Kaushalya Madhawa

Use custom test-time augmentations

2

Thank you for creating this library. I slightly modified `BaseCAM` to pass a `ttach.Compose` object so we can use a custom set of test-time augmentations in `forward_augmentation_smoothing()`.

ValueError: GraphDef cannot be larger than 2GB.

1

Thank you very much for sharing the code publicly. I tweaked the code to run it on GPU without requiring much memory. (can send a PR) When I ran calculate_hessian_inverse.py...

The current implementation performs a brute-force search to find images with embeddings in the k-neighborhood for a text embedding. This can be made efficient using an external library such as...

- [ ] Mention how and where to download ROCO dataset. - [ ] Mention details on the dataset.json files

Perform zero-shot classification on multiple datasets. - [ ] Test set of ROCO dataset with class labels - [ ] Find additional medical image classification datasets

LIPIS-based content loss is missing

2

The paper claims to use a content loss based on a pre-trained VGG-19 model. However, only an L1-loss is used in the training code. How was the performance mentioned in...