Samples of eval always be —— pixAcc: 100.000, mIoU: 50.000

[YangYangGirl]

[YangYangGirl]

When I tried to fix the bugs, I had tried：

1. Visualize
   Found the demo_vis.png filled with all red pixels.
   2.Train only two imgs, and eval them.
   Found the bugs alive.
2. Print items of datasets.
   
   then I got:
   
   
3. Check the calculation of Acc
   When input imgs of [100, 100], got pixAcc = 8186/8186 = 1.

It seemed to be perfect when I checked, so I still be confused about the terrible results :(

[YangYangGirl]

Here is my code:
https://github.com/YangYangGirl/SegmenTron/blob/master/segmentron/data/dataloader/mots.py

[YangYangGirl]

SegmenTron/segmentron/data/dataloader/ade.py

Line 73 in 4bc605e

return torch.LongTensor(np.array(mask).astype('int32') - 1)

Why do you minus 1? Maybe it's the key to solve the problem.

[YangYangGirl]

By getting rid of the minus one operation, the model works !

[LikeLy-Journey]

sorry for wasting you so much time. Minus 1 may because the label of ade dataset not start from 0.

[YangYangGirl]

Thank you for your reply. For further verification, I downloaded the data set of ade and checked that its mask started from 0. Your explanation may not be correct.