search

hustvl / WeakTr

WeakTr: Exploring Plain Vision Transformer for Weakly-supervised Semantic Segmentation
MIT License
120 stars 2 forks source link

Problems with reproduction #2

Closed siyueyu closed 1 year ago

siyueyu commented 1 year ago

Thanks for the awesome work! It is really interesting and powerful!

I followed the steps in README to reproduce the results reported in the paper. However, for step 1, I can only get a performance of around 64% for VOC validation. Could you please share some possible reasons with me for this circumstance?

Unrealluver commented 1 year ago

Thanks for your interest. We run the WeakTr CAM training (step 1) on a single A4000 GPU. And we noticed that the CAM training process would be different when launched on different kinds of GPUs. We have some suggestions for you to get a satisfactory result:

  1. fix the random seeds and slightly adjust the learning rate.
  2. adjust the total training epochs or use an early stop to prevent overfitting.