DataLoader Epoch에 따라 Random Subset Sampling

https://stackoverflow.com/questions/47432168/taking-subsets-of-a-pytorch-dataset

Taking subsets of a pytorch dataset

 

 

Epoch에 따라 홀수, 짝수에 맞춰서 DataLoader 내 Random Sampling을 수행하며, 전체 데이터를 절반으로 나눈뒤,  짝수, 홀수 Index에 맞춰서 학습시키게끔 수정

 

data_loader = build_data_loader(cfg.data.train)
    
    # for odd, even sampling
    indicies_odds = list(range(1, len(data_loader), 2))
    indicies_evens = list(range(0, len(data_loader), 2))
    
    odd_sampler = SubsetRandomSampler(indicies_odds)
    evens_sampler = SubsetRandomSampler(indicies_evens)
    
    # Odd Data loader
    train_odds_loader = torch.utils.data.DataLoader(
        data_loader,
        batch_size=cfg.data.batch_size,
        sampler = odd_sampler,
        num_workers=8,
        drop_last=True,
        pin_memory=True
    )
    # Even Data loader
    train_evens_loader = torch.utils.data.DataLoader(
        data_loader,
        batch_size=cfg.data.batch_size,
        sampler = evens_sampler,
        num_workers=8,
        drop_last=True,
        pin_memory=True
    )

 

밑에 Train

for epoch in range(start_epoch, cfg.train_cfg.epoch):
        print('\nEpoch: [%d | %d]' % (epoch + 1, cfg.train_cfg.epoch))

        if epoch%2 == 0:
            epoch_avg_rec_acc, writer = train(train_evens_loader, model, optimizer, epoch, start_iter, cfg, train_logger, writer)
        
        elif epoch%2 == 1:
            epoch_avg_rec_acc, writer = train(train_odds_loader, model, optimizer, epoch, start_iter, cfg, train_logger, writer)
        lr_scheduler.step()