Skip to content
/ IBA Public
forked from BioroboticsLab/IBA

Information Bottlenecks for Attribution

License

Notifications You must be signed in to change notification settings

hiworldwzj/IBA

 
 

Repository files navigation

IBA: Informational Bottlenecks for Attribution

[Paper Arxiv] | [Paper Code] | [Reviews] | [API Documentation] | [Examples] | [Installation]

Build Status Documentation Status

Example GIF
Iterations of the Per-Sample Bottleneck

This repository contains an easy-to-use implementation for the IBA attribution method. Our methods minimizes the amount of transmitted information while retaining a high classifier score for the explained class. In our paper, we run this optimization per single sample (Per-Sample Bottleneck) and trained a neural network to predict the relevant areas (Readout Bottleneck). See our paper for a in-depth description: "Restricting the Flow: Information Bottlenecks for Attribution".

Generally, we advise using the Per-Sample Bottleneck over the Readout Bottleneck. We saw it to perform better and is more flexible as it only requires to estimate the mean and variance of the feature map. The Readout Bottleneck has the advantage of producing attribution maps with a single forward pass once trained.

For the code to reproduce our paper, see IBA-paper-code.

This library provides a TensorFlow v1 and a PyTorch implementation.

PyTorch

Examplary usage for the Per-Sample Bottleneck:

from IBA.pytorch import IBA, tensor_to_np_img, get_imagenet_folder, imagenet_transform
from IBA.utils import plot_saliency_map, to_unit_interval, load_monkeys

from torch.utils.data import DataLoader
from torchvision.models import vgg16
import torch

# imagenet_dir = /path/to/imagenet/validation

# Load model
dev = 'cuda:0' if  torch.cuda.is_available() else 'cpu'
model = vgg16(pretrained=True)
model.to(dev)

# Add a Per-Sample Bottleneck at layer conv4_1
iba = IBA(model.features[17])

# Estimate the mean and variance of the feature map at this layer.
val_set = get_imagenet_folder(imagenet_dir)
val_loader = DataLoader(val_set, batch_size=64, shuffle=True, num_workers=4)
iba.estimate(model, val_loader, n_samples=5000, progbar=True)

# Load Image
monkeys, target = load_monkeys(pil=True)
monkeys_transform = imagenet_transform()(monkeys)

# Closure that returns the loss for one batch
model_loss_closure = lambda x: -torch.log_softmax(model(x), dim=1)[:, target].mean()

# Explain class target for the given image
saliency_map = iba.analyze(monkeys_transform.unsqueeze(0).to(dev), model_loss_closure, beta=10)

# display result
model_loss_closure = lambda x: -torch.log_softmax(model(x), 1)[:, target].mean()
heatmap = iba.analyze(monkeys_transform[None].to(dev), model_loss_closure )
plot_saliency_map(heatmap, tensor_to_np_img(monkeys_transform))

We provide a notebook with the Per-Sample Bottleneck and the Readout Bottleneck.

Tensorflow

from IBA.tensorflow_v1 import IBACopyInnvestigate, model_wo_softmax, get_imagenet_generator
from IBA.utils import load_monkeys, plot_saliency_map
from keras.applications.vgg16 import VGG16, preprocess_input

# imagenet_dir = /path/to/imagenet/validation

# load model & remove the final softmax layer
model_softmax = VGG16(weights='imagenet')
model = model_wo_softmax(model_softmax)

# after layer block4_conv1 the bottleneck will be added
feat_layer = model.get_layer(name='block4_conv1')

# add the bottleneck by coping the model
iba = IBACopyInnvestigate(
    model,
    neuron_selection_mode='index',
    feature_name=feat_layer.output.name,
)

# estimate feature mean and std
val_gen = get_imagenet_generator(imagenet_dir)
iba.fit_generator(val_gen, steps_per_epoch=50)

# load image
monkeys, target = load_monkeys()
monkeys_scaled =  preprocess_input(monkeys)

# get the saliency map and plot
saliency_map = iba.analyze(monkeys_scaled[None], neuron_selection=target)
plot_saliency_map(saliency_map, img=monkeys)

Table: Overview over the different tensorflow classes. (Task) type of task (i.e. regression, classification, unsupervised). (Layer) requires you to add a layer to the explained model. (Copy) copies the tensorflow graph.

Class Task Layer Copy Remarks
IBALayer Any Recommended
IBACopy Any Very flexible
IBACopy Classification Nice API for classification

Documentation

[PyTorch API] | [TensorFlow API]

The API documentation is hosted here.

Table: Examplary jupyter notebooks

Notebook Description
pytorch_IBA_per_sample.ipynb Per-Sample Bottleneck
pytorch_IBA_train_readout.ipynb Train a Readout Bottleneck
tensorflow_IBALayer_cifar.ipynb Train a CIFAR model containing an IBALayer
tensorflow_IBACopy_imagenet.ipynb Explains a ImageNet model
tensorflow_IBACopyInnvestigate_imagenet.ipynb innvestigate api wrapper

Installation

You can install it directly from git:

$ pip install git+https://github.com/BioroboticsLab/IBA

To install the dependencies for torch, tensorflow, tensorflow-gpu or developement dev, use the following syntax:

$ pip install git+https://github.com/BioroboticsLab/IBA[torch, dev]

For development, you can also clone the repository locally and then install in development mode:

$ git clone https://github.com/BioroboticsLab/IBA
$ cd per-sample-bottlneck
$ pip install -e .

Table: Supported versions

Package From To
TensorFlow 1.12.0 1.15.0
PyTorch 1.1.0 1.4.0

Reference

If you use this software for a scientific publication, please cite our paper:

@inproceedings{
Schulz2020Restricting,
title={Restricting the Flow: Information Bottlenecks for Attribution},
author={Karl Schulz and Leon Sixt and Federico Tombari and Tim Landgraf},
booktitle={International Conference on Learning Representations},
year={2020},
url={https://openreview.net/forum?id=S1xWh1rYwB}
}

About

Information Bottlenecks for Attribution

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 59.0%
  • Jupyter Notebook 40.2%
  • Shell 0.8%