LRP Toolbox 1.3.1

This release disentangles the formula types 11 and 99 of the Caffe LRP implementation.
Formula 11 now implements the vanilla backprop gradient and only formula 99 implements Sensitivity Analysis, i.e., the pixel-wise euclidean norm of the vanilla gradient. The singularity image def-file has been updated slightly.

LRP Toolbox 1.3.0

New in 1.3.0:
Standalone Python implementation:

update to python 3
updated treatment of softmax and target class
lrp_aware option for efficient calculation of multiple backward passes (at the cost of a more expensive forward pass)
custom colormaps in render.py
gpu support when cupy is installed. this is an optional feature. without the cupy package, the python code will execute using the cpu/numpy.

Caffe implementation

updated the installation config
new recommended formula types 100, 102, 104
support for Guided Backprop via formula type 166
new python wrapper to use lrp in pycaffe
pycaffe demo file
bugfixes
singularity image definition for building a hassle-free OS-agnostic command line executable

This will with high probability be the final release for the LRP Toolbox.

Release Candidate for the LRP Toolbox 1.3.0rc2

Now with optional GPU support for python, using cupy!
Just install cupy and you are ready to use your GPU.
Do not install cupy and keep using numpy on your CPU.

Release Candidate for the LRP Toolbox 1.3.0

Working on Code is done, functionality complete.
Documentation still needs to be updated for version 1.3.0

New in this version:

For python:

• update to python 3
• update treatment of softmax and target class
• lrp_aware option for efficient calculation of multiple backward passes
• custom colormaps (used in publications) available in render.py

For caffe:

• new formula types 100, 102, 104
• new python wrapper to use lrp in pycaffe
• convenience functions for the python wrapper + demo file (lrp_python_demo.py)
• updates for the installation config
• bugfixes

Version 1.2.0

Version 1.2.0 is done. Pretty much the same as 1.2.0-rc1, but better!

New in this version:
For python and matlab:

• Conv-Layer support and sum- and maxpooling have been implemented
• rudimentary network training support for all existing layers
• pre-setting lrp decomposition parameters per layer implemented
• w² and flat weight decompositions implemented for all layers (where applicable)
• demo code demonstrating LRP on LeNet-5, showing off the new features
• a slight change in the plain text model description. Backwards-compatibility for older toolbox versions is ensured.

Caffe:

• heatmapping applications producing minimal outputs have been implemented. Instead of writing out the network input and render heatmaps as images, only plain text relevance maps and the top-10 network predictions are produced per sample.
• w² and flat weight decomposition methods implemented, including a layer selection index whilch allows for a treatment with either w² or flat decomposition for all layers blow (inclusive) this layer index, and either eps- or alpha-decomposition for all upper layers. This can be used to control the resolution and semantics of the explanatory heatmaps. (See http://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=7532763)

Release Candidate for the LRP Toolbox 1.2.0

Code work is done and functionality is complete.
Future changes, up to version 1.2.0 will be concerned with documentation.

New in this version:
For python and matlab:

• Conv-Layer support and sum- and maxpooling have been implemented
• rudimentary network training support for all existing layers
• pre-setting lrp decomposition parameters per layer implemented
• w² and flat weight decompositions implemented for all layers (where applicable)
• demo code demonstrating LRP on LeNet-5, showing off the new features
• a slight change in the plain text model description. Backwards-compatibility for older toolbox versions is ensured.

Caffe:

• heatmapping applications producing minimal outputs have been implemented. Instead of writing out the network input and render heatmaps as images, only plain text relevance maps and the top-10 network predictions are produced per sample.
• w² and flat weight decomposition methods implemented, including a layer selection index whilch allows for a treatment with either w² or flat decomposition for all layers blow (inclusive) this layer index, and either eps- or alpha-decomposition for all upper layers. This can be used to control the resolution and semantics of the explanatory heatmaps. (See http://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=7532763)

Upgrading the Caffe-Addon

New in this release:
Zeiler's Deconvolution method for Caffe
Simonyan's Sensitivity Analysis method for Caffe

Functionality as been added for both sequential heatmapping and batch-processing code.
The Batch-processing lrp binaries compiled from caffe-master-lrp/demonstrator/compute_heatmaps_parallel.cpp via the build.sh now actually work, without manually adapting batch size in the code to the defined value in the deploy.prototxt of the model.
Output images are now png instead of jpg to avoid compression artefacts.

No changes to the standalone matlab/python code. This will be a separate update.
The Caffe code this is based on is still old though. Updating this will also be a future update.