the parent test method
return getattr(ModelTesterMixin, self._testMethodName)(self)
def test_sdpa_can_dispatch_composite_models(self):
super().test_sdpa_can_dispatch_composite_models()
class CLIPTextModelTester:
def __init__(
self,
parent,
batch_size=12,
seq_length=7,
is_training=True,
use_input_mask=True,
use_labels=True,
vocab_size=99,
hidden_size=32,
projection_dim=32,
num_hidden_layers=2,
num_attention_heads=4,
intermediate_size=37,
dropout=0.1,
attention_dropout=0.1,
max_position_embeddings=512,
initializer_range=0.02,
scope=None,
):
self.parent = parent
self.batch_size = batch_size
self.seq_length = seq_length
self.is_training = is_training
self.use_input_mask = use_input_mask
self.use_labels = use_labels
self.vocab_size = vocab_size
self.hidden_size = hidden_size
self.projection_dim = projection_dim
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.intermediate_size = intermediate_size
self.dropout = dropout
self.attention_dropout = attention_dropout
self.max_position_embeddings = max_position_embeddings
self.initializer_range = initializer_range
self.scope = scope
def prepare_config_and_inputs(self):
input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
input_mask = None
if self.use_input_mask:
input_mask = random_attention_mask([self.batch_size, self.seq_length])
if input_mask is not None:
batch_size, seq_length = input_mask.shape
rnd_start_indices = np.random.randint(1, seq_length - 1, size=(batch_size,))
for batch_idx, start_index in enumerate(rnd_start_indices):
input_mask[batch_idx, :start_index] = 1
input_mask[batch_idx, start_index:] = 0
config = self.get_config()
return config, input_ids, input_mask
def get_config(self):
return CLIPTextConfig(
vocab_size=self.vocab_size,
hidden_size=self.hidden_size,
projection_dim=self.projection_dim,
num_hidden_layers=self.num_hidden_layers,
num_attention_heads=self.num_attention_heads,
intermediate_size=self.intermediate_size,
dropout=self.dropout,
attention_dropout=self.attention_dropout,
max_position_embeddings=self.max_position_embeddings,
initializer_range=self.initializer_range,
)
def create_and_check_model(self, config, input_ids, input_mask):
model = CLIPTextModel(config=config)
model.to(torch_device)
model.eval()
with torch.no_grad():
result = model(input_ids, attention_mask=input_mask)
result = model(input_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
self.parent.assertEqual(result.pooler_output.shape, (self.batch_size, self.hidden_size))
def create_and_check_model_with_projection(self, config, input_ids, input_mask):
model = CLIPTextModelWithProjection(config=config)
model.to(torch_device)
model.eval()
with torch.no_grad():
result = model(input_ids, attention_mask=input_mask)
result = model(input_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
self.parent.assertEqual(result.text_embeds.shape, (self.batch_size, self.projection_dim))
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
config, input_ids, input_mask = config_and_inputs
inputs_dict = {"input_ids": input_ids, "attention_mask": input_mask}
return config, inputs_dict
@require_torch
class CLIPTextModelTest(CLIPModelTesterMixin, unittest.TestCase):
all_model_classes = (CLIPTextModel, CLIPTextModelWithProjection) if is_torch_available() else ()
fx_compatible = True
test_pruning = False
test_head_masking = False
model_split_percents = [0.5, 0.8, 0.9]
def setUp(self):
self.model_tester = CLIPTextModelTester(self)
self.config_tester = ConfigTester(self, config_class=CLIPTextConfig, hidden_size=37)
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_model_with_projection(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_model_with_projection(*config_and_inputs)
@unittest.skip
def test_training(self):
pass
@unittest.skip
def test_training_gradient_checkpointing(self):
pass
@unittest.skip(
reason="This architecture seem to not compute gradients properly when using GC, check: https://github.com/huggingface/transformers/pull/27124"
)
