analysis_handle

Description

Handle for conducting a concept embedding analysis.

For details on the workflow of a concept analysis see ConceptAnalysis.analysis(). In short:

Input

All of

• The concept (defined via concept data)

• The main model

• The layers to analyse and compare

Output

All of

• The layer hosting the best embedding,

• The best embedding,

• The quality metric values for the best embedding

Classes

ConceptAnalysis	Handle for conducting a concept embedding analysis.
EmbeddingReduction	Aggregator callables to get the mean from a list of embeddings.

Functions

data_for_concept_model([concept_model, ...])	Data handles with activation maps for and ground truth from concept.
default_act_cache_root(wrapped_data, ...)	Determine the default activation cache root as <cache_root>/<img_base>_<mainmodelclassname>-<layerid>.
default_cache_roots(wrapped_data[, ...])	Given a dataset to be wrapped and a concept model, provide standard cache roots paths.
default_mask_cache_root(wrapped_data[, ...])	Determine the default activation data mask cache root.
model_hash(model[, truncate])	Return a hex md5-hash of the main model topology for comparison purposes.