torch.nn.Conv3D on MPS backend #77818
Comments
|
bump |
|
Hi, Thanks for the report. |
|
Thanks for the issue, @tfmoraes , we indeed don't have Conv3D support, but should do a better job of reporting error. |
|
@kulinseth is there an intention to support Conv3D and other 3D layers? At InVesalius we use Conv3D in our 3D implementation of Unet. |
|
Hi there, it looks like MPS is only allowing rank 4 tensors, is this something we can (want to?) work around, or do we need MPS to natively support 3D (or nd) convolutions? |
|
any updates on this? |
@tfmoraes , that’s right we are planning to support Conv3D in MPS layer . Unfortunately I can’t discuss the time line but will update the Issue with PR enabling the support . |
|
Hey @kulinseth can I be assigned to the issue? |
We are still waiting on the MPS support to be available in OS and then I can assign the issue . Thanks for your interest @mirceatlx |
|
I am using |
|
This would quite helpful - I'm working through Deep Learning With PyTorch which has a good in-depth case study relying on |
|
After torch 2.0 release the problem is still here. When can users expect conv3s will be compatible with mps on M1 Mac machines? |
|
I just added a pull request that implements 3D convolutions and its backward in MPS. Seems pretty straightforward with certain limitations (no transpose 3D convolution available in MPSGraph). I have also started working on Pooling operations (Max/Avg) in 3D, which are slightly more tricky since MPS only provides 4D operators, so some reshaping is required. |
|
@mattiaspaul I've been using your implementation of the mps 3d convolutions for the last week. Thanks for that! I also have a fully working implementation of mps MaxPool3D locally that uses MPS's maxPooling4DWithSourceTensor function as a workaround. I will open an issue so it can get tagged in a future PR. Are you far along with your implementation of MaxPool3D? Otherwise, I can create a PR. |
|
Here's the issue I created for supporting maxpool3d on mps: #100674 |
|
Hello, when will the 3D convolutions and maxpool3d be incorporated? I saw that both are under review. When can we expect them? Both would be really useful. |
Fixes pytorch#77818 I saw that PR pytorch#99246 was approved, but no one fixed the rebase conflicts, so I am bringing this up again to be merged. I am leveraging @mattiaspaul work. Quoting the description here: > * this pull request enables 3D convolutions (forward/backward) for MPS (Apple Silicon) within the same Convolution.mm file as conv2d. > * does not support channel_last (since pytorch doesn't implement channel_last for 3D tensors) > * does not support conv3d_transpose and treats depth-separable convolutions not as normal case (there are no MPS kernels available for either of those so far) > * requires MacOS >=13.2 (Ventura) Please, let me know if there are any other changes needed and I'll be happy to implement them. Pull Request resolved: pytorch#114183 Approved by: https://github.com/malfet
Fixes pytorch#77818 I saw that PR pytorch#99246 was approved, but no one fixed the rebase conflicts, so I am bringing this up again to be merged. I am leveraging @mattiaspaul work. Quoting the description here: > * this pull request enables 3D convolutions (forward/backward) for MPS (Apple Silicon) within the same Convolution.mm file as conv2d. > * does not support channel_last (since pytorch doesn't implement channel_last for 3D tensors) > * does not support conv3d_transpose and treats depth-separable convolutions not as normal case (there are no MPS kernels available for either of those so far) > * requires MacOS >=13.2 (Ventura) Please, let me know if there are any other changes needed and I'll be happy to implement them. Pull Request resolved: pytorch#114183 Approved by: https://github.com/malfet
🐛 Describe the bug
Using
Conv3Don MPS backend, like in this sample code:Python process are being aborted with this error:
This is the full report:
Versions
cc @kulinseth @albanD @malfet @DenisVieriu97 @razarmehr @abhudev
The text was updated successfully, but these errors were encountered: