Y. Cai, et al., "Transformer-based Contrastive Prototypical Clustering for Multimodal Remote Sensing Data", under consideration at Inf. Sci.
pip install torch==1.10.x torchvision==0.11.x spectral PyYAML scikit-learn
python train.py
Configurations are defined in config.yaml.
# general
seed: 42
workers: 2
dataset: "Trento"
dataset_root: 'HSI-Lidar-Trento\\'
model_path: "save/Trento"
# train options
batch_size: 256
image_size: 7
joint_train_epoch: 20
dim_emebeding: 512
lr_scale: 10
is_labeled_pixel: False
is_pretrain: False
# loss options
learning_rate: 0.0005 #0.00002
weight_decay: 0.0005
contrastive_param: 0.5 # tau
weight_clu_loss: 1 # lambda
regularizer_coef: 0.00001 # gamma