Remove the need for to_homogeneous(dataset.train_node_ids) for Labeled homogeneous inputs
#115
Conversation
|
/unit_test |
|
/e2e_test |
|
/integration_test |
GiGL Automation@ 18:32:23UTC : 🔄 |
GiGL Automation@ 18:32:27UTC : 🔄 @ 20:00:36UTC : ✅ Workflow completed successfully. |
GiGL Automation@ 18:32:32UTC : 🔄 @ 19:20:47UTC : ✅ Workflow completed successfully. |
|
/unit_test |
GiGL Automation@ 20:14:09UTC : 🔄 |
kmontemayor2-sc
changed the title
to_homogeneous(dataset.train_node_ids) for Labeled homogeneous inputs
| elif isinstance(input_nodes, abc.Mapping): | ||
| if len(input_nodes) != 1: | ||
| raise ValueError( | ||
| f"If input_nodes is provided as a mapping, it must contain exactly one key/value pair. Received: {input_nodes}. This may happen if you call Loader(node_ids=dataset.node_ids) with a heterogeneous dataset." | ||
| ) | ||
| node_type, node_ids = next(iter(input_nodes.items())) | ||
| is_labeled_homoogeneous = node_type == DEFAULT_HOMOGENEOUS_NODE_TYPE |
There was a problem hiding this comment.
FYI I'm ok with not adding this change - we can just add the util if there's pushback here.
|
|
||
|
|
||
| @dataclass(frozen=True) | ||
| class _ResolvedNodeSamplerInput: |
There was a problem hiding this comment.
I'd prefer not to add anotherNodeSamplerInput derivative if we can avoid it for the sake of reducing the complexity of our codebase -- is the value here primarily the is_labeled_homogeneous field? Is there any way we can just have resolve_node_sampler_input_from_user_input return a Tuple[NodeSamplerInput, bool] or even a Tuple[NodeType, torch.Tensor, bool], since it seems like the only place this is used in in the __init__ of the ABLPLoader?
There was a problem hiding this comment.
Yeah I wanted to do this to avoid the three sized tuple.
I understand the apprehension here about the new dataclass though - do you think renaming it could help? Something like _ParsedInputs or equivalent?
|
/unit_test |
GiGL Automation@ 20:15:15UTC : 🔄 @ 20:52:38UTC : ✅ Workflow completed successfully. |
Headline change here is to remove the need to do:
As demonstrated in
_run_cora_supervised1 and_run_distributed_neighbor_loader_labeled_homogeneous2.Also:
neighborloader.resolve_node_sampler_input_from_user_inputNodeSamplerInputtype alias to be shared between DistABLPLoader and DistNeighborLoader.Added unit tests for new util.