Skip to content

Latest commit

 

History

History
27 lines (18 loc) · 751 Bytes

README.md

File metadata and controls

27 lines (18 loc) · 751 Bytes
Raw

How to run the proof-of-concept experiment

cd experiments/poc/

  1. Generate the proof-of-concept dataset (2D and 3D)

    python generate_poc_data.py

  2. Train the 2D UNet on 2D dataset

    python train_2d.py

  3. Locate where the 2D model checkpoint is saved (in ./tmp/noise0.5/shape/.../model.dat and then copy the path to POCVoxelEnv.shape_checkpoint in poc_config.py

  4. Train the 3D UNet on 3D volumesset, with or without 2D pretraining

    python train_3d.py

The default 3D model is ACSUNet p.. To change ACSConv to Conv3d / Conv2_5d or random initialization, modify POCVoxelConfig.conv and POCVoxelConfig.pretrained in poc_config.py.