prediction failure on Cityscapes images

[remtav]

While testing the code with supplied Cityscapes config file and pretrained weights, about 1 prediction out of 80 fails completely. See attached pictures. Did not occur on original tensorflow model.

[zzllc3]

I met the same problem on the other dataset. Do you have any idea to solve this problem?

[sunlibocs]

I also meet this problem. Has anyone found a solution?

[nyoki-mtl]

Sorry for late reply.
This problem is because the default value of tensorflow batch normalization eps is different.
So, it can be solved by calling model.update_bn_eps ().

pytorch-segmentation/src/eval_cityscapes.py

Line 60 in 7a30cab

model.update_bn_eps()

[PierfrancescoArdino]

Sorry @nyoki-mtl when you have to call model.update_bn_eps() in the evaluation/prediction?

[nyoki-mtl]

@PierfrancescoArdino
If you want to use the trained model converted from tensorflow, you should call update_bn_eps.
But, when you train or predict on your model, you don't need to do.

[PierfrancescoArdino]

I've tried only the "eval_cityscapes.py" script without changing anything with the converted trained model from TF and I have the same problems of the first post.