Open jiequanz opened 4 years ago
Hi @Jiequannnnnnnnnn,
Apologies for the delayed reply - I've been busy with personal stuff over the past month.
Great to see FlashTorch applied to images of brain (I think?). For my own curiosity, can you tell me a bit more about your project?
I will have some time to address the existing issues/questions in the next week or so, so please expect some update on this soon.
Sorry again for the wait & thanks for your patience!
Hi, I am trying to modify this repo to make it fit on 3D images (64 by 64 by 64)
I have a couple questions.
1, Why did you only take the first Conv2d layer here?
2, I noticed that if I reset "backprop = Backprop(model)", and call "backprop.visualize" again, the "_record_gradients(module, grad_in, grad_out)" will be called twice. If I repeat again, the " _record_gradients(module, grad_in, grad_out)" will be called 3 times. This continue to grow each time I reset "backprop = Backprop(model)". Do you know why? This is shown below. ------------------------------------------------------------------------------- ------------------------------------------------------------------------------- ------------------------------------------------------------------------------- ------------------------------------------------------------------------------- -------------------------------------------------------------------------------
3, I had to remove the line self.model.eval() in the initiation of Backprop to get values for gradients; otherwise the gradient is going to be zero everywhere.
Thanks! Please let me know if you know what happened.