SV-CoDe

PyTorch implementation of nine-view image demixing and 2D reconstruction using coordinate-gated networks with positional encoding.

Setup

Use an NVIDIA GPU with CUDA-enabled PyTorch. Run commands from the repository root.

conda create -n SV-CoDe python=3.11 -y
conda activate SV-CoDe

Install matching torch and torchvision packages using the PyTorch installation selector, then:

python -m pip install numpy pandas tifffile pytorch-msssim tensorboardX
python -c "import torch; print('CUDA available:', torch.cuda.is_available())"

Dependency versions are not pinned; this setup has not been validated end to end. Evaluation currently requires CUDA and does not support CPU or Apple MPS.

Data

Supply a directory containing matching TIFF files for each sample:

File	Contents / shape
meas_N.tif	2D measurement containing nine lens views
demix_N.tif	Ground-truth views: (9, 2400, 2400) or (2400, 2400, 9)
gt_N.tif	Reconstruction target: (2400, 2400)

Training selects indices 1, 4, 7, ..., with dataset length equal to the number of measurement files divided by three (rounded down). Evaluation reads indices 1 through --num_samples; all three files are required for each index. Data is supplied separately. Lens centers in train_amp.py and result.py are configured for the v16/v17 geometry; update both for other acquisitions.

Training

Edit the args.dir_data path in the scc branch of train_amp.py first. Training does not accept a --dir_data option.

python train_amp.py --task_name SV_CoDe --computer scc

Defaults: 35 epochs, batch size 3, learning rate 1e-4, BF16 mixed precision, and MSE + MS-SSIM loss. Use --precision fp32 or --precision amp_fp16 if needed. Checkpoints, logs, and validation previews are saved under train/SV_CoDe/.

Evaluation

Provide the checkpoint and replace the test data path:

python result.py --task_name SV_CoDe --epoch 35 \
  --checkpoint checkpoints/model_epoch0035.pth \
  --dir_data /path/to/test_data --num_samples 300

Use --num_samples 1 for an initial check. Reconstructed and demixed TIFFs plus metrics.csv (PSNR, SSIM, and timing) are saved to test/simulation/SV_CoDe_35/. Use python train_amp.py --help or python result.py --help for available options.

Cluster jobs

The qsub scripts activate SV-CoDe. Adjust their project, resource requests, and data paths for your cluster, then submit from the repository root:

qsub -cwd train_SV_CoDe.qsub
qsub -cwd test_SV_Code.qsub

These jobs are independent. The test job uses the explicit epoch-35 checkpoint. The current .gitignore excludes *.qsub, so include these scripts separately when sharing the materials.

Code

• model.py: network architecture and Fourier loss.
• dataset.py: TIFF loading, cropping, and patch extraction.
• train_amp.py / result.py: training and evaluation.
• utils.py: argument logging and checkpoint saving.