Diiferent result when delete K-Fold Cross Validation

1

I experiment a CNN model with the following path:

 init model -> K-fold CV with 60% data A -> Validate with 20% data B  -> Train with 80% data A+B -> Test with 20% data C

I initiate model once in every experiment, then Inside CV process I deepcopy model:

net_cv = deepcopy(model)

And in the last train (which using 80% data A+B), I deepcopy model again:

net_app = deepcopy(model)

So, net_cv and net_app should be independent. I did this to make both model have same initial weight.

But, when I run, the result of net_app with and without CV Process will give a different result. I don’t know how this is happening. I’ve seed my model once in every experiment using:

torch.manual_seed(11317)
torch.cuda.manual_seed(11317)
np.random.seed(11317)
torch.backends.cudnn.enabled=False

I hope somebody here can help me find the problem.
Thank You

UPDATE
I’ve found that the problem is only occur when the architecture/model is using dropout function, what happens?

2

I’ve found the solution, you should put seed before train phase. So, it will looks like:

init model
     v
K-fold CV with 60% data A
     v
Validate with 20% data B 
     v
Seed model again
     v
Train with 80% data A+B 
     v
Test with 20% data C

I think it happens because dropout randomize run every fold but seed only run in the first time of initiation