def test_training_gradient_checkpointing_use_reentrant(self):
pass
@unittest.skip(
reason="This architecture seem to not compute gradients properly when using GC, check: https://github.com/huggingface/transformers/pull/27124"
)
def test_training_gradient_checkpointing_use_reentrant_false(self):
pass
@unittest.skip(reason="CLIP does not use inputs_embeds")
def test_inputs_embeds(self):
pass
@slow
def test_model_from_pretrained(self):
model_name = "openai/clip-vit-base-patch32"
model = CLIPTextModel.from_pretrained(model_name)
self.assertIsNotNone(model)
@slow
def test_model_with_projection_from_pretrained(self):
model_name = "openai/clip-vit-base-patch32"
model = CLIPTextModelWithProjection.from_pretrained(model_name)
self.assertIsNotNone(model)
self.assertTrue(hasattr(model, "text_projection"))
@parameterized.expand(TEST_EAGER_MATCHES_SDPA_INFERENCE_PARAMETERIZATION)
@slow
@is_flaky()
def test_eager_matches_sdpa_inference(self, *args):
# adding only flaky decorator here and call the parent test method
return getattr(ModelTesterMixin, self._testMethodName)(self)
def test_sdpa_can_dispatch_composite_models(self):
super().test_sdpa_can_dispatch_composite_models()
def test_sdpa_can_dispatch_on_flash(self):
self.skipTest(reason="CLIPTextModel has two attention masks: `causal_attention_mask` and `attention_mask`")
class CLIPModelTester:
def __init__(self, parent, text_kwargs=None, vision_kwargs=None, is_training=True):
if text_kwargs is None:
text_kwargs = {}
if vision_kwargs is None:
vision_kwargs = {}
self.parent = parent
self.text_model_tester = CLIPTextModelTester(parent, **text_kwargs)
self.vision_model_tester = CLIPVisionModelTester(parent, **vision_kwargs)
self.batch_size = self.text_model_tester.batch_size # need bs for batching_equivalence test
self.is_training = is_training
def prepare_config_and_inputs(self):
text_config, input_ids, attention_mask = self.text_model_tester.prepare_config_and_inputs()
vision_config, pixel_values = self.vision_model_tester.prepare_config_and_inputs()
config = self.get_config()
return config, input_ids, attention_mask, pixel_values
def get_config(self):
return CLIPConfig(
text_config=self.text_model_tester.get_config().to_dict(),
vision_config=self.vision_model_tester.get_config().to_dict(),
projection_dim=64,
)
def create_and_check_model(self, config, input_ids, attention_mask, pixel_values):
model = CLIPModel(config).to(torch_device).eval()
with torch.no_grad():
result = model(input_ids, pixel_values, attention_mask)
self.parent.assertEqual(
result.logits_per_image.shape, (self.vision_model_tester.batch_size, self.text_model_tester.batch_size)
)
self.parent.assertEqual(
result.logits_per_text.shape, (self.text_model_tester.batch_size, self.vision_model_tester.batch_size)
)
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
config, input_ids, attention_mask, pixel_values = config_and_inputs
inputs_dict = {
"input_ids": input_ids,
"attention_mask": attention_mask,
"pixel_values": pixel_values,
"return_loss": True,
}
return config, inputs_dict
@require_torch
class CLIPModelTest(CLIPModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
all_model_classes = (CLIPModel,) if is_torch_available() else ()
pipeline_model_mapping = (
{"feature-extraction": CLIPModel, "image-feature-extraction": CLIPVisionModel} if is_torch_available() else {}
)
additional_model_inputs = ["pixel_values"]
fx_compatible = True
test_head_masking = False
test_pruning = False
test_resize_embeddings = False
test_attention_outputs = False
_is_composite = True
def setUp(self):
self.model_tester = CLIPModelTester(self)
common_properties = ["projection_dim", "logit_scale_init_value"]
self.config_tester = ConfigTester(
self, config_class=CLIPConfig, has_text_modality=False, common_properties=common_properties
)
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_config(self):
self.config_tester.run_common_tests()
@unittest.skip(reason="Hidden_states is tested in individual model tests")
