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

Improve vision models (#17731)

Browse Source 
* Improve vision models

* Add a lot of improvements

* Remove to_2tuple from swin tests

* Fix TF Swin

* Fix more tests

* Fix copies

* Improve more models

* Fix ViTMAE test

* Add channel check for TF models

* Add proper channel check for TF models

* Apply suggestion from code review

* Apply suggestions from code review

* Add channel check for Flax models, apply suggestion

* Fix bug

* Add tests for greyscale images

* Add test for interpolation of pos encodigns

Co-authored-by: Niels Rogge <nielsrogge@Nielss-MacBook-Pro.local>
This commit is contained in:
NielsRogge
2022-06-24 11:34:51 +02:00
committed by GitHub
parent 893ab12452
commit 0917870510
39 changed files with 801 additions and 916 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
tests/models/beit/test_modeling_beit.py
View File

@@ -153,6 +153,16 @@ class BeitModelTester:
result = model(pixel_values, labels=labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.type_sequence_label_size))
# test greyscale images
config.num_channels = 1
model = BeitForImageClassification(config)
model.to(torch_device)
model.eval()
pixel_values = floats_tensor([self.batch_size, 1, self.image_size, self.image_size])
result = model(pixel_values, labels=labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.type_sequence_label_size))
def create_and_check_for_semantic_segmentation(self, config, pixel_values, labels, pixel_labels):
config.num_labels = self.num_labels
model = BeitForSemanticSegmentation(config)

8
tests/models/beit/test_modeling_flax_beit.py
View File

@@ -105,7 +105,6 @@ class FlaxBeitModelTester(unittest.TestCase):
return config, pixel_values, labels
def create_and_check_model(self, config, pixel_values, labels):
model = FlaxBeitModel(config=config)
result = model(pixel_values)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
@@ -121,6 +120,13 @@ class FlaxBeitModelTester(unittest.TestCase):
result = model(pixel_values)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.type_sequence_label_size))
# test greyscale images
config.num_channels = 1
model = FlaxBeitForImageClassification(config)
pixel_values = floats_tensor([self.batch_size, 1, self.image_size, self.image_size])
result = model(pixel_values)
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
(

116
tests/models/data2vec/test_modeling_data2vec_vision.py
View File

@@ -37,10 +37,7 @@ if is_torch_available():
Data2VecVisionForSemanticSegmentation,
Data2VecVisionModel,
)
from transformers.models.data2vec.modeling_data2vec_vision import (
DATA2VEC_VISION_PRETRAINED_MODEL_ARCHIVE_LIST,
to_2tuple,
)
from transformers.models.data2vec.modeling_data2vec_vision import DATA2VEC_VISION_PRETRAINED_MODEL_ARCHIVE_LIST
if is_vision_available():
@@ -94,6 +91,10 @@ class Data2VecVisionModelTester:
self.out_indices = out_indices
self.num_labels = num_labels
# in BeiT, 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])
@@ -131,9 +132,7 @@ class Data2VecVisionModelTester:
model.eval()
result = model(pixel_values)
# expected sequence length = num_patches + 1 (we add 1 for the [CLS] token)
image_size = to_2tuple(self.image_size)
patch_size = to_2tuple(self.patch_size)
num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
num_patches = (self.image_size // self.patch_size) ** 2
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, num_patches + 1, self.hidden_size))
def create_and_check_for_image_classification(self, config, pixel_values, labels, pixel_labels):
@@ -286,109 +285,6 @@ class Data2VecVisionModelTest(ModelTesterMixin, unittest.TestCase):
msg=f"Parameter {name} of model {model_class} seems not properly initialized",
)
def test_attention_outputs(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.return_dict = True
# in Data2VecVision, the seq_len equals the number of patches + 1 (we add 1 for the [CLS] token)
image_size = to_2tuple(self.model_tester.image_size)
patch_size = to_2tuple(self.model_tester.patch_size)
num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
seq_len = num_patches + 1
encoder_seq_length = getattr(self.model_tester, "encoder_seq_length", seq_len)
encoder_key_length = getattr(self.model_tester, "key_length", encoder_seq_length)
chunk_length = getattr(self.model_tester, "chunk_length", None)
if chunk_length is not None and hasattr(self.model_tester, "num_hashes"):
encoder_seq_length = encoder_seq_length * self.model_tester.num_hashes
for model_class in self.all_model_classes:
inputs_dict["output_attentions"] = True
inputs_dict["output_hidden_states"] = False
config.return_dict = True
model = model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(**self._prepare_for_class(inputs_dict, model_class))
attentions = outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions
self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
# check that output_attentions also work using config
del inputs_dict["output_attentions"]
config.output_attentions = True
model = model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(**self._prepare_for_class(inputs_dict, model_class))
attentions = outputs.attentions
self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
self.assertListEqual(
list(attentions[0].shape[-3:]),
[self.model_tester.num_attention_heads, encoder_seq_length, encoder_key_length],
)
out_len = len(outputs)
# Check attention is always last and order is fine
inputs_dict["output_attentions"] = True
inputs_dict["output_hidden_states"] = True
model = model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(**self._prepare_for_class(inputs_dict, model_class))
self.assertEqual(out_len + 1, len(outputs))
self_attentions = outputs.attentions
self.assertEqual(len(self_attentions), self.model_tester.num_hidden_layers)
self.assertListEqual(
list(self_attentions[0].shape[-3:]),
[self.model_tester.num_attention_heads, encoder_seq_length, encoder_key_length],
)
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 = 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(len(hidden_states), expected_num_layers)
# Data2VecVision has a different seq_length
image_size = to_2tuple(self.model_tester.image_size)
patch_size = to_2tuple(self.model_tester.patch_size)
num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
seq_length = num_patches + 1
self.assertListEqual(
list(hidden_states[0].shape[-2:]),
[seq_length, 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)
def check_pt_tf_outputs(self, tf_outputs, pt_outputs, model_class, tol=2e-4, name="outputs", attributes=None):
# We override with a slightly higher tol value, as semseg models tend to diverge a bit more
super().check_pt_tf_outputs(tf_outputs, pt_outputs, model_class, tol, name, attributes)

