RuntimeError: Given groups=1, weight of size [16, 3, 1, 1], expected input[8, 4, 256, 256] to have 3 channels, but got 4 channels instead about hed-unet HOT 10 CLOSED

This sounds like you are inputting images with 4 channels while having the model configured to take 3 channels as inputs.

As you were mentioning png inputs earlier, I'm assuming that you're reading the data as RGBA instead of RGB.

If you do not need the fourth channel, you can strip it away (img = img[:3]). If you do indeed need the fourth channel, you can set input_channels: 4 in the config.yml file.

from hed-unet.

from hed-unet.

from hed-unet.

from hed-unet.

Can you make sure that the loaded image array has only 3 channels? I.e. by adding print(img.shape) to the full_forward method in train.py?

from hed-unet.

from hed-unet.

Assuming you're using the provided INRIA data loading code, that will tile each input image into tiles of 256x256 (see the TILESIZE = 256 constant in deep_learning/utils/data.py.

I still believe that the fourth channel comes from the alpha channel in the png, could you simply cut it off by changing data = raster.read() to data = raster.read()[:3] in deep_learning/utils/data.py (and then also removing the cached .npz files?)

from hed-unet.

from hed-unet.

Yes but png files can have 4 channels, and when you convert them to tiff that 4th channel is probably kept. Could you please simply try the solution I suggested? Or what do you expect me to do here?

from hed-unet.

from hed-unet.