def test_hidden_states_output(self):
pass
@unittest.skip(reason="Inputs_embeds is tested in individual model tests")
def test_inputs_embeds(self):
pass
@unittest.skip(reason="Retain_grad is tested in individual model tests")
def test_retain_grad_hidden_states_attentions(self):
pass
@unittest.skip(reason="CLIPModel does not have input/output embeddings")
def test_model_get_set_embeddings(self):
pass
# override as the `logit_scale` parameter initialization is different for CLIP
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:
# check if `logit_scale` is initialized as per the original implementation
if name == "logit_scale":
self.assertAlmostEqual(
param.data.item(),
np.log(1 / 0.07),
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",
)
def _create_and_check_torchscript(self, config, inputs_dict):
if not self.test_torchscript:
self.skipTest(reason="test_torchscript is set to False")
configs_no_init = _config_zero_init(config) # To be sure we have no Nan
configs_no_init.torchscript = True
configs_no_init.return_dict = False
for model_class in self.all_model_classes:
model = model_class(config=configs_no_init)
model.to(torch_device)
model.eval()
try:
input_ids = inputs_dict["input_ids"]
pixel_values = inputs_dict["pixel_values"] # CLIP needs pixel_values
traced_model = torch.jit.trace(model, (input_ids, pixel_values))
except RuntimeError:
self.fail("Couldn't trace module.")
with tempfile.TemporaryDirectory() as tmp_dir_name:
pt_file_name = os.path.join(tmp_dir_name, "traced_model.pt")
try:
torch.jit.save(traced_model, pt_file_name)
except Exception:
self.fail("Couldn't save module.")
try:
loaded_model = torch.jit.load(pt_file_name)
except Exception:
self.fail("Couldn't load module.")
model.to(torch_device)
model.eval()
loaded_model.to(torch_device)
loaded_model.eval()
model_state_dict = model.state_dict()
loaded_model_state_dict = loaded_model.state_dict()
non_persistent_buffers = {}
for key in loaded_model_state_dict:
if key not in model_state_dict:
non_persistent_buffers[key] = loaded_model_state_dict[key]
loaded_model_state_dict = {
key: value for key, value in loaded_model_state_dict.items() if key not in non_persistent_buffers
}
self.assertEqual(set(model_state_dict.keys()), set(loaded_model_state_dict.keys()))
model_buffers = list(model.buffers())
for non_persistent_buffer in non_persistent_buffers.values():
found_buffer = False
for i, model_buffer in enumerate(model_buffers):
if torch.equal(non_persistent_buffer, model_buffer):
found_buffer = True
break
self.assertTrue(found_buffer)
model_buffers.pop(i)
models_equal = True
for layer_name, p1 in model_state_dict.items():
p2 = loaded_model_state_dict[layer_name]
if p1.data.ne(p2.data).sum() > 0:
models_equal = False
self.assertTrue(models_equal)
def test_load_vision_text_config(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
# Save CLIPConfig and check if we can load CLIPVisionConfig from it
with tempfile.TemporaryDirectory() as tmp_dir_name:
config.save_pretrained(tmp_dir_name)
vision_config = CLIPVisionConfig.from_pretrained(tmp_dir_name)
self.assertDictEqual(config.vision_config.to_dict(), vision_config.to_dict())
# Save CLIPConfig and check if we can load CLIPTextConfig from it
with tempfile.TemporaryDirectory() as tmp_dir_name:
config.save_pretrained(tmp_dir_name)
text_config = CLIPTextConfig.from_pretrained(tmp_dir_name)
self.assertDictEqual(config.text_config.to_dict(), text_config.to_dict())
@slow
def test_model_from_pretrained(self):
model_name = "openai/clip-vit-base-patch32"
model = CLIPModel.from_pretrained(model_name)
self.assertIsNotNone(model)
@parameterized.expand(TEST_EAGER_MATCHES_SDPA_INFERENCE_PARAMETERIZATION)
@slow
@is_flaky()
def test_eager_matches_sdpa_inference(self, *args):
# adding only flaky decorator here and call the parent test method
return getattr(ModelTesterMixin, self._testMethodName)(self)
def test_sdpa_can_dispatch_composite_models(self):
super().test_sdpa_can_dispatch_composite_models()
