Selecting best checkpoints based on accuracy not loss in eval.py
When we use cross-entropy as loss function, it is possible to have higher classification accuracy with the higher loss. Especially in the beginning of training. For instance, with a batch size of 10 and 0.5 as softmax probability in the binary classifier, we will get 6.931471805599453 as the loss with 50% accuracy however if just for one sample with 1 as a true class we mistakenly get 0.0001 as softmax probability the loss will be 9.210340371976182 with 90% accuracy. In this case for saving the best checkpoints, this is better to investigate the accuracy, not loss.