PyTorch code accompanying the ECCV 2020 paper

Making Sense of CNNs: Interpreting Deep Representations & Their Invariances with Invertible Neural Networks
Robin Rombach*, Patrick Esser*, Björn Ommer
* equal contribution

arXiv | BibTeX | Project Page

• Requirements
• Demos
• Training
  • Data
  • Invariances of Classifiers
    • ResNet
    • AlexNet
• Evaluation
• Pretrained Models
• BibTeX

A suitable conda environment named invariances can be created and activated with:

conda env create -f environment.yaml
conda activate invariances

Optionally, you can then also conda install tensorflow-gpu=1.14 to speed up FID evaluations.

To get started you can directly dive into some demos. After installing the requirements as described above, simply run

streamlit run invariances/demo.py

Please note that checkpoints will be downloaded on demand, which can take a while. You can see the download progress displayed in the terminal running the streamlit command.

We provide demonstrations on

• visualization of adversarial attacks
• visualization of network representations and their invariances
• revealing the texture bias of ImageNet-CNNs
• visualizing invariances from a video (resulting in image to video translation)
• image mixing via their network representations

Note that all of the provided demos can be run without a dataset, and you can add your own images into data/custom.

If not present on your disk, all required datasets (ImageNet, AnimalFaces and ImageNetAnimals) will be downloaded and prepared automatically. The data processing and loading rely on the autoencoders package and are described in more detail here.

Note: If you already have one or more of the datasets present, follow the instructions linked above to avoid downloading them again.

To recover invariances of an ResNet classifier trained on the AnimalFaces dataset, run

edflow -b configs/resnet/animalfaces/base.yaml configs/resnet/animalfaces/train/<layer>.yaml -t

where <layer> is one of input, maxpool, layer1, layer2, layer3, layer4, avgpool, fc, softmax. To enable logging to wandb, adjust configs/project.yaml and add it to above command:

edflow -b configs/resnet/animalfaces/base.yaml configs/resnet/animalfaces/train/<layer>.yaml configs/project.yaml -t

To reproduce the training procedure from the paper, run

edflow -b configs/alexnet/base_train.yaml configs/alexnet/train/<layer>.yaml -t

where <layer> is one of conv5, fc6, fc7, fc8, softmax. To enable logging to wandb, adjust configs/project.yaml and add it to above command:

edflow -b configs/alexnet/base_train.yaml configs/alexnet/train/<layer>.yaml configs/project.yaml -t

Evaluations run automatically after each epoch of training. To start an evaluation manually, run

edflow -p logs/<log_folder>/configs/<config>.yaml

and, optionally, add -c <path to checkpoint> to evaluate a specific checkpoint instead of the last one.

Pretrained models (e.g. autoencoders and classifiers) will be downloaded automatically on their first use in a demo, training or evaluation script.

@inproceedings{rombach2020invariances,
  title={Making Sense of CNNs: Interpreting Deep Representations \& Their Invariances with INNs},
  author={Rombach, Robin and Esser, Patrick and Ommer, Bj{\"o}rn},
  booktitle={Proceedings of the European Conference on Computer Vision},
  year={2020}
}