README.md

DANet

A PyTorch implementation of DANet based on CVPR 2019 paper Dual Attention Network for Scene Segmentation.

Requirements

• Anaconda
• PyTorch

conda install pytorch torchvision cudatoolkit=10.1 -c pytorch

• opencv

pip install opencv-python

• tensorboard

pip install tensorboard

• pycocotools

pip install git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI

• fvcore

pip install git+https://github.com/facebookresearch/fvcore

• cityscapesScripts

pip install git+https://github.com/mcordts/cityscapesScripts.git

• detectron2

pip install git+https://github.com/facebookresearch/detectron2.git@master

Datasets

For a few datasets that detectron2 natively supports, the datasets are assumed to exist in a directory called datasets/, under the directory where you launch the program. They need to have the following directory structure:

Expected dataset structure for Cityscapes:

cityscapes/
  gtFine/
    train/
      aachen/
        color.png, instanceIds.png, labelIds.png, polygons.json,
        labelTrainIds.png
      ...
    val/
    test/
  leftImg8bit/
    train/
    val/
    test/

run ./datasets/prepare_cityscapes.py to creat labelTrainIds.png.

Training

To train a model, run

python train_net.py --config-file <config.yaml>

For example, to launch end-to-end DANet training with ResNet-50 backbone on 8 GPUs, one should execute:

python train_net.py --config-file configs/r50.yaml --num-gpus 8

Evaluation

Model evaluation can be done similarly:

python train_net.py --config-file configs/r50.yaml --num-gpus 8 --eval-only MODEL.WEIGHTS checkpoints/model.pth

Results

There are some difference between this implementation and official implementation:

1. No Multi-Grid and Multi-Scale Testing;
2. The image sizes of Multi-Scale Training are (800, 832, 864, 896, 928, 960);
3. Training step is set to 24000;
4. Learning rate policy is WarmupMultiStepLR;
5. Position Attention Module (PAM) uses the similar mechanism as Channel Attention Module (CAM), just uses the tensor and its transpose to compute attention.

Name	train time (s/iter)	inference time (s/im)	train mem (GB)	PA %	mean PA %	mean IoU %	FW IoU %	download link
R50	0.49	0.12	27.12	94.19	75.31	66.64	89.54	model | ga7k
R101	0.65	0.16	28.81	94.29	76.08	67.57	89.69	model | xnvs