Add EfficientNet Image PreProcessor (#37055)
* added efficientnet image preprocessor but tests fail * ruff checks pass * ruff formatted * properly pass rescale_offset through the functions * - corrected indentation, ordering of methods - reshape test passes when casted to float64 - equivalence test doesn't pass * all tests now pass - changes order of rescale, normalize acc to slow - rescale_offset defaults to False acc to slow - resample was causing difference in fast and slow. Changing test to bilinear resolves this difference * ruff reformat * F.InterpolationMode.NEAREST_EXACT gives TypeError: Object of type InterpolationMode is not JSON serializable * fixes offset not being applied when do_rescale and do_normalization are both true * - using nearest_exact sampling - added tests for rescale + normalize * resolving reviews --------- Co-authored-by: Yoni Gozlan <74535834+yonigozlan@users.noreply.github.com>
This commit is contained in:
Zeeshan Khan Suri
committed by
GitHub
GitHub
parent
32eca7197a
commit
a7d2bbaaa8
@@ -17,15 +17,26 @@ import unittest
|
||||
|
||||
import numpy as np
|
||||
|
||||
from transformers.image_utils import PILImageResampling
|
||||
from transformers.testing_utils import require_torch, require_vision
|
||||
from transformers.utils import is_vision_available
|
||||
from transformers.utils import (
|
||||
is_torch_available,
|
||||
is_torchvision_available,
|
||||
is_vision_available,
|
||||
)
|
||||
|
||||
from ...test_image_processing_common import ImageProcessingTestMixin, prepare_image_inputs
|
||||
|
||||
|
||||
if is_torch_available():
|
||||
import torch
|
||||
|
||||
if is_vision_available():
|
||||
from transformers import EfficientNetImageProcessor
|
||||
|
||||
if is_torchvision_available():
|
||||
from transformers import EfficientNetImageProcessorFast
|
||||
|
||||
|
||||
class EfficientNetImageProcessorTester:
|
||||
def __init__(
|
||||
@@ -41,6 +52,10 @@ class EfficientNetImageProcessorTester:
|
||||
do_normalize=True,
|
||||
image_mean=[0.5, 0.5, 0.5],
|
||||
image_std=[0.5, 0.5, 0.5],
|
||||
do_rescale=True,
|
||||
rescale_offset=True,
|
||||
rescale_factor=1 / 127.5,
|
||||
resample=PILImageResampling.BILINEAR, # NEAREST is too different between PIL and torchvision
|
||||
):
|
||||
size = size if size is not None else {"height": 18, "width": 18}
|
||||
self.parent = parent
|
||||
@@ -54,6 +69,7 @@ class EfficientNetImageProcessorTester:
|
||||
self.do_normalize = do_normalize
|
||||
self.image_mean = image_mean
|
||||
self.image_std = image_std
|
||||
self.resample = resample
|
||||
|
||||
def prepare_image_processor_dict(self):
|
||||
return {
|
||||
@@ -62,6 +78,7 @@ class EfficientNetImageProcessorTester:
|
||||
"do_normalize": self.do_normalize,
|
||||
"do_resize": self.do_resize,
|
||||
"size": self.size,
|
||||
"resample": self.resample,
|
||||
}
|
||||
|
||||
def expected_output_image_shape(self, images):
|
||||
@@ -83,6 +100,7 @@ class EfficientNetImageProcessorTester:
|
||||
@require_vision
|
||||
class EfficientNetImageProcessorTest(ImageProcessingTestMixin, unittest.TestCase):
|
||||
image_processing_class = EfficientNetImageProcessor if is_vision_available() else None
|
||||
fast_image_processing_class = EfficientNetImageProcessorFast if is_torchvision_available() else None
|
||||
|
||||
def setUp(self):
|
||||
super().setUp()
|
||||
@@ -93,30 +111,80 @@ class EfficientNetImageProcessorTest(ImageProcessingTestMixin, unittest.TestCase
|
||||
return self.image_processor_tester.prepare_image_processor_dict()
|
||||
|
||||
def test_image_processor_properties(self):
|
||||
image_processing = self.image_processing_class(**self.image_processor_dict)
|
||||
self.assertTrue(hasattr(image_processing, "image_mean"))
|
||||
self.assertTrue(hasattr(image_processing, "image_std"))
|
||||
self.assertTrue(hasattr(image_processing, "do_normalize"))
|
||||
self.assertTrue(hasattr(image_processing, "do_resize"))
|
||||
self.assertTrue(hasattr(image_processing, "size"))
|
||||
for image_processing_class in self.image_processor_list:
|
||||
image_processing = image_processing_class(**self.image_processor_dict)
|
||||
self.assertTrue(hasattr(image_processing, "image_mean"))
|
||||
self.assertTrue(hasattr(image_processing, "image_std"))
|
||||
self.assertTrue(hasattr(image_processing, "do_normalize"))
|
||||
self.assertTrue(hasattr(image_processing, "do_resize"))
|
||||
self.assertTrue(hasattr(image_processing, "size"))
|
||||