33
tests/models/deit/test_modeling_deit.py
View File

@@ -131,6 +131,25 @@ class DeiTModelTester:
result = model(pixel_values)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_for_masked_image_modeling(self, config, pixel_values, labels):
model = DeiTForMaskedImageModeling(config=config)
model.to(torch_device)
model.eval()
result = model(pixel_values)
self.parent.assertEqual(
result.logits.shape, (self.batch_size, self.num_channels, self.image_size, self.image_size)
)
# test greyscale images
config.num_channels = 1
model = DeiTForMaskedImageModeling(config)
model.to(torch_device)
model.eval()
pixel_values = floats_tensor([self.batch_size, 1, self.image_size, self.image_size])
result = model(pixel_values)
self.parent.assertEqual(result.logits.shape, (self.batch_size, 1, self.image_size, self.image_size))
def create_and_check_for_image_classification(self, config, pixel_values, labels):
config.num_labels = self.type_sequence_label_size
model = DeiTForImageClassification(config)
@@ -139,6 +158,16 @@ class DeiTModelTester:
result = model(pixel_values, labels=labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.type_sequence_label_size))
# test greyscale images
config.num_channels = 1
model = DeiTForImageClassification(config)
model.to(torch_device)
model.eval()
pixel_values = floats_tensor([self.batch_size, 1, self.image_size, self.image_size])
result = model(pixel_values, labels=labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.type_sequence_label_size))
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
(
@@ -208,6 +237,10 @@ class DeiTModelTest(ModelTesterMixin, unittest.TestCase):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_model(*config_and_inputs)
def test_for_masked_image_modeling(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_for_masked_image_modeling(*config_and_inputs)
def test_for_image_classification(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_for_image_classification(*config_and_inputs)

