search

zbf1991 / RRM

Reliability Does Matter: An End-to-End Weakly Supervised Semantic Segmentation Approach
94 stars 11 forks source link

Can people get the result in the paper if they train it from 'init'? #10

Closed johnnylu305 closed 4 years ago

johnnylu305 commented 4 years ago

Hi, nice to meet you. There are two pretrain weights about your classification network. One is "ilsvrc-cls_rna-a1_cls1000_ep-0001.params" and another one is "res38_cls.pth". My question is that can people get the same result in your paper if they train from "ilsvrc-cls_rna-a1_cls1000_ep-0001.params"?

zbf1991 commented 4 years ago

yes, no matter train from 'init' or from 'cls', the final performance is the same. I have tested on my computer. A few days ago, another researcher told me he/she cannot get the result using the 'init', I have found some bugs and fixed it. If you cannot get the result, please let me know, I will check it again.

johnnylu305 commented 4 years ago

@zbf1991 Thank you for your explanation. Sorry, I forgot to ask another minor question. Why you set densecrfloss to 1e-7? Doesn't it disable your energy loss?

johnnylu305 commented 4 years ago

I think densecrfloss is just a hyperparameter to adjust energy loss. This is because the value of energy loss is very small(ex: -4700377).