/black

/black

Hi @Nic-Ma, what's the motivation of this one? I thought we previously said (link) said it wasn't a good idea as the batch wouldn't be contiguously copied to the device?

Hi @Nic-Ma, what's the motivation of this one? I thought we previously said (link) said it wasn't a good idea as the batch wouldn't be contiguously copied to the device?

Hi @rijobro ,

Thanks for your review.
With this transform, we can cache the data in GPU directly then avoid duplicated copying from CPU to GPU in every epoch, and execute following transforms with GPU tensor to accelerate the training progress.
I tested locally with the spleen segmentation tutorial, it can help improve the training speed by 30%.
You can check the unit test of ToDeviced, it's a typical usage example.
And I think this PR is a non-breaking enhancement, if you want to move a batch data to GPU together, you still can use the regular method and use DataLoader num_worker > 0.

Thanks.

Thanks for the reminder @rijobro #1489 Provides more options such as non-blocking... Shall we consider those? Or reopen the PR and work on that one

If you guys think non_blocking is necessary, I can also add it.
Thanks for the great discussion!

Hi @wyli @rijobro ,

Do you guys have other concerns or comments here?

Thanks in advance.

If you guys think non_blocking is necessary, I can also add it.
Thanks for the great discussion!

I'm not familiar with this option, @dongyang0122 any idea?

If you guys think non_blocking is necessary, I can also add it.
Thanks for the great discussion!

I'm not familiar with this option, @dongyang0122 any idea?

The transform is helpful and necessary to enable GPU-based data pre-processing or augmentation. The computing efficiency will be greatly improved.

thanks @dongyang0122 , the .to() API of pytorch has a non_blocking=False by default, https://pytorch.org/docs/stable/generated/torch.Tensor.to.html do you have experience with this flag?

Hi @wyli ,

As the doc said: "When non_blocking, tries to convert asynchronously with respect to the host if possible, e.g., converting a CPU Tensor with pinned memory to a CUDA Tensor. "
Maybe let me add a **kwargs to this ToDevice transform?

Thanks.

thanks @Nic-Ma, sounds good

/black