Awesome

🦸🏻‍♀️ DIGA - Dynamically Instance-Guided Adaptation: A Backward-free Approach for Test-Time Domain Adaptive Semantic Segmentation

We are still working on this repo now.

Environment Setup

Before everything start, please make sure the following environment variables(UDATADIR, UPRJDIR, UOUTDIR) are setup. The following operation would only modifying the above three folders on your devices.

example:

# e.g. 
export UDATADIR=~/data # dir for dataset
export UPRJDIR=~/code # dir for code
export UOUTDIR=~/output # dir for output such as logs
export WANDB_API_KEY="xxx360492802218be41f76xxxx" # your wandb key
export NUM_WORKERS=0 # number of works used
mkdir -p $UDATADIR $UDATADIR $UOUTDIR # create dir if it does not exist

ps. Wandb is a wonderful tool for visualization similar to Tensorboard but offer more functions (link).

Code Setup

Copy or clone this project.
Organize the files in the following structure.

Code for this project should be put in $UPRJDIR with the following structure:

$UPRJDIR
    ├── DIGA
            ├── ...
            ├── Readme.md
            ├── ...

Then you can use cd $UPRJDIR/DIGA to get in the project directory.

Dataset Setup

Init file structure by running cp -r $UPRJDIR/DIGA/src/utils/advent_list_lib/* $UDATADIR/
download dataset

Please read the following before get start.

The environment variable $UDATADIR setted in last step would be used to locate the dataset files when program is running. Each dataset is contained in one folder, the structure in $UDATADIR should be like this:

$UDATADIR
    ├── GTA5 # (GTA5 in paper)
    ├── synthia # (Synthia in paper)
    ├── Cityscapes # (CityScapes in paper)
    ├── BDD # (BDD100K in paper)
    ├── IDD # (IDD in paper)
    ├── NTHU # (CrossCity in paper)
    ├── Mapillary # (Mapillary in paper)

In most dataset folders, there is one folder called “advent_list”, which contains the list for training, validation, testing set [following implementation in ADVENT].

We provide the lists file in this project, the folder is $UPRJDIR/DIGA/src/utils/advent_list_lib

Running the follows would set up the folders for you automatically:

# copy DIGA/src/utils/advent_list_lib/* to $UDATADIR/
cp -r $UPRJDIR/DIGA/src/utils/advent_list_lib/* $UDATADIR/

Now, the structure should look like this:

$UDATADIR
    ├── GTA5 # (GTA5 in paper)
    │   ├── advent_list
    │   ├── ...
    ├── synthia # (Synthia in paper)
    │   ├── ...
    │   ├── ...
    ├── Cityscapes # (CityScapes in paper)
    │   ├── advent_list
    │   ├── ...
    ├── BDD # (BDD100K in paper)
    │   ├── advent_list
    │   ├── ...
    ├── IDD # (IDD in paper)
    │   ├── ...
    │   ├── ...
    ├── NTHU # (CrossCity in paper)
    │   ├── advent_list
    │   ├── ...
    ├── Mapillary # (Mapillary in paper)
    │   ├── advent_list
    │   ├── ...

The next step is downloading each dataset. It would cost too much space to offer instructions about downloading and there are many tutorial online. Instead, we offer an illustration as follows about the final dataset structure. You can download the datasets and organize them like follows. Some instructions can be find here (link1, link2)

$UDATADIR
├── BDD
│   ├── advent_list
│   ├── images
│   ├── labels
├── Cityscapes
│   ├── README
│   ├── advent_list
│   ├── gtFine
│   ├── leftImg8bit
├── GTA5
│   ├── advent_list
│   ├── images
│   ├── labels
├── IDD
│   ├── gtFine
│   ├── iddscripts
│   ├── leftImg8bit
├── Mapillary
│   ├── LICENSE
│   ├── README
│   ├── advent_list
│   ├── config.json
│   ├── small
│   ├── testing
│   ├── training
│   └── validation
├── NTHU
│   ├── Rio
│   ├── Rome
│   ├── Taipei
│   ├── Tokyo
│   └── advent_list
├── synthia
│   ├── Depth
│   ├── GT
│   ├── RGB

Pre-trained model Setup

Download pre-trained models
Organize the files in the following structure.

