Using GradCAM for model volo_d2_224

[stevenlimcorn]

I have noticed the available support for ViT model for GradCAM and the exmaple is well structured.

I was wondering if it is possible to use other architectures such as volo_d2_224. In the paper they mentioned about using outlook attention. Unlike ViT which has 192 channels, this model has 512 channels as the output. Using timm library it will look something like this.

timm_model = timm.create_model(model_name='volo_d2_224', pretrained=True)
target_layers = [timm_model.network[-1][-1].norm1]

Did I use the wrong layer for this? or is this still not supported?

Thanks!