def test_sdpa_can_dispatch_on_flash(self):
self.skipTest(reason="CLIP text tower has two attention masks: `causal_attention_mask` and `attention_mask`")
@pytest.mark.torch_compile_test
def test_sdpa_can_compile_dynamic(self):
self.skipTest(reason="CLIP model can't be compiled dynamic, error in clip_loss`")
class CLIPForImageClassificationModelTester(CLIPModelTester):
def __init__(self, parent):
super().__init__(parent)
self.batch_size = self.vision_model_tester.batch_size
self.num_hidden_layers = self.vision_model_tester.num_hidden_layers
self.hidden_size = self.vision_model_tester.hidden_size
self.seq_length = self.vision_model_tester.seq_length
def prepare_config_and_inputs(self):
_, pixel_values = self.vision_model_tester.prepare_config_and_inputs()
config = self.get_config()
return config, pixel_values
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
config, pixel_values = config_and_inputs
inputs_dict = {"pixel_values": pixel_values}
return config, inputs_dict
@require_torch
class CLIPForImageClassificationModelTest(CLIPModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
all_model_classes = (CLIPForImageClassification,) if is_torch_available() else ()
pipeline_model_mapping = {"image-classification": CLIPForImageClassification} if is_torch_available() else {}
fx_compatible = False
test_head_masking = False
test_pruning = False
test_resize_embeddings = False
test_attention_outputs = False
_is_composite = True
def setUp(self):
self.model_tester = CLIPForImageClassificationModelTester(self)
@unittest.skip(reason="CLIPForImageClassification does not support inputs_embeds")
def test_inputs_embeds(self):
pass
@unittest.skip(reason="CLIPForImageClassification does not support inputs_embeds")
def test_model_get_set_embeddings(self):
pass
@unittest.skip(reason="CLIPForImageClassification does not support gradient checkpointing yet")
def test_training_gradient_checkpointing(self):
pass
@unittest.skip(reason="CLIPForImageClassification does not support gradient checkpointing yet")
def test_training_gradient_checkpointing_use_reentrant(self):
pass
@unittest.skip(reason="CLIPForImageClassification does not support gradient checkpointing yet")
def test_training_gradient_checkpointing_use_reentrant_false(self):
pass
@unittest.skip(reason="CLIP uses a non-standard initialization scheme")
def test_initialization(self):
pass
@parameterized.expand(TEST_EAGER_MATCHES_SDPA_INFERENCE_PARAMETERIZATION)
@slow
@is_flaky()
def test_eager_matches_sdpa_inference(self, *args):
# adding only flaky decorator here and call the parent test method
return getattr(ModelTesterMixin, self._testMethodName)(self)
def test_sdpa_can_dispatch_composite_models(self):
super().test_sdpa_can_dispatch_composite_models()
# We will verify our results on an image of cute cats
def prepare_img():
url = "http://images.cocodataset.org/val2017/000000039769.jpg"
im = Image.open(requests.get(url, stream=True).raw)
return im
@require_vision
@require_torch
class CLIPModelIntegrationTest(unittest.TestCase):
@slow
def test_inference(self):
model_name = "openai/clip-vit-base-patch32"
model = CLIPModel.from_pretrained(model_name, attn_implementation="sdpa").to(torch_device)
processor = CLIPProcessor.from_pretrained(model_name)
image = prepare_img()
inputs = processor(
text=["a photo of a cat", "a photo of a dog"], images=image, padding=True, return_tensors="pt"
).to(torch_device)
# forward pass
with torch.no_grad():
outputs = model(**inputs)
# verify the logits
self.assertEqual(
outputs.logits_per_image.shape,
torch.Size((inputs.pixel_values.shape[0], inputs.input_ids.shape[0])),
)
self.assertEqual(
outputs.logits_per_text.shape,
torch.Size((inputs.input_ids.shape[0], inputs.pixel_values.shape[0])),
)
expected_logits = torch.tensor([[24.5701, 19.3049]], device=torch_device)
torch.testing.assert_close(outputs.logits_per_image, expected_logits, rtol=1e-3, atol=1e-3)
@slow
def test_inference_interpolate_pos_encoding(self):
