📜 Changelog

Version 3.0.0

Added in version 3.0.0:

• Support for visualizing voxels.

• Support in utils for generating energy images.

• transforms.voxels’s, providing support for transforming voxels.

• modules.voxels’s, providing support for deep learning on voxels.

Changed in version 3.0.0:

• Split transforms into transforms.points and transforms.voxels.

• Split modules into modules.points and modules.voxels.

• Renamed PCDDataset to Dataset.

• Renamed Collator to PCDCollator.

• Renamed PCDDataModule to DataModule.

• Renamed PCDLit to LitModule.

Version 2.0.0

Notable changes

• torch is now used instead of numpy for controlling the randomness in transforms.

• New LighthningModule PCDLit, which supports customization for optimizer and scheduler.

• New storing scheme for point clouds, stored as plain .npy files instead of a single .npz file.

Added in version 2.0.0:

• Support for visualizing a generic point cloud (#69).

• Commands aidsorb create and aidsorb prepare now support configuration files.

• Option for ignoring critical indices in PointNetBackbone.forward().

• Support for visualizing .npy file with draw_pcd_from_file() and aidsorb visualize.

• RandomSample and RandomFlip transformations.

• Support for erasing a fraction of points and local patches in RandomErase.

• Support for configuring activation function in conv1d_block() and dense_block().

• Support to .load_from_checkpoint without arguments for PCDLit and PCDDataModule.

• Support for unlabeled data in PCDDataset, PCDDataModule and Collator.

• Option drop_last for PCDDataModule.

• PCDLit which supports customization for optimizer and scheduler (#25).

• center_pcd() as a functional interface of Center.

Changed in version 2.0.0:

• Renamed get_elements() to get_atom_names().

• Columns in PCDDataset.Y now follow the order specified by the user (#67).

• Renamed litmodels to litmodules (#63).

• Collator and pad_pcds() now accept keyword-only arguments.

• PointNet moved from models to modules.

• get_names() now returns tuple instead of list.

• Bumped Lightning version to >=2.5.0 (#29).

• upsample_pcd() moved from data to transforms and now raises error if target size is not greater than the original.

• split_pcd() moved from utils to transforms and no longer copies data.

• transforms now use torch instead of numpy (#32).

• Removed defaults for transforms since there is no consensus on “good” defaults.

• Point clouds are now stored as plain .npy files under a directory files instead of a single .npz (#3).

Removed in version 2.0.0:

• Jitter, use RandomJitter instead.

• models to simplify codebase and improve project structure.

• PointLit, use PCDLit instead.

• Identity since it is equivalent to torch.nn.Identity (and thus redundant).

Version 1.0.0

🎂 First release for public usage.