Skip to content

Latest commit

 

History

History
34 lines (24 loc) · 2.53 KB

TRAIN.md

File metadata and controls

34 lines (24 loc) · 2.53 KB
Raw

Train

Please make sure you follow the instructions in INSTALL.md to build necessary dependencies.

To run the training code, you also need to download the mesh_release.zip in our NBA2K dataset and unzip it on your local machine.

You might need to change the following options in img_to_mesh/src/experiments/mesh/mesh_train.yaml:

  • data_root_dir: Set it to the mesh_release dataset path on your machine.
  • vis_gpu: Set it as your CUDA_VISIBLE_DEVICES.
  • base_log_dir: The root dir of the mesh log folder.
  • log_name: Name of the mesh log folder. You can left it as None and img_to_mesh/src/trainers/train_mesh.py can handle it automatically. You can also manually identify it.
  • custom_postfix: Postfix added to the end of log_name, default is ''. If you set it to a non-empty string, the final name of the mesh log folder should be ${log_name}-${custom_postfix}.

You also need to change the following lines in img_to_mesh/src/experiments/mesh/mesh_run.sh:

cfg_path=experiments/mesh/mesh_train.yaml

After you finish above steps, run the training code:

cd img_to_mesh/src
bash experiments/mesh/mesh_run.sh

Play with hyperparameters

If you use the default setting in img_to_mesh/src/experiments/mesh/mesh_train.yaml and img_to_mesh/src/constants.py. You should get similar results as our pretrained checkpoints. You can also play with the hyperparameters if you want.

You can change the training settings in the training subsection of img_to_mesh/src/experiments/mesh/mesh_train.yaml.

You can change the loss weights in the loss subsection of img_to_mesh/src/experiments/mesh/mesh_train.yaml.

For the Spiral Convolution network architecture, you should change the relevant parts in img_to_mesh/src/constants.py. Please see comments in constants.py for more details. You can also check the original implementation of Spiral Conv.

Note: If you want to change DS_FACTORS in constants.py, you need to make sure you have installed opendr and MPI-MESH following INSTALL.md. You also need to change dataset.meshpackage in mesh_train.yaml to 'mpi-mesh'. For the first run, it will compute some meta data for the Spiral Conv and save it to a local pkl file. For the following runs, it will directly load the saved meta data and you can use 'trimesh' as the dataset.meshpackage. For the default DS_FACTORS, we already provided the pkl files in img_to_mesh/data/mesh/.