# CLIP models have an `interpolate_pos_encoding` argument in their forward method,
# allowing to interpolate the pre-trained position embeddings in order to use
# the model on higher resolutions. The DINO model by Facebook AI leverages this
# to visualize self-attention on higher resolution images.
model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32").to(torch_device)
processor = CLIPProcessor.from_pretrained(
"openai/clip-vit-base-patch32", size={"height": 180, "width": 180}, crop_size={"height": 180, "width": 180}
)
image = Image.open("./tests/fixtures/tests_samples/COCO/000000039769.png")
inputs = processor(text="what's in the image", images=image, return_tensors="pt").to(torch_device)
# interpolate_pos_encodiung false should return value error
with self.assertRaises(ValueError, msg="doesn't match model"):
with torch.no_grad():
model(**inputs, interpolate_pos_encoding=False)
# forward pass
with torch.no_grad():
outputs = model(**inputs, interpolate_pos_encoding=True)
# verify the logits
expected_shape = torch.Size((1, 26, 768))
self.assertEqual(outputs.vision_model_output.last_hidden_state.shape, expected_shape)
expected_slice = torch.tensor(
[[-0.1538, 0.0322, -0.3235], [0.2893, 0.1135, -0.5708], [0.0461, 0.1540, -0.6018]]
).to(torch_device)
torch.testing.assert_close(
outputs.vision_model_output.last_hidden_state[0, :3, :3], expected_slice, rtol=6e-3, atol=4e-4
)

168
transformers/tests/models/clip/test_processing_clip.py Normal file
View File

@@ -0,0 +1,168 @@
# 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 shutil
import tempfile
import unittest
import pytest
from transformers import AutoTokenizer, CLIPTokenizer, CLIPTokenizerFast
from transformers.testing_utils import require_vision
from transformers.utils import is_vision_available
from ...test_processing_common import ProcessorTesterMixin
if is_vision_available():
from transformers import CLIPImageProcessor, CLIPProcessor
TEST_MODEL_PATH = "openai/clip-vit-base-patch32"
@require_vision
class CLIPProcessorTest(ProcessorTesterMixin, unittest.TestCase):
processor_class = CLIPProcessor
@classmethod
def setUpClass(cls):
cls.tmpdirname = tempfile.mkdtemp()
tokenizer = AutoTokenizer.from_pretrained(TEST_MODEL_PATH)
image_processor = CLIPImageProcessor.from_pretrained(TEST_MODEL_PATH)
processor = CLIPProcessor(
image_processor=image_processor,
tokenizer=tokenizer,
)
processor.save_pretrained(cls.tmpdirname)
@classmethod
def get_tokenizer(cls, **kwargs):
return CLIPTokenizer.from_pretrained(cls.tmpdirname, **kwargs)
@classmethod
def get_rust_tokenizer(cls, **kwargs):
return CLIPTokenizerFast.from_pretrained(cls.tmpdirname, **kwargs)
@classmethod
def get_image_processor(cls, **kwargs):
return CLIPImageProcessor.from_pretrained(cls.tmpdirname, **kwargs)
@classmethod
def tearDownClass(cls):
shutil.rmtree(cls.tmpdirname)
def test_save_load_pretrained_default(self):
tokenizer_slow = self.get_tokenizer()
tokenizer_fast = self.get_rust_tokenizer()
image_processor = self.get_image_processor()
with tempfile.TemporaryDirectory() as tmpdir:
processor_slow = CLIPProcessor(tokenizer=tokenizer_slow, image_processor=image_processor)
processor_slow.save_pretrained(tmpdir)
processor_slow = CLIPProcessor.from_pretrained(tmpdir, use_fast=False)
