Awesome

UDA-Bench: Revisiting Common Assumptions in Unsupervised Domain Adaptation Using a Standardized Framework

ECCV 2024

UDA-Bench is a comprehensive and standardized PyTorch framework for training and evaluating Unsupervised Domain Adaptation (UDA) methods. This repository provides a foundation for researchers to:

• Benchmark UDA methods: Easily compare different UDA techniques on a standardized platform.
• Implement new UDA methods: Extend the framework to incorporate novel methods with minimal effort.
• Reproduce experiments: Replicate the comparative studies conducted in our ECCV 2024 paper.

Highlights:

• Consolidated Framework: A unified structure for training and evaluating UDA methods, streamlining the research process.
• Comprehensive Method Support: Includes implementations of leading UDA algorithms like DANN, CDAN, MCC, MDD, MemSAC, ILADA, SAFN, BSP, MCD, AdaMatch, DALN, and ToAlign.
• Modular Architecture: Easily extendable to accommodate new UDA algorithms, loss functions, and backbones.
• Reproducibility: Facilitates replicating experimental setups and results from our paper, enhancing research transparency.

Installation

Install the necessary dependencies using the requirements.txt file:

pip install -r requirements.txt

Datasets

The repository includes dataset loaders for DomainNet, visDA, CUB200, and OfficeHome. The train and test files are available in the data directory as .txt files. You can download the images from the official sources of these datasets.

UDA Methods

The following UDA methods are currently implemented in this framework:

1. DANN
2. CDAN
3. MCC
4. MDD
5. MemSAC
6. ILADA
7. SAFN
8. BSP
9. MCD
10. AdaMatch
11. DALN
12. ToAlign

Training

Train an existing UDA method using the following command:

#!/bin/bash

export trainer=cdan # Select a UDA method
export dataset=DomainNet # Choose a dataset
export n_class=345 # Number of classes
export data_root=/data # Path to data directory
export source=real # Source domain
export target=clipart # Target domain

python3 train.py --config configs/$trainer.yml \
    --source data/$dataset/${source}_train.txt \
    --target data/$dataset/${target}_train.txt \
    --num_class $n_class --data_root $data_root \
    --num_iter 90000 --exp_name test --trainer $trainer

Replace trainer with any of the implemented UDA methods.

Comparative Studies

Reproduce the comparative studies conducted in our paper:

Changing Backbone Architecture

The framework supports the following backbones:

• ResNet50: resnet50
• ConvNext: timm_convnext
• SWIN: timm_swin
• ResMLP: timm_resmlp
• DeiT: timm_deit

Change the default ResNet-50 architecture using the following command:

#!/bin/bash

export trainer=cdan
export dataset=DomainNet
export n_class=345
export data_root=/data
export source=real
export target=clipart
export backbone=timm_deit # Select a backbone

python3 train.py --config configs/$trainer.yml \
    --source data/$dataset/${source}_train.txt \
    --target data/$dataset/${target}_train.txt \
    --num_class $n_class --data_root $data_root \
    --num_iter 90000 --exp_name test --trainer $trainer \
    --backbone $backbone

Effect of Amount of Data

Reproduce results with reduced target unlabeled data:

#!/bin/bash

export trainer=cdan
export dataset=DomainNet
export n_class=345
export data_root=/data
export source=real
export target=clipart
export backbone=resnet50
export tgt_data_vol=50 # Percentage of target unlabeled data

python3 train.py --config configs/$trainer.yml \
    --source data/$dataset/${source}_train.txt \
    --target data/$dataset/${target}_train.txt \
    --num_class $n_class --data_root $data_root \
    --num_iter 90000 --exp_name test --trainer $trainer \
    --backbone $backbone --target_imb_factor ${tgt_data_vol}

Implementing New UDA Methods

Extend the framework to add new UDA methods:

1. Add a new config file: Create configs/<method>.yaml.
2. Implement the forward pass module: Create UDA_trainer/<method>.py.
3. Implement new loss functions: Create new loss functions in losses/.

You can also modify the architecture, dataloader, or training strategy if needed.

Citation

If you use this code or our work, please cite our paper:

@article{kalluri2024lagtran,
       author   = {Kalluri, Tarun and Ravichandran, Sreyas and Chandraker, Manmohan},
       title    = {UDA-Bench: Revisiting Common Assumptions in Unsupervised Domain Adaptation Using a Standardized Framework},
       journal  = {ECCV},
       year     = {2024},
       url      = {},
     },

Contact

For any questions or inquiries, please contact Tarun Kalluri.