|
||||
def test_image_processor_from_dict_with_kwargs(self):
|
||||
image_processor = self.image_processing_class.from_dict(self.image_processor_dict)
|
||||
self.assertEqual(image_processor.size, {"height": 18, "width": 18})
|
||||
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, {"height": 18, "width": 18})
|
||||
|
||||
image_processor = self.image_processing_class.from_dict(self.image_processor_dict, size=42)
|
||||
self.assertEqual(image_processor.size, {"height": 42, "width": 42})
|
||||
image_processor = image_processing_class.from_dict(self.image_processor_dict, size=42)
|
||||
self.assertEqual(image_processor.size, {"height": 42, "width": 42})
|
||||
|
||||
def test_rescale(self):
|
||||
# EfficientNet optionally rescales between -1 and 1 instead of the usual 0 and 1
|
||||
image = np.arange(0, 256, 1, dtype=np.uint8).reshape(1, 8, 32)
|
||||
|
||||
image_processor = self.image_processing_class(**self.image_processor_dict)
|
||||
for image_processing_class in self.image_processor_list:
|
||||
image_processor = image_processing_class(**self.image_processor_dict)
|
||||
if image_processing_class == EfficientNetImageProcessorFast:
|
||||
image = torch.from_numpy(image)
|
||||
|
||||
rescaled_image = image_processor.rescale(image, scale=1 / 127.5)
|
||||
expected_image = (image * (1 / 127.5)).astype(np.float32) - 1
|
||||
self.assertTrue(np.allclose(rescaled_image, expected_image))
|
||||
# Scale between [-1, 1] with rescale_factor 1/127.5 and rescale_offset=True
|
||||
rescaled_image = image_processor.rescale(image, scale=1 / 127.5, offset=True)
|
||||
expected_image = (image * (1 / 127.5)) - 1
|
||||
self.assertTrue(torch.allclose(rescaled_image, expected_image))
|
||||
|
||||
rescaled_image = image_processor.rescale(image, scale=1 / 255, offset=False)
|
||||
expected_image = (image / 255.0).astype(np.float32)
|
||||
self.assertTrue(np.allclose(rescaled_image, expected_image))
|
||||
# Scale between [0, 1] with rescale_factor 1/255 and rescale_offset=True
|
||||
rescaled_image = image_processor.rescale(image, scale=1 / 255, offset=False)
|
||||
expected_image = image / 255.0
|
||||
self.assertTrue(torch.allclose(rescaled_image, expected_image))
|
||||
|
||||
else:
|
||||
rescaled_image = image_processor.rescale(image, scale=1 / 127.5, dtype=np.float64)
|
||||
expected_image = (image * (1 / 127.5)).astype(np.float64) - 1
|
||||
self.assertTrue(np.allclose(rescaled_image, expected_image))
|
||||
|
||||
rescaled_image = image_processor.rescale(image, scale=1 / 255, offset=False, dtype=np.float64)
|
||||
expected_image = (image / 255.0).astype(np.float64)
|
||||
self.assertTrue(np.allclose(rescaled_image, expected_image))
|
||||
|
||||
@require_vision
|
||||
@require_torch
|
||||
def test_rescale_normalize(self):
|
||||
if self.image_processing_class is None or self.fast_image_processing_class is None:
|
||||
self.skipTest(reason="Skipping slow/fast equivalence test as one of the image processors is not defined")
|
||||
|
||||
image = torch.arange(0, 256, 1, dtype=torch.uint8).reshape(1, 8, 32).repeat(3, 1, 1)
|
||||
image_mean_0 = (0.0, 0.0, 0.0)
|
||||
image_std_0 = (1.0, 1.0, 1.0)
|
||||
image_mean_1 = (0.5, 0.5, 0.5)
|
||||
image_std_1 = (0.5, 0.5, 0.5)
|
||||
|
||||
image_processor_fast = self.fast_image_processing_class(**self.image_processor_dict)
|
||||
|
||||
# Rescale between [-1, 1] with rescale_factor=1/127.5 and rescale_offset=True. Then normalize
|
||||
rescaled_normalized = image_processor_fast.rescale_and_normalize(
|
||||
image, True, 1 / 127.5, True, image_mean_0, image_std_0, True
|
||||
)
|
||||
expected_image = (image * (1 / 127.5)) - 1
|
||||
expected_image = (expected_image - torch.tensor(image_mean_0).view(3, 1, 1)) / torch.tensor(image_std_0).view(
|
||||
3, 1, 1
|
||||
)
|
||||
self.assertTrue(torch.allclose(rescaled_normalized, expected_image, rtol=1e-3))
|
||||
|
||||
# Rescale between [0, 1] with rescale_factor=1/255 and rescale_offset=False. Then normalize
|
||||
rescaled_normalized = image_processor_fast.rescale_and_normalize(
|
||||
image, True, 1 / 255, True, image_mean_1, image_std_1, False
|
||||
)
|
||||
expected_image = image * (1 / 255.0)
|
||||
expected_image = (expected_image - torch.tensor(image_mean_1).view(3, 1, 1)) / torch.tensor(image_std_1).view(
|
||||
3, 1, 1
|
||||
)
|
||||
self.assertTrue(torch.allclose(rescaled_normalized, expected_image, rtol=1e-3))
|
||||
|
||||
Reference in New Issue
Block a user