cls

MIC for Domain-Adaptive Image Classification

Getting started

Installation

For this project, we used python 3.8.5. We recommend setting up a new virtual environment:

python -m venv ~/venv/mic-cls
source ~/venv/mic-cls/bin/activate

In that environment, the requirements can be installed with:

pip install -r requirements.txt

We use Weights and Biases (wandb) to track our experiments and results. For that purpose, please create a new project MIC with your account.

Datasets

The datasets used in the repository can be downloaded from the following links:

• Office-Home
• VisDA-2017 (under classification track)

The datasets are automatically downloaded to the examples/data/ folder if it is not available.

Training

For experiments in our paper, we use a script to automatically generate the different configurations and train them:

python run_experiments.py --exp <ID>

More information about the available experiments and their assigned IDs, can be found in experiments.py.

The experiment progress is logged on Weights&Biases.

For VisDA-2017, the mean over the class accuracies is reported. This value is denoted as 'mean correct' in the logs as explained in val-iisc/SDAT#1.

Where to find MIC in the code?

The most relevant files for MIC are:

• experiment.py: Definition of the experiment configurations in our paper.
• examples/cdan_mcc_sdat_masking.py: Training script for UDA with MIC(SDAT).
• dalib/modules/masking.py: Implementation of MIC.
• dalib/modules/teacher.py: Implementation of the EMA teacher.

Acknowledgements

MIC for classification is based on the following open-source projects. We thank their authors for making the source code publicly available.

• SDAT
• Transfer Learning Library