search

princeton-vl / RAFT

BSD 3-Clause "New" or "Revised" License
3.27k stars 633 forks source link

Trian model with masked custom dataset #100

Open Roypic opened 3 years ago

Roypic commented 3 years ago

Hi！

Thanks for sharing the work！

There‘s a problem when I try to train a model with custom dataset; When I train the origin dataset which composed just like public dataset you listed below, the model works well and the loss drops quickly. But, when I add the mask to the both input and label, things got changed, the loss keeps still after 100 epoch and seems really big compared to the one without mask. Even if I add the mask to the output, problem still exists. But this problem did not happened to other models like flownet. Could you please give some suggests about what probably the problem is and let me figure out？

Really Thanks.

zhilaAI commented 3 years ago

how did you train on a custom dataset?

zhilaAI commented 3 years ago

I could train the model on my own dataset but gives me all nan. Did you face the same problem?

ting-liu-nycu commented 1 year ago

Hi！

Thanks for sharing the work！

There‘s a problem when I try to train a model with custom dataset; When I train the origin dataset which composed just like public dataset you listed below, the model works well and the loss drops quickly. But, when I add the mask to the both input and label, things got changed, the loss keeps still after 100 epoch and seems really big compared to the one without mask. Even if I add the mask to the output, problem still exists. But this problem did not happened to other models like flownet. Could you please give some suggests about what probably the problem is and let me figure out？

Really Thanks.

Hi, I also met same problem!! How did you solve it in the end? Thank you so much.

swjtuer0762 commented 8 months ago

how did you train on a custom dataset?