fhvilshoj / lrp

Implementation of the LRP rules.

LRP Framework

This project aims to implement the LRP rules for any tensorflow:1.4 graph consisting of simple components such as linear layers, convolutions, LSTMs and simple pooling layers. See Status to see how far we have come.

Status

(alpha, beta)-rule, epsilon-rule, W^2-rule and flat-rule

• Linear layers
• Convolutions
• Max pooling
• Nonlinearities
• LSTM
• Concatenates
• Splits
• Tile
• Reshaping (Reshape, Expand_dims, Squeeze)
• Sparse matrix multiplication
• Sparse reshape

Usage

A simple usage of the framework:

from lrp import lrp

with g.as_default():
    inp = ...
    pred = ...
    config = LRPConfiguration()
    # Set propagation rule for, e.g., linear layers
    config.set(LAYER.LINEAR, AlphaBetaConfiguration(alpha=2, beta=-1))

    # Calculate the relevance scores using lrp
    expl = lrp.lrp(inp, pred, config)

    with tf.Session() as sess:
        # Compute prediction and explanation
        prediction, explanation = sess.run([pred, expl], feed_dict={inp: ...})