How bu-cisl/sv-fouriernet works
Project type and purpose
• SV-FourierNet is a research-focused, Python-based deep-learning application for computational microscopy.
• It implements a Fourier Neural Network to reconstruct wide-field, high-resolution images from multi-aperture miniature microscope measurements.
• Key use cases: training on physics-based simulated data, running inference (reconstruction) on test captures, and visualizing results.
File structure
• Root files:
• README.md – project description, dependencies, usage
• dataset.py – data loader / physics-based simulator interface
• model.py – SV-FourierNet architecture definition
• train.py – training loop, loss, optimizer, checkpointing, TensorBoardX logging
• test.py – inference script, loads model + pre-trained weights, runs reconstruction on test images
• utils.py – helper routines (e.g., metrics, image I/O, pre-/post-processing)
• Directories:
• test/ – sample measurement input (meas.tif) and possibly other test data
README for additional insights
• Architecture: novel multi-channel Fourier neural network with learned global receptive field
• Workflow:
a. Physics simulator generates training pairs
b. train.py trains model.py, logs via TensorBoardX
c. test.py loads pretrained weights, processes test/meas.tif, produces reconstructions
• Dependencies: Python 3.10, PyTorch 2.0, TensorBoardX, numpy, scikit-image, tifffile, pandas, pytorch_msssim
• External asset: Google Drive link for downloading pretrained weights
Main components and interactions
a. Data Preparation / Simulator
• Block: “Physics-based Simulator / dataset.py”
• Outputs synthetic measurement–ground-truth pairs
b. Training Pipeline
• Block: “train.py”
• Consumes simulator output, invokes model.py, computes losses, writes checkpoints & logs
• Connects to TensorBoardX for logging
• Saves weights (e.g., .pt) to root directory
c. Model Definition
• Block: “SV-FourierNet (model.py)”
• Shows internal modules (e.g., Fourier layers, global receptive field) if desired
d. Utilities
• Block: “utils.py”
• Shared services: data augmentation, PSNR/SSIM metrics, file I/O
e. Inference Pipeline
• Block: “test.py”
• Loads pretrained weights, reads test/meas.tif, runs forward pass, writes reconstructed images
f. External Storage & Visualization
• Block: “Google Drive (pretrained weights)”
• Block: “images/” for static assets and output snapshots
• Block: “TensorBoardX” for visualizing training curves
Architectural patterns and design principles
• Modular separation: data loader, model, training script, inference script, utilities
• Single-entry point scripts (train.py, test.py) controlling workflows
• Checkpointing and experiment tracking (TensorBoardX) for reproducibility
• Loose coupling via standard file formats (TIFF, PyTorch .pt, numpy arrays)
Detailed element checklist
• Module blocks: dataset.py, model.py, utils.py
• Workflow blocks: train.py, test.py
• Storage blocks: checkpoint files, test/meas.tif, images/ outputs
• External service blocks: TensorBoardX, Google Drive
• Simulator output → Training script
• Training script ↔ Model definition
• Training script → Checkpoint storage
• Inference script → Checkpoint loading
• Inference script → Output images
• Annotations for technologies: Python 3.10, PyTorch 2.0, numpy, etc.