Closed pierreablin closed 2 years ago
This is a WIP
@tomMoral yes this is what I'll try to do next.
An idea:
why not compute a matching first between the datasets in torch
and tf
, store it a file (potentially on OSF but I am fine in the repo to start with), and fix this order when loading the data?
that way we can do the train/valid
split in the objective
. This also solves the issue with the slow data consistency check as the order should always be the same, so no need to recompute it anymore.
WDYT @zaccharieramzi ?
I think it's a good idea
All green in local
all green in local, just this problem of the CIFAR DL again, submitting a tentative PR to fix the robustness
no time to wait for the macos workers
This PR adds a validation split to the train / val / test. The test data should only be used to report final perf, validation set is to be used to early stop.