Issue with saving and loading

I am facing a unique problem related to saving and loading the load. I am using the pytorch transfer learning tutorial transfer_learning_tut as the base. On top this code I defined a function so as to take a screen-shot of the model anytime the validation accuracy surpassed the previous best.

Further, at the end of all epochs the tutorial indicates that model is loaded with best_model_wts. Ideally the performance of the model saved best_model_wts and the final_model_saved using the function checkpoint should yeild the same result. However, this is not the case. Am I missing something trivial?

Thanks in advance for your help