Contrastive Learning for Cross-Domain Open World Recognition

1. Requirements

Packages

You can install required packages via pip:

pip install -r requirements.txt

The last dependency is torchlars which has to be installed after the others:

pip install torchlars

Datasets

Data should be placed under ~/data/.

ROD, synROD, ARID

Follow OWR Visualdomains benchmark instructions for datasets preparation.

COSDA-HR

Download data from the official site and untar it.

CORe50

Download data from the official site. You will need the 'cropped_128x128_images.zip' file.

2. Training

We perform training on a single GPU.

The basic training command for ROD:

python train.py --dataset OWR --source ROD --dataorder 0

The dataorder param is necessary to choose one of the 5 available class orders for incremental learning. The training output is stored in logs/OWR_ROD_do_0/

To train on synROD:

python train.py --dataset OWR --source synROD --dataorder 0

To train on COSDA-HR:

python train.py --dataset COSDA-HR --dataorder 0

To train on CORe50:

python train.py --dataset CORe50 --dataorder 0

3. Evaluation

To perform an eval on a single experiment and episode one can use the eval.py script:

python eval.py --source ROD --target arid --load_path logs/OWR_ROD_0/episode_0.model --mode OWR_eval --eval_episode 0 --dataorder 0

After performing a complete training for all 5 class orders for a given dataset one can also test all models together.

For OWR datasets:

python eval_OWR.py ROD arid logs/OWR_ROD

The listed paramerts represent in order: "source", "target" and "train output path". Note that the output path should not include the suffix _do_* as it is automatically added by the eval script.

Similarly to test a complete COSDA-HR experiment:

python eval_COSDA.py logs/COSDA_source

And for CORe50:

python eval_CORe50.py logs/CORe50_source