search

yu-changqian / TorchSeg

Fast, modular reference implementation and easy training of Semantic Segmentation algorithms in PyTorch.
MIT License
1.41k stars 253 forks source link

About performance of PSPNet-ResNet50 #80

Open dontLoveBugs opened 5 years ago

dontLoveBugs commented 5 years ago

Hi, I use your code to train PSPNet-ResNet50 in ade dataset and my result is:

Method backbone Mean IoU Accuracy
PSPNet resnet50v1c 38.98% 78.52%

The result is worse than yours. My environment is pytorch 1.0, cuda 10.

yu-changqian commented 5 years ago

Could you give me more details? What's the performance results with this checkpoint?

dontLoveBugs commented 5 years ago

I have tested the checkpoint and it has the same checkpoint as you shows. However, when I train the model from scratch，all my trained results are worse than yours. At fact, I has changed your code in order to read datasets without the train/val.txt files and added tensorboard to visulize training process. But I think these changes should not make a great difference. This is my changed code.

yu-changqian commented 5 years ago

Maybe you can use the lr=2e-2. I will also test the setting with the distributed-parallel api if I have the free machines.

dontLoveBugs commented 5 years ago

At fact, I test various parameters combinations, such as lr 0.1 to 0.2, weght_decay 0.0001 to 0.0005, with or without using apex api.