This is the official implementation of LKD-Net.
- Clone our repository.
git clone https://github.com/SWU-CS-MediaLab/LKD-Net.git
cd LKD-Net
- Make conda environment.
conda create -n LKD python=3.7.0
conda activate LKD
- Install dependencies.
conda conda install pytorch==1.10.1 torchvision==0.11.2 torchaudio==0.10.1 cudatoolkit=10.2 -c pytorch
pip install -r requirement.txt
RESIDE official website here.
DATASETS FILE STRUCTURE
|-- data
|-- ITS
|-- hazy
|-- *.png
|-- gt
|-- *.png
|-- OTS
|-- hazy
|-- *.jpg
|-- gt
|-- *.jpg
|-- SOTS
|-- indoor
|-- hazy
|-- *.png
|-- gt
|-- *.png
|-- outdoor
|-- hazy
|-- *.jpg
|-- gt
|-- *.png
Run the following script to train you own model.
python train.py \
--model LKD-t \
--model_name LKD.py \
--num_workers 8 \
--save_dir ./result \
--datasets_dir ./data \
--train_dataset ITS \
--valid_dataset SOTS \
--exp_config indoor \
--gpu 0 \
--exp_name train_my_model \
Run the following script to test the trained mode.
python test.py --model (model name) --model_weight (model weight dir) --data_dir (your dataset dir) --save_dir (path to save test result) --dataset (dataset name) --subset (subset name)
For example, we test the LKD-t on the SOTS indoor set.
python test.py --model LKD-t --model_weight ./result/RESIDE-IN/LKD-t/LKD-t.pth --data_dir ./data --save_dir ./result --dataset SOTS --subset indoor
Thanks for the work of Yuda Song et al. Our code is heavily borrowed from the implementation of Dehazeformer.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent