Deeplab_ResNet101 with CEloss can not reach 69.5%

meng-tang / rloss

Regularized Losses (rloss) for Weakly-supervised CNN Segmentation

MIT License

207 stars 47 forks source link

Deeplab_ResNet101 with CEloss can not reach 69.5% #2

Closed luwanxuan closed 5 years ago

luwanxuan commented 5 years ago

Hi,

We use the code and hyperparameters you provided in Deeplab_ResNet101 with CEloss, but 20,000 training steps can only reach 62% mIOU. Is there any other tricks? Or does the hyperparameters need to be changed?

I look forward to hearing from you. Also look forward to your code for pytorch or tensorflow.