benchmarks
benchmarks copied to clipboard
Different kinds of preprocessing
Inside preprocessing.py you are using inception preprocessing to train images and vgg preprocessing for evaluation, according to slim. It's a little bit confusing. If you want to train a different network from inception, then this is not the optimized preprocessing in terms of accuracy. Are you planning to integrate different kinds of preprocessing?
We do plan on looking into the preprocessing eventually, as part of an effort to converge to good accuracies. See #104 for more information.
Assigning to @bignamehyp for now.