processor_fast = CLIPProcessor(tokenizer=tokenizer_fast, image_processor=image_processor)
processor_fast.save_pretrained(tmpdir)
processor_fast = CLIPProcessor.from_pretrained(tmpdir)
self.assertEqual(processor_slow.tokenizer.get_vocab(), tokenizer_slow.get_vocab())
self.assertEqual(processor_fast.tokenizer.get_vocab(), tokenizer_fast.get_vocab())
self.assertEqual(tokenizer_slow.get_vocab(), tokenizer_fast.get_vocab())
self.assertIsInstance(processor_slow.tokenizer, CLIPTokenizer)
self.assertIsInstance(processor_fast.tokenizer, CLIPTokenizerFast)
self.assertEqual(processor_slow.image_processor.to_json_string(), image_processor.to_json_string())
self.assertEqual(processor_fast.image_processor.to_json_string(), image_processor.to_json_string())
self.assertIsInstance(processor_slow.image_processor, CLIPImageProcessor)
self.assertIsInstance(processor_fast.image_processor, CLIPImageProcessor)
def test_save_load_pretrained_additional_features(self):
with tempfile.TemporaryDirectory() as tmpdir:
processor = CLIPProcessor(tokenizer=self.get_tokenizer(), image_processor=self.get_image_processor())
processor.save_pretrained(tmpdir)
tokenizer_add_kwargs = CLIPTokenizer.from_pretrained(tmpdir, bos_token="(BOS)", eos_token="(EOS)")
image_processor_add_kwargs = CLIPImageProcessor.from_pretrained(
tmpdir, do_normalize=False, padding_value=1.0
)
processor = CLIPProcessor.from_pretrained(
tmpdir, bos_token="(BOS)", eos_token="(EOS)", do_normalize=False, padding_value=1.0
)
self.assertEqual(processor.tokenizer.get_vocab(), tokenizer_add_kwargs.get_vocab())
self.assertIsInstance(processor.tokenizer, CLIPTokenizerFast)
self.assertEqual(processor.image_processor.to_json_string(), image_processor_add_kwargs.to_json_string())
self.assertIsInstance(processor.image_processor, CLIPImageProcessor)
def test_image_processor(self):
image_processor = self.get_image_processor()
tokenizer = self.get_tokenizer()
processor = CLIPProcessor(tokenizer=tokenizer, image_processor=image_processor)
image_input = self.prepare_image_inputs()
input_image_proc = image_processor(image_input, return_tensors="np")
input_processor = processor(images=image_input, return_tensors="np")
for key in input_image_proc:
self.assertAlmostEqual(input_image_proc[key].sum(), input_processor[key].sum(), delta=1e-2)
def test_tokenizer(self):
image_processor = self.get_image_processor()
tokenizer = self.get_tokenizer()
processor = CLIPProcessor(tokenizer=tokenizer, image_processor=image_processor)
input_str = "lower newer"
encoded_processor = processor(text=input_str)
encoded_tok = tokenizer(input_str)
for key in encoded_tok:
self.assertListEqual(encoded_tok[key], encoded_processor[key])
def test_processor(self):
image_processor = self.get_image_processor()
tokenizer = self.get_tokenizer()
processor = CLIPProcessor(tokenizer=tokenizer, image_processor=image_processor)
input_str = "lower newer"
image_input = self.prepare_image_inputs()
inputs = processor(text=input_str, images=image_input)
self.assertSetEqual(set(inputs.keys()), {"input_ids", "attention_mask", "pixel_values"})
# test if it raises when no input is passed
with pytest.raises(ValueError):
processor()
def test_tokenizer_decode(self):
