https://github.com/acmi-lab/OnlineLabelShift

Code accompanying our paper titled Online Label Shift: Optimal Dynamic Regret meets Practical Algorithms
https://github.com/acmi-lab/OnlineLabelShift

distribution-shift domain-adaptation machine-learning

Last synced: 7 days ago
JSON representation

Code accompanying our paper titled Online Label Shift: Optimal Dynamic Regret meets Practical Algorithms

• Host: GitHub
• URL: https://github.com/acmi-lab/OnlineLabelShift
• Owner: acmi-lab
• License: apache-2.0
• Created: 2023-05-31T05:31:06.000Z (over 1 year ago)
• Default Branch: main
• Last Pushed: 2023-06-01T10:48:06.000Z (over 1 year ago)
• Last Synced: 2024-08-02T15:32:45.023Z (3 months ago)
• Topics: distribution-shift, domain-adaptation, machine-learning
• Language: Python
• Homepage: https://arxiv.org/abs/2305.19570
• Size: 91.8 KB
• Stars: 4
• Watchers: 3
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
# Online Label Shift

`OnlineLabelShift` is the official implementation of the accompanying paper [Online Label Shift: Optimal Dynamic Regret meets Practical Algorithms](https://arxiv.org/pdf/2305.19570.pdf). 

For more details, please refer to the paper. 



  



## Setup Environment

```

conda env update --file environment.yml

```

## Run Online Label Shift Experiment

The following command runs the online label shift experiment.

It expects the base model has been trained and saved under `/model`

```

python scripts/run_ols.py -d synthetic -m logreg --do-all 1 -t 1000 --save 1

```

To see all the options

```

python scripts/run_ols.py -h

```

## Train model

This script supports model training for synthetic, cifar10, and mnist datasets.

```

python scripts/train_model.py -d  -m  -e 

```

The corresponding models are:

| Data      | Model    |

| --------- | -------- |

| synthetic | logreg   |

| mnist     | fcn      |

| cifar10   | resnet18 |

## Generate Synthetic Data

To run experiments on synthetic data, one should first generate the data with the following command:

```

python scripts/gen_synth_data.py

```