jiuzhuanzhuan / DSSLIC

DSSLIC: Deep Semantic Segmentation-based Layered Image Compression

DSSLIC

Pytorch implmementation of DSSLIC: Deep Semantic Segmentation-based Layered Image Compression

Paper: https://arxiv.org/abs/1806.03348

Original Requirements:

• Ubuntu 16.04
• Python 2.7
• Cuda 8.0
• Pyorch 0.3.0

My Environments:

• GTX 1080 Ti (with 11GB graphic memory)
• Ubuntu 16.04
• Python 3.5
• Cuda 9.0
• Pytorch 0.4.1

Training

ADE20K Dataset

python3 train.py --name ADE20K_model --dataroot ./datasets/ADE20K/ --label_nc 151 --loadSize 256 --resize_or_crop resize --batchSize 2

This takes ~7 seconds per epoch on my system.

Cityscapes Dataset

python3 train.py --name Cityscapes_model --dataroot ./datasets/cityscapes/ --label_nc 35 --loadSize 512 --resize_or_crop scale_width --batchSize 1

This take ~30 seconds per epoch on my system.

Testing

ADE20K testset

python3 test.py --name ADE20K_model --dataroot ./datasets/ADE20K/ --label_nc 151 --loadSize 384 --resize_or_crop resize --batchSize 1 --how_many 50

Kodak testset

python3 test.py --name ADE20K_model --dataroot ./datasets/Kodak/ --label_nc 151 --loadSize 384 --resize_or_crop resize --batchSize 1 --how_many 24

Cityscapes testset

python3 test.py --name Cityscapes_model --dataroot ./datasets/cityscapes/ --label_nc 35 --loadSize 1536 --resize_or_crop scale_width --batchSize 1 --how_many 50