Got errors in pyramid_gaussian function

[pz0910]

Got error :
pyramid = list(pyramid_gaussian(img, n_scales-1, multichannel=True))
File "/home/jsgx/.conda/envs/pytorch/lib/python3.6/site-packages/skimage/transform/pyramids.py", line 197, in pyramid_gaussian
image = img_as_float(image)
File "/home/jsgx/.conda/envs/pytorch/lib/python3.6/site-packages/skimage/util/dtype.py", line 378, in img_as_float64
return convert(image, np.float64, force_copy)
File "/home/jsgx/.conda/envs/pytorch/lib/python3.6/site-packages/skimage/util/dtype.py", line 244, in convert
raise ValueError("Images of type float must be between -1 and 1.")
ValueError: Images of type float must be between -1 and 1.

python main.py --n_GPUs 2 --batch_size 16
It seems image tensors have to be normlized to [-1,1]. Should I add --rgb_range=1 to solve this problem

[SeungjunNah]

@pz-AOE
Strange. Which version for scikit-image are you using?
I tested with scikit-image v0.16.2 and I always have [0, 255] input -> [0, 255] output for float32 type in the pyramid_gaussian function. This is as expected: see the image_as_float function description here.
--rgb_range 255 should not have any problems.

[pz0910]

@pz-AOE
Strange. Which version for scikit-image are you using?
I tested with scikit-image v0.16.2 and I always have [0, 255] input -> [0, 255] output for float32 type in the pyramid_gaussian function. This is as expected: see the image_as_float function description here.
--rgb_range 255 should not have any problems.

Thanks for replying.
I am using scikit-image v.014.0. I will try to upgrade the dependencies to check if it works.
By the way, could you please share your yaml env file?

[SeungjunNah]

My main conda environment is for multiple purposes and it includes many unrelated packages to this project.
Here are the versions I used for main dependencies.

imageio 2.8.0
numpy 1.18.1
matplotlib 3.2.1
tqdm 4.46.1
scikit-image 0.16.2
readline 8.0

If you want to create a new conda environment, the following would suffice.

conda install -y tqdm imageio scikit-image matplotlib readline
conda install -y pytorch torchvision cudatoolkit=10.2 -c pytorch

[pz0910]

Problem solved by upgrading scikit-image to v0.16.2.
Thanks

[davidvct]

if you are using RTX 4090 like me, cuda 10.2 is not compatible with it.
I am using cuda 11.3, training seems to work (still ongoing).
installation of cuda 11.3:
conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.3 -c pytorch