This repo contains the experiments for the ICLR 2021 paper:

@inproceedings{xie2021innout,
  author = {Sang Michael Xie and Ananya Kumar and Robbie Jones and Fereshte Khani and Tengyu Ma and Percy Liang},
  booktitle = {International Conference on Learning Representations (ICLR)},
  title = {In-N-Out: Pre-Training and Self-Training using Auxiliary Information for Out-of-Distribution Robustness},
  year = {2021},
}

The main script is in innout/main.py. The script requires a YAML config file - an example is configs/landcover/CNN1D.yaml. To dynamically change values of the config file with command line arguments, simply add new arguments of the form --key=val where the key can be any string of multiple keys separated by periods. This is to allow for changing nested components of the config file. For example --model.args.depth=3 changes the config dictionary in this way: config['model']['args']['depth'] = 3. It is important that the key and value are separated by an equals sign.

Examples of how to run the main script on our datasets are in scripts/.

The first time you run this project, create a virualenv in the current directory (which contains the README):

python3 -m venv .env
source .env/bin/activate
pip install -e .

In subsequent runs you only need to activate the environment:

source .env/bin/activate

In-N-Out does better than all other methods on two real-world remote sensing datasets: Landcover and Cropland, and one standard ML benchmark dataset, CelebA. In this CodaLab worksheet, we show all our code and runs for these experiments for reproducibility. The data can also be downloaded from the CodaLab worksheet.

Here is an example run of the baseline model for Landcover (see others on the CodaLab worksheet):

python innout/main.py 
    --dataset.args.unlabeled_prop=0.9
    --epochs=400
    --scheduler.num_epochs=400
    --seed=112
    --dataset.args.seed=1
    --group_name=landcover
    --dataset.args.include_ERA5=False
    --model.args.in_channels=8
    --config=configs/landcover/CNN1D.yaml
    --model_dir=models/landcover_unlabeledprop_0.9/landcover_baseline_unlabeledprop0.9_trial1
    --run_name=landcover_baseline_unlabeledprop0.9_trial1
    --no_wandb
    --return_best