This repo provides the pipeline for working with RF datasets, labeling them and training both IQ and spectrogram based models. The SigMF standard is used for managing RF data and the labels/annotations on the data. It also uses the Torchsig framework for performing RF related augmentation of the data to help make the trained models more robust and functional in the real world.
Verify install with GPU support
Follow the instructions here to install Poetry: https://python-poetry.org/docs/#installation
https://github.com/miek/inspectrum
This utility is useful for inspecting sigmf files and the annotations that the auto label scripts make.
https://github.com/vietanhdev/anylabeling
This program is used for image annotation and offers AI-assisted labelling.
This project uses Poetry for dependency management and packaging. Poetry can be used with external virtual environments. If using a non-Poetry virtual environment, start by activating the environment before running Poetry commands. See note in Poetry docs for more info.
To activate the Poetry virtual environment with all of the Python modules configured, run the following:
poetry shellSee Poetry docs for more information.
git clone https://github.com/IQTLabs/rfml.git
cd rfml
git submodule update --init --recursive
poetry install$ python -c 'import torch; print(torch.cuda.is_available())'
TrueIf the output does not match or errors occur, try installing Pytorch manually (current version or previous versions).
pip install torch==2.0.1 torchvision==0.15.2Our current approach is to capture examples of signals of interest to create labeled datasets. There are many methods for doing this and many challenges to consider. One practical method for accomplishing this is to isolate signals of interest and compare those to a specific background RF environment. For simplicity we apply the same label to all the signals present in the background environment samples. We use this to essentially teach the model to ignore those signals. For this to work, it is important that the signals of interest are isolated from the background RF environment. Since it is really tough these days to find an RF free environment, we have build a mini-faraday cage enclosure by lining the inside of a pelican case with foil. There are lots of instructions, like this one, available online if you want to build your own. With this, the signal will be very strong, so make sure you adjust the SDR's gain appropriately.
The scripts in label_scripts use signal processing to automatically label IQ data. The scripts looks at the signal power to detect when there is a signal present in the IQ data and estimate the occupied bandwidth of the signal.
In the labeling scripts, the settings for autolabeling need to be tuned for the type of signals that were collected.
annotation_utils.annotate(
rfml.data.Data(filename),
avg_window_len=256, # The window size to use when averaging signal power
power_estimate_duration=0.1, # Process the file in chunks of power_estimate_duration seconds
debug_duration=0.25, # If debug==True, then plot debug_duration seconds of data in debug plots
debug=False, # Set True to enable debugging plots
verbose=False, # Set True to eanble verbose messages
dry_run=False, # Set True to disable annotations being written to SigMF-Meta file.
bandwidth_estimation=True, # If set to True, will estimate signal bandwidth using Gaussian Mixture Models. If set to a float will estimate signal bandwidth using spectral thresholding.
force_threshold_db=None, # Used to manually set the threshold used for detecting a signal and creating an annotation. If None, then the automatic threshold calculation will be used instead.
overwrite=True, # If True, any existing annotations in the .sigmf-meta file will be removed
max_annotations=None, # If set, limits the number of annotations to add.
dc_block=None, # De-emphasize the DC spike when trying to calculate the frequencies for a signal
time_start_stop=None, # Sets the start/stop time for annotating the recording (must be tuple or list of length 2).
n_components = None, # Sets the number of mixture components to use when calculating signal detection threshold. If not set, then automatically calculated from labels.
n_init=1, # Number of initializations to use in Gaussian Mixture Method. Increasing this number can significantly increase run time.
fft_len=256, # FFT length used in calculating bandwidth
labels = { # The labels dictionary defines the annotations that the script will attempt to find.
"mavic3_video": { # The dictionary keys define the annotation labels. Only a key is necessary.
"bandwidth_limits": (8e6, None), # Optional. Set min/max bandwidth limit for a signal. If None, no min/max limit.
"annotation_length": (10000, None), # Optional. Set min/max annoation length in number of samples. If None, no min/max limit.
