search

hszhao / PSPNet

Pyramid Scene Parsing Network, CVPR2017.
https://hszhao.github.io/projects/pspnet
Other
1.59k stars 546 forks source link

Figuring ways to train this network ? #60

Open jaypatravali opened 7 years ago

jaypatravali commented 7 years ago

Hello all,

I have been going through the issues, most of it seem not less than 3 months old. So has anyone succesfully trained this network achieving near equal scores as reported in the paper.

My application is a little different. I attempted on the Cityscapes dataset ( for my use case, only 5 labels instead of 19 as given in the paper) gives me a mIoU of 44.9. My current setup is a 2x12GB TitanX pascal, batch size two, using the pspnet50_ADE20K_473.prototxt (added softmaxloss and datapaths) and the pspnet50_ADE20K.caffemodel as pretrained weights. My inputs were resized as 473x473. I know this is not the right way, but just wanted to give it a try.

I have been trying to wrap my head around running this one. Any help is most appreciated.

Thanks!

bhadresh74 commented 7 years ago

@jaypatravali I tried to train the network of my own dataset by adding softmaxloss and datapath. I have 360X480 input size and just three classes. I cant get good training accuracy and the loss is around ~87. Do you think I am missing something?

Please check the below issue link, https://github.com/hszhao/PSPNet/issues/63

Appreciate your help. Bee

ThienAnh commented 7 years ago

@jaypatravali @bhadresh74 Can you public script for train?

bhadresh74 commented 7 years ago

@ThienAnh I have shared it in one of my comments at (https://github.com/hszhao/PSPNet/issues/63) under name train_val.txt

ThienAnh commented 7 years ago

@bhadresh74 Thank you so much. I have new problem: I have same issue with you: loss value alway 87.3365. So how can i initial model to fix it? Thank for your help.

ThienAnh commented 7 years ago

Hi . Where can i get pretrain model for python?