Calculation of FID metric made use of classification probabilities instead of feature vectors

Based on my understanding, the calculation of fid should make use of feature vector(dim=2048) from the max pool layers of InceptionNet.

However, the FID() metric in fid.py is initiliased as below:

if num_features is None and feature_extractor is None:
    num_features = 1000
    feature_extractor = InceptionModel(return_features=False, device=device)

The InceptionModel return prediction probabilities(dim=1000) if return_features is set to False by specification. update() then make use of this this feature_extractor to get the probabilities instead of feature vectors.

I think feature_extractor should be InceptionModel(return_features=True, device=device) instead?

Start in fid.py by tracing FID() initialization and update() through InceptionModel. Check how return_features and num_features determine the extracted representation, then verify the metric uses 2048-dimensional feature vectors rather than 1000-dimensional classification probabilities and run the relevant existing FID tests.