(Training working based off the below command)
Create new virtual env, and install from requirements.txt.
nB: batch dimension
nO: object dimension (always 1, for now. Will squeeze away this dimension often)
num_points: number of points in pointcloud
- Example (replace the last two paths with the paths to your training and test samples folder):
python supervised_training/train_relativerotation.py --randomize_z_canonical __--stats_json="/home/richard/Dropbox (MIT)/bandu_project_data/train/fps_randomizenoiseTrue_numfps10_samples/rr_pn_stats.json" "/home/richard/Dropbox (MIT)/bandu_project_data/train/fps_randomizenoiseTrue_numfps10_samples" "/home/richard/Dropbox (MIT)/bandu_project_data/train/fps_randomizenoiseTrue_numfps10_samples"
- Changing or utilizing the following lines should be necessary for implementing implicit PDF:
In supervised_training/train_relativerotation.py:
- model = DGCNNCls(num_class=4)
- predicted_quaternions = model(batch['rotated_pointcloud'].squeeze(1).permute(0, 2, 1))
The dataset class file is here: supervised_training/dataset.py
You can initialize each dataset by feeding in the path to the samples directory, e.g. "/home/richard/Dropbox (MIT)/bandu_project_data/train/fps_randomizenoiseTrue_numfps10_samples"
Each batch in the dataset is a dictionary with multiple values.
batch['rotated_pointcloud']:
- These are the pointclouds in the starting orientation, on the table.
batch['canonical_pointcloud']:
- These are the pointclouds that are in the canonical/identity/stacked orientation.
The reason why the identity orientation pointclouds are stacked, is because each object's
identity orientation when loading into pyBullet is the stacked orientation.
After dropping the object randomly on the table, we log a new starting orientation and 'rotated_pointcloud' in PyBullet.
batch['rotated_quat']:
- These are the orientations of the above 'rotated_pointcloud' in PyBullet.
Since we want to apply a world-frame rotation to this rotated_pointcloud, such that it becomes
the canonical_pointcloud (and this rotation corresponds to the object rotating from the starting orientation to the stacked orientation),
the regression target for our neural network is the *inverse* of this 'rotated_quat':
-- R.from_quat(batch['rotated_quat']).inv()
The default keyword arguments in supervised_training/dataset.py are the ones used in the paper, so there is no need to change the max_shear etc.
helper_scripts/visualize_sample_pkl.py is a template for reading the .pkl files in the samples folders.