mertyg / debug-mistakes-cce Goto Github PK

Hello,
I'm a PhD student working on XAI. And I came across your paper so I decided to test the provided implementation, which I thank you for. The code is very clear and easy to understand !

While studying the code, I noticed the following:
In cce_utils.py, line 68, shouldn't W_clamp_min be computed as

• (W_clamp_min / (min_margins * concept_norms)).T
• instead of (W_clamp_min / (max_margins * concept_norms)).T ?

I'm saying this based on equation (6) provided in section 3.2 of your paper. Please correct me if I'm wrong or if I misunderstood something.

Thank you in advance!

Excellent work! Can I know when you will release the code?

First of all, thank you for your work.

I am re-implementing your codes for spurious detection experiments.

By the way, I found that the wrong evaluation images are used.

Load image list

from dataset import MetashiftManager

dataset_name = "bear-bird-cat-dog-elephant:dog(snow)"
manager = MetashiftManager()
classes, train_domain = dataset_name.split(":")
classes = classes.split("-")
num_classes = len(classes)
shift_class = train_domain.split("(")[0]
spurious_concept = train_domain.split("(")[1][:-1].lower()
print(f"Shift Class: {shift_class}, Spurious Concept: {spurious_concept}")

Result:

Shift Class: dog, Spurious Concept: snow

Get image list for all classes

class_images = manager.get_class_ims(classes)
len(class_images["dog"])

Result:

2580

Check images in "dog" class

from PIL import Image
Image.open(class_images["dog"][2077])

from PIL import Image
Image.open(class_images["dog"][1616])

For classification, there is no "dog" in an image cropped from this image.

Many images in class "dog" are unrelated to class "dog".

Please let me know the reason.

Is your experiment wrong?

I found that the wrong images contain hotdogs...