Issue with PILMask in segmentation

Part 1 (2020)
1

So here is the issue:

I first tried to use MaskBlock in my DataBlock and I got CUDA runtime error when I started training. So I thought maybe there is something wrong with my labels. When I checked the output from PILMask.create(which is used inside MaskBlock) I saw this:

image
image810×348 13.9 KB

It is completely changing my labels and they don’t mean anything now! (Compare it with when I open them simply by Image.open)

Any solution to the problem?

2

Okay I think I found the problem:

image
image1431×789 199 KB

I wrote my own transform to get rid of any value greater than 20 (because it would be out of the labels defined for dataset) and I created a new block with my new Transform.
It is working now and there is no more error!

1 Like