Lrp tutorial added

[JohannesK14]

Reference Issues

664

Changes

Addition of an LRP tutorial and a sample image. The tutorial replicates the results of the basic PyTorch LRP implementation provided by the LRP authors using the LRP class of captum.

[facebook-github-bot]

Hi @JohannesK14!

Thank you for your pull request and welcome to our community.

Action Required

In order to merge any pull request (code, docs, etc.), we require contributors to sign our Contributor License Agreement, and we don't seem to have one on file for you.

Process

In order for us to review and merge your suggested changes, please sign at https://code.facebook.com/cla. If you are contributing on behalf of someone else (eg your employer), the individual CLA may not be sufficient and your employer may need to sign the corporate CLA.

Once the CLA is signed, our tooling will perform checks and validations. Afterwards, the pull request will be tagged with CLA signed. The tagging process may take up to 1 hour after signing. Please give it that time before contacting us about it.

If you have received this in error or have any questions, please contact us at cla@fb.com. Thanks!

[facebook-github-bot]

Thank you for signing our Contributor License Agreement. We can now accept your code for this (and any) Facebook open source project. Thanks!

[NarineK]

Thank you for the tutorial, @JohannesK14! We will provide review comments soon!

A quick question: Where is the castle.jpeg from ? We need to know the copyright permissions and the attribution required for that image.

[NarineK]

Thank you very much for working on this tutorial, @JohannesK14 ! Overall looks great to me.

One ask:
Since we are adding more tutorials to Captum I'm thinking that perhaps we can combine this with the resnet one and call it Resnet_VGG_TorchVision_Interpret.ipynb, if we want to showcase LRP example explicitly for VGG. The content of this tutorial can be added to the bottom of the Resnet_VGG_TorchVision_Interpret.ipynb file.
In the top of theResnet_VGG_TorchVision_Interpret file we can briefly describe that we are showcasing LRP for VGG as well.

Feel free to add an example of LRP for Resnet as well in that same notebook.
https://github.com/pytorch/captum/blob/master/tutorials/Resnet_VGG_TorchVision_Interpret.ipynb

If you combine those tutorials, I'd also suggest that you do minor adjustments for the website text as well:
https://github.com/pytorch/captum/blob/master/website/tutorials.json#L17
https://github.com/pytorch/captum/blob/master/website/pages/tutorials/index.js#L70

Couple more minor comments:
"As one see" -> "As one can see" ?
"Currently implemented in captum are " -> "Currently implemented rules in captum are " ? 
 Also, do you mind adding here a link to the documentation for the recommended rules ?

[JohannesK14]

Thank you for the tutorial, @JohannesK14! We will provide review comments soon!

A quick question: Where is the castle.jpeg from ? We need to know the copyright permissions and the attribution required for that image.

Hello @NarineK ,

I took the image from here: https://git.tu-berlin.de/gmontavon/lrp-tutorial

I could ask the author for the copyright permission.

[JohannesK14]

Regarding your second comment:

Of course, I can add the tutorial to the Resnet_TorchVision_Interpret.ipynb. But then I would suggest omitting VGG and just use Resnet and also the swan picture. Otherwise, it will look a little bit thrown together, I think. And Resnet is also more interesting than VGG, because there you need the implementation with back hooks for the skip connections.

And you would not need copyright permission for the castle anymore.

I will have a look into that in the next few days.

[NarineK]

Regarding your second comment:

Of course, I can add the tutorial to the Resnet_TorchVision_Interpret.ipynb. But then I would suggest omitting VGG and just use Resnet and also the swan picture. Otherwise, it will look a little bit thrown together, I think. And Resnet is also more interesting than VGG, because there you need the implementation with back hooks for the skip connections.

And you would not need copyright permission for the castle anymore.

I will have a look into that in the next few days.

Hi @JohannesK14, agree, let's use Resnet and the swan image instead.

[NarineK]

@JohannesK14 would you have time to address the comments in this PR as discussed previously ?

[HeinrichAD]

Unfortunately, the example notebook does not work anymore.

RuntimeError: Output 0 of BackwardHookFunctionBackward is a view and is being modified inplace. This view was created inside a custom Function (or because an input was returned as-is) and the autograd logic to handle view+inplace would override the custom backward associated with the custom Function, leading to incorrect gradients. This behavior is forbidden. You can fix this by cloning the output of the custom Function.

---

I like the example. Does anybody has an idea how to solve the issue?
I mean, since it's a pre-trained model, we can't just change the inline relu layer to a "normal" relu layer, for example.

[JohannesK14]

@HeinrichAD I just set the inplace flag to false. Why do you see a problem there? To my knowledge, the inplace functionality could save some memory, but should deliver the same results in a VGG.

@NarineK Sorry for my long absence. I tried the switch to the ResNet architecture this morning. Unfortunately, the design of the composition of the rules for the ResNet18 architecture is more complicated than I thought it would be. I don't have the time to tinker with the rules to create a acceptable result for the visualization of the attributions, which I am very sorry for. Therefore, I switched to the picture of the swan, which also gives a pleasant result and avoids copyright issues.

Do you think I should still merge it with the ResNet tutorial as you suggested earlier?

[HeinrichAD]

@JohannesK14 I looked into the source code [1]. In this case it will be the same. So no problems here.

[1] https://pytorch.org/vision/main/_modules/torchvision/models/vgg.html#vgg16