Meta Domain Shift

Enviroment

• Python 3.7
• Pytorch 1.7
• CUDA 10

Getting started

Git clone the repo:

git clone git@github.com:ebennequin/meta-domain-shift.git

Then cd meta-domain-shift and create a virtual environment (if you don't already have it, use pip install virtualenv):

virtualenv venv --python=python3.7
source venv/bin/activate

Then install dependencies: pip install -r requirements.txt Some perturbations used in CIFAR-100-C-FewShot and miniImageNet-C use Wand: sudo apt-get install libmagickwand-dev

Data

To install the datasets to your machine, please follow this walkthrough.

Run an experiment

Configure your experiment by changing the values in configs/*.py, then launch your experiment. python -m scripts.run_experiment

On some machines, the src module will not be found by Python. If this happens to you, run export PYTHONPATH=$PYTHONPATH:path/to/meta-domain-shift to tell Python where you're at.

All outputs of the experiment (explicit configuration, logs, trained model state and TensorBoard logs) can then be found in the directory specified in configs/experiment_config.py. By default, an error will be risen if the specified directory already exists (in order to not harm the results of previous experiments). You may change this behaviour in configs/experiment_config.py by setting OVERWRITE = True.

Reproducing results

See the detailed documentation here.

Track trainings with Tensorboard

We log the loss and validation accuracy during the training for visualization in Tensorboard. The logs of an experiment can be found in the output directory (events.out.tfevents.[...]). To visualize them in Tensorboard, run:

tensorboard --logdir=output_dir

Contribute

1. Make sure you start from an up to date master branch

   git checkout master && git pull
   

2. Create a new branch

   git checkout -b my-branch
   

3. Add your changes

   git add -p
   

4. Commit your changes

   git commit -m "Add FEMNIST dataset"
   

5. Push your changes to the origin remote

   git push --set-upstream origin my-branch
   

6. Open a pull request on GitHub and ask a teammate to review your code
7. Once reviewed, merge your branch to master

Using a new library

If your changes to the code need a new library (not in the requirements), please add it to requirements.txt

References

Meta-learning code is modified from https://github.com/sicara/FewShotLearning Image perturbations are modified from https://github.com/hendrycks/robustness