68
tests/models/swin/test_modeling_swin.py
View File

@@ -14,6 +14,7 @@
# limitations under the License.
""" Testing suite for the PyTorch Swin model. """
import collections
import inspect
import os
import pickle
@@ -33,7 +34,7 @@ if is_torch_available():
from torch import nn
from transformers import SwinForImageClassification, SwinForMaskedImageModeling, SwinModel
from transformers.models.swin.modeling_swin import SWIN_PRETRAINED_MODEL_ARCHIVE_LIST, to_2tuple
from transformers.models.swin.modeling_swin import SWIN_PRETRAINED_MODEL_ARCHIVE_LIST
if is_vision_available():
from PIL import Image
@@ -141,6 +142,25 @@ class SwinModelTester:
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, expected_seq_len, expected_dim))
def create_and_check_for_masked_image_modeling(self, config, pixel_values, labels):
model = SwinForMaskedImageModeling(config=config)
model.to(torch_device)
model.eval()
result = model(pixel_values)
self.parent.assertEqual(
result.logits.shape, (self.batch_size, self.num_channels, self.image_size, self.image_size)
)
# test greyscale images
config.num_channels = 1
model = SwinForMaskedImageModeling(config)
model.to(torch_device)
model.eval()
pixel_values = floats_tensor([self.batch_size, 1, self.image_size, self.image_size])
result = model(pixel_values)
self.parent.assertEqual(result.logits.shape, (self.batch_size, 1, self.image_size, self.image_size))
def create_and_check_for_image_classification(self, config, pixel_values, labels):
config.num_labels = self.type_sequence_label_size
model = SwinForImageClassification(config)
@@ -149,6 +169,16 @@ class SwinModelTester:
result = model(pixel_values, labels=labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.type_sequence_label_size))
# test greyscale images
config.num_channels = 1
model = SwinForImageClassification(config)
model.to(torch_device)
model.eval()
pixel_values = floats_tensor([self.batch_size, 1, self.image_size, self.image_size])
result = model(pixel_values)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.type_sequence_label_size))
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
(
@@ -198,6 +228,14 @@ class SwinModelTest(ModelTesterMixin, unittest.TestCase):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_model(*config_and_inputs)
def test_for_masked_image_modeling(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_for_masked_image_modeling(*config_and_inputs)
def test_for_image_classification(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_for_image_classification(*config_and_inputs)
def test_inputs_embeds(self):
# Swin does not use inputs_embeds
pass
@@ -299,7 +337,11 @@ class SwinModelTest(ModelTesterMixin, unittest.TestCase):
self.assertEqual(len(hidden_states), expected_num_layers)
# Swin has a different seq_length
patch_size = to_2tuple(config.patch_size)
patch_size = (
config.patch_size
if isinstance(config.patch_size, collections.abc.Iterable)
else (config.patch_size, config.patch_size)
)
num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
@@ -323,7 +365,11 @@ class SwinModelTest(ModelTesterMixin, unittest.TestCase):
def test_hidden_states_output(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
image_size = to_2tuple(self.model_tester.image_size)
image_size = (
self.model_tester.image_size
if isinstance(self.model_tester.image_size, collections.abc.Iterable)
else (self.model_tester.image_size, self.model_tester.image_size)
)
for model_class in self.all_model_classes:
inputs_dict["output_hidden_states"] = True
@@ -339,8 +385,16 @@ class SwinModelTest(ModelTesterMixin, unittest.TestCase):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.patch_size = 3
image_size = to_2tuple(self.model_tester.image_size)
patch_size = to_2tuple(config.patch_size)
image_size = (
self.model_tester.image_size
if isinstance(self.model_tester.image_size, collections.abc.Iterable)
else (self.model_tester.image_size, self.model_tester.image_size)
)
patch_size = (
config.patch_size
if isinstance(config.patch_size, collections.abc.Iterable)
else (config.patch_size, config.patch_size)
)
padded_height = image_size[0] + patch_size[0] - (image_size[0] % patch_size[0])
padded_width = image_size[1] + patch_size[1] - (image_size[1] % patch_size[1])
@@ -354,10 +408,6 @@ class SwinModelTest(ModelTesterMixin, unittest.TestCase):
config.output_hidden_states = True
self.check_hidden_states_output(inputs_dict, config, model_class, (padded_height, padded_width))
def test_for_image_classification(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_for_image_classification(*config_and_inputs)
@slow
def test_model_from_pretrained(self):
for model_name in SWIN_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:

