imagenet_and_broden_fetcher.py only able to fetch texture concepts about tcav HOT 2 CLOSED

Color-based concept: we simply created a tensor (3D for RGB) and filled in the right channel for each color with some added noise. For example, for red color concept image, you make an empty tensor, then only fill in R channel with full value (whatever your network expects) with some uniform noise! Make sure you visualize this made-up image to ensure the noise level is appropriate (ig you would expect to get varying shades of red).

from tcav.

It looks like that file was only used to create texture concepts from broaden. As you noticed, I think you could create a similar function to download_texture_to_working_folder to have it fetch color-based concepts. https://github.com/tensorflow/tcav/blob/master/tcav/tcav_examples/image_models/imagenet/imagenet_and_broden_fetcher.py#L220 shows that the current logic ignores any color-related images when looking for texture images.

@BeenKim what process did you go through to create color-based concepts from datasets? Was it from broaden?

from tcav.