Skip to content

preshanth/SAM-RFI

BranchesTags

Repository files navigation

SAM-RFI: Radio Frequency Interference Detection with SAM2

CI Python 3.10+ Code style: black Ruff License: MIT

Authors: Preshanth Jagannathan (pjaganna@nrao.edu), Srikrishna Sekhar (ssekhar@nrao.edu), Derod Deal (dealderod@gmail.com)

SAM-RFI is a Python package that applies Meta's Segment Anything Model 2 (SAM2) for Radio Frequency Interference (RFI) detection and flagging in radio astronomy data. The system processes CASA measurement sets and generates precise segmentation masks for contaminated visibilities. In order to do that it leverages the rfi_toolbox package which presents general purpose RFI simulation and measurement set handling

Overview

SAM-RFI leverages the state-of-the-art SAM2 vision transformer for RFI segmentation in radio astronomy visibility data. The package provides a complete pipeline from data generation to trained models capable of detecting and flagging RFI with superior accuracy compared to traditional statistical methods.

Key Features:

  • SAM2-based segmentation using Hiera transformer architecture
  • Physically realistic synthetic data generation with exact ground truth
  • Complete training pipeline with validation tracking
  • Iterative flagging for progressive RFI cleaning
  • GPU-accelerated training and inference
  • Command-line interface for all operations
  • Modular Python API for custom workflows

Installation

Prerequisites

  • Python 3.10, 3.11, or 3.12
  • CUDA-capable GPU (recommended for training)
  • CASA tools (optional, for measurement set operations)

Basic Installation

# Clone repository
git clone https://github.com/preshanth/SAM-RFI.git
cd SAM-RFI

# Create conda environment
conda create -n samrfi python=3.12 -y
conda activate samrfi

# Install core dependencies (CPU-only, no GPU/CASA)
pip install pandas>=2.2.0 numpy>=1.26.0 --only-binary :all:
pip install -e .

Installation Options

SAM-RFI supports modular installation based on your needs:

# GPU support (training and inference)
pip install -e .[gpu]

# CASA tools (measurement set operations)
pip install -e .[casa]

# GPU + CASA (complete functionality)
pip install -e .[gpu,casa]

# Development (all dependencies + testing tools)
pip install -e .[dev]

# Install pre-commit hooks (for development)
pre-commit install

Installation extras:

  • Core (default): Data preprocessing, synthetic data generation, evaluation metrics
  • [gpu]: PyTorch, transformers, SAM2 models (required for training/inference)
  • [casa]: CASA tools for measurement set I/O
  • [viz]: Interactive visualization tools (HoloViews, Bokeh, Datashader)
  • [dev]: All dependencies plus testing and linting tools
  • [ci]: Minimal dependencies for continuous integration

Verify Installation

# Check CLI availability
samrfi --help

# Test core imports (no GPU/CASA required)
python -c "from samrfi.data import Preprocessor; from samrfi.data_generation import SyntheticDataGenerator; print('Core installation successful')"

# Test GPU functionality (requires [gpu])
python -c "from samrfi.training import SAM2Trainer; from samrfi.inference import RFIPredictor; print('GPU installation successful')"

# Test CASA functionality (requires [casa])
python -c "from samrfi.data.ms_loader import MSLoader; print('CASA installation successful')"

Quick Start

1. Generate Synthetic Training Data

Generate physically realistic training data with exact ground truth masks:

samrfi generate-data \
  --source synthetic \
  --config configs/synthetic_train_4k.yaml \
  --output ./datasets/train_4k

Configuration (configs/synthetic_train_4k.yaml):

synthetic:
  num_samples: 4000
  num_channels: 1024
  num_times: 1024
  num_baselines: 2
  num_pols: 4

  # Physical scales (milli-Jansky and Jansky)
  noise_mjy: 1.0                 # 1 mJy Gaussian noise
  rfi_power_min: 1000.0          # 1000 Jy RFI minimum
  rfi_power_max: 10000.0         # 10000 Jy RFI maximum

  # RFI types per sample
  rfi_type_counts:
    narrowband_persistent: 2
    broadband_persistent: 1
    frequency_sweep: 1
    narrowband_bursty: 2
    broadband_bursty: 1

  # Bandpass effects
  enable_bandpass_rolloff: true
  bandpass_polynomial_order: 8
  polarization_correlation: 0.8

processing:
  patch_size: 1024
  stretch: null                  # No stretch for synthetic (preserves physical scales)
  enable_augmentation: true      # 4-way rotation augmentation
  normalize_before_stretch: false
  normalize_after_stretch: false

This generates batched datasets saved to ./datasets/train_4k/exact_masks/ with perfect ground truth masks.

