To start the app:
docker-compose up
Make sure to have Docker daemon running.
Before you start development: download the test and validation datasets at http://www.image-net.org/download-images (ISVRC 2012 dataset) or follow https://github.com/onnx/models/blob/master/vision/classification/imagenet_prep.md
Download the datasets for teacher, poisoned teacher and student model:
- Teacher: https://drive.google.com/file/d/19tfqPEx8jzRaj5qw_tF4lbQ2aUWMyZqH/view?usp=sharing
- Poisoned teacher: https://drive.google.com/file/d/1_pXhJrT1L0ksiV0_hIqUFKKP1x9PP29Z/view?usp=sharing
- Student: https://drive.google.com/file/d/1bFI6Np2L6HzMCu3nZa3425wfDgojpKSt/view?usp=sharing
Start the app: docker-compose up and then follow the link shown in the console. In the notebooks folder you will find a jupyter notebook called train_mobilenet.ipynb. To use a trained model, see the eval_mobilenet.ipynb notebook instead. This will also produce a my-model.onnx file, which should be stored in /home/checkpoints. You can then find the model on your own machine in this repo in the /latent-backdoor/params folder.
- ensure that the GPU version of mxnet is enabled in
latent-backdoor/train/setup-dependencies.sh - clone the repository into your home directory on peregrine
- run
pip install gdown - Update the files
ensure-data.sh,train-teacher.py,train-trigger.pyandtrain-student.pyto point to your own/data/<s or p number>folder. Then run it to download and extract the data into the appropriate folder - submit the batch for the network that you want to train:
sbatch train-peregrine.txt