shadowxzt / rgb-d-segmentation.pytorch Goto Github PK

The dataset I downloaded from the official website is in mat format. Would it be convenient for you to provide the dataset you use?

Hi, I was trying the RedNet code which is very similar to your work here. I found in their code in the following lines, they actually simply clone the targets to targets_m while here you add 0.5. So I was wondering why you add this and how you determine this 0.5. So I did find that their way only works for PyTorch 1.2 or lower, otherwise, it causes an IndexError: target -1 out of bounds when calculating the crossentrophyloss. By adding 0.5 this can be resolved and can work with higher version of PyTorch. It puzzles me a lot. Is it because of the design of the later verison? I appreciate it if you could provide some insights.

utils.py -> class CrossEntropyLoss2d(nn.Module):

for inputs, targets in zip(inputs_scales, targets_scales):
mask = targets > 0
targets_m = targets.clone() + 0.5
targets_m[mask] -= 1.0