laura-rieger/deep-explanation-penalization

Code for using CDEP from the paper "Interpretations are useful: penalizing explanations to align neural networks with prior knowledge" https://arxiv.org/abs/1909.13584

deep-explanation-penalization's Issues

How do I get segmentation files of ISIC dataset?

Hi, I appreciate your awesome code!!
Btw how do I get segmentation files of the ISIC dataset?
I can't find any code in 02_sort_imgs.py that save
segmentation files in "segmentation_path = os.path.join(data_path, "segmentation")"

I found in another issue that I can download SONIC study in ISIC website
but I still don't know if this data is already included in the dataset I can download with
01_download_imgs.py file.

So which data of the study in the ISIC dataset did you exactly use for the ISIC result in your paper?
There were the following studies.

2018 JID Editorial Images (0 / 100)
Dermoscopedia (CC-BY) (0 / 5)
HAM10000 (0 / 10015)
MSK-1 (0 / 1100)
MSK-2 (0 / 1535)
MSK-3 (0 / 225)
MSK-4 (0 / 947)
MSK-5 (0 / 111)
SONIC (9251)
UDA-1 (0 / 557)
UDA-2 (0 / 60)

error in dataset creation of decoyMNIST

The dataset creation and loading for decoy has a bug (getting the error below)
Could it be looked into @laura-rieger @csinva ?

python train_mnist_decoy.py

Traceback (most recent call last):
  File "train_mnist_decoy.py", line 107, in <module>
    complete_dataset = utils.TensorDataset(train_x_tensor,train_y_tensor) # create your datset
  File "/home/avani.gupta/miniconda3/envs/py3.8/lib/python3.8/site-packages/torch/utils/data/dataset.py", line 365, in __init__
    assert all(tensors[0].size(0) == tensor.size(0) for tensor in tensors), "Size mismatch between tensors"
AssertionError: Size mismatch between tensors

Why not fine-tune the weights of conv layers in the experiment of ISIC

Hi,

I notice that you choose to freeze the conv layers and only train the fully connected layers when using a pretrained VGG16 model in the ISIC experiment. Could you please, if convenient, tell me why not retrain the whole model since this seems to be an extra accuracy gain compared to training only fc layers.

All best.

Downloading data for the ISIC experiments

Hi!
I am trying to download the data for the reported ISIC experiments using the scripts you provided. I encountered the following issues:

• 00_download_metadata.py prints Fetching metadata for 23906 images. The paper states The ISIC dataset consists of 21,654 images (19,372 benign). How can I figure out which images you used?
• 01_download_imgs.py only downloads 50 images although in the config the value is set to 25000. Am I expected to change start_offset = 0 or the limit parameter in the URL somehow?

Thanks a lot for the great work!
Best regards
Verena

Hyperparameter of ColorMNIST

Hi, Laura.
I found that the performance(Test accuracy=23.41%) of CDEP that the jupyter notebook in this repo says in ColorMNIST dataset is different from the one in ICML paper(Test accuracy=31.0%).
So could you tell me the hyperparameters that you use to get the performance? (lr, lambda, batch_size, seed)
Thanks a lot!

Performance in isic-skin-cancer/analyze_isic.ipynb is too different from the one in the paper

Hi, long time no see
Could you tell me in more detail that how you get the result of isic datset in the paper?
For example, hyperparameters like batch_size, learning_rate, reg_rate extra.
Cause the performance in the paper and the ipynb file is quite different.
Thanks in advance!

Extension to natural images?

First, really interesting work -- I happened to chance upon this repo, as Github recommended it.

For computer vision e.g., ColorMNIST have you considered natural images? Is there a bottleneck in terms of memory or compute, that required an MNIST variant? (Or is it just more difficult to find an obvious bias in natural images?)

The text gender bias is really interesting. Will have take a closer read before asking questions though!

There is no gradcam file in pytorch-cnn-visualizations/src

Hi, again.
While I'm using analyze_isic.ipynb I found out that there is no gradcam related python file in pytorch-cnn-visualizations/src path
Could you upload the file?

ISIC images

Since the ISIC archive has changed, it's not clear anymore which images you used to train the models for the ISIC experiments. Specifically, it's unclear which images without patches were used. Would you be able to provide a table with the names of the included ISIC images?

Thanks!

Identifying Colorful Patches

Hello all,

First of all, I want to say great job on this paper and thank you so much for making the code available in such a usable fashion. I am trying to reproduce some of your results for the ISIC dataset. Specifically, I am trying to generate labels for benign images that identify the colorful patches. Is this something that can be done using your codebase? Let me know, thanks!

How come Gradient Sum and EG do two gradient steps?

Hi! Thanks for sharing this repo!

In the MNIST Decoy code, for method 1 and 2 (gradient_sum and eg), there are two gradient steps per batch. The first step uses gradients from just explanation_penalty, and the second step uses the gradients from both the explanation_penalty and the log loss. I was wondering what the reason for this was?

Reference in code:

deep-explanation-penalization/mnist/DecoyMNIST/train_mnist_decoy.py

Line 168 in 8249af7

elif args.grad_method ==1:

Thanks!

What's the performance when CNN is also trained?

Hi, Laura.
I found that with the vgg16 you trained with the ISIC dataset, actual training only happens on the FC layer at the end.
Have you tried training all the layers as well?
Thanks in advance!

About segmentation imags

I can't find segmentation imags in the segmentation data file. Could you provide this seg data?