After completing the warm-up epochs, Torch prints the following warning:

lib/python3.9/site-packages/torch/optim/lr_scheduler.py:131: UserWarning: Detected call of lr_scheduler.step() before optimizer.step(). In PyTorch 1.1.0 and later, you should call them in the opposite order: optimizer.step() before lr_scheduler.step(). Failure to do this will result in PyTorch skipping the first value of the learning rate schedule. See more details at https://pytorch.org/docs/stable/optim.html#how-to-adjust-learning-rate
warnings.warn("Detected call of lr_scheduler.step() before optimizer.step(). "

For anyone recreating this, the img_aug.py file should have this alteration to work correctly I think?

for i in range(len(folders)):
    fd = folders[i]
    tfd = '../../../../' + target_folders[i]

when before it was...

for i in range(len(folders)):
    fd = folders[i]
    tfd = target_folders[i]

Hi there,

What is the total number of pararms for the ProtoPNet with VGG19 and ResNet50 as a base?

Hello,
Would it be possible to share the versions for the packages needed to run the project? I am having trouble figuring out which version to use for each one (Python, Pytorch, Numpy, OpenCV and Augmentor).

Thank you in advance.

In my lab, one epoch uses about 800s, 1000 epochs are too large.

After the training is completed, I understand the process. However, I cannot understand how the prototypes are learned at the beginning? Is it user defined? The paper doesn't say so.

Hi,

I am not sure if there is anything wrong with the training. ProtoPNet is able to achieve 100% accuracy (push an no_push) on CUB_200 with VGG-19 backbone around 10 epochs. However, the paper said that the best accuracy is 78.0 ± 0.2. I am using the settings.py for model training.

Could anyone help with this?

First of all, thank you for sharing the source code of this inspiring work. It's really appreciated!

After training the model, the results I was getting making predictions with local_analysis.py didn’t match the accuracy calculated by the method test() in train_and_test.py.

Please correct if I am wrong, but I believe it’s because in local_analysis.py, model.eval() is missing after loading the model.

Does anyone reproduce the accuracy that is claimed in the paper?

Dear authors,

Thank you so much for releasing the code of this fascinating work!
Is there any chance you could also release the pre-trained ProtoPNet models used in your paper?

This would be very useful for us and the community, I believe. Thanks!

Anh

Hi,

Thank you for the repo :)

I am interested in your work, and I was wondering if you plan on releasing any pre-trained models ?

Thanks,
Elias

Dear authors,

may I ask how did you train the baseline resnet34 model to reach 82 acc in CUB-2000-2011 as reported in the paper?

I used cropped images and pretrained res34 but can still only reach 70...

Thanks a lot for any possible hint in advance!

I tried to reproduce this code. However, an error occurred in "img_aug.py". The error is listed as follows.

raise IndexError("There are no images in the pipeline. "
IndexError: There are no images in the pipeline. Add a directory using add_directory(), pointing it to a directory containing images.

Hi all,

In local_analysis a trained model should be specified, is the intention to use a push or no_push model?

Best regards,

Hi there,

I have followed the instruction and intended to reproduce the result. However, after 100 epochs, even though the training accuracy has around 99% but the testing accuracy is never over 1%. Any suggestions or tips? It is quite odd for me.

Thank you very much.

The rotate step of the pipeline throws an error on the image "Ring_Billed_Gull_0100_52779.jpg"

  File "lib/python3.9/site-packages/Augmentor/Operations.py", line 843, in do
     image = image.crop((int(round(E)), int(round(A)), int(round(X - E)), int(round(Y - A))))
  File "lib/python3.9/site-packages/PIL/Image.py", line 1206, in crop
     raise ValueError("Coordinate 'lower' is less than 'upper'")
ValueError: Coordinate 'lower' is less than 'upper'

Replacing the call to rotate on line 23 with rotate_without_crop fixes the problem:

p.rotate_without_crop(probability=1, max_left_rotation=15, max_right_rotation=15)

The work is very interesting.

The work is very interesting. I follow the tutorial to train the model. When I use local_analysis.py, I get an error in the model where the picture cannot be input.

The error code is here:

At the same time, normalize cannot process a single picture.

Hi there, my machine trains the model one epoch using 1200s, it's too long, do you know how to continue training with the last weights? Do model_dir need to be changed? Thanks!