Pydgn

Added

- A documentation (it was about time!!!)
- Possibility of specifying inner and outer validation ratio
- We can now use a specific data loader and specify its arguments in the configuration file
- We can now force a metric to compute node-based or graph-based metrics, rather than looking at the ground truth's shape.
- Possibility of evaluating on validation (and test) every `n` epochs
- Use entrypoints to simplify usage of the library
- All arguments must be now specified in the config file. There is a template one can use in the doc.

Changed (IMPORTANT!)

- Removed any backward compatibilities with very old versions (<=0.6.2)
- Substituted Loss and Score classes with a single Metric class, to avoid code redundancy
- Pre-computed random outer validation splits (extracted from 10% of outer training set) for data splits from "A Fair Comparison on Graph Neural Networks for Graph Classification". Note that this does not impact the test splits.

Changed

- The setup installation files now work with Pytorch 1.10 and Pytorch Geometric 2.0.3, and the library assumes Python >= 3.8

Fixed
- Fixed minor bug in experiment. The function create_unsupervised_model looked for supervised_config, rather than unsupervised_config, when looking for the readout
- Feature request: loss, score, and additiveloss now take a parameter `use_nodes_batch_size` to force computation w.r.t. input nodes rather than target dimension (the default)

Added

- Minor refactoring of the engines to avoid redundant flow of information

Fixed
- Fixed a bug in EventHandler. If one extends EventHandler with new events, which are triggered by a training engine, make sure that callbacks that implement the EventHandler interface do not break when the new events are triggered.
- Refactored Profiler to abstract from the EventHandler. This created problems when a callback implmenented an interface that extends EventHandler. If the callback does not implement a particular method, nothing happens and the dispatcher moves on.

Added

- PyDGN Temporal: Support for `single graph sequence` tasks, the most common use case at the moment
(tested for supervised experiments only)

Fixed

- Minor in `cgmm_incremental` experiment
- loss/score now considers the case of reduction=mean (default) and sum when computing the epoch's loss/score

Fixed

- When using checkpoints, we can now switch devices without getting a deserialization error