Thanks for sharing the code. Can you share the train.log for each dataset with us? Training the model from scratch takes too much time.