DualAttention_for_Segmentation

This is the repository for re-implementation of dual attention network with pytorch.

Requirements

*python 3.x

• pytorch >= 1.0
• torchvision
• pandas
• numpy
• Pillow
• tqdm
• PyYAML
• addict
• tensorboardX
• adabound

Dataset

PASCAL VOC(2007/2012)

You can download from this link

Training

If you want to train a model, please run python utils/build_dataset.py to make csv_files for training and validation.

Then, just run python train.py ./PATH_TO_CONFIG_FILE

For example, when running python train.py ./result/danet_drn_d_22/config.yaml, the configuration described in ./result/danet_drn_d_22/config.yaml will be used .

If you want to set your own configuration, please make config.yaml like this:

model: drn_d_22
attention: True       # if you use dual attention modules or not

class_weight: True    # if you use class weight to calculate cross entropy or not
writer_flag: True     # if you use tensorboardx or not

n_classes: 21         # including background class
batch_size: 32
crop_height: 300
crop_width: 300
height: 256
width: 256
num_workers: 4
max_epoch: 300

optimizer: AdaBound
learning_rate: 0.001
lr_patience: 10       # Patience of LR scheduler
momentum: 0.9         # momentum of SGD
dampening: 0.0        # dampening for momentum of SGD
weight_decay: 0.001   # weight decay for SGD
nesterov: True        # enables Nesterov momentum
final_lr: 0.1         # final learning rate for AdaBound

dataset_dir: /xxxx/xxxx/xxxx/VOCdevkit
year: 2012 # pascal voc 2007 or 2012
result_path: ./result/drn_d_22

References

Dual Attention Network for Scene Segmentation,
Jun Fu, Jing Liu, Haijie Tian, Yong Li, Yongjun Bao, Zhiwei Fang,and Hanqing Lu,
in CVPR2019
arXiv
Github