GPU-aware-converter

System converting pytorch-based model network to minimize gpu.
Currently optimized for changing unet.

run system demo

• requirements
  • make virtual environment as follow (python 3.8, pytorch 1.8.0, cuda 11.1, memcnn)

  conda create -n demo python=3.8 anaconda -y
  conda activate demo
  pip install torch==1.8.0+cu111 torchvision==0.9.0+cu111 torchaudio==0.8.0 -f https://download.pytorch.org/whl/torch_stable.html
  pip install memcnn

• Note that we provide (1) the implementation with only-python (master branch) and (2) the implementation with bind.so to use cudnnConvolutionBackwardFilter (cpp_binding branch). If you want to run another, git checkout master or git checkout cpp_binding.
• see the effectiveness of a stitchable method
  • plain : python demo-stitchable.py --use_stitch 0
    • Upload images with sizes (64,4096,4096) and expect GPU assignment to be fail

    conda activate demo
    python demo-stitchable.py --use_stitch 0 # OOM (not working)

  • stitchable(proposed) : python demo-stitchable.py --use_stitch 1
    • Upload images with sizes (64,4096,4096) and expect run slowly but successfully

    conda activate demo
    python demo-stitchable.py --use_stitch 1 # working

• see the effectiveness of layer convert method
  • run plain model (without our system converter)

    conda activate demo
    python demo-plain.py # require approximately 11GB

  • run converted model with our system converter

    conda activate demo
    python demo-our_system.py # require approximately 4GB

run full-type running

• requirements
  • memcnn : https://memcnn.readthedocs.io/en/latest/installation.html
  • fastmri : https://github.com/facebookresearch/fastMRI
  • See requirements.txt for environments and package version
• download dataset
  • visit https://fastmri.med.nyu.edu/ and submit Name, Email, Istitution
  • download dataset knee + singlecoil + validation
• run "example-fastmri.ipynb"