How commreteris/nind-denoise works
Project Type and Purpose
• This is a command-line image-denoising tool (nind-denoise) built on PyTorch.
• It wraps a workflow combining Darktable CLI + gmic for preprocessing and an RL-deblur Lua script, then passes the result through a neural network (Unet/UtNet variants) for denoising.
• It also includes training utilities (dataset builders, trainers, config files) to reproduce or fine-tune models.
File-Structure Analysis
• Top-level:
• denoise.py — CLI entry point for inference
• pyproject.toml, requirements.in — Python build/config files
• src/config/operations.yaml — high-level pipeline definition
• src/nind_denoise/:
• common/libs/… — shared helper modules (I/O, JSON, image ops, PyTorch utilities)
• configs/*.yaml — default & per-experiment training/validation configs
• datasets & dataset_torch_3.py — CSV-based dataset loader for training/validation
• denoise_dir.py, denoise_image.py — batched & single-image inference drivers
• loss.py, nnModules.py, nn_common.py, graph_utils.py — training/inference logic
• networks/… — neural-network architectures (Unet, UtNet, third-party nets)
• nn_train.py — training orchestrator (loads config, dataset, model, runs training loop)
• tools/… — scripts to download/crop datasets, pick validation set, graph results
• models/… — timestamped model checkpoints + training metadata
• src/lua-scripts/ — RL-deblur implementation in Lua, invoked via gmic or Darktable CLI
README Insights
• Describes end-user workflow: python3 src/denoise.py <raw_image> [options]
• External dependencies: Darktable CLI, gmic, Lua script, GPU/CPU acceleration via torch.accelerator
• Automatic download of pretrained models from a Backblaze B2 bucket
• Training requirements and dataset generation steps
Detail Is Key
• Every major step: preprocessing scripts, config load, model fetch, inference call, output save
• In training flow, show where common/libs utilities plug in (loss calculation, graphing)
• Don’t lump “all code” into a single monolith—split by directory/function (common, networks, tools, trainer)