2. Train SAM2 Model

SAM2 models automatically download from HuggingFace on first use. Models are cached at ~/.cache/huggingface/hub/.

samrfi train \
  --config configs/gpu_v100_training.yaml \
  --dataset ./datasets/train_4k/exact_masks \
  --validation-dataset ./datasets/val_1k/exact_masks

Training configuration (configs/gpu_v100_training.yaml):

model:
  sam_checkpoint: large          # Options: tiny, small, base_plus, large
  device: cuda

training:
  num_epochs: 10
  batch_size: 12
  learning_rate: 1.0e-5
  weight_decay: 0.0
  save_best_only: true

output:
  output_dir: ./samrfi_data
  save_plots: true

Available SAM2 models:

  • tiny (40 MB) - Fastest, lower accuracy
  • small (180 MB) - Balanced performance
  • base_plus (330 MB) - Good accuracy
  • large (850 MB) - Best accuracy, recommended for production

GPU memory requirements:

  • 11 GB VRAM: tiny, batch_size=2
  • 32 GB VRAM: base_plus, batch_size=12
  • 40+ GB VRAM: large, batch_size=8-12

3. Apply Trained Model

Single-pass prediction:

samrfi predict \
  --model ./samrfi_data/sam2_rfi_best.pth \
  --input observation.ms \
  --patch-size 1024

Iterative prediction (recommended for deep cleaning):

samrfi predict \
  --model ./samrfi_data/sam2_rfi_best.pth \
  --input observation.ms \
  --iterations 3 \
  --patch-size 1024

Iterative flagging progressively finds fainter RFI by masking already-flagged regions in each iteration. Typically converges in 2-3 passes.

Prediction options:

  • --iterations N - Number of flagging passes (default: 1)
  • --num-antennas N - Limit number of antennas loaded
  • --patch-size SIZE - Must match training patch size
  • --stretch {SQRT,LOG10,null} - Must match training configuration
  • --threshold FLOAT - Probability threshold (default: adaptive/mean)
  • --no-save - Preview only, do not write flags to MS

CLI Reference

Data Generation

# Generate synthetic training data
samrfi generate-data \
  --source synthetic \
  --config configs/synthetic_train_4k.yaml \
  --output ./datasets/train_4k

# Generate data from measurement set
samrfi generate-data \
  --source ms \
  --config configs/ms_data.yaml \
  --output ./datasets/vla_pband

Training

# Train with validation dataset
samrfi train \
  --config configs/gpu_v100_training.yaml \
  --dataset ./datasets/train_4k/exact_masks \
  --validation-dataset ./datasets/val_1k/exact_masks

# Resume training from checkpoint
samrfi train \
  --config configs/gpu_v100_training.yaml \
  --dataset ./datasets/train_4k/exact_masks \
  --resume ./samrfi_data/sam2_rfi_best.pth

Inference

# Single-pass prediction
samrfi predict \
  --model ./samrfi_data/sam2_rfi_best.pth \
  --input observation.ms

# Iterative prediction (3 passes)
samrfi predict \
  --model ./samrfi_data/sam2_rfi_best.pth \
  --input observation.ms \
  --iterations 3

Configuration

# Create default configuration
samrfi create-config \
  --type {training|data|validation} \
  --output config.yaml

# Validate configuration
samrfi validate-config --config config.yaml

Python API

Core Data Operations (No GPU/CASA Required)

from samrfi.data import Preprocessor, TorchDataset
from samrfi.data_generation import SyntheticDataGenerator
from samrfi.evaluation import compute_iou, compute_ffi

# Generate synthetic data
generator = SyntheticDataGenerator(config_path='configs/synthetic_train_4k.yaml')
dataset = generator.generate(num_samples=1000, output_dir='./datasets/synthetic')

