fixing issue with opacus
A bug has been found when calling Opacus dataloader batch_size
attribute: it returns None
, instead of the value of the current batch_size (see https://gitlab.inria.fr/fedbiomed/fedbiomed/-/blob/feature/fixing-fedprox-loss-display/fedbiomed/common/training_plans/_torchnn.py#L510 )
Actually, this opacus dataloader DPDataLoader
returns batch of several size, but in average should equal the size of the batch specified in the training_data_loader
This branch should be merged after feature/fixing-fedprox-loss-display
MR
Edited by BOUILLARD Yannick