Are there any particular differences with pytorch imagenet loader (which is typically implemented using ImageFolder) and the torch loader?
I ask because I am trying to use the weights from soumith inception_v1 repo (https://github.com/soumith/inception.torch) and port them in pytorch. I have implemented the model and loaded the weights from the h5py files (https://github.com/antspy/inception_v1.pytorch) , but accuracy on imagenet is terrible.
I am not sure what is wrong, but some spot checks suggest that the labels are not in the same order. So I was wondering if this is a known issue and that we need to permute the weights of the last layer.