Matthew Muckley

So my install process is as follows in a few bash commands: ```bash conda create -n memory_test_py39 python=3.9 conda activate memory_test_py39 conda install anaconda conda install pytorch torchvision torchaudio cudatoolkit=11.3...

@soumickmj @834799106 I can now reproduce this with the minimal install environment. Reproduction environment here: https://gist.github.com/mmuckley/838289a388bc65a7adb23d67908635c9.

I do think it is related to `SliceDataset` itself as I see similar characteristics with `VarNetDataTransform`.

Actually I have to take that back. With the minimal environment and no transforms, I see no issue.

Okay I think I found the issue: it is due to the `h5py` from `pip`. It's a relatively recent issue that has been documented here: https://forum.hdfgroup.org/t/h5py-memory-leak-in-combination-with-pytorch-dataset-with-multiple-workers/9114 If you install the...

With the `h5py` from `conda` if I `print(h5py.version.info)` then I get the following: ```bash Summary of the h5py configuration --------------------------------- h5py 3.6.0 HDF5 1.10.6 Python 3.9.7 (default, Sep 16 2021,...

Okay I opened #217 to do this. Feel free to propose any changes. For what it's worth, I did some small tests on adding extra copy commands to `SliceDataset` to...

Hello @soumickmj I do not observe HDF5 being updated on `conda`, at least for Python version 3.8 or 3.9.

Hello @abishpius, I think Google Colab would be great. The only restriction I can think of is that we're not allowed to redistribute the fastMRI data, and the tutorial notebook...

Hello @Xieyyyy, did you install ignoring dependencies? I saw this issue with PyTorch 1.11. https://github.com/mmuckley/torchkbnufft/blob/4e5acbf3d946d21a60a8c516f123762cf7a625c0/setup.py#L22