# Preprocess data
import numpy as np
data = np.random.randn(2, 4, 1024, 1024) + 1j * np.random.randn(2, 4, 1024, 1024)
preprocessor = Preprocessor(data)
dataset = preprocessor.create_dataset(patch_size=1024, stretch=None)

# Evaluate predictions
iou = compute_iou(predicted_mask, ground_truth_mask)
ffi = compute_ffi(data, flags=predicted_mask)

Measurement Set Operations (Requires [casa])

from samrfi.data.ms_loader import MSLoader

# Load measurement set
loader = MSLoader('observation.ms')
loader.load(num_antennas=5, mode='DATA')

# Access data
data = loader.data              # Complex visibilities: (baselines, pols, channels, times)
magnitude = loader.magnitude    # Magnitude
flags = loader.load_flags()     # Existing flags

# Save new flags
loader.save_flags(predicted_flags)

Training (Requires [gpu])

from samrfi.training import SAM2Trainer
from samrfi.data import TorchDataset

# Load batched dataset
dataset = TorchDataset.from_directory('./datasets/train_4k/exact_masks')

# Create trainer
trainer = SAM2Trainer(dataset, device='cuda')

# Train model
trainer.train(
    num_epochs=10,
    batch_size=12,
    sam_checkpoint='large',
    learning_rate=1e-5,
    output_dir='./samrfi_data',
    save_best_only=True
)

Inference (Requires [gpu])

from samrfi.inference import RFIPredictor

# Load predictor with trained model
predictor = RFIPredictor(
    model_path='./samrfi_data/sam2_rfi_best.pth',
    device='cuda'
)

# Single-pass prediction
flags = predictor.predict_ms(
    ms_path='observation.ms',
    patch_size=1024,
    save_flags=True
)

# Iterative prediction (3 passes)
flags = predictor.predict_iterative(
    ms_path='observation.ms',
    num_iterations=3,
    patch_size=1024,
    save_flags=True
)

print(f"Flagged {flags.sum() / flags.size * 100:.2f}% of data")

Single Array Validation

from samrfi.inference import RFIPredictor
import numpy as np

# Load model
predictor = RFIPredictor(model_path='model.pth', device='cuda')

# Predict on arbitrary-sized array
data = np.load('baseline_data.npy')  # Any shape, e.g., (2048, 511)
flags, probabilities = predictor.predict_array(
    data,
    threshold=None,  # Adaptive (uses mean of probabilities)
    return_probabilities=True
)

# Save probabilities for custom thresholding
np.save('probabilities.npy', probabilities)

# Apply custom threshold
custom_flags = probabilities > 0.1  # More aggressive flagging

Architecture

Data Pipeline

[Measurement Set or Synthetic Generator]
        ↓
    MSLoader.load()
        ├─ Complex visibilities (baselines, pols, channels, times)
        └─ Combine spectral windows
        ↓
    Preprocessor.create_dataset()
        ├─ 4-way rotation augmentation (optional)
        ├─ Patchify into patch_size × patch_size
        ├─ Extract 3-channel features:
        │   • Channel 1: Spatial gradient (edge detection)
        │   • Channel 2: Log amplitude (intensity, [-3, 4])
        │   • Channel 3: Phase ([-π, π] → [0, 1])
        ├─ Apply optional stretch (SQRT/LOG10 for real, None for synthetic)
        └─ ImageNet normalization
        ↓
    BatchedDataset (streaming)
        ├─ Batch files: batch_*.pt + metadata.json
        ├─ On-demand loading in DataLoader workers
        └─ OS filesystem cache for efficiency

Training Pipeline

    BatchedDataset
        ↓
    SAMDataset wrapper
        ├─ Extract bounding boxes from ground truth
        ├─ Add random perturbation (±20 pixels)
        └─ Format: {pixel_values, input_boxes, ground_truth_mask}
        ↓
    SAM2Trainer.train()
        ├─ Load SAM2Model from HuggingFace
        ├─ Freeze vision + prompt encoders
        ├─ Train mask decoder only (~10% of parameters)
        ├─ Loss: DiceCELoss (Dice + Cross-Entropy)
        └─ Save: sam2_rfi_best.pth