image_processor = self.get_image_processor()
tokenizer = self.get_tokenizer()
processor = CLIPProcessor(tokenizer=tokenizer, image_processor=image_processor)
predicted_ids = [[1, 4, 5, 8, 1, 0, 8], [3, 4, 3, 1, 1, 8, 9]]
decoded_processor = processor.batch_decode(predicted_ids)
decoded_tok = tokenizer.batch_decode(predicted_ids)
self.assertListEqual(decoded_tok, decoded_processor)

187
transformers/tests/models/clip/test_tokenization_clip.py Normal file
View File

@@ -0,0 +1,187 @@
# 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 json
import os
import unittest
from transformers import CLIPTokenizer, CLIPTokenizerFast
from transformers.models.clip.tokenization_clip import VOCAB_FILES_NAMES
from transformers.testing_utils import require_ftfy, require_tokenizers
from ...test_tokenization_common import TokenizerTesterMixin
@require_tokenizers
class CLIPTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
from_pretrained_id = "openai/clip-vit-base-patch32"
tokenizer_class = CLIPTokenizer
rust_tokenizer_class = CLIPTokenizerFast
test_rust_tokenizer = True
from_pretrained_kwargs = {}
test_seq2seq = False
@classmethod
def setUpClass(cls):
super().setUpClass()
vocab = ["l", "o", "w", "e", "r", "s", "t", "i", "d", "n", "lo", "l</w>", "w</w>", "r</w>", "t</w>", "low</w>", "er</w>", "lowest</w>", "newer</w>", "wider", "<unk>", "<|startoftext|>", "<|endoftext|>"] # fmt: skip
vocab_tokens = dict(zip(vocab, range(len(vocab))))
merges = ["#version: 0.2", "l o", "lo w</w>", "e r</w>"]
cls.special_tokens_map = {"unk_token": "<unk>"}
cls.vocab_file = os.path.join(cls.tmpdirname, VOCAB_FILES_NAMES["vocab_file"])
cls.merges_file = os.path.join(cls.tmpdirname, VOCAB_FILES_NAMES["merges_file"])
with open(cls.vocab_file, "w", encoding="utf-8") as fp:
fp.write(json.dumps(vocab_tokens) + "\n")
with open(cls.merges_file, "w", encoding="utf-8") as fp:
fp.write("\n".join(merges))
@classmethod
def get_tokenizer(cls, pretrained_name=None, **kwargs):
kwargs.update(cls.special_tokens_map)
pretrained_name = pretrained_name or cls.tmpdirname
return CLIPTokenizer.from_pretrained(pretrained_name, **kwargs)
@classmethod
def get_rust_tokenizer(cls, pretrained_name=None, **kwargs):
kwargs.update(cls.special_tokens_map)
pretrained_name = pretrained_name or cls.tmpdirname
return CLIPTokenizerFast.from_pretrained(pretrained_name, **kwargs)
def get_input_output_texts(self, tokenizer):
input_text = "lower newer"
output_text = "lower newer"
return input_text, output_text
def test_full_tokenizer(self):
tokenizer = CLIPTokenizer(self.vocab_file, self.merges_file, **self.special_tokens_map)
text = "lower newer"
bpe_tokens = ["lo", "w", "er</w>", "n", "e", "w", "er</w>"]
tokens = tokenizer.tokenize(text)
self.assertListEqual(tokens, bpe_tokens)
input_tokens = tokens + [tokenizer.unk_token]
input_bpe_tokens = [10, 2, 16, 9, 3, 2, 16, 20]
self.assertListEqual(tokenizer.convert_tokens_to_ids(input_tokens), input_bpe_tokens)
@require_ftfy
def test_check_encoding_slow_fast(self):
for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"):
tokenizer_s = self.get_tokenizer(pretrained_name, **kwargs)
tokenizer_r = self.get_rust_tokenizer(pretrained_name, **kwargs)
text = "A\n'll 11p223RF☆ho!!to?'d'd''d of a cat to-$''d."
