v0.2.6

• Fix bug where an asterisk path value to net_names didn't work
• Update test_asterisk_in_net_names_fetches_filenames to test for empty list
• Update documentation to make it clearer that a string can be passed to net_name if it has an asterisk on the end
• Closes #52

v0.2.5

• Improves preprocessing execution times (closes #42)
  • For a single network of 17804 nodes and 308081 edges, _create_pyg_graphs execution time was reduced from 2614 seconds to 0.75 seconds - an almost 3500x speedup.
• Adds better support for loading pretrained models
• Exposes neighborhood sample size to the config as neighbor_sample_size key
• Enables saving of training loss data to a .tsv file

v0.2.4

• Adds model parallelism to BIONIC, allowing use of multiple GPUs for training and inference
• Removed the need for initial node feature SVD approximation for large networks - initial one-hot node features for networks of any size can now be efficiently represented

v0.2.3

• Fixes neighborhood sampling size
• Improves test robustness
• Removes unused files
• Corrects semi-supervised config lambda value

v0.2.2

• Adds TensorBoard training loss tracking and embedding projector support

v0.2.1

• Fixes a bug where sampler edge weights were not properly aligned to edge_index (see #16).
• Fixes a bug in the classification output formatting (see #17).
• Adds an additional linear layer to semi-supervised classifier to improve optimization.

v0.2.0

• If gene/protein labels are available, they can be used to train the BIONIC features through a semi-supervised classification objective
• A "shared encoder" feature is now available, allowing all networks to be encoded using the same GAT
• Updates BIONIC to use the newest PyTorch and PyTorch Geometric distributions

v0.1.0

The first release of BIONIC.