tlc.integration.detectron2.umap_reduce_embeddings_hook

Hook to apply UMAP reduction to embeddings generated during a run.

Module Contents

Classes

Class	Description
UMAPReduceEmbeddingsHook	Hook to apply UMAP reduction to embeddings generated during a run.

API

class tlc.integration.detectron2.umap_reduce_embeddings_hook.UMAPReduceEmbeddingsHook(run_url: str, n_components: int = 3, delete_source_tables: bool = True)

Bases: detectron2.engine.hooks.HookBase

Hook to apply UMAP reduction to embeddings generated during a run.

Will use the most recently written metrics table in the run to fit a UMAP model. This model will then be used to apply dimensionality reduction to all other viable metrics tables collected during training. A metrics table is considered viable for reduction if it contains at least one column where the schema defines a value with a number role of NUMBER_ROLE_NN_EMBEDDING.

Note: because of the current design, metrics collection hooks registered with the run will need to ensure that the most recent metrics table written to the run is the one that should be used to fit the UMAP model. This can be done by ensuring the order of hooks is correct and their collection frequencies are set appropriately. This limitation will be made more explicit in the future.

Parameters:

• run_url – The URL of the run.

• n_components – The number of components to reduce the embeddings to.

Create a new UMAP reduction hook.

after_train() → None

Perform UMAP reduction on metrics generated during training.