* New baseline: Gram! [Detecting Out-of-Distribution Examples with Gram Matrices](https://proceedings.mlr.press/v119/sastry20a.html), ICML 2020. Check out the tutorial to know more :).
* FeatureExtractor.predict() and FeatureExtractor.predict_tensor() now takes a new argument as input: postproc_fns. This argument is a list of Callable functions that applies to the batch features obtained after each (batch-wise) forward. It allows us to perform postprocessing on internal features on the fly rather than after a forward pass on the whole dataset that would require saving the features of the whole dataset in GPU memory, likely leading to an OOM (especially for feature maps).