PointNet AutoEncoder training on ModelNet40

Training script can be found in src.polar.train.train.py with the function main. It can be imported from polar as train_ae.

Minimal example

Typically, a training would be done as follows:

1. Create a file train_ae.py with the following:

   from polar import train_ae
   
   if __name__ == '__main__':
       train_ae()
   

2. Then, run

python train_ae.py --name demo --shuffle --sigma 0.05

Parameters

It accepts the following parameters:

• Base

  • name (Required)
  • log_dir (str, default='logs/ae')
  • batch_size (int, default=64)
  • num_workers (int, default=4)

• Dataset

  • rootdir (str, default='modelnet')
  • classes (str, default=None)
  • exclude_classes (str, default=None)
  • samples_per_class (int, default=None)

• Preprocessing

  • shuffle (bool, default=False)
  • num_points (int, default=1024)
  • max_angle (int, default=180)
  • max_trans (float, default=0.0)

• Augmentations

  • sigma (float, default=0.0)
  • min_scale (float, default=1.0)
  • keep_ratio (float, default=1.0)
  • p (float, default=0.5)

• Autoencoder

  • first_stage_widths (int, default=(64, 64))
  • second_stage_widths (int, default=(64, 128, 1024))
  • decoder_widths (int, default=(1024, 1024))
  • dropout (float, default=0.1)

• Training

  • lr (float, default=0.001)
  • resume_optimizer (bool, default=False)
  • checkpoint (str, default=None)
  • freeze_decoder (bool, default=False)
  • epochs (int, default=150)

• Loss

  • norm (int, default=2)
  • density_weight (float, default=0.0)
  • density_radius (float, default=0.1)