hangfrieddays / cigformer

CIGformer

CIGformer

---

This is the official implementation of ["CIGformer: A Transformer-Based Pansharpening Network Through Continuous Information Guidance"]

The original github address of this code is hangfrieddays/CIGformer (github.com)

Overview of CIGformer

Architecture of ISBlock

Architecture of IGBlock

Requirements

This environment is mainly based on python=3.6 with CUDA=10.2

conda create -n CIGformer python=3.10
conda activate CIGformer
conda install pytorch=1.7.1 torchvision=0.2.2 cudatoolkit=10.2
pip install mmcv==1.2.7
conda install gdal=3.1.0 -c conda-forge
conda install scikit-image=0.17.2
pip install scipy==1.5.3
pip install gpustat==0.6.0
pip install numba==0.53.1 
pip install einops==0.3.0 
pip install timm==0.3.2
pip install sewar==0.4.4

Create Dataset

cd ./utils
python handle_raw.py
python clip_patch.py

Train CIGformer

Due to the large size of the dataset, we only provide some samples in './data' to verify the code.

conda activate CIGformer
export CUDA_VISIBLE_DEVICES='0';
# train TransferNetwork
python transfertrain.py

# train CIGformer
python fusiontrain.py

you can pass hyper-parameters below:

• pretrained = False
• log_pth = 'path to save your training log'
• log_name = 'CIGformer'
• config_pth = 'records/CIGformer/config.yml'

You can modify the config file 'models/model.py' for different purposes.

Citing CIGformer

Consider cite CIGformer in your publications if it helps your research.