search

holyseven / PSPNet-TF-Reproduce

Training PSPNet in Tensorflow. Reproduce the performance from the paper.
MIT License
125 stars 30 forks source link

structure_in_paper #9

Closed FengLoveBella closed 6 years ago

FengLoveBella commented 6 years ago

Hi @holyseven, nice work. I have a problem when using 'structure_in_paper' setting in your released paper, because of the initial model ResNet_v1_101 is different to the code in structure_in_paper. Could you share the initial resnet101 model which can be used in the 'structure_in_paper' mode? Thank you very much!

holyseven commented 6 years ago

I have no such initial model. Discussion with the author of PSPNet:

For the beginning blocks of ResNet, the original implementation use kernel size as 7 while we use 3, these two both use stride as 2 to downsize the input size as half of the original input image for faster training. It does not make much difference regards the final accuracy, while kernel size as 3 is faster than 7 for training.

There are some trained models from https://github.com/hszhao/PSPNet, but on ADE, VOC or Cityscapes.

FengLoveBella commented 6 years ago

@holyseven Thank you very much. And another question, did you try class weight in your code? does it help? (Because the classes are inbalance)

holyseven commented 6 years ago

FCN tried to do the class balancing on VOC but found unnecessary.

Although our labels are mildly unbalanced (about 3/4 are background), we find class balancing unnecessary.

I tried the simplest (giving each class a weight, the inverse of pixel number rate) and focal loss with gamma=2, but found no help. Maybe more hyper-parameter tuning or more advanced techniques will help.

FengLoveBella commented 6 years ago

@holyseven OK, I got it, 3ks.