To get started, please simply fork this GitLab repository and follow the structure, testing and submissions guidelines below.
Define your dependencies in a requirements.txt (or a conda environment.yml) file.
Please make sure to always have a reference to the latest adversarial-vision-challenge like so:
git+https://github.com/bethgelab/adversarial-vision-challenge
To run a model server, use the model_server utility method from the above defined adversarial-vision-challenge library, as follows:
from adversarial_vision_challenge import model_server
foolbox_model = load_your_model()
model_server(foolbox_model)
Define the following properties in the crowdai.json:
- challenge_id: "nips-2018-avc-robust-model-track"
- grader_id: "nips-2018-avc-robust-model-track"
- authors: your crowdai username
- description: a description of your model
Define the entrypoint command that should be run when the built docker image is started, i.e. in this case: python main.py.
For further fully functional model examples, check the following repos:
- ResNet18 TensorFlow Model: https://gitlab.crowdai.org/adversarial-vision-challenge/resnet18_model_baseline
- ResNet18 Frozen Noise TensorFlow Model: https://gitlab.crowdai.org/adversarial-vision-challenge/resnet18_frozen_noise_model_baseline
- TODO: Link 3
- TODO: Link 4
- TODO: Link 5
Within the root-folder of this repository, simple run:
avc-test-model .
Please always make sure that you have the newest version of the adversarial-vision-challenge library installed.
Submissions are done by simply running:
avc-submit /path/to/your/repo
or simply avc-submit . from within the root folder of this repo.
You can define your own submisson tag via --tag your_tag,
otherwise a random one will be generated.
If you define your own, please use a new tag for every new submission. Every submission will create a new GitLab issue, where you can track the progress.