Awesome
CoNMix for Source-free Single and Multi-target Domain Adaptation
Official pytorch implementation for CoNMix (website)
Accepted at IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)-2023, Waikoloa, Hawaii
Vikash Kumar, Rohit Lal, Himanshu Patil, Anirban Chakraborty
Abstract
This work introduces the novel task of Source-free Multi-target Domain Adaptation and proposes adaptation framework comprising of Consistency with Nuclear-Norm Maximization and MixUp knowledge distillation CoNMix as a solution to this problem.
Getting Started
Download Data
Manually download the dataset Office-Home
<details> <summary>Click to see full directory tree</summary> data
├── office-home
├── Art
├── Art.txt
├── Clipart
├── Clipart.txt
├── Product
├── Product.txt
├── Real_World
└── RealWorld.txt
</details>
Download Models
Get models in this link: R50-ViT-B_16, ViT-B_16, ViT-L_16...
wget https://storage.googleapis.com/vit_models/imagenet21k/R50+ViT-B_16.npz
mkdir -p model/vit_checkpoint/imagenet21k
mv R50+ViT-B_16.npz model/vit_checkpoint/imagenet21k/R50+ViT-B_16.npz
Prerequisites:
- See requirements.sh
- Install dependencies using
sh scripts/requirements.sh
Training
Install the dependencies and run scripts.
Stage 1: Source only Training
We train a simple source free model using supervised loss.
sh scripts/source_train.sh
Stage 2: STDA training
Source Free Single Target domain adaptaion training
sh scripts/STDA.sh
Stage 3: KD-MTDA training
Source Free Multi Target Domain adaptation. This must be trained after completing STDA training.
sh scripts/MTDA.sh
Testing
For testing any model use the test_model_acc.py code. There are two function
multi_domain_avg_acc()
: Gives average acc across all domains (equal weight to all domains irrespective of images)test_model()
: Testing on single domain
Changes to be done in code
- saved model path
- model type
- domain/list of domain to be used in function
multi_domain_avg_acc()
ortest_model()
- update
bottleneck_dim
variable as per your training model
Code Reference
Citation
@inproceedings{kumar2023conmix,
title={CoNMix for Source-free Single and Multi-target Domain Adaptation},
author={Kumar, Vikash and Lal, Rohit and Patil, Himanshu and Chakraborty, Anirban},
booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
pages={4178--4188},
year={2023}
}