Accuracy decrease after load the saved resnet50 model

[saluei (sa_lu)]

Hi
This Problem reported in

Accuracy decrease after load the saved model

My guess is here model accuracy was depend on the batches that go through the testing phase so instead of performing once test for 5 time and average it then save the model and then redo it

However, does not solve.
I have the same problem,

validation and train accuracy of ResNet50 decreased 70% when I loaded the model with load_state_dict() function !!!
However, when I save the training results everything was OK
I use torchvision.models.resnet50 train it with CFAR100 and get 96%, 76% train and test accuracy, save the model satte and then load it with the same hyperparameter.
However I get 40%, 3% train, and test accuracy,
any help is appreciated.

[ptrblck]

Could you compare the outputs of the model for a static input (e.g torch.ones) and after calling model.eva() before saving the state_dict and after loading it?
If the result is equal (up to floating point precision) then I would recommend to check the data loading and make sure both scripts use the same data processing steps.

[saluei (sa_lu)]

Thank,s ptrblck
I do what you suggest and find my bug, in fact, it is my fault in the forward function,
in the first training I use torch.vision forward function,
but the second time when I load the saved model I implemented a custom forward function,
It is my error in custom forward function implementation ,