Skip to content

Type cast before normalize segformer #8

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
amyeroberts merged 3 commits into from
Sep 2, 2022
Merged

Type cast before normalize segformer #8

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
41c16ea
Cast images to numpy arrays in call to enable consistent behaviour wi…
amyeroberts Aug 5, 2022
4aa8246
Remove accidental clip changes
amyeroberts Aug 5, 2022
ba8eefa
Merge branch 'type-cast-before-normalize-update-methods' into type-ca…
amyeroberts Aug 5, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
22 changes: 15 additions & 7 deletions src/transformers/models/segformer/feature_extraction_segformer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -112,13 +112,15 @@ def __call__(
segmentation_maps (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `List[PIL.Image.Image]`, `List[np.ndarray]`, `List[torch.Tensor]`, *optional*):
Optionally, the corresponding semantic segmentation maps with the pixel-wise annotations.

return_tensors (`str` or [`~utils.TensorType`], *optional*, defaults to `'np'`):
If set, will return tensors of a particular framework. Acceptable values are:
return_tensors (`str` or [`~utils.TensorType`], *optional*, defaults to `None`):
If set, will return a tensor of a particular framework.

- `'tf'`: Return TensorFlow `tf.constant` objects.
- `'pt'`: Return PyTorch `torch.Tensor` objects.
- `'np'`: Return NumPy `np.ndarray` objects.
- `'jax'`: Return JAX `jnp.ndarray` objects.
Acceptable values are:
- `'tf'`: Return TensorFlow `tf.constant` object.
- `'pt'`: Return PyTorch `torch.Tensor` object.
- `'np'`: Return NumPy `np.ndarray` object.
- `'jax'`: Return JAX `jnp.ndarray` object.
- None: Return list of `np.ndarray` objects.

Returns:
[`BatchFeature`]: A [`BatchFeature`] with the following fields:
Expand Down Expand Up @@ -193,8 +195,14 @@ def __call__(
self.resize(map, size=self.size, resample=Image.NEAREST) for map in segmentation_maps
]

# if do_normalize=False, the casting to a numpy array won't happen, so we need to do it here
make_channel_first = True if isinstance(images[0], Image.Image) else images[0].shape[-1] in (1, 3)
images = [self.to_numpy_array(image, rescale=False, channel_first=make_channel_first) for image in images]

if self.do_normalize:
images = [self.normalize(image=image, mean=self.image_mean, std=self.image_std) for image in images]
images = [
self.normalize(image=image, mean=self.image_mean, std=self.image_std, rescale=True) for image in images
]

# return as BatchFeature
data = {"pixel_values": images}
Expand Down
38 changes: 38 additions & 0 deletions tests/models/segformer/test_feature_extraction_segformer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -19,6 +19,7 @@
import numpy as np
from datasets import load_dataset

from parameterized import parameterized
from transformers.testing_utils import require_torch, require_vision
from transformers.utils import is_torch_available, is_vision_available

Expand Down Expand Up @@ -333,3 +334,40 @@ def test_reduce_labels(self):
encoding = feature_extractor(image, map, return_tensors="pt")
self.assertTrue(encoding["labels"].min().item() >= 0)
self.assertTrue(encoding["labels"].max().item() <= 255)

@parameterized.expand(
[
("do_resize_True_do_normalize_True", True, True),
("do_resize_True_do_normalize_False", True, False),
("do_resize_True_do_normalize_True", True, True),
("do_resize_True_do_normalize_False", True, False),
("do_resize_False_do_normalize_True", False, True),
("do_resize_False_do_normalize_False", False, False),
("do_resize_False_do_normalize_True", False, True),
("do_resize_False_do_normalize_False", False, False),
]
)
def test_call_flags(self, _, do_resize, do_normalize):
# Initialize feature_extractor
feature_extractor = self.feature_extraction_class(**self.feat_extract_dict)
feature_extractor.do_resize = do_resize
feature_extractor.do_normalize = do_normalize
# create random PIL images
image_inputs = prepare_image_inputs(self.feature_extract_tester, equal_resolution=False)

expected_shapes = [(3, *x.size[::-1]) for x in image_inputs]
if do_resize:
expected_shapes = [
(
self.feature_extract_tester.num_channels,
self.feature_extract_tester.size,
self.feature_extract_tester.size,
)
for _ in range(self.feature_extract_tester.batch_size)
]

pixel_values = feature_extractor(image_inputs, return_tensors=None)["pixel_values"]
self.assertEqual(len(pixel_values), self.feature_extract_tester.batch_size)
for idx, image in enumerate(pixel_values):
self.assertEqual(image.shape, expected_shapes[idx])
self.assertIsInstance(image, np.ndarray)