Inference Pipeline

    [Trained Model] + [Measurement Set]
        ↓
    RFIPredictor.predict_ms() or predict_iterative()
        ↓
    MSLoader.load() → Preprocessor → Patches
        ↓
    SAM2Model.forward()
        ├─ Vision encoder: Extract features
        ├─ Prompt encoder: Encode bounding boxes
        └─ Mask decoder: Predict segmentation
        ↓
    Reconstruction
        ├─ Sigmoid(logits) > threshold
        ├─ Reverse rotations
        ├─ Combine patches → full waterfall
        └─ Boolean flags: (baselines, pols, channels, times)
        ↓
    MSLoader.save_flags() → Write to MS FLAG column

Iterative Flagging

Iterative flagging progressively discovers deeper RFI by masking known contamination in each pass, revealing fainter interference hidden beneath brighter sources.

Algorithm

Iteration 1: Raw data → Model → Flags_1 (finds bright RFI)
Iteration 2: Data with Flags_1 masked → Model → Flags_2 (finds hidden RFI)
Iteration 3: Data with Flags_1|2 masked → Model → Flags_3 (final cleanup)

Final: Flags_cumulative = Flags_1 | Flags_2 | Flags_3

Guidelines

  • Single pass (N=1): Fast, suitable for mild contamination (5-10% flagging)
  • 2-3 iterations: Recommended for deep cleaning (15-30% flagging)
  • >3 iterations: Diminishing returns, increased risk of over-flagging

Example

# Compare single vs iterative
samrfi predict --model model.pth --input obs.ms
# Output: Flagged 12.5% of data

samrfi predict --model model.pth --input obs.ms --iterations 3
# Output: Iteration 1: 12.5%, Iteration 2: 4.2%, Iteration 3: 1.1%
# Total: Flagged 17.8% of data

Synthetic Data Generation

RFI Types

The synthetic data generator produces physically realistic RFI signatures:

  1. Narrowband Persistent - Continuous narrowband signals (GPS, satellites)
  2. Broadband Persistent - Continuous wideband interference (power lines, harmonics)
  3. Narrowband Bursty - Intermittent narrowband pulses (radar, transmitters)
  4. Broadband Bursty - Transient wideband events (lightning, arcing)
  5. Frequency Sweeps - Linear and quadratic chirps (scanning radar)

Physical Parameters

  • Noise: 1 mJy (milli-Jansky) Gaussian, matches typical system noise
  • RFI Power: 1000-10000 Jy (Jansky), 10^6-10^7 dynamic range
  • Bandpass: 8th-order polynomial edge rolloff
  • Polarization: Correlated RFI across XX/YY feeds (0.8 correlation)

Ground Truth Advantage

Synthetic data provides exact ground truth masks, enabling supervised training with 100% accurate labels. This is not possible with real observations, where RFI locations are only estimates from statistical flaggers.

Preservation of Physical Scales:

processing:
  normalize_before_stretch: false  # Critical for synthetic data
  normalize_after_stretch: false
  stretch: null  # Preserves 10^6-10^7 dynamic range

Model Management

Automatic Downloads

SAM2 models are automatically downloaded from HuggingFace Hub on first use and cached locally:

from samrfi.training import SAM2Trainer

# Model downloads automatically if not cached
trainer = SAM2Trainer(dataset, device='cuda')
trainer.train(num_epochs=10, sam_checkpoint='large')

Cache location: ~/.cache/huggingface/hub/

Custom Cache Directory

export HF_HOME=/path/to/custom/cache
samrfi train --config config.yaml --dataset ./datasets/train

Pre-download Models

from samrfi.utils.model_cache import ModelCache

cache = ModelCache()
cache.download_model('large', show_progress=True)

Or via command line:

python -c "from samrfi.utils.model_cache import ModelCache; ModelCache().download_model('large')"

HuggingFace Hub Integration

SAM-RFI supports seamless integration with HuggingFace Hub for sharing and downloading trained models and datasets.