Source models are required to be put in specific directory for running. The pre-trained models are available here (link).

ps. We use consistent models from previous work of this repo link.

$UDATADIR
├── models # create a new folder called "models" under $UDATADIR 
    ├── DA_Seg_models # downloaded from above link
        ├── GTA5
        ├── GTA5_baseline.pth
        ├── SYNTHIA
        ├── SYNTHIA_source.pth
        ├── ...
├── ...
├── BDD

The following cmd would do this automatically:

mkdir -p $UDATADIR/models
cd $UDATADIR/models
wget -O download.zip https://www.dropbox.com/s/gpzm15ipyt01mis/DA_Seg_models.zip?dl=0https://www.dropbox.com/s/gpzm15ipyt01mis/DA_Seg_models.zip?dl=0
unzip download.zip -d ./

ps. adding, deleting, editing path could be done at configs/model/net/gta5_source.yaml

Python Environment

We provided 3 ways to setup the environments.

Using Develop Environment in VS Code (most Recommended)
Using Docker (Recommended)
Using Pip/Conda

Using Develop Environment in VS Code (Recommended)

If you are using VS Code, this is the most recommended way. You can set up all the environment for this project in just one step.

Open this project in VS Code
Install extension “Dev Containers”
Press Cmd/Ctrl+Shift+P → Dev Container: Rebuild and Reopen in Container

ps. The config folder .devcontainer has been included in our project, you can edit it as if you would introduce or remove some libraries.

ps. Details and Instructions about Dev Container in VS Code can be found here (link).

Using Docker

Dockerfile is at .devcontainer/Dockerfile

Using Pip/Conda

Install pytorch form official website (link).

pip install \
    tensorboard
    pandas \
    opencv-python \
    pytorch-lightning \
    hydra-core \
    hydra-colorlog \
    hydra-optuna-sweeper \
    torchmetrics \
    pyrootutils \
    pre-commit \
    pytest \
    sh \
    omegaconf \
    rich \
    fiftyone \
    jupyter \
    wandb \
    grad-cam \
    tensorboardx \
    ipdb \
    hydra-joblib-launcher

ps. We use pytorch/pytorch:1.12.1-cuda11.3-cudnn8-devel while other versions should also work.

Running the code

Get Start

The following command will use test the performance of a model pre-trained on GTA5 with target as the validation set of Cityscapes. Hyper-parameters are set as default.

python src/train.py experiment=ttda

Output Example:

wandb: Run summary:
...
wandb: test/acc/dataloaderr13_idx_0 55.01257 # mIoU of 13 class
wandb: test/acc/dataloaderr16_idx_0 49.24151 # mIoU of 16 class
wandb:   test/acc/dataloaderr_idx_0 45.81422 # mIoU of 19 class
...

Custom

You can use the following to start an experiment with custom source model, target dataset and hyper-parameters.

# e.g.
python src/train.py \
    experiment=ttda \
    model/net=gta5_source \
    datamodule/test_list=cityscapes \
    model.cfg.bn_lambda=0.8 \
    model.cfg.proto_lambda=0.8 \
    model.cfg.fusion_lambda=0.8 \
    model.cfg.confidence_threshold=0.9 \
    model.cfg.proto_rho=0.1

The available choices are listed as following:

Source model:

model/net=gta5_source,synthia_source,gta5_synthia_source
Target dataset:

datamodule/test_list=idd,crosscity,cityscapes,bdd,mapillary
Hyper-Paramerters:

bn_lambda: 0-1 (default 0.8)

proto_lambda: 0-1 (default 0.8)

fusion_lambda: 0-1 (default 0.8)

confidence_threshold: 0-1 (default 0.9)

proto_rho: 0-1 (default 0.1)

ps. File based customization is also supported by modifying configs/model/diga.yaml . Note that the cmd line has higher priority if there are conflicting options.