search

NVlabs / DeepInversion

Official PyTorch implementation of Dreaming to Distill: Data-free Knowledge Transfer via DeepInversion (CVPR 2020)
Other
485 stars 77 forks source link

feature_statistics not used #5

Closed cyang-cityu closed 4 years ago

cyang-cityu commented 4 years ago

Why you define a list feature_statistics in imagenet_inversion.py but never use it in the code?

pamolchanov commented 4 years ago

Great observation! In the beginning, we implemented the r_feature loss via feature_statistics list and were outputting mean and variance of channels via the forward function. Later we found that it can be implemented more efficiently with forward hooks but forgot to remove feature_statistics.