How sreyas-mohan/deepfreq works
Project Type and Purpose
• This is a research‐code repository for “Data‐driven Estimation of Sinusoid Frequencies.” It provides data‐generation, model‐training (DeepFreq), baseline comparison (PSnet, CBLasso), pretrained networks, and evaluation scripts/notebooks.
File‐Structure Analysis
Top‐level scripts:
• generate_dataset.py – creates synthetic test data
• train.py – trains the DeepFreq networks from scratch
• test.py – runs benchmarks (DeepFreq vs. PSnet vs. CBLasso)
• example_notebook.ipynb – demo of pretrained‐model inference & visualization
• modules.py + util.py – core PyTorch model definitions and helpers
Packages:
• cblasso/ – MATLAB‐based CBLasso interface (cbl_via_matlab.py + .m files)
• data/ – dataset classes, data‐loading logic, noise/loss functions
Resources:
• test_dataset/ – .npy test signals & precomputed CBLasso results
• pretrained_models/ – DeepFreq and PSnet model checkpoints
• figures/, example_notebook.html – visualization output
Config & deps:
• requirements.txt, _config.yml, .gitignore
Key Components & Their Relationships
a. Data Generation
– generate_dataset.py outputs noisy sinusoids → saved in data/ or custom test_dataset/
b. Data Module
– data/dataset.py, data/data.py, data/noise.py, etc.
– Loads .npy signals, applies noise, constructs PyTorch DataLoader
c. Model Module
– modules.py defines DeepFreq frequency‐representation (FR) & frequency‐counting (FC) architectures
– util.py contains shared utilities (e.g., loss functions in data/loss.py)
d. Training Pipeline (train.py)
– Uses Data Module + Model Module
– Produces FR & FC model checkpoints → saved under output_dir
e. Pretrained Models
– pretrained_models/DeepFreq & /PSnet contain .pth files for FR, FC, PSnet baseline
f. Inference & Benchmarking (test.py + example_notebook.ipynb)
– Loads test_dataset/ + pretrained_models/
– Runs:
• DeepFreq inference: FR → FC → predicted frequencies
• PSnet inference: PSnet modules
• CBLasso baseline via cblasso/cbl_via_matlab.py + MATLAB/CVX
– Aggregates metrics (Chambers, FNR) → outputs to results/ or figures/
g. Visualization
– example_notebook shows signal plots, error curves
– generate_dataset.py + figures/ for static images
Architectural Patterns & Design Principles
• Modular separation:
• Data generation vs. data‐loading vs. modeling vs. evaluation
• Pluggable baselines (DeepFreq, PSnet, CBLasso) via uniform interface
• Pipeline stages:
1. Data generation → 2. Data loading → 3. Training → 4. Model artifacts → 5. Inference → 6. Benchmarking & visualization
• Language interoperability:
• Python ↔ MATLAB for CBLasso baseline