reproducibility with dropout

[arakoto]

Hi all,

I am training a CNN model with a dropout and I would like the model I obtain to be reproducible (when rerunning the code).

I have seeded the random generator manually at the beginning of the code  

 torch.manualSeed(1)

just after importing torch but I still do not get reproducible results.

can someone tell me if  there is something else I need to do for this purpose?

thanks a lot in advance

A.

[Vislab]

use mode:evaluate() to disable the dropouts and you will get what you are looking for when evaluating your model

[Francisco Vitor Suzano Massa]

If you are using cuda, you also need to set
cutorch.manualSeed

[arakoto]

Thanks a lot for all your answers.  

seeding cutorch seems to work ok.

A.