type object 'MinkowskiEngineBackend._C.RegionType' has no attribute 'HYBRID' in v0.5

[RyanRuihanG]

Describe the bug Hello! ME.RegionType.HYBRID, which existed in v0.4.3, disappears in v0.5. I wonder about the changed name of this attribute. (The members of RegionType in v0.5 only contain HYPER_CUBE, HYPER_CROSS and CUSTOM.)

Thank you!

[chenxl124578]

type object 'MinkowskiEngineBackend._C.RegionType' has no attribute 'HYPERCUBE'. I also get the bug. how can i deal with? Thanks!

[LinXiLuo]

Get the same bug

[mbanani]

@chenxl124578 Some of the layers include a simple name change; eg, HYPERCUBE -> HYPER_CUBE, while others are deprecated.

[sandeepnmenon]

Hi @mbanani
I am trying to run the code in presumably written with the 0.4.3 https://github.com/chrischoy/SpatioTemporalSegmentation/blob/4afee296ebe387d9a06fc1b168c4af212a2b4804/models/modules/common.py#L58

Which uses ME.RegionType.HYBRID. How to use the ME.RegionType.CUSTOM type in replacement to this? I am getting the following error

    kernel_generator = ME.KernelGenerator(
  File "/home/anaconda3/envs/py3-mink/lib/python3.8/site-packages/MinkowskiEngine/MinkowskiKernelGenerator.py", line 303, in __init__
    self.kernel_volume = get_kernel_volume(
  File "/home/anaconda3/envs/py3-mink/lib/python3.8/site-packages/MinkowskiEngine/MinkowskiKernelGenerator.py", line 91, in get_kernel_volume
    region_offset.numel() > 0
AttributeError: 'NoneType' object has no attribute 'numel'

While trying to create an ME.KernelGenerator using

ME.KernelGenerator(
      3, 1, 1, region_type=ME.RegionType.CUSTOM, axis_types=[<RegionType.HYPER_CUBE: 0>, <RegionType.HYPER_CUBE: 0>, <RegionType.HYPER_CUBE: 0>, <RegionType.HYPER_CROSS: 1>], dimension=4)

It is because I am not providing region_offsets. Can you help me understand offsets to give in this case?

Else, how to install MinkowskiEngine 0.4.3. I dont see an archive.

Thank you

[mbanani]

@sandeepnmenon, I am not sure. I faced a similar issue a while ago, and this issue wasn't really discussed in the documentation I saw or the other issues. I wasn't relying on pre-trained weights for that part of the project, so I simply used a different region type.

As for installation, you can probably clone the repo with the tag 4.3 (https://github.com/NVIDIA/MinkowskiEngine/tree/v0.4.3) and build that one.

If you're simply looking for the Minkowski ResUNet architecture, check out FCGF (https://github.com/chrischoy/FCGF) from Chris Choy. It has another model that you can use, and it's been updated with the 0.4 -> 0.5 changes.

[sandeepnmenon]

@mbanani Thank you for the info.

I am not trying to use pre-trained weights, I am looking to use the SpatioTemporal ResNetUnet Architecture from https://github.com/chrischoy/SpatioTemporalSegmentation. I will try with the older tagged version.

Thank you

[suyunzzz]

anyone has solved this problem?

[emilyemliyM]

@mbanani谢谢你的信息。

我没有尝试使用预训练的权重，我希望使用来自https://github.com/chrischoy/SpatioTemporalSegmentation的 SpatioTemporal ResNetUnet 架构。 我将尝试使用较旧的标记版本。

谢谢

Could I get a link about the "https://github.com/chrischoy/SpatioTemporalSegmentation的 SpatioTemporal ResNetUnet ?"?

Really thanks!