EPFL-VILAB / omnidata-paper-code-dump

19 stars 3 forks source link

problems of WeightedRandomSampler in multi-gpu training #5

Open icoz69 opened 2 years ago

icoz69 commented 2 years ago

hi, i am running train_depth with multi gpus, and find a problem that each gpu processes the same subset from the dataloader. I am using the default multi-gpu mode, ie strategy=‘ddp’ . However, by printing the loaded data of training batches, I found the inputs to all gpu nodes are the same. I think the problem lies in that the original WeightedRandomSampler does not apply to DistributedDataParallel.