graphstorm
graphstorm copied to clipboard
awslabs
Enterprise graph machine learning framework for billion-scale graphs for ML scientists and data scientists.
The current GraphStorm `Evaluator`s have misleading naming convention and lack of default metric. The new `Evaluator`s will be task-specific. Sample code ``` evaluator = gs.eval.GSgnnClassificationEvaluator(eval_frequency=100) ``` **Requested changes** 1. Create...
For a classification task, cannot use all metrics at the same time. Examples: Accuracy and auc_roc requires different prediction return values. Therefore they cannot be used in one training/evaluation epoch....
We should implement a proper link prediction inference script. Users can provide a list of nodes and the link prediction inference script should return top K nodes that are likely...
We need to support following user experience: - [ ] #513 - [ ] Resuming a distillation task from a saved checkpoint - [ ] Provide an end2end experience of...
We need a gconstruct config validator in general.
We are now using list[dict] for lots of input type. Actually we can upgrade that to a list[Mapping] if the input should be invariant.
*Issue #, if available:* *Description of changes:* - Implemented hits@k metric for link-prediction - Refactored `GSgnnMrrLPEvaluator` and `GSgnnPerEtypeMrrLPEvaluator` to work with both mrr and hits@k By submitting this pull request,...
We should use a field to store the valid length of a token list instead of using an attention mask.
Currently, gconstruct doesn't enforce the balancing between training/validation/test sets in graph partitioning for node prediction tasks. If training/validation/test nodes are not evenly split across graph partitions, the node split algorithm...