37
tests/models/swin/test_modeling_tf_swin.py
View File

@@ -21,7 +21,7 @@ import unittest
import numpy as np
from transformers import SwinConfig
from transformers.testing_utils import require_tf, require_vision, slow
from transformers.testing_utils import require_tf, require_vision, slow, to_2tuple
from transformers.utils import cached_property, is_tf_available, is_vision_available
from ...test_configuration_common import ConfigTester
@@ -36,7 +36,6 @@ if is_tf_available():
TFSwinForImageClassification,
TFSwinForMaskedImageModeling,
TFSwinModel,
to_2tuple,
)
@@ -141,12 +140,34 @@ class TFSwinModelTester:
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, expected_seq_len, expected_dim))
def create_and_check_for_masked_image_modeling(self, config, pixel_values, labels):
model = TFSwinForMaskedImageModeling(config=config)
result = model(pixel_values)
self.parent.assertEqual(
result.logits.shape, (self.batch_size, self.num_channels, self.image_size, self.image_size)
)
# test greyscale images
config.num_channels = 1
model = TFSwinForMaskedImageModeling(config)
pixel_values = floats_tensor([self.batch_size, 1, self.image_size, self.image_size])
result = model(pixel_values)
self.parent.assertEqual(result.logits.shape, (self.batch_size, 1, self.image_size, self.image_size))
def create_and_check_for_image_classification(self, config, pixel_values, labels):
config.num_labels = self.type_sequence_label_size
model = TFSwinForImageClassification(config)
result = model(pixel_values, labels=labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.type_sequence_label_size))
# test greyscale images
config.num_channels = 1
model = TFSwinForImageClassification(config)
pixel_values = floats_tensor([self.batch_size, 1, self.image_size, self.image_size])
result = model(pixel_values)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.type_sequence_label_size))
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
config, pixel_values, labels = config_and_inputs
@@ -192,6 +213,14 @@ class TFSwinModelTest(TFModelTesterMixin, unittest.TestCase):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_model(*config_and_inputs)
def test_for_masked_image_modeling(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_for_masked_image_modeling(*config_and_inputs)
def test_for_image_classification(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_for_image_classification(*config_and_inputs)
@unittest.skip(reason="Swin does not use inputs_embeds")
def test_inputs_embeds(self):
pass
@@ -336,10 +365,6 @@ class TFSwinModelTest(TFModelTesterMixin, unittest.TestCase):
config.output_hidden_states = True
self.check_hidden_states_output(inputs_dict, config, model_class, (padded_height, padded_width))
def test_for_image_classification(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_for_image_classification(*config_and_inputs)
@slow
def test_model_from_pretrained(self):
for model_name in TF_SWIN_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:

26
tests/models/vit/test_modeling_flax_vit.py
View File

@@ -91,8 +91,7 @@ class FlaxViTModelTester(unittest.TestCase):
return config, pixel_values
def create_and_check_model(self, config, pixel_values, labels):
def create_and_check_model(self, config, pixel_values):
model = FlaxViTModel(config=config)
result = model(pixel_values)
# expected sequence length = num_patches + 1 (we add 1 for the [CLS] token)
@@ -101,6 +100,19 @@ class FlaxViTModelTester(unittest.TestCase):
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))
def create_and_check_for_image_classification(self, config, pixel_values):
config.num_labels = self.type_sequence_label_size
model = FlaxViTForImageClassification(config=config)
result = model(pixel_values)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.type_sequence_label_size))
# test greyscale images
config.num_channels = 1
model = FlaxViTForImageClassification(config)
pixel_values = floats_tensor([self.batch_size, 1, self.image_size, self.image_size])
result = model(pixel_values)
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
(
@@ -123,7 +135,15 @@ class FlaxViTModelTest(FlaxModelTesterMixin, unittest.TestCase):
def test_config(self):
self.config_tester.run_common_tests()
# We neeed to override this test because ViT's forward signature is different than text models.
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_classification(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_for_image_classification(*config_and_inputs)
# We need to override this test because ViT's forward signature is different than text models.
def test_forward_signature(self):
config, _ = self.model_tester.prepare_config_and_inputs_for_common()

7
tests/models/vit/test_modeling_tf_vit.py
View File

@@ -133,6 +133,13 @@ class TFViTModelTester:
result = model(pixel_values, interpolate_pos_encoding=True, training=False)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.type_sequence_label_size))
# test greyscale images
config.num_channels = 1
model = TFViTForImageClassification(config)
pixel_values = floats_tensor([self.batch_size, 1, self.image_size, self.image_size])
result = model(pixel_values)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.type_sequence_label_size))
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
config, pixel_values, labels = config_and_inputs

