search

fuxiAIlab / xrl-bench

MIT License
12 stars 2 forks source link

Issue with Saliency Maps Becoming Zeros After Certain Training Steps #3

Open Andy-Mielk opened 2 months ago

Andy-Mielk commented 2 months ago

I have successfully run your project and attempted to apply the imagePerturbationSaliency and imageSarfa methods provided by you to my own pre-trained DQN model. I've noticed that as the model's performance improves, the saliency maps become increasingly clear. However, there's a peculiar issue: after a certain number of training steps, the saliency maps turn completely to zeros. I'm curious if you are aware of the reason behind this and if there's a solution to this problem. Thanks a lot!

yuxiong0521 commented 1 month ago

I have successfully run your project and attempted to apply the imagePerturbationSaliency and imageSarfa methods provided by you to my own pre-trained DQN model. I've noticed that as the model's performance improves, the saliency maps become increasingly clear. However, there's a peculiar issue: after a certain number of training steps, the saliency maps turn completely to zeros. I'm curious if you are aware of the reason behind this and if there's a solution to this problem. Thanks a lot!

I haven't encountered this problem, maybe you can find out why by debugging with breakpoints and let me know!