Skip to content

ardizzone/ibinn_imagenet

 
 

Repository files navigation

ibinn_imagenet

This repository contains the code for:

"Generative Classifiers as a Basis for Trustworthy Computer Vision"
Radek Mackowiak, Lynton Ardizzone, Ullrich Köthe, Carsten Rother
https://arxiv.org/abs/2007.15036

Requirements

  • Python 3.x
  • CUDA
  • cuDNN

Installation

For the installation we recommend a python venv.

git clone [email protected]:RayDeeA/ibinn_imagenet.git
cd ibinn_imagenet/
python3 -m venv .
source bin/activate
pip install -r requirements.txt

Clone and install FrEIA 0.2 (Note: the newest version is currently incompatible):

git clone https://github.com/VLL-HD/FrEIA.git
cd FrEIA
git checkout v0.2
python setup.py develop

In this way, any changes made to FrEIA are reflected immediately without reinstallation of the package.

Usage

Run a training by executing:

python -m ibinn_imagenet.train.ibinn_imagenet_classifier [OPTIONS]

All options are visible and explained when executing the ibinn_imagenet_classifier with the --help parameter:

python -m ibinn_imagenet.train.ibinn_imagenet_classifier --help

If you aim to reproduce our results consider installing xptl (https://github.com/titus-leistner/xptl) from Titus Leistner and use the *.ini-files provided under ibinn_imagenet/config

With xptl installed start a training by executing:

xptl-schedule ibinn_imagenet/config/beta_[VALUE].ini [QUEUEING_COMMAND] ./SLURM_train_ibinn_imagenet_classifier.sh [OUTPUT_FOLDER]

e.g.:

xptl-schedule ibinn_imagenet/config/beta_0_0.ini "" ./SLURM_train_ibinn_imagenet_classifier.sh /mnt/data/output

to run the training on a local machine, or:

xptl-schedule ibinn_imagenet/config/beta_0_0.ini sbatch ./SLURM_train_ibinn_imagenet_classifier.sh /mnt/data/output

in case you want to run the trainings on a cluster system utilizing a scheduler (e.g. SLURM, LSF)

Evaluation for a learned model can be performed by executing the following command. Note however that currently only "accuracy" and "feed_forward_accuracy" are supported values for the field EVALUATION_METHOD.

xptl-schedule ibinn_imagenet/config/beta_[VALUE].ini [QUEUEING_COMMAND] "./SLURM_eval_ibinn_imagenet_classifier.sh [PATH_TO_MODEL_FILE] [EVALUATION_METHOD]"

e.g.:

xptl-schedule ibinn_imagenet/config/beta_inf.ini "" "./SLURM_eval_ibinn_imagenet_classifier.sh /path/to/models/directory/beta_inf.avg accuracy"

The other evaluation methods described in the paper, will be made available soon.

Trained models can be downloaded here:
https://heibox.uni-heidelberg.de/d/e7b5ba0d30f24cdca416/

These models can be easily loaded, e.g.:

from ibinn_imagenet.model.classifiers.invertible_imagenet_classifier import trustworthy_gc_beta_0, trustworthy_gc_beta_32
beta_0_model = trustworthy_gc_beta_0(pretrained=True)
beta_32_model = trustworthy_gc_beta_32(pretrained=True)

or:

from ibinn_imagenet.model.classifiers.invertible_imagenet_classifier import trustworthy_gc_beta_0, trustworthy_gc_beta_32
beta_0_model = trustworthy_gc_beta_0(pretrained=True, pretrained_model_path="/path/to/beta_0.avg.pt")
beta_32_model = trustworthy_gc_beta_32(pretrained=True, pretrained_model_path="/path/to/beta_32.avg.pt")

in case you want to load a model from your hard drive.

Project Organization

├── LICENSE                                 <- The License
│
├── README.md                               <- The top-level README for developers using this project.
│
├── requirements.txt                        <- The requirements file for reproducing the analysis environment
│
├── ibinn_imagenet                          <- Source code
│   │
│   ├── config                              <- Configuration files
│   │
│   ├── data                                <- Code defining the data
│   │
│   ├── model                               <- Code defining the models
│   │
│   ├── train                               <- Code defining the training logic
│   │
│   ├── utils                               <- Utilities

About

No description, website, or topics provided.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 96.7%
  • Shell 3.3%