60
tests/models/vit/test_modeling_vit.py
View File

@@ -120,6 +120,25 @@ class ViTModelTester:
result = model(pixel_values)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_for_masked_image_modeling(self, config, pixel_values, labels):
model = ViTForMaskedImageModeling(config=config)
model.to(torch_device)
model.eval()
result = model(pixel_values)
self.parent.assertEqual(
result.logits.shape, (self.batch_size, self.num_channels, self.image_size, self.image_size)
)
# test greyscale images
config.num_channels = 1
model = ViTForMaskedImageModeling(config)
model.to(torch_device)
model.eval()
pixel_values = floats_tensor([self.batch_size, 1, self.image_size, self.image_size])
result = model(pixel_values)
self.parent.assertEqual(result.logits.shape, (self.batch_size, 1, self.image_size, self.image_size))
def create_and_check_for_image_classification(self, config, pixel_values, labels):
config.num_labels = self.type_sequence_label_size
model = ViTForImageClassification(config)
@@ -128,6 +147,16 @@ class ViTModelTester:
result = model(pixel_values, labels=labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.type_sequence_label_size))
# test greyscale images
config.num_channels = 1
model = ViTForImageClassification(config)
model.to(torch_device)
model.eval()
pixel_values = floats_tensor([self.batch_size, 1, self.image_size, self.image_size])
result = model(pixel_values)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.type_sequence_label_size))
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
(
@@ -197,6 +226,10 @@ class ViTModelTest(ModelTesterMixin, unittest.TestCase):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_model(*config_and_inputs)
def test_for_masked_image_modeling(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_for_masked_image_modeling(*config_and_inputs)
def test_for_image_classification(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_for_image_classification(*config_and_inputs)
@@ -240,3 +273,30 @@ class ViTModelIntegrationTest(unittest.TestCase):
expected_slice = torch.tensor([-0.2744, 0.8215, -0.0836]).to(torch_device)
self.assertTrue(torch.allclose(outputs.logits[0, :3], expected_slice, atol=1e-4))
@slow
def test_inference_interpolate_pos_encoding(self):
# ViT 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 = ViTModel.from_pretrained("facebook/dino-vits8").to(torch_device)
feature_extractor = ViTFeatureExtractor.from_pretrained("facebook/dino-vits8", size=480)
image = prepare_img()
inputs = feature_extractor(images=image, return_tensors="pt")
pixel_values = inputs.pixel_values.to(torch_device)
# forward pass
with torch.no_grad():
outputs = model(pixel_values, interpolate_pos_encoding=True)
# verify the logits
expected_shape = torch.Size((1, 3601, 384))
self.assertEqual(outputs.last_hidden_state.shape, expected_shape)
expected_slice = torch.tensor(
[[4.2340, 4.3906, -6.6692], [4.5463, 1.8928, -6.7257], [4.4429, 0.8496, -5.8585]]
).to(torch_device)
self.assertTrue(torch.allclose(outputs.last_hidden_state[0, :3, :3], expected_slice, atol=1e-4))

