Issue with ViT in BioClip Visual Part: ViT Returns CLS Token Instead of Logits

[Link7808]

We are using the visual part (ViT) of BioClip to process images. However, there is an issue with the forward method in BaseCAM.
In the following line of code:
self.outputs = outputs = self.activations_and_grads(input_tensor)
target_categories = np.argmax(outputs.cpu().data.numpy(), axis=-1)
The outputs in this case is the CLS token embedding, which is a high-dimensional vector used to represent the global semantic information of the input image. This embedding is not a classification result or logits, but rather a feature vector.

[johnbradley]

@Link7808 Do you have some example code that reproduces this issue? Is the BaseCAM class from https://github.com/jacobgil/pytorch-grad-cam ? If so I have a first attempt jupyter notebook that uses pytorch-grad-cam with pybioclip.

[Link7808]

@johnbradley
Let me clarify the issue. In this notebook, we are using:
classifier = TreeOfLifeClassifier() model = classifier.model.visual targets = None
BaseCAM attempts to automatically get the index of the predicted class with:
if targets is None: target_categories = np.argmax(outputs.cpu().data.numpy(), axis=-1) targets = [ClassifierOutputTarget(category) for category in target_categories]
However, the issue arises because outputs is the CLS token embedding, which is a feature vector representing the global semantic information of the input image. It’s not logits, so it can’t be used to correctly get target_categories.
This causes Grad-CAM to not correctly select the target for visualization.

[johnbradley]

@Link7808 After looking through the gradcam code I see what you mean about it expecting a different output.
I updated the notebook in in the gradcam branch to use a custom model.
This follows a pattern I found on a gradcam PR.

---

The changes I made where:

class ImageClassifier(nn.Module):
    def __init__(self):
        super(ImageClassifier, self).__init__()
        classifier = TreeOfLifeClassifier()
        self.clip = classifier.model
        self.txt_features = classifier.txt_features
        self.txt_names = classifier.txt_names

    def forward(self, x):
        img_features = self.clip.visual(x)
        img_features = F.normalize(img_features, dim=-1)
        logits = (self.clip.logit_scale.exp() * img_features @ self.txt_features)
        result = F.softmax(logits, dim=1)

        # Print out the target found
        for target in np.argmax(result.cpu().data.numpy(), axis=-1):
            print(target, self.txt_names[target])

        return result

The above code will print out the target number and label associated with it.

Outside of adding a couple imports the only other changes were to instantiate the new model and set target_layers.

model = ImageClassifier()
...

target_layers = [model.clip.visual.transformer.resblocks[-1].ln_1]

[Link7808]

I noticed you’re using eigencam. Do you have any thoughts on why Grad-CAM is performing poorly?

[johnbradley]

@Link7808 I've found setting eigen_smooth=True when calling create_grad_cam_image() with method="gradcam" results in more reasonable outputs.

Help for eigen_smooth in vit_example.py says

Reduce noise by taking the first principle componet of cam_weights*activations

This pytorch_grad_cam code seems to be where the eigen_smooth flag takes effect.

[johnbradley]

I'm not sure why without eigen_smooth=True the results don't look very good.