text_tokenized_s = tokenizer_s.tokenize(text)
text_tokenized_r = tokenizer_r.tokenize(text)
self.assertListEqual(text_tokenized_s, text_tokenized_r)
# Test that the tokenization is identical on an example containing a character (Latin Small Letter A
# with Tilde) encoded in 2 different ways
text = "xa\u0303y" + " " + "x\xe3y"
text_tokenized_s = tokenizer_s.tokenize(text)
text_tokenized_r = tokenizer_r.tokenize(text)
self.assertListEqual(text_tokenized_s, text_tokenized_r)
# Test that the tokenization is identical on unicode of space type
spaces_unicodes = [
"\u0009", # (horizontal tab, '\t')
"\u000b", # (vertical tab)
"\u000c", # (form feed)
"\u0020", # (space, ' ')
"\u200e", # (left-to-right mark):w
"\u200f", # (right-to-left mark)
]
for unicode_seq in spaces_unicodes:
text_tokenized_s = tokenizer_s.tokenize(unicode_seq)
text_tokenized_r = tokenizer_r.tokenize(unicode_seq)
self.assertListEqual(text_tokenized_s, text_tokenized_r)
# Test that the tokenization is identical on unicode of line break type
line_break_unicodes = [
"\u000a", # (line feed, '\n')
"\r\n", # (carriage return and line feed, '\r\n')
"\u000d", # (carriage return, '\r')
"\r", # (carriage return, '\r')
"\u000d", # (carriage return, '\r')
"\u2028", # (line separator)
"\u2029", # (paragraph separator)
# "\u0085", # (next line)
]
# The tokenization is not identical for the character "\u0085" (next line). The slow version using ftfy transforms
# it into the Horizontal Ellipsis character "…" ("\u2026") while the fast version transforms it into a
# space (and thus into an empty list).
for unicode_seq in line_break_unicodes:
text_tokenized_s = tokenizer_s.tokenize(unicode_seq)
text_tokenized_r = tokenizer_r.tokenize(unicode_seq)
self.assertListEqual(text_tokenized_s, text_tokenized_r)
def test_offsets_mapping_with_different_add_prefix_space_argument(self):
# Test which aims to verify that the offsets are well adapted to the argument `add_prefix_space`
for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"):
text_of_1_token = "hello" # `hello` is a token in the vocabulary of `pretrained_name`
text = f"{text_of_1_token} {text_of_1_token}"
tokenizer_r = self.get_rust_tokenizer(
pretrained_name,
use_fast=True,
)
encoding = tokenizer_r(text, return_offsets_mapping=True, add_special_tokens=False)
self.assertEqual(encoding.offset_mapping[0], (0, len(text_of_1_token)))
self.assertEqual(
encoding.offset_mapping[1],
(len(text_of_1_token) + 1, len(text_of_1_token) + 1 + len(text_of_1_token)),
)
text = f" {text}"
tokenizer_r = self.get_rust_tokenizer(
pretrained_name,
use_fast=True,
)
encoding = tokenizer_r(text, return_offsets_mapping=True, add_special_tokens=False)
self.assertEqual(encoding.offset_mapping[0], (1, 1 + len(text_of_1_token)))
self.assertEqual(
encoding.offset_mapping[1],
(1 + len(text_of_1_token) + 1, 1 + len(text_of_1_token) + 1 + len(text_of_1_token)),
)
def test_log_warning(self):
# Test related to the breaking change introduced in transformers v4.17.0
# We need to check that an error in raised when the user try to load a previous version of the tokenizer.
with self.assertRaises(TypeError) as context:
self.get_rust_tokenizer("robot-test/old-clip-tokenizer")
self.assertTrue(
context.exception.args[0].startswith(
"The `backend_tokenizer` provided does not match the expected format."
)
)
@require_ftfy
def test_tokenization_python_rust_equals(self):
super().test_tokenization_python_rust_equals()
@unittest.skip(reason="CLIP always lower cases letters")
def test_added_tokens_do_lower_case(self):
pass