135
tests/models/vit_mae/test_modeling_tf_vit_mae.py
View File

@@ -38,7 +38,6 @@ if is_tf_available():
import tensorflow as tf
from transformers import TFViTMAEForPreTraining, TFViTMAEModel
from transformers.models.vit_mae.modeling_tf_vit_mae import to_2tuple
if is_vision_available():
@@ -67,6 +66,7 @@ class TFViTMAEModelTester:
type_sequence_label_size=10,
initializer_range=0.02,
num_labels=3,
mask_ratio=0.6,
scope=None,
):
self.parent = parent
@@ -85,8 +85,14 @@ class TFViTMAEModelTester:
self.attention_probs_dropout_prob = attention_probs_dropout_prob
self.type_sequence_label_size = type_sequence_label_size
self.initializer_range = initializer_range
self.mask_ratio = mask_ratio
self.scope = scope
# in ViTMAE, the expected sequence length = (num_patches + 1) * (1 - config.mask_ratio), rounded above
# (we add 1 for the [CLS] token)
num_patches = (image_size // patch_size) ** 2
self.seq_length = int(math.ceil((1 - mask_ratio) * (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])
@@ -116,29 +122,21 @@ class TFViTMAEModelTester:
attention_probs_dropout_prob=self.attention_probs_dropout_prob,
is_decoder=False,
initializer_range=self.initializer_range,
mask_ratio=self.mask_ratio,
)
def create_and_check_model(self, config, pixel_values, labels):
model = TFViTMAEModel(config=config)
result = model(pixel_values, training=False)
# expected sequence length = (num_patches + 1) * (1 - config.mask_ratio), rounded above
# (we add 1 for the [CLS] token)
image_size = to_2tuple(self.image_size)
patch_size = to_2tuple(self.patch_size)
num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
expected_seq_len = int(math.ceil((1 - config.mask_ratio) * (num_patches + 1)))
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, expected_seq_len, self.hidden_size))
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_for_pretraining(self, config, pixel_values, labels):
model = TFViTMAEForPreTraining(config)
result = model(pixel_values, training=False)
# expected sequence length = num_patches
image_size = to_2tuple(self.image_size)
patch_size = to_2tuple(self.patch_size)
num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
expected_seq_len = num_patches
num_patches = (self.image_size // self.patch_size) ** 2
expected_num_channels = self.patch_size**2 * self.num_channels
self.parent.assertEqual(result.logits.shape, (self.batch_size, expected_seq_len, expected_num_channels))
self.parent.assertEqual(result.logits.shape, (self.batch_size, num_patches, expected_num_channels))
# test greyscale images
config.num_channels = 1
@@ -147,7 +145,7 @@ class TFViTMAEModelTester:
pixel_values = floats_tensor([self.batch_size, 1, self.image_size, self.image_size])
result = model(pixel_values, training=False)
expected_num_channels = self.patch_size**2
self.parent.assertEqual(result.logits.shape, (self.batch_size, expected_seq_len, expected_num_channels))
self.parent.assertEqual(result.logits.shape, (self.batch_size, num_patches, expected_num_channels))
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
@@ -179,7 +177,6 @@ class TFViTMAEModelTest(TFModelTesterMixin, unittest.TestCase):
@unittest.skip(reason="ViTMAE does not use inputs_embeds")
def test_inputs_embeds(self):
# ViTMAE does not use inputs_embeds
pass
def test_model_common_attributes(self):
@@ -266,114 +263,6 @@ class TFViTMAEModelTest(TFModelTesterMixin, unittest.TestCase):
output_for_kw_input = model(**inputs_np, noise=noise)
self.assert_outputs_same(output_for_dict_input, output_for_kw_input)
def test_attention_outputs(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.return_dict = True
# in ViTMAE, the seq_len equals (number of patches + 1) * (1 - mask_ratio), rounded above
image_size = to_2tuple(self.model_tester.image_size)
patch_size = to_2tuple(self.model_tester.patch_size)
num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
seq_len = int(math.ceil((1 - config.mask_ratio) * (num_patches + 1)))
encoder_seq_length = getattr(self.model_tester, "encoder_seq_length", seq_len)
encoder_key_length = getattr(self.model_tester, "key_length", encoder_seq_length)
chunk_length = getattr(self.model_tester, "chunk_length", None)
if chunk_length is not None and hasattr(self.model_tester, "num_hashes"):
encoder_seq_length = encoder_seq_length * self.model_tester.num_hashes
for model_class in self.all_model_classes:
inputs_dict["output_attentions"] = True
inputs_dict["output_hidden_states"] = False
config.return_dict = True
model = model_class(config)
outputs = model(**self._prepare_for_class(inputs_dict, model_class), training=False)
attentions = outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions
self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
# check that output_attentions also work using config
del inputs_dict["output_attentions"]
config.output_attentions = True
model = model_class(config)
outputs = model(**self._prepare_for_class(inputs_dict, model_class), training=False)
attentions = outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions
self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
if chunk_length is not None:
self.assertListEqual(
list(attentions[0].shape[-4:]),
[self.model_tester.num_attention_heads, encoder_seq_length, chunk_length, encoder_key_length],
)
else:
self.assertListEqual(
list(attentions[0].shape[-3:]),
[self.model_tester.num_attention_heads, encoder_seq_length, encoder_key_length],
)
out_len = len(outputs)
# Check attention is always last and order is fine
inputs_dict["output_attentions"] = True
inputs_dict["output_hidden_states"] = True
model = model_class(config)
outputs = model(**self._prepare_for_class(inputs_dict, model_class), training=False)
if hasattr(self.model_tester, "num_hidden_states_types"):
added_hidden_states = self.model_tester.num_hidden_states_types
elif self.is_encoder_decoder:
added_hidden_states = 2
else:
added_hidden_states = 1
self.assertEqual(out_len + added_hidden_states, len(outputs))
self_attentions = outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions
self.assertEqual(len(self_attentions), self.model_tester.num_hidden_layers)
if chunk_length is not None:
self.assertListEqual(
list(self_attentions[0].shape[-4:]),
[self.model_tester.num_attention_heads, encoder_seq_length, chunk_length, encoder_key_length],
)
else:
self.assertListEqual(
list(self_attentions[0].shape[-3:]),
[self.model_tester.num_attention_heads, encoder_seq_length, encoder_key_length],
)
def test_hidden_states_output(self):
def check_hidden_states_output(inputs_dict, config, model_class):
model = model_class(config)
outputs = model(**self._prepare_for_class(inputs_dict, model_class))
hidden_states = 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(len(hidden_states), expected_num_layers)
# ViTMAE has a different seq_length
image_size = to_2tuple(self.model_tester.image_size)
patch_size = to_2tuple(self.model_tester.patch_size)
num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
seq_length = int(math.ceil((1 - config.mask_ratio) * (num_patches + 1)))
self.assertListEqual(
list(hidden_states[0].shape[-2:]),
[seq_length, 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)
# overwrite from common since TFViTMAEForPretraining has random masking, we need to fix the noise
# to generate masks during test
def check_pt_tf_models(self, tf_model, pt_model, tf_inputs_dict):

