domain-generalization-platform

This project is implemented based on dassl (v0.5.0) official project and adopts inheritance design and registration mechanism, so that developers and researchers can quickly complete the project construction and deployment of local cross domain classification tasks, and can also be more easily integrated into their own projects.

• 基于 DASSL 的跨域分类任务模板。
• 本项目基于 DASSL(v0.5.0) 官方项目实现，采用继承设计和注册机制，使得开发人员和研究人员能够快速完成本地跨域分类任务的项目搭建和部署，同时也能够更方便的集成到自己的项目中。

Installation & Prerequisite

• The installation of DASSL based on pytorch, See the official tutorial for details;
• It requires the following packages:

Python 3.7
PyTorch >= 1.6
torchvision 0.7.0
CUDA 10.1
DASSL >= v0.5.0
flake8==3.7.9
yapf==0.29.0
isort==4.3.21
yacs
gdown
tb-nightly
future
scipy
scikit-learn
tqdm

Usage

Data preparation

Download and prepare data following https://github.com/KaiyangZhou/Dassl.pytorch/blob/master/DATASETS.md. The PACS dataset structure as example looks like below:

pacs/
|–– images/
|–– splits/

Training and testing

• You can run training and testing directly by the linux command (PACS dataset as example) like below:

python train.py --root ${DATASET_PATH} --trainer ${TRAINER} --source-domains art_painting --target-domains cartoon photo sketch --dataset-config-file ${DASSL_PATH}/configs/datasets/dg/pacs.yaml --config-file ${DASSL_PATH}/configs/trainers/dg/vanilla/pacs.yaml --output-dir ${OUTPUT_DIR} MODEL.BACKBONE.NAME resnet18

• Or you can cd script and run bash res18.sh to train and test like below:

#!/bin/bash

cd ..

DATA=~/dataset/
DASSL=~/Dassl.pytorch

D1=art_painting
D2=cartoon
D3=photo
D4=sketch

################### leave one domain out setting
DATASET=pacs
TRAINER=Vanilla
NET=resnet18

for SEED in $(seq 1 2)
do
    for SETUP in $(seq 1 4)
    do
        if [ ${SETUP} == 1 ]; then
            S1=${D2}
            S2=${D3}
            S3=${D4}
            T=${D1}
        elif [ ${SETUP} == 2 ]; then
            S1=${D1}
            S2=${D3}
            S3=${D4}
            T=${D2}
        elif [ ${SETUP} == 3 ]; then
            S1=${D1}
            S2=${D2}
            S3=${D4}
            T=${D3}
        elif [ ${SETUP} == 4 ]; then
            S1=${D1}
            S2=${D2}
            S3=${D3}
            T=${D4}
        fi
        
        python train.py \
        --root ${DATA} \
        --seed ${SEED} \
        --trainer ${TRAINER} \
        --source-domains ${S1} ${S2} ${S3}  \
        --target-domains ${T} \
        --dataset-config-file ${DASSL}/configs/datasets/dg/${DATASET}.yaml \
        --config-file ${DASSL}/configs/trainers/dg/vanilla/${DATASET}.yaml \
        --output-dir baseline/${DATASET}/${TRAINER}/${NET}_nodetach/${T}/seed${SEED} \
        MODEL.BACKBONE.NAME ${NET}
    done
done

• For only test, you can use --eval-only. Meanwhile, you also need to provide the trained model and specify which model file (i.e. saved at which epoch) to use. For example, to use model.pth.tar-20 saved at output/source_only_office31/model, you can add following command:

--eval-only \
--model-dir output/source_only_office31 \
--load-epoch 20

Custom settings

You can define your own trainer, training config and dataset config by following step:

Custom trainer

• To creat your own trainer, you can refer to the template in the folder trainer/Vanilla_freezen.py;
• It mainly realizes the call of new classes through inheritance design and registration mechanism like below:

@TRAINER_REGISTRY.register()
class name_of_your_trainer(TrainerX):
    """Your own trainer."""
    def __init__(self, cfg):
    ### ... ###

• For more details and code implementation, please refer to the official tutorial.

Custom training config & dataset config

You can set your own training or dataset parameters by changing the xxx.yaml template in the folder config/trainer or config/dataset.

VSCode debugger launch config

{
    "name": "Python: train.py",
    "type": "python",
    "request": "launch",
    "program": "${workspaceRoot}/train.py",
    "console": "integratedTerminal",
    "justMyCode": false,
    "args": [
        "--root","~/dataset/",
        "--seed", "1",
        "--trainer", "Vanilla_freezen",
        "--source-domains", "art_painting",
        "--target-domains", "cartoon", "photo", "sketch",
        "--dataset-config-file", "configs/datasets/shape_task1_pacs.yaml",
        "--config-file", "~/Dassl.pytorch/configs/trainers/dg/vanilla/pacs.yaml",       
        "--output-dir", "shapetask1/pacs/Vanilla_singles/resnet18_nodetach/random/art_painting/seed1",
        "MODEL.BACKBONE.NAME", "resnet18" 
    ]
}