How to use InceptionV3 and densenet161？

kazuto1011 / grad-cam-pytorch

PyTorch re-implementation of Grad-CAM (+ vanilla/guided backpropagation, deconvnet, and occlusion sensitivity maps)

MIT License

787 stars 174 forks source link

How to use InceptionV3 and densenet161？ #25

Closed tuji-sjp closed 4 years ago

tuji-sjp commented 4 years ago

I find the "target_layer_names" of these two models, but when I run the modified code, I get the following error:

RuntimeError: size mismatch, m1: [1 x 277248], m2: [768 x 1000] at /opt/conda/conda-bld/pytorch_1535490206202/work/aten/src/THC/generic/THCTensorMathBlas.cu:249

How do I solve it? Please help me, thank you very much!

kazuto1011 commented 4 years ago

I cannot guess what caused the error because:

There's no hint related to this repository in the reported error.
This repository does not contain the keyword target_layer_names.

About another repository? https://github.com/jacobgil/pytorch-grad-cam/blob/master/grad-cam.py#L76

tuji-sjp commented 4 years ago

Oh, this is indeed the code of another repository, I made a mistake, so sorry. And "target_layer_names" is yours "target_layer".

kazuto1011 commented 4 years ago

In this repository, you can get various maps as follows:

$ python main.py demo1 --arch "inception_v3" --target-layer "Mixed_7c" -i "samples/cat_dog.png"
$ python main.py demo1 --arch "densenet161" --target-layer "features" -i "samples/cat_dog.png"

Please note that:

For Grad-CAM, we want to extract the activation maps from the last convolution to be exact but in densenet161 case the "features" output is not activated (currently cannot access this line).
The gradient maps generated by Deconvnet and Guided Backpropagation are not valid because torchvision's Inception v3 and DenseNet-161 includes F.relu not only nn.ReLU which my code cannot handle.

tuji-sjp commented 4 years ago

Hello, I fused the logits of the three models. Maybe you know how does this ensemble model gets its GradCAM?