C-Mixup: Improving Generalization in Regression

Official code of C-Mixup.

If you find this repository useful in your research, please cite the following paper:

@inproceedings{yao2022cmix,
  title={C-Mixup: Improving Generalization in Regression},
  author={Yao, Huaxiu and Wang, Yiping and Zhang, Linjun and Zou, James and Finn, Chelsea},
  booktitle={Proceeding of the Thirty-Sixth Conference on Neural Information Processing Systems},
  year={2022}
}

Prerequisites

• python 3.7.13
• matplotlib 3.3.4
• numpy 1.20.1
• pandas 1.2.3
• pillow 9.0.1
• pytorch 1.11.0
• pytorch_transformers 1.2.0
• torchvision 0.9.0
• wilds 2.0.0

Datasets and Scripts

We put all code except Echo and PovertyMap on the src folder. Echo and PovertyMap datasets are built upon different codebase, which are put in the echo and povertymap folders, respectively.

Airfoil

This dataset can be downloaded via the link in the Google Drive. Please put the corresponding datafolder to src/data

The command to run C-Mixup on Airfoil is:

python main.py --dataset Airfoil --mixtype kde --kde_bandwidth 1.75 --use_manifold 1 --store_model 1 --read_best_model 0

NO2

This dataset can be downloaded via the link in the Google Drive. Please put the corresponding datafolder to src/data

The command to run C-Mixup on NO2 is:

python main.py --dataset NO2 --mixtype kde --kde_bandwidth 1.2 --use_manifold 0 --store_model 1 --read_best_model 0

Exchange_rate

This dataset can be downloaded via the link in the Google Drive. Please put the corresponding datafolder to src/data

The command to run C-Mixup on Exchange_rate is:

python main.py --dataset TimeSeries --data_dir ./data/exchange_rate/exchange_rate.txt --ts_name exchange_rate --mixtype kde --kde_bandwidth 5e-2 --use_manifold 1 --store_model 1 --read_best_model 0

Electricity

This dataset can be downloaded via the link in the Google Drive. Please put the corresponding datafolder to src/data

The command to run C-Mixup on Electricity is:

python main.py --dataset TimeSeries --data_dir ./data/electricity/electricity.txt --ts_name electricity --mixtype kde --kde_bandwidth 0.5 --use_manifold 0 --store_model 1 --read_best_model 0

RCF-MNIST

This dataset can be downloaded via the link in the Google Drive. Please put the corresponding datafolder to src/data

The command to run C-Mixup on RCF-MNIST is:

python main.py --dataset RCF_MNIST --data_dir ./data/RCF_MNIST --mixtype random --batch_type 1 --kde_bandwidth 0.2 --use_manifold 1 --store_model 1 --read_best_model 0

Crime

This dataset can be downloaded via the link in the Google Drive. Please put the corresponding datafolder to src/data

The command to run C-Mixup on Crime is:

python main.py --dataset CommunitiesAndCrime --mixtype kde --kde_bandwidth 4.0 --use_manifold 1 --store_model 1 --read_best_model 0

Skillcraft

This dataset can be downloaded via the link in the Google Drive. Please put the corresponding datafolder to src/data

The command to run C-Mixup on Skillcraft is:

python main.py --dataset SkillCraft --mixtype kde --kde_bandwidth 1.0 --use_manifold 0 --store_model 1 --read_best_model 0

DTI

This dataset can be downloaded via the link in the Google Drive. Please put the corresponding datafolder to src/data

The command to run C-Mixup on DTI is:

python main.py --dataset Dti_dg --data_dir ./data/dti --mixtype kde --kde_bandwidth 20.0 --use_manifold 1 --store_model 1 --read_best_model 0

PovertyMap

To get detailed information of the datasets, please refer to Appendix E of the paper or original paper.

This code is built upon LISA and Wilds.

Before running, please cd PovertyMap

The datasets will be automatically downloaded when running the scripts provided below.

python main.py --dataset poverty --algorithm mixup --data-dir ../../datasets/ --experiment_dir .. --is_kde 1 --kde_bandwidth 0.5

EchoNet

To get detailed information of the datasets, please refer to the website.

This code is built upon EchoNet.

Before running, please cd EchoNet.

You need to follow the guideline from the website and download the dataset into ../../EchoNet-Dynamic/ directory first.

For the preparation you need to install the echonet environment and complete segmentation tasks by running the commands:

pip install --upgrade --user . 
python echonet/__main__.py segmentation --save_video

The command to run C-Mixup on EchoNet is:

echonet video --batch_size 10 --device cuda --num_workers 0 --num_epochs 20 --mixtype kde --bandwidth 50.0 --run_test True