TruLens Explain_Improvements

Jump to bottom

Piotr Mardziel edited this page May 30, 2024 · 1 revision

API Design

Should InputInfluence (emphasis on "input") handle embeddings as if they are "inputs"? Alternatively can create new class for such the situation or just suggest to use InternalInfluence.

Visualizations

Add option for different output formats: html, latex, ...
Add some configuration of output formatting within a format: colors, etc.

Backends Parity

Implement all common options for all backends; most relevant method is fprop.
- keyword argument handling (see pytorch fprop)
- rnn time index handling (see pytorch fprop)
- numpy return (see tf2 fprop)

Features

Interventions on multiple layers.

Code Quality

Factor out redundancies especially between the backends.

Documentation/Examples

Example of NLP model from tfhub.
Examples of NLP model internal attributions (considering "embeddings" to be input here).

Fix Known bugs

(Issue) Non-deterministic outputs even when using CPU only.