150
tests/models/vit_mae/test_modeling_vit_mae.py
View File

@@ -35,7 +35,7 @@ if is_torch_available():
from torch import nn
from transformers import ViTMAEForPreTraining, ViTMAEModel
from transformers.models.vit.modeling_vit import VIT_PRETRAINED_MODEL_ARCHIVE_LIST, to_2tuple
from transformers.models.vit.modeling_vit import VIT_PRETRAINED_MODEL_ARCHIVE_LIST
if is_vision_available():
@@ -64,6 +64,7 @@ class ViTMAEModelTester:
type_sequence_label_size=10,
initializer_range=0.02,
num_labels=3,
mask_ratio=0.6,
scope=None,
):
self.parent = parent
@@ -82,8 +83,14 @@ class ViTMAEModelTester:
self.attention_probs_dropout_prob = attention_probs_dropout_prob
self.type_sequence_label_size = type_sequence_label_size
self.initializer_range = initializer_range
self.mask_ratio = mask_ratio
self.scope = scope
# in ViTMAE, the expected sequence length = (num_patches + 1) * (1 - config.mask_ratio), rounded above
# (we add 1 for the [CLS] token)
num_patches = (image_size // patch_size) ** 2
self.seq_length = int(math.ceil((1 - mask_ratio) * (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])
@@ -109,6 +116,7 @@ class ViTMAEModelTester:
attention_probs_dropout_prob=self.attention_probs_dropout_prob,
is_decoder=False,
initializer_range=self.initializer_range,
mask_ratio=self.mask_ratio,
)
def create_and_check_model(self, config, pixel_values, labels):
@@ -116,26 +124,16 @@ class ViTMAEModelTester:
model.to(torch_device)
model.eval()
result = model(pixel_values)
# expected sequence length = (num_patches + 1) * (1 - config.mask_ratio), rounded above
# (we add 1 for the [CLS] token)
image_size = to_2tuple(self.image_size)
patch_size = to_2tuple(self.patch_size)
num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
expected_seq_len = int(math.ceil((1 - config.mask_ratio) * (num_patches + 1)))
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, expected_seq_len, self.hidden_size))
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_for_pretraining(self, config, pixel_values, labels):
model = ViTMAEForPreTraining(config)
model.to(torch_device)
model.eval()
result = model(pixel_values)
# expected sequence length = num_patches
image_size = to_2tuple(self.image_size)
patch_size = to_2tuple(self.patch_size)
num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
expected_seq_len = num_patches
num_patches = (self.image_size // self.patch_size) ** 2
expected_num_channels = self.patch_size**2 * self.num_channels
self.parent.assertEqual(result.logits.shape, (self.batch_size, expected_seq_len, expected_num_channels))
self.parent.assertEqual(result.logits.shape, (self.batch_size, num_patches, expected_num_channels))
# test greyscale images
config.num_channels = 1
@@ -145,7 +143,7 @@ class ViTMAEModelTester:
pixel_values = floats_tensor([self.batch_size, 1, self.image_size, self.image_size])
result = model(pixel_values)
expected_num_channels = self.patch_size**2
self.parent.assertEqual(result.logits.shape, (self.batch_size, expected_seq_len, expected_num_channels))
self.parent.assertEqual(result.logits.shape, (self.batch_size, num_patches, expected_num_channels))
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
@@ -175,8 +173,8 @@ class ViTMAEModelTest(ModelTesterMixin, unittest.TestCase):
def test_config(self):
self.config_tester.run_common_tests()
@unittest.skip(reason="ViTMAE does not use inputs_embeds")
def test_inputs_embeds(self):
# ViTMAE does not use inputs_embeds
pass
def test_model_common_attributes(self):
@@ -208,126 +206,6 @@ class ViTMAEModelTest(ModelTesterMixin, unittest.TestCase):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_for_pretraining(*config_and_inputs)
def test_attention_outputs(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.return_dict = True
# in ViTMAE, the seq_len equals (number of patches + 1) * (1 - mask_ratio), rounded above
image_size = to_2tuple(self.model_tester.image_size)
patch_size = to_2tuple(self.model_tester.patch_size)
num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
seq_len = int(math.ceil((1 - config.mask_ratio) * (num_patches + 1)))
encoder_seq_length = getattr(self.model_tester, "encoder_seq_length", seq_len)
