Using the consensus of features from multiple layers to attain robust image classification.
This is the core, polished code used to implement our DeepConsensus paper.
The frozen original, experimental code to reproduce the results is found in this repository.
To see how DeepConsensus improves the robustness of an arbitrary CNN, run:
python3 mnist_train.pyThis will train the CNN and its DeepConsensus version on a 64x64 MNIST dataset, where only the test set is perturbed with 20 pixel translations in both x and y axes. Note the massive improvement in test scores.
Make sure you have Python 3. Clone the repo, then run:
pip3 install numpy torch torchvision tqdm | File | Description |
|---|---|
deepconsensus.py |
Implementation of a DeepConsensus layer |
cnn_example.py |
Sample CNN without DeepConsensus |
deepconsensus_example.py |
Sample DeepConsensus that inherits from cnn_example.py |
mnist.py |
Translation pertubations of MNIST. Run python3 mnist.py to see samples |
mnist_train.py |
cnn_example.py vs deepconsensus_example.py on perturbed MNIST |