Fix 'CUDA out of memory'

[atxxfs]

The whole validation dataset (batch_size = 6,854) was passed in after a training epoch. Then it caused 'CUDA out of memory' ERROR.
This is how to fix the problem in main.py. ( The original lines are annotated. )

with torch.no_grad():

    net.eval()
    val_input = val_input.to(device=args.device)
    val_target = val_target.to(device=args.device)

    # out = net(A_wave, val_input)
    # --------------------------------
    tmp_val_losses = []
    tmp_maes = []
    for i in range(0, val_input.shape[0], batch_size):
        out = net(A_wave, val_input[i:i+batch_size, ...])
        loss = loss_criterion(out, val_target[i:i+batch_size, ...]).to(device="cpu")
        tmp_val_losses.append(np.ndarray.item(loss.detach().numpy()))
        out_unnormalized = out.detach().cpu().numpy() * stds[0] + means[0]
        target_unnormalized = val_target[i:i+batch_size, ...].detach().cpu().numpy() * stds[0] + means[0]
        mae = np.mean(np.absolute(out_unnormalized - target_unnormalized))
        tmp_maes.append(mae)
    val_loss = sum(tmp_val_losses) / len(tmp_val_losses)
    validation_losses.append(val_loss)
    mae = sum(tmp_maes) / len(tmp_maes)
    validation_maes.append(mae)
    # --------------------------------
    # val_loss = loss_criterion(out, val_target).to(device="cpu")
    # validation_losses.append(np.ndarray.item(val_loss.detach().numpy()))
    # out_unnormalized = out.detach().cpu().numpy() * stds[0] + means[0]
    # target_unnormalized = val_target.detach().cpu().numpy() * stds[0] + means[0]
    # mae = np.mean(np.absolute(out_unnormalized - target_unnormalized))
    # validation_maes.append(mae)
    out = None
    val_input = val_input.to(device="cpu")
    val_target = val_target.to(device="cpu")

print("Training loss: {}".format(training_losses[-1]))
print("Validation loss: {}".format(validation_losses[-1]))
print("Validation MAE: {}".format(validation_maes[-1]))