deeplabv3-pytorch

DeepLabV3

Rethinking Atrous Convolution for Semantic Image Segmentation - DeepLabV3

Implementation of DeepLabV3 using PyTorch

Available architectures

DeepLabV3	Backbone	mean IoU
deeplabv3_resnet50	resnet50	58.6(after 5 epochs)
deeplabv3_resnet101	resnet101	-
deeplabv3_mobilenetv3	mobilenetv3_large	-

Dataset

├── COCO 
    ├── annotations
        ├──instances_train2014.json
        ├──instances_val2014.json
    ├── train2014
        ├── COCO_train2014_00000000000.png
        ├── COCO_train2014_00000000001.png
    ├── val2014
        ├── COCO_val2014_00000000000.png
        ├── COCO_val2014_00000000001.png

Train

Note: Modify these arguments according to your data and model in train.py

parser.add_argument('--data-path', default='../../Datasets/COCO', help='dataset path') 
parser.add_argument('--dataset', default='coco', help='dataset name')                  
parser.add_argument('--model', default='deeplabv3_resnet50', help='model')             

Distributed Data Parallel:

1. deeplabv3_resnet50:

python -m torch.distributed.launch --nproc_per_node=2 --use_env train.py --lr 0.02 --dataset coco -b 8 --model deeplabv3_resnet50

1. deeplabv3_resnet101:

python -m torch.distributed.launch --nproc_per_node=2 --use_env train.py --lr 0.02 --dataset coco -b 8 --model deeplabv3_resnet101

1. deeplabv3_mobilenet_v3_large:

python -m torch.distributed.launch --nproc_per_node=2 --use_env train.py --lr 0.02 --dataset coco -b 8 --model deeplabv3_mobilenet_v3_large

Without Distributed Data Parallel:

python train.py