search

AustinDoolittle / Pytorch-Gain

An implementation of GAIN heatmap network in pytorch. Original paper: https://arxiv.org/abs/1802.10171
30 stars 16 forks source link

For training details #2

Closed xunge closed 4 years ago

xunge commented 4 years ago

Hi:) Really appreciate for your work~ I'm trying to train the GAIN model on ImageNet dataset based on resnet18 for 40 epoches, but the heatmap generated is not satisfied, so I have some questions.

  1. I find that both your code and paper point that the batch size for training is 1, is it necessary?
  2. The pretrained epoch default is 100, but it cost too much for me. Could I use the pretrained models?
  3. Can you show me the heatmap generated yourself?

Thank you for a lot~

AustinDoolittle commented 4 years ago

Hi @xunge! This is a project that I haven't actively worked on for a couple years now, so unfortunately I don't have the pretrained model or any example outputs of this implementation. It is also important to mention that I was not able to recreate the results of the paper, but that could be because my implementation was incorrect. I did have some luck contacting the original author of the paper when I was implementing this, that may be a good route forward if you are interested in further pursuing this algorithm.

Thanks!

AustinDoolittle commented 4 years ago

And per batch size, I do not specifically remember the reason that the batch size was restricted to 1, but I think it had something to do with the online gradcam calculation... not totally sure

AustinDoolittle commented 4 years ago

You may also find some luck in discussing the project with some of the smart folks who have forked this project and extended the work that I have started!

xunge commented 4 years ago

Thanks for your reply, I think your work is correct and I need to try some other models. Appreciate for your work again.

paganpasta commented 3 years ago

@xunge Would you have the modified code for ImageNet dataset? I am trying to work with cifar10 using resnet18 and and having difficulty producing good top-1 accuracy.