encoder_key_length = getattr(self.model_tester, "key_length", encoder_seq_length)
chunk_length = getattr(self.model_tester, "chunk_length", None)
if chunk_length is not None and hasattr(self.model_tester, "num_hashes"):
encoder_seq_length = encoder_seq_length * self.model_tester.num_hashes
for model_class in self.all_model_classes:
inputs_dict["output_attentions"] = True
inputs_dict["output_hidden_states"] = False
config.return_dict = True
model = model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(**self._prepare_for_class(inputs_dict, model_class))
attentions = outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions
self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
# check that output_attentions also work using config
del inputs_dict["output_attentions"]
config.output_attentions = True
model = model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(**self._prepare_for_class(inputs_dict, model_class))
attentions = outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions
self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
if chunk_length is not None:
self.assertListEqual(
list(attentions[0].shape[-4:]),
[self.model_tester.num_attention_heads, encoder_seq_length, chunk_length, encoder_key_length],
)
else:
self.assertListEqual(
list(attentions[0].shape[-3:]),
[self.model_tester.num_attention_heads, encoder_seq_length, encoder_key_length],
)
out_len = len(outputs)
# Check attention is always last and order is fine
inputs_dict["output_attentions"] = True
inputs_dict["output_hidden_states"] = True
model = model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(**self._prepare_for_class(inputs_dict, model_class))
if hasattr(self.model_tester, "num_hidden_states_types"):
added_hidden_states = self.model_tester.num_hidden_states_types
elif self.is_encoder_decoder:
added_hidden_states = 2
else:
added_hidden_states = 1
self.assertEqual(out_len + added_hidden_states, len(outputs))
self_attentions = outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions
self.assertEqual(len(self_attentions), self.model_tester.num_hidden_layers)
if chunk_length is not None:
self.assertListEqual(
list(self_attentions[0].shape[-4:]),
[self.model_tester.num_attention_heads, encoder_seq_length, chunk_length, encoder_key_length],
)
else:
self.assertListEqual(
list(self_attentions[0].shape[-3:]),
[self.model_tester.num_attention_heads, encoder_seq_length, encoder_key_length],
)
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 = 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(len(hidden_states), expected_num_layers)
# ViTMAE has a different seq_length
image_size = to_2tuple(self.model_tester.image_size)
patch_size = to_2tuple(self.model_tester.patch_size)
num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
seq_length = int(math.ceil((1 - config.mask_ratio) * (num_patches + 1)))
self.assertListEqual(
list(hidden_states[0].shape[-2:]),
[seq_length, 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)
# overwrite from common since ViTMAEForPretraining has random masking, we need to fix the noise
# to generate masks during test
def check_pt_tf_models(self, tf_model, pt_model, pt_inputs_dict):

6
tests/models/yolos/test_modeling_yolos.py
View File

@@ -31,7 +31,7 @@ if is_torch_available():
from torch import nn
from transformers import YolosForObjectDetection, YolosModel
from transformers.models.yolos.modeling_yolos import YOLOS_PRETRAINED_MODEL_ARCHIVE_LIST, to_2tuple
from transformers.models.yolos.modeling_yolos import YOLOS_PRETRAINED_MODEL_ARCHIVE_LIST
if is_vision_available():
@@ -86,9 +86,7 @@ class YolosModelTester:
self.num_detection_tokens = num_detection_tokens
# we set the expected sequence length (which is used in several tests)
# expected sequence length = num_patches + 1 (we add 1 for the [CLS] token) + num_detection_tokens
image_size = to_2tuple(self.image_size)
patch_size = to_2tuple(self.patch_size)
num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
num_patches = (image_size[1] // patch_size) * (image_size[0] // patch_size)
self.expected_seq_len = num_patches + 1 + self.num_detection_tokens
def prepare_config_and_inputs(self):