Sovereign 8s.is mirror of standardgalactic/PKDiff
Find a file
Repository files (latest commit first)
Filename Latest commit message Latest commit date
2025-10-27 16:48:35 +08:00
networks Create network.py 2025-10-27 15:17:33 +08:00
utils Add files via upload 2025-10-27 15:16:59 +08:00
eval_patch_dataset.py Add files via upload 2025-10-27 15:04:28 +08:00
README.md Update README.md 2025-10-27 16:48:35 +08:00
requirements.txt Add files via upload 2025-10-27 15:45:41 +08:00
train.py Update train.py 2025-10-27 15:47:22 +08:00
trainer.py Add files via upload 2025-10-27 15:04:28 +08:00
utils1.py Add files via upload 2025-10-27 15:04:28 +08:00

PKDiff-for-Multimodal-Remote-Sensing-Segmentation

The code is currently being organized and supplemented.

# Multimodal Remote Sensing Segmentation (DDP) Training
Based on PyTorch DistributedDataParallel (DDP), this project provides scripts for multi-GPU training and evaluation,
supporting single-machine multi-GPU and multi-machine multi-GPU setups, Linear Warmup → Cosine learning rate scheduling,
parameter-grouped learning rates, distributed evaluation aggregation, and NVML GPU information printing.
## Project Structure

data_root/
├── tmp                       
    ├──Vaihingen
    ├──Potsdam

repo_root/
├── train.py                       # Training entry point
├── networks/network.py            # NetworkConfig, Network
├── trainer.py                     # Trainer, TrainerConfig
├── utils1.py                      # Data paths/constants (train_ids/test_ids/BATCH_SIZE/...)
├── utils/evaluation.py            # Evaluator, denoise, metrics, accuracy
├── eval_patch_dataset.py          # EvalPatchDataset
└── requirements.txt               # Dependency file

1. Create Environment

# ① Create & activate environment
conda create -n pkdiff python=3.10 -y
conda activate pkdiff
# ② Install heavy dependencies (with binaries)
conda install -c conda-forge rasterio affine -y

2. Install requirements.txt Dependencies

Run the following command in the project root directory:

pip install -r requirements.txt

3. Start Training

Single Machine Multi-GPU (e.g., 4 GPUs)

CUDA_VISIBLE_DEVICES=0,1,2,3 \
torchrun --nproc_per_node=4 train.py