RuntimeError: Conv3D is not supported on MPS

[tlnagy]

Attempting to segment a 3D image using plant-seg master gives RuntimeError: Conv3D is not supported on MPS.

I believe this is due to the v2.2.2 version of pytorch that is currently being installed. From my sleuthing, it looks like this is potentially resolved in pytorch >2.3.0 because pytorch/pytorch#114183 was merged.

[qin-yu]

If you install PlantSeg with conda/mamba, did you try mamba install pytorch==2.3.1 torchvision==0.18.1 torchaudio==2.3.1 -c pytorch [ref]? I don't have access to a Mac at the moment.

[tlnagy]

Unfortunately, I keep running into conflict resolution issues. I first ran tnagy in plant-seg on  qy/fix-pred-advanced [!] ➜ mamba create -n plant-seg2 -c pytorch -c conda-forge python=3.11 pytorch::pytorch plant-seg --no-channel-priority followed by mamba install pytorch==2.3.1 torchvision==0.18.1 torchaudio==2.3.1 -c pytorch and I got the following error:

tnagy in plant-seg on  qy/fix-pred-advanced [!] ➜  mamba create -n plant-seg2 -c pytorch -c conda-forge python=3.11 pytorch::pytorch plant-seg --no-channel-priority
tnagy in plant-seg on  qy/fix-pred-advanced [!] ➜  mamba install pytorch==2.3.1 torchvision==0.18.1 torchaudio==2.3.1 -c pytorch
pytorch/noarch                                              Using cache
pytorch/osx-arm64                                           Using cache
DEPRECATION: --no-python-version-warning is deprecated. pip 25.1 will enforce this behaviour change. A possible replacement is to remove the flag as it's a no-op. Discussion can be found at https://github.com/pypa/pip/issues/13154

Pinned packages:

  - python=3.11

error    libmamba Could not solve for environment specs
    The following packages are incompatible
    ├─ pytorch ==2.3.1 * is not installable because there are no viable options
    │  ├─ pytorch 2.3.1 would require
    │  │  └─ python >=3.10,<3.11.0a0 *, which does not exist (perhaps a missing channel);
    │  ├─ pytorch 2.3.1 would require
    │  │  └─ llvm-openmp <16 *, which does not exist (perhaps a missing channel);
    │  ├─ pytorch 2.3.1 would require
    │  │  └─ python >=3.12,<3.13.0a0 *, which does not exist (perhaps a missing channel);
    │  ├─ pytorch 2.3.1 would require
    │  │  └─ python >=3.8,<3.9.0a0 *, which does not exist (perhaps a missing channel);
    │  └─ pytorch 2.3.1 would require
    │     └─ python >=3.9,<3.10.0a0 *, which does not exist (perhaps a missing channel);
    ├─ torchaudio ==2.3.1 * is not installable because there are no viable options
    │  ├─ torchaudio 2.3.1 would require
    │  │  └─ pytorch ==2.3.1 *, which cannot be installed (as previously explained);
    │  ├─ torchaudio 2.3.1 would require
    │  │  └─ python >=3.10,<3.11.0a0 *, which does not exist (perhaps a missing channel);
    │  ├─ torchaudio 2.3.1 would require
    │  │  └─ python >=3.12,<3.13.0a0 *, which does not exist (perhaps a missing channel);
    │  ├─ torchaudio 2.3.1 would require
    │  │  └─ python >=3.8,<3.9.0a0 *, which does not exist (perhaps a missing channel);
    │  └─ torchaudio 2.3.1 would require
    │     └─ python >=3.9,<3.10.0a0 *, which does not exist (perhaps a missing channel);
    └─ torchvision ==0.18.1 * is not installable because there are no viable options
       ├─ torchvision 0.18.1 would require
       │  └─ pytorch ==2.3.1 *, which cannot be installed (as previously explained);
       ├─ torchvision 0.18.1 would require
       │  └─ python >=3.10,<3.11.0a0 *, which does not exist (perhaps a missing channel);
       ├─ torchvision 0.18.1 would require
       │  └─ python >=3.12,<3.13.0a0 *, which does not exist (perhaps a missing channel);
       ├─ torchvision 0.18.1 would require
       │  └─ python >=3.8,<3.9.0a0 *, which does not exist (perhaps a missing channel);
       └─ torchvision 0.18.1 would require
          └─ python >=3.9,<3.10.0a0 *, which does not exist (perhaps a missing channel).
critical libmamba Could not solve for environment specs

[qin-yu]

Maybe install PyTorch first:

mamba create -n plant-seg2 pytorch==2.3.1 torchvision==0.18.1 torchaudio==2.3.1 -c pytorch
mamba install -c conda-forge plant-seg

[qin-yu]

Hi @tlnagy I tried to use PyTorch 2.3.1, but there is a new error saying MPS doesn't support yet another operation: 'The operator 'aten::max_pool3d_with_indices' is not currently implemented for the MPS device. and this is not solved by PyTorch yet: pytorch/pytorch#77764. I doubt if it make sense to set a PyTorch or Python version just for Conv3D. But thank you for reporting this!

