CPU-heavy operation slowing training about time-series-library

[harbingerarbitrager422]

singled it out to test_loss = self.vali(test_data, test_loader, criterion) that's causing problems. for some reason it doesnt run on the gpu while everything else does?

test data is being loaded with batch_size of 1, so if your dataset is huge, the testing phase is essentially a long for loop, which is going to be slow. It does run on GPU but with a batch_size of 1 I suppose the GPU memory footprint is too small for you to notice. To speed up the training, simply comment out this line. It shouldn't pose any issue since you do still perform inference on the validation set

[ashpoolresearch]

singled it out to test_loss = self.vali(test_data, test_loader, criterion) that's causing problems. for some reason it doesnt run on the gpu while everything else does?

[lizhuoq]

singled it out to test_loss = self.vali(test_data, test_loader, criterion) that's causing problems. for some reason it doesnt run on the gpu while everything else does?

test data is being loaded with batch_size of 1, so if your dataset is huge, the testing phase is essentially a long for loop, which is going to be slow. It does run on GPU but with a batch_size of 1 I suppose the GPU memory footprint is too small for you to notice. To speed up the training, simply comment out this line. It shouldn't pose any issue since you do still perform inference on the validation set

I think this is correct for very large datasets, as it indeed wastes time and the test_loss during the training phase, besides being logged, serves no purpose, so I also agree with commenting out this line of code.