Quick Start

Download and use a published model:

# Automatically downloads model from HuggingFace Hub
samrfi predict --model polarimetic/sam-rfi/large --input observation.ms

Publish your trained model:

# Upload model to HuggingFace Hub
samrfi publish --type model \
  --input ./samrfi_data/sam2_rfi_best.pth \
  --repo-id polarimetic/sam-rfi

Publish a dataset:

# Upload training dataset
samrfi publish --type dataset \
  --input ./datasets/train_4k/exact_masks \
  --repo-id polarimetic/sam-rfi-dataset

Features

  • Automatic Model Downloads: Models are downloaded and cached on first use
  • Smart Path Detection: CLI accepts both local paths and HuggingFace repo IDs
  • Private Repositories: Support for private models with token authentication
  • Model Cards: Auto-generated documentation with training metrics
  • Latest Versioning: Simple "latest" approach per model size

Python API

from samrfi.inference import RFIPredictor

# Initialize with HuggingFace model (auto-downloads if needed)
predictor = RFIPredictor(
    model_path="polarimetic/sam-rfi/large",
    device="cuda"
)

# Use normally
flags = predictor.predict_ms("observation.ms")

Cache Management

Models are cached at ~/.cache/huggingface/hub/ after first download. Set custom location:

export HF_HOME=/path/to/custom/cache

Authentication

For private repositories, set your HuggingFace token:

export HF_TOKEN=hf_xxxxx
samrfi publish --type model --input model.pth --repo-id user/private-repo --private

Get your token from: https://huggingface.co/settings/tokens

Complete Guide

For detailed documentation including troubleshooting, batch publishing, and advanced usage, see HuggingFace Integration Guide.

Training

Resume Training

Training can be resumed from any checkpoint to continue where you left off:

# Initial training
samrfi train --config config.yaml --dataset ./datasets/train --epochs 10

# Resume and extend to 20 epochs
samrfi train --config config.yaml --dataset ./datasets/train --epochs 20 \
  --resume ./samrfi_data/sam2_rfi_best.pth

Restored state:

  • Model weights
  • Optimizer state (momentum, learning rates)
  • Training/validation loss history
  • Epoch counter

Checkpoints:

  • sam2_rfi_best.pth - Best validation loss (updated during training)
  • model_sam2-large_YYYYMMDD_HHMMSS.pth - Final checkpoint with full state

Hyperparameter Tuning

Fast iteration (debugging):

model:
  sam_checkpoint: tiny
training:
  num_epochs: 3
  batch_size: 8
  learning_rate: 1.0e-4

Production quality:

model:
  sam_checkpoint: large
training:
  num_epochs: 20
  batch_size: 4
  learning_rate: 1.0e-5
  weight_decay: 0.0

Loss Convergence

Expected behavior: Loss decreases from approximately 1.0 to below 0.3 within 10 epochs.

Troubleshooting stalled training (loss >0.8 after 5 epochs):

  • Adjust learning rate (try 5e-6 or 2e-5)
  • Verify data quality (visualize sample patches)
  • Modify batch size (try 2 or 8)
  • Check for NaN values in input data

Evaluation

Metrics Module

from samrfi.evaluation import (
    compute_iou,           # Intersection over Union
    compute_precision,     # True Positive Rate
    compute_recall,        # Sensitivity
    compute_f1,            # Harmonic mean of precision/recall
    compute_dice,          # Dice coefficient
    evaluate_segmentation  # All metrics
)

# Evaluate predictions
metrics = evaluate_segmentation(predicted_mask, ground_truth_mask)
# Returns: {'iou': 0.85, 'precision': 0.90, 'recall': 0.82, 'f1': 0.86}

Statistical Validation

from samrfi.evaluation import (
    compute_statistics,              # Before/after statistics
    compute_ffi,                     # Flagging Fidelity Index
    print_statistics_comparison      # Formatted output
)

# Compute Flagging Fidelity Index
ffi_metrics = compute_ffi(data, flags=predicted_mask)
# Returns: {'ffi': 0.65, 'mad_reduction': 0.45, 'std_reduction': 0.52}

# Print comparison
print_statistics_comparison(data, predicted_mask)

Flagging Fidelity Index (FFI): Measures flagging quality by balancing noise reduction against over-flagging penalty. Higher values indicate better flagging performance.

