Pydgn

Changed

- Refactored the toml file of the library to remove legacy files `setup.py` and `setup.cfg`

Added

- Utility functions in `pydgn.evaluation.util` to retrieve configuration files in a model selection folder and filter them for post-hoc analyses.
- Updated tutorial in documentation accordingly

Fixed

- Solved some deprecation warnings when using `torch.tensor` to clone an existing tensor in DataProvider

[1.4.1] Support for Python 3.8-3.11

Changed

- Commented the single temporal graph learning example to remove the dependency from torch-geometric package. The code to support temporal learning stays.
- Removed the configuration sample for the temporal setting
- PyG requirements is <= 2.3.0, so we can remove all the other dependencies
- Modified the `setup/create_environment.sh` file to use Python 3.9 and PyG 2.3.0

Added

- You can specify weights for loss in `AdditiveLoss` by passing a dictionary of (loss name, loss weight) entries as an argument.
See the documentation of `AdditiveLoss` for more info or the example in `examples/MODEL_CONFIGS/config_SupToyDGN.yml`.

Fixed

- Better handling of `len()` in `TUDatasetInterface`
- Fixed minor bug when shuffle was set to false that triggered an assertion in training engine

Changed

- Package requirements now specify an upper bound on some packages, including Pytorch and PyG to ensure compatibility. Better be safe than sorry :)

Changed

- Updates to tests to make the fake datasets compatible with PyG 2.1.0

Fixed

- IDLE ray workers not deallocating GPUs

- Now we sort the data list returned by training engine as if samples were not shuffled.

Meaning the returned data list is consistent with the original ordering of the dataset.

Comments

We tried to provide support for creating an environment with PyG 2.2.0, but importing the library seems to cause
`segmentation fault` in certain cases. Therefore, we will wait until the issue is fixed and then update the script.

Added

- You can now specify a specific subset of gpus to use in the configuration file.

Just add the optional field `gpus_subset: 1,2,3` if you want to only use GPUs with index 1,2 and 3.

Changed

- Ray 2.0.0 seems to have a problem with killing `IDLE` processes and releasing their resources, i.e. OOM on GPU.
We are reverting to a version that we were using before and did not have this problem.

Fixed

- Minor check in splitter
- Minor fix in link prediction splitter, one evaluation link was being left out
- Minor fix in early stopper, `epoch_results` dict was overwritten after applying early stopping. Does not affect performances since the field is re-initialized the subsequent epoch by the training engine.
- Removed setting random seed for map-style dataset. It was not useful (see Torch doc on reproducibility) and could cause transforms based on random sampling (e.g. negative sampling) to behave always in the same way

Changed

- Changed semantics of gradient clipper, as there are not many alternatives out there