Explain Scruples

https://github.com/allenai/scruples

Features

• explain anecdotes with LIME
• report generation with statistics on explanation features

What is where?

File	features
anecdotes.py	generates LIME explanations for a random anecdotes subset and creates pickled explanations as well as html report + plots
anecdotes_utils.py	interface functions that call the scruples REST API, reads anecdotes data from /data/ directory
anecdotes_shap.py	explanations with SHAPley values, not functional currently
anecdotes_LIME.py	explanations with LIME, functional
anecdotes_anchors.py	explanations with anchors, functional but slow

Findings

• model considers connotated verbs and personal pronouns

Run scruples

• clone XAI_scruples

• create venv from requirements.txt

• clone and install scruples in venv from requirements.txt: https://github.com/allenai/scruples#setup

• download scruples model & config: https://github.com/allenai/scruples/blob/master/docs/demos.md#norms

• move model (.json & .bin) to ./model/anecdotes or ./models/dilemmas

• update path and config: start_server.sh (for GPU; CPU: start_server_CPU.sh)

    ./start_server.sh
  

Run scruples in Docker

• clone XAI_scruples

• create venv from requirements.txt

• clone scruples: https://github.com/allenai/scruples

• build and run docker container from Dockerfile

    docker build --tag scruples .
    docker run --name scruples_instance -publish 5050:8000 scruples
  

Run

• download scruples data: https://github.com/allenai/scruples#data
• move data to XAI_scruples/data/anecdotes and/or XAI_scruples/data/dilemmas
• configure params in ancedotes.py
• run ancedotes.py
• html and pngs will be created in root dir

Runtime LIME on sample anecdote

• single explanation with 10 pertubations: 14s CPU
• single explanation with 5000 pertubations: 6532s CPU, 348s GPU
• CPU i7, GPU GTX1080

Other explanation techniques

• Anchors: runtime extremly high, ~3hours per anecdote on GPU

• SHAP not compatible with embedding layers: shap/shap#595