"annotation_seconds": (0.0001, 0.0025), # Optional. Set min/max annotation length in seconds. If None, no min/max limit.
"set_bandwidth": (-8.5e6, 9.5e6) # Optional. Ignore bandwidth estimation, set bandwidth manually. Limits are in relation to center frequency.
}
}
)
If you see annotations where harmonics or lower power, unintentional signals are getting selected, try setting the force_threshold_db. The automatic threshold calculation maybe selecting a value that is too low. Find a value for force_threshold_db where it is selecting the intended signals and ignoring the low power ones.
If the frequency bounds are not lining up with the top or bottom part of a signal, make the spectral_energy_threshold higher. Sometime a setting as high as 0.99 is required
Some tuning is needed for signals that have a short transmission duration and/or limited bandwidth. Here are a couple things to try if they are getting skipped:
annotation_lengthsets the minimum and maximum number of samples for an annotation. If the signal does not satisfy these constraints, it will not be annotated. Try modifying this.annotation_secondsfollows the same concept asannotation_lengthexcept with units being changed to seconds instead of samples.bandwidth_limitssets the minimum and maximum bandwidth (in Hz) for a signal to be annotated. These limits will be compared against the estimated bandwidth. Try modifying these limits or the bandwidth estimation method.
After you have finished labeling your data, the next step is to train a model on it. This repo makes it easy to train both IQ and Spectrogram based models from sigmf data.
This repo provides an automated script for training and evaluating models. To do this, configure the mixed_experiments.py file or create your own experiment to point to the data you want to use and set the training parameters:
"experiment_0": { # A name to refer to the experiment
"class_list": ["mavic3_video","mavic3_remoteid","environment"], # The labels that are present in the sigmf-meta files
"train_dir": ["data/samples/mavic-30db", "data/samples/mavic-0db", "data/samples/environment"], # Directory with SigMF files
"iq_epochs": 10, # Number of epochs for IQ training, if it is 0 or None, it will be skipped
"spec_epochs": 10, # Number of epochs for spectrogram training, if it is 0 or None, it will be skipped
"notes": "DJI Mavic3 Detection" # Notes to your future self
}Once you have an experiments file configured, run it:
python3 mixed_experiments.pyOnce the training has completed, it will print out the logs location, model accuracy, and the location of the best checkpoint:
I/Q TRAINING COMPLETE
Find results in experiment_logs/experiment_1/iq_logs/08_08_2024_09_17_32
Total Accuracy: 98.10%
Best Model Checkpoint: experiment_logs/experiment_1/iq_logs/08_08_2024_09_17_32/checkpoints/checkpoint.ckptOnce you have a trained model, you need to convert it into a portable format that can easily be served by TorchServe. To do this, use export_model.py:
python3 rfml/export_model.py --model_name=drone_detect --checkpoint=experiment_logs/experiment_1/iq_logs/08_08_2024_09_17_32/checkpoints/checkpoint.ckpt --index_to_name=experiment_logs/experiment_1/iq_logs/08_08_2024_09_17_32/index_to_name.jsonThis will create a _torchscript.pt and _torchserve.pt file in the weights folder.
A .mar file will also be created in the models/ folder. GamutRF can run this model and use it to classify signals.
annotation_utils.py - DSP based automated labelling tools
auto_label.py - CV based automated labelling tools
data.py - RF data operations tool
experiment.py - Class to manage experiments
export_model.py - Convert and export model checkpoints to Torchscript/Torchserve/MAR format.
models.py - Class for I/Q models (based on TorchSig)
experiments/ - Experiment configurations and run script
sigmf_pytorch_dataset.py - PyTorch style dataset class for SigMF data (based on TorchSig)
spectrogram.py - Spectrogram tools
test_data.py - Test for data.py (might be outdated)
train_iq.py - Training script for I/Q models
train_spec.py - Training script for spectrogram models
zst_parse.py - ZST file parsing tool, for GamutRF-style filenames
The notebooks/ directory contains various experiments we have conducted during development.