search

MisaOgura / flashtorch

Visualization toolkit for neural networks in PyTorch! Demo -->
https://youtu.be/18Iw4qYqfPo
MIT License
734 stars 87 forks source link

Can I use my own model? #32

Closed zxupstar closed 4 years ago

zxupstar commented 4 years ago

If so, are there any requirements for the model?

MisaOgura commented 4 years ago

Hi @zxupstar,

FlashTorch is fairly agnostic to various CNN architectures.

For creating saliency maps, your model needs to have at least one conv2d layer as it will find the first conv2d layer and grabs gradients from it. For activation maximisation, you have the control to specify which layers/filters to visualise.

I'll leave this open for a week in case you have any further questions. After this I will close this so please feel free to open another issue with details.

Many thanks,

SarfarazHabib commented 4 years ago

Hi @MisaOgura I am having problems using activation maximization on my own model. Have mentioned the details in a new issue. Can you please have a look ?

MisaOgura commented 4 years ago

I'm closing this for now - @zxupstar please re-open this if/when needed.

Thanks @SarfarazHabib I will reply on the thread in #34