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

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]: