Concept Analysis Tooling

The tooling for concept analysis is collected in the module hybrid_learning.concepts. For details have a look at the API Reference.

Analysis Handles

A convenient analysis handle with functions to conduct concept analysis steps and store results is ConceptAnalysis.

Concept and Concept Embedding Modelling

A concept is defined by sample data points and is modeled as an instance of Concept. A concept can be used to train a model for predicting/detecting it properly. Here, concepts are used to train a linear concept model (e.g. an instance of ConceptDetectionModel2D), which predicts a concept from the intermediate output of a DNN. The parameters of a trained concept model describe a linear concept embedding of a concept into a layer of a DNN. Embeddings and operations thereon are modeled in ConceptEmbedding. For translation between concept model and concept embedding use

• to_embedding() for model to embedding, and

• from_embedding() for embedding to model.

The relevant modelling classes are:

Concept	Representation of a concept with data and meta information.
SegmentationConcept2D	Concept with segmentation data.
ConceptEmbedding	Representation of an embedding of a concept within a DNN.
ConceptDetectionModel2D	Pytorch model implementation of a concept embedding for 2D conv layers.
ConceptSegmentationModel2D	A concept model that segments the concept in an image.
ConceptClassificationModel2D	A concept model that classifies whether an image-level concept is recognized in an activation map.

The concept model classes are accompanied by custom handles for training and testing derived from TrainEvalHandle(). The handles are:

ConceptDetection2DTrainTestHandle	Train test handle for concept localization models.
ConceptSegmentation2DTrainTestHandle	Train and test handle for concept segmentation providing loss and metric defaults.
ConceptClassification2DTrainTestHandle	Train and test handle for concept classification providing loss and metric defaults.

Intermediate Output Retrieval and Model Extension

To retrieve the intermediate output of pytorch models, the pytorch hooks mechanism is used. Wrappers for adding/retrieving intermediate output of DNNs are defined in the module model_extension:

ActivationMapGrabber	Wrapper class to obtain intermediate outputs from models.
ExtendedModelStump	Optionally apply a modification to the model stump output in the forward method.
HooksHandle	Wrapper that registers and unregisters hooks from model that save intermediate output.
ModelExtender	This class wraps a given model and extends its output.
ModelStump	Obtain the intermediate output of a sub-module of a complete NN.

Training and Validation

Training and Validation Interface

The following handle classes from hybrid_learning.concepts.train_eval are used to provide a generic training and validation interface for pytorch models:

EarlyStoppingHandle	Handle encapsulating early stopping checks.
ResettableOptimizer	Wrapper around torch optimizers to enable reset and automatic learning rate handling.
TrainEvalHandle	Handle for training and evaluation of pytorch models.

Furthermore, a set of callback handles is pre-defined that can be added to a training or evaluation run:

Callback	A callback base class that eases implementing a custom callback handle.
CsvLoggingCallback	Extract the values stored in matplotlib figures and store them as CSV.
LoggingCallback	Log batch and epoch KPI results.
ProgressBarUpdater	Update the progress bar postfix after each batch and epoch.
TensorboardLogger	Write batch and epoch KPI results to a tensorboard log directory.

Training KPIs

Some loss and metric functions are available for different training setups of the concept models. They all inherit from torch.nn.Module.

aggregating_kpis	Implementation of KPIs that cannot be computed on a batch level.
batch_kpis	Loss and metric functions and classes that can be calculated per batch.

ConceptDetection2DTrainTestHandle	Train test handle for concept localization models.
ConceptSegmentation2DTrainTestHandle	Train and test handle for concept segmentation providing loss and metric defaults.
ConceptClassification2DTrainTestHandle	Train and test handle for concept classification providing loss and metric defaults.