search

utkuozbulak / pytorch-cnn-visualizations

Pytorch implementation of convolutional neural network visualization techniques
MIT License
7.85k stars 1.49k forks source link

How to use the backprop on densenet and other multi-channel models? #26

Closed nessieyang closed 6 years ago

nessieyang commented 6 years ago

Hi, I want to move this code to densenet and other models. Here is the problem: When I run the code 'guided_backprop' based on DenseNet. Should I also replace the Relu inside the DenseBlock and Dense Layers? I know that we need to replace the Relu layers in VGG, but in Dense Block, the connection is a little bit different and complex. So do we need to replace them as well?

utkuozbulak commented 6 years ago

Its tricky, I would try giving a shot on replacing every ReLU, then only inside DenseBlock then only in DenseLayers to see what it produces.

nessieyang commented 6 years ago

I have tried but the results seems like not so good. I will try it more.

On 09/18/2018 21:01, Utku Ozbulak wrote:

Its tricky, I would try giving a shot on replacing every ReLU, then only inside DenseBlock then only in DenseLayers to see what it produces.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.

LukeJaffe commented 6 years ago

Hey @nessieyang, any progress on getting this to work with DenseNet?

dontLoveBugs commented 5 years ago

Densenet has a bn layer, does it affect the results?

utkuozbulak commented 5 years ago

@dontLoveBugs I'm not sure, perhaps there are studies that analyzed the effects of BN for backprop-related techniques.

Let us know if you find any.