VainF
question about ignore_index 255？
hi, i have a question about the ignore index 255 when i trainning the dataset of citysacpes. while training, the background will be set to label 255, and in the loss function it will not be calculated. So, my questions is: when i test the new image, it will predict a 255 label? in the cityscapes.py, the decode_target function can transfer 255 to 19?? Looking forward your answer, Thank you!
Hi @hezichuanqi, ignoring backgroud pixels is a common practice in scene parsing. These background pixels are typically outliers which can not appropriately annotated.
So, my questions is: when i test the new image, it will predict a 255 label?
No. During inference, the network will predict a [0, 18] label for all pixels. And during evaluation, background pixels with label 255 will be masked and ignored.
in the cityscapes.py, the decode_target function can transfer 255 to 19??
The decode_fn transfers 255 to 19 to visualize the groud truth.
Thank you very much for your answers.I'm still have a question: ------During inference, the network will predict a [0, 18] label for all pixels. I can understand the class number is 19. but, if we want to inference a new picture, how those pixels belong to the background can be recognized.In other words, the backgound pixels doesn't belong to [0,18] labels.
The model will output the most likely class of [0, 18], that's why you have the following codes
https://github.com/VainF/DeepLabV3Plus-Pytorch/blob/3f84cd27d5e845c9d92aed6639b0685ef22ed964/main.py#L174
Since the background was never learned, the probabilities (namely the logits) of the background class should be relatively small.
Hi, there, what if my background class is 0, but others is from 1 to 13? How to ignore the background prediction? because at the test part, I got the background predicted as some else foreground class like 12. It has disturbed me for such long time. Thanks a lot.
Hi, there, what if my background class is 0, but others is from 1 to 13? How to ignore the background prediction? because at the test part, I got the background predicted as some else foreground class like 12. It has disturbed me for such long time. Thanks a lot.
Don't quite understand your question. If you want to neglect class 0 during training, you can set ignore_index=0 for the loss. Or if you want it during validating, then just use something like
preds = outputs.detach()[1:].max(dim=1)[1].cpu().numpy()
Hope this helps.
Hi, there, what if my background class is 0, but others is from 1 to 13? How to ignore the background prediction? because at the test part, I got the background predicted as some else foreground class like 12. It has disturbed me for such long time. Thanks a lot.
Don't quite understand your question. If you want to neglect class 0 during training, you can set ignore_index=0 for the loss. Or if you want it during validating, then just use something like
preds = outputs.detach()[1:].max(dim=1)[1].cpu().numpy()
Hope this helps.
I wondered if setting the ignorable index is necessary, cause my code goes wrong with a bug. Now got fixed, but the performance is not that good. Thanks a lot!