Full log:

INFO: P [MainThread] 2025-01-29 00:08:06,718 plantseg - Logger configured at initialisation. PlantSeg logger name: plantseg
INFO: P [MainThread] 2025-01-29 00:08:07,202 plantseg.utils - You are using the latest version of PlantSeg: 2.0.0b2.
INFO: N [MainThread] 2025-01-29 00:08:58,897 Run viewer - Plantseg is ready!
2025-01-29 00:08:59.234 python3.12[87312:117850972] +[IMKClient subclass]: chose IMKClient_Modern
2025-01-29 00:08:59.234 python3.12[87312:117850972] +[IMKInputSession subclass]: chose IMKInputSession_Modern
2025-01-29 00:09:03.098 python3.12[87312:117850972] The class 'NSOpenPanel' overrides the method identifier.  This method is implemented by class 'NSWindow'
INFO: N [MainThread] 2025-01-29 00:09:21,341 Task - import_image_task started
INFO: N [MainThread] 2025-01-29 00:09:22,128 Task - import_image_task complete in 0.79s
INFO: N [MainThread] 2025-01-29 00:09:35,929 Task - unet_prediction_task started
INFO: P [Dummy-1] 2025-01-29 00:09:36,862 plantseg.functionals.prediction.prediction - Zoo prediction: Running model from PlantSeg official zoo.
INFO: P [Dummy-1] 2025-01-29 00:09:36,898 plantseg.functionals.prediction.prediction - Computing theoretical minimum halo from model.
WARNING: P [Dummy-1] 2025-01-29 00:09:39,259 plantseg.functionals.prediction.utils.size_finder - Encountered 'The operator 'aten::max_pool3d_with_indices' is not currently implemented for the MPS device. If you want this op to be added in priority during the prototype phase of this feature, please comment on https://github.com/pytorch/pytorch/issues/77764. As a temporary fix, you can set the environment variable `PYTORCH_ENABLE_MPS_FALLBACK=1` to use the CPU as a fallback for this op. WARNING: this will be slower than running natively on MPS.' at patch shape (416, 416, 416), unexpected but continuing.
WARNING: P [Dummy-1] 2025-01-29 00:10:04,776 plantseg.functionals.prediction.utils.size_finder - Encountered 'The operator 'aten::max_pool3d_with_indices' is not currently implemented for the MPS device. If you want this op to be added in priority during the prototype phase of this feature, please comment on https://github.com/pytorch/pytorch/issues/77764. As a temporary fix, you can set the environment variable `PYTORCH_ENABLE_MPS_FALLBACK=1` to use the CPU as a fallback for this op. WARNING: this will be slower than running natively on MPS.' at patch shape (208, 208, 208), unexpected but continuing.
WARNING: P [Dummy-1] 2025-01-29 00:10:05,524 plantseg.functionals.prediction.utils.size_finder - Encountered 'The operator 'aten::max_pool3d_with_indices' is not currently implemented for the MPS device. If you want this op to be added in priority during the prototype phase of this feature, please comment on https://github.com/pytorch/pytorch/issues/77764. As a temporary fix, you can set the environment variable `PYTORCH_ENABLE_MPS_FALLBACK=1` to use the CPU as a fallback for this op. WARNING: this will be slower than running natively on MPS.' at patch shape (112, 112, 112), unexpected but continuing.
WARNING: P [Dummy-1] 2025-01-29 00:10:06,802 plantseg.functionals.prediction.utils.size_finder - Encountered 'The operator 'aten::max_pool3d_with_indices' is not currently implemented for the MPS device. If you want this op to be added in priority during the prototype phase of this feature, please comment on https://github.com/pytorch/pytorch/issues/77764. As a temporary fix, you can set the environment variable `PYTORCH_ENABLE_MPS_FALLBACK=1` to use the CPU as a fallback for this op. WARNING: this will be slower than running natively on MPS.' at patch shape (64, 64, 64), unexpected but continuing.
WARNING: P [Dummy-1] 2025-01-29 00:10:07,113 plantseg.functionals.prediction.utils.size_finder - Encountered 'The operator 'aten::max_pool3d_with_indices' is not currently implemented for the MPS device. If you want this op to be added in priority during the prototype phase of this feature, please comment on https://github.com/pytorch/pytorch/issues/77764. As a temporary fix, you can set the environment variable `PYTORCH_ENABLE_MPS_FALLBACK=1` to use the CPU as a fallback for this op. WARNING: this will be slower than running natively on MPS.' at patch shape (32, 32, 32), unexpected but continuing.
INFO: P [Dummy-1] 2025-01-29 00:10:07,136 plantseg.functionals.prediction.prediction - For raw in shape (263, 1024, 1024): set patch shape (-56, -56, -56), set halo shape (44, 44, 44)
WARNING: P [Dummy-1] 2025-01-29 00:10:07,187 plantseg.functionals.prediction.utils.size_finder - Encountered 'The operator 'aten::max_pool3d_with_indices' is not currently implemented for the MPS device. If you want this op to be added in priority during the prototype phase of this feature, please comment on https://github.com/pytorch/pytorch/issues/77764. As a temporary fix, you can set the environment variable `PYTORCH_ENABLE_MPS_FALLBACK=1` to use the CPU as a fallback for this op. WARNING: this will be slower than running natively on MPS.' at batch size 1, unexpected but continuing with halved batch size.