texturize enhance: pic should be PIL Image or ndarray

[terrainer]

Full traceback:
Traceback (most recent call last): File "<frozen runpy>", line 198, in _run_module_as_main File "<frozen runpy>", line 88, in _run_code File "C:\tools\miniconda3\envs\texturize\Scripts\texturize.exe\__main__.py", line 7, in <module> File "C:\tools\miniconda3\envs\texturize\Lib\site-packages\texturize\__main__.py", line 154, in main cmd = commands.Enhance(target_arr, source_arr, zoom=zoom) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "C:\tools\miniconda3\envs\texturize\Lib\site-packages\texturize\commands.py", line 105, in __init__ self.target = load_tensor_from_image(target, device="cpu") ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "C:\tools\miniconda3\envs\texturize\Lib\site-packages\texturize\io.py", line 58, in load_tensor_from_image return V.to_tensor(image).unsqueeze(0).to(device, dtype) ^^^^^^^^^^^^^^^^^^ File "C:\tools\miniconda3\envs\texturize\Lib\site-packages\torchvision\transforms\functional.py", line 140, in to_tensor raise TypeError(f"pic should be PIL Image or ndarray. Got {type(pic)}") TypeError: pic should be PIL Image or ndarray. Got <class 'tuple'>

Getting this when trying to use the enhance command with two RGB, 8 bits-per-channel, PNG files. Unsure what I'm doing wrong.
Command used: texturize enhance tests/lowres.png with tests/highres.png --zoom=4

[POD666]

Just got the same issue:

$ texturize remake sample.jpg like sample2.jpg --weight 0.5 -o test.jpg
...
pic should be PIL Image or ndarray. Got <class 'tuple'>

[htoyryla]

I get the same error and it persists regardless of which PyTorch & torchvision version is in use.

However, I have used texturize successfully two years ago. I still have the env I used then and it works (texturize version 0.12).

So you could try to install that

pip install texturize==0.12

and see if it works. It is not the latest version though.

[htoyryla]

Seems to me that the errors result as follows

First the source and the target are extracted from the command line here

texturize/src/texturize/__main__.py

Lines 141 to 142 in 637a213

	source_arr, source_props = io.load_tensor_from_files(filename)
	target_arr = io.load_tensor_from_files(target) if target else None

using this function

texturize/src/texturize/io.py

Lines 15 to 26 in 637a213

	def load_tensor_from_files(glob_pattern, device='cpu', mode=None) -> tuple:
	arrays, props = [], []
	for filename in sorted(glob.glob(glob_pattern)):
	img = load_image_from_file(filename, mode)
	arr = load_tensor_from_image(img, device)
	arrays.append(arr)
	prop = ''.join([s[2] for s in difflib.ndiff(glob_pattern, filename) if s[0]=='+'])
	props.append(prop + ":" + str(arr.shape[1]))
	assert all(a.shape[2:] == arrays[0].shape[2:] for a in arrays[1:])
	return torch.cat(arrays, dim=1), props

which returns a tuple (a tensor, a list of props). For the source, these are saved in separate variables, but the target the tuple is stored.

In Remake

texturize/src/texturize/commands.py

Lines 120 to 124 in 637a213

	class Remake(Command):
	def __init__(self, target, source, weights=[1.0]):
	self.source = load_tensor_from_image(source, device="cpu")
	self.target = load_tensor_from_image(target, device="cpu")
	self.weights = torch.tensor(weights, dtype=torch.float32).view(-1, 1, 1, 1)

then this target is passed here

texturize/src/texturize/io.py

Lines 47 to 59 in 637a213

	def load_tensor_from_image(image, device, dtype=torch.float32):
	# NOTE: torchvision incorrectly assumes that I;16 means signed, but
	# Pillow docs say unsigned.
	if isinstance(image, PIL.Image.Image) and image.mode == "I;16":
	import numpy
	arr = numpy.frombuffer(image.tobytes(), dtype=numpy.uint16)
	arr = arr.reshape((image.height, image.width))
	assert arr.min() >= 0 and arr.max() < 65536
	image = arr.astype(numpy.float32) / 65536.0
	if not isinstance(image, torch.Tensor):
	return V.to_tensor(image).unsqueeze(0).to(device, dtype)
	return image

where it checks if it is a tensor, it is not because it is tuple (tensor, list), and assuming it is a PIL image tries to convert it to a tensor which of course results in an error.

I have not tried to modify the code but a simple solution might be to modify this line

texturize/src/texturize/__main__.py

Line 142 in 637a213

target_arr = io.load_tensor_from_files(target) if target else None

to this

target_arr, _ = io.load_tensor_from_files(target) if target else None

Edits: due to copy-paste errors the links to code were not correct in the original posting, now corrected.

[htoyryla]

Tried to test the change I proposed but not being familiar with poetry tool did not find a way to run my version.