reproducing TRIM from the paper "Transformation Importance with Applications to Cosmology" 🌌 (ICLR Workshop 2020) (original) (raw)

Official code for using / reproducing TRIM from the paper Transformation Importance with Applications to Cosmology (ICLR 2020 Workshop). This code shows examples and provides useful wrappers for calculating importance in a transformed feature space.

This repo is actively maintained. For any questions please file an issue.

examples/documentation

dependencies: depends on the pip-installable acd package
examples: different folders (e.g. ex_cosmology, ex_fake_news, ex_mnist, ex_urban_sound contain examples for using TRIM in different settings)
src: the core code is in the trim folder, containing wrappers and code for different transformations
requirements: tested with python 3.7 and pytorch > 1.0

Attribution to different scales in cosmological images	Fake news attribution to different topics

Attribution to different NMF components in MNIST classification	Attribution to different frequencies in audio classification

sample usage

import torch import torch.nn as nn from trim import TrimModel from functools import partial

setup a trim model

model = nn.Sequential(nn.Linear(10, 10), nn.ReLU(), nn.Linear(10, 1)) # orig model transform = partial(torch.rfft, signal_ndim=1, onesided=False) # fft inv_transform = partial(torch.irfft, signal_ndim=1, onesided=False) # inverse fft model_trim = TrimModel(model=model, inv_transform=inv_transform) # trim model

get a data point

x = torch.randn(1, 10) s = transform(x)

can now use any attribution method on the trim model

get (input_x_gradient) attribution in the fft space

s.requires_grad = True model_trim(s).backward() input_x_gradient = s.grad * s

see notebooks for more detailed usage

ACD (ICLR 2019 pdf, github) - extends CD to CNNs / arbitrary DNNs, and aggregates explanations into a hierarchy
CDEP (ICML 2020 pdf, github) - penalizes CD / ACD scores during training to make models generalize better
DAC (arXiv 2019 pdf, github) - finds disentangled interpretations for random forests
PDR framework (PNAS 2019 pdf) - an overarching framewwork for guiding and framing interpretable machine learning

reference

feel free to use/share this code openly
if you find this code useful for your research, please cite the following:

@article{singh2020transformation, title={Transformation Importance with Applications to Cosmology}, author={Singh, Chandan and Ha, Wooseok and Lanusse, Francois, and Boehm, Vanessa, and Liu, Jia and Yu, Bin}, journal={arXiv preprint arXiv:2003.01926}, year={2020}, url={https://arxiv.org/abs/2003.01926}, }