def image_to_internal(self, image: Imagelike) -> torch.Tensor:
"""Transform an image into a torch Tensor.
"""
if isinstance(image, torch.Tensor):
image_tensor = image
elif isinstance(image, np.ndarray):
# at this point we need to know the range (if further
# preprocessing, e.g., normalization, is required ...)
if False and (0 <= image).all():
if (image <= 1).all(): # image is range 0.0 - 1.0
pass
elif (image <= 255).all(): # image is range 0 - 255
pass
# Question: channel first or channel last?
# H X W X C ==> C X H X W
# image = np.transpose(image, (2, 0, 1))
# preprocess_numpy expects numpy.ndarray of correct size,
# dtype float and values in range [0.0, 1.0].
# It performs the following operations:
# 1. [no resizing]
# 2. numpy.ndarray -> torch.Tensor
# 3. normalization [0.0, 1.0] -> torch.imagenet_range
image_tensor = self.preprocess_numpy(image)
# old: explicit transformation:
# H X W X C ==> C X H X W
# image = np.transpose(image, (2, 0, 1))
#
# image = torch.from_numpy(image)
# image = image.add(-self.imagenet_mean_.view(3, 1, 1)).\
# div(self.imagenet_std_.view(3, 1, 1))
#
# add batch dimension: C X H X W ==> B X C X H X W
# image = image.unsqueeze(0)
else:
# the _image_to_internal function expects as input a PIL image!
image = Image.as_pil(image)
# image should be PIL Image. Got <class 'numpy.ndarray'>
image_tensor = self._image_to_internal(image)
if image_tensor.dim() == 4:
# image is already batch
image_batch = image_tensor
elif image_tensor.dim() == 3:
# create a mini-batch as expected by the model
# by adding a batch dimension: C X H X W ==> B X C X H X W
image_batch = image_tensor.unsqueeze(0)
else:
raise ValueError(f"Data of invalid shape {image.shape} cannot "
"be transformed into an internal torch image.")
# move the input and model to GPU for speed if available
image_batch = image_batch.to(self._device)
return image_batch
def test_pillow(self):
"""Test :py:class:`Size` type.
"""
module = importlib.import_module('dltb.thirdparty.pil')
self.assertIn('pil', Image.supported_formats())
image = Image(self.example_image_filename)
pil = Image.as_pil(image)
self.assertIsInstance(pil, module.PIL.Image.Image)