Development

Testing

# Run all tests
pytest tests/ -v

# Unit tests only
pytest tests/unit -v

# Integration tests
pytest tests/integration -v

# With coverage
pytest tests/ --cov=samrfi --cov-report=html

# Skip slow tests
pytest -m "not slow"

Code Quality

Pre-commit hooks are configured to run automatically on git commit:

# Install hooks
pre-commit install

# Run manually
pre-commit run --all-files

Checks performed:

  • Black (code formatting, line length 100)
  • Ruff (linting and auto-fixes)
  • isort (import sorting)
  • Trailing whitespace, EOF, YAML/JSON/TOML validation
  • Large file detection (>5MB)

Manual formatting:

# Format code
black src/ tests/ --line-length 100

# Lint code
ruff check src/ tests/ --fix

# Sort imports
isort src/ tests/ --profile black --line-length 100

Type Checking

# Type check (optional)
mypy src/ --ignore-missing-imports

Project Structure

SAM-RFI/
├── src/samrfi/
│   ├── cli.py                      # Command-line interface
│   ├── config/                     # Configuration management
│   │   ├── config_loader.py        # YAML configuration loading
│   │   └── validators.py           # Configuration validation
│   ├── data/                       # Data loading and preprocessing
│   │   ├── ms_loader.py            # CASA measurement set I/O
│   │   ├── preprocessor.py         # Waterfall to patches pipeline
│   │   ├── sam_dataset.py          # PyTorch dataset wrapper
│   │   ├── torch_dataset.py        # Batched streaming datasets
│   │   └── gpu_transforms.py       # Kornia-based GPU transforms
│   ├── data_generation/            # Dataset generators
│   │   ├── synthetic_generator.py  # Physics-based RFI simulation
│   │   └── ms_generator.py         # MS to dataset converter
│   ├── training/
│   │   └── sam2_trainer.py         # SAM2 training loop
│   ├── inference/
│   │   └── predictor.py            # RFI prediction (single/iterative)
│   ├── evaluation/                 # Metrics and validation
│   │   ├── metrics.py              # Segmentation metrics
│   │   └── statistics.py           # Flagging quality statistics
│   └── utils/                      # Utilities
│       ├── logger.py               # Logging configuration
│       ├── model_cache.py          # HuggingFace model downloads
│       └── errors.py               # Custom exceptions
│
├── tests/                          # Test suite
│   ├── unit/                       # Unit tests
│   ├── integration/                # Integration tests
│   └── conftest.py                 # Shared fixtures
│
├── configs/                        # Example configurations
│   ├── gpu_*.yaml                  # GPU-specific training configs
│   ├── synthetic_*.yaml            # Synthetic data configs
│   └── validation.yaml             # Validation config
│
├── .github/workflows/              # CI/CD
│   └── ci.yml                      # GitHub Actions workflow
│
├── pyproject.toml                  # Package definition
├── .pre-commit-config.yaml         # Pre-commit hooks
└── README.md                       # This file

Citation

A paper describing SAM-RFI is in preparation. In the meantime, if you use this software in your research, please cite the repository:

@software{samrfi2025,
  title = {SAM-RFI: Radio Frequency Interference Detection with SAM2},
  author = {Deal, Derod and Jagannathan, Preshanth},
  year = {2025},
  url = {https://github.com/preshanth/SAM-RFI}
}

Please check back for the updated citation once the paper is published.

License

MIT License - see LICENSE for details.

Acknowledgments

  • Meta AI - SAM2 architecture and pre-trained models
  • HuggingFace - Transformers library and model hosting
  • NRAO - Radio astronomy expertise and computational resources
  • NAC - National Astronomy Consortium support and funding

Support

About

This is an implementation of Segmentation based Radio Frequency Intereference (RFI) mitigation.

Topics

radio radio-astronomy rfi segment-anything

Resources

Readme

License

MIT license
Activity

Stars

10 stars

Watchers

1 watching

Forks

1 fork
Report repository

Releases 1

SAM-RFI First Release Latest
Dec 30, 2025

Packages

No packages published

Contributors 4

  •  
  •  
  •  
  •  

Languages