classification using best model gives strange results

[Gabriele Campanella]

Hi there!

I'm using facebook's 18 layer resnet model for a 6 class classification task. The validation error was around 1%. I saved the best model and used that to classify again the validation set to get the labels and build a confusion matrix. However, when pushing the images forward through the network, I obtained completely wrong labels (mostly 1s and a few 2s on a 6 class task). I didn't calculate accuracy, but it is nowhere near 99%.

The gist below is the code I used for classifying the images given my best model.
https://gist.github.com/gabricampanella/b7f6941b05b087a557c202a519ee0b11

Thank you.