[Feature] Generated Spectrogram lacks support vector for x and y axis

[tapyu]

Is there an existing feature already?

• Yes, I have checked the existing features.

Description

It follows an MWE that generates a spectrogram dataset from TorchSig pipeline

import h5py
from matplotlib.ticker import EngFormatter, MultipleLocator
from torchsig.datasets.datasets import TorchSigIterableDataset
from torchsig.utils.data_loading import WorkerSeedingDataLoader
from torchsig.signals.builders.fm import FMSignalGenerator
from torchsig.utils.writer import DatasetCreator, default_collate_fn
from torchsig.transforms.transforms import Spectrogram
from torchsig.transforms.impairments import Impairments

import matplotlib.pyplot as plt
from matplotlib.patches import Rectangle

import numpy as np

dataset_metadata = {
    "num_iq_samples_dataset": 108_032,
    "fft_size": 512,
    "num_signals_min": 1,
    "num_signals_max": 1,
    "sample_rate": 20e6,
    "noise_power_db": 0,
    "snr_db_min": 50,
    "snr_db_max": 50,
    "signal_duration_min": 0.00378112,
    "signal_duration_max": 0.00378112,
    "bandwidth_min": 500e3,
    "bandwidth_max": 1e6,
    "signal_duration_in_samples_min": 75622,
    "signal_duration_in_samples_max": 75622,
    "frequency_min": -10e6,
    "frequency_max": 10e6,
    "signal_center_freq_min": -9e6,
    "signal_center_freq_max": 9e6,
    "class_list": "fm",
    "class_distribution": "1",
    "fft_stride": 512,
    "cochannel_overlap_probability": 0,
}

samp_period = 1 / dataset_metadata["sample_rate"]
rf_min = 90e6
fc = 100e6
rf_max = 110e6
ex_duration = 0.0054016


def get_dataset():

    impairments = Impairments(level=0) # noiseless channel

    
    iter_dataset = TorchSigIterableDataset(
        signal_generators=[FMSignalGenerator(**dataset_metadata)], # NOTE: added in next step
        metadata=dataset_metadata,
        transforms=[impairments.dataset_transforms, Spectrogram(fft_size=512)],
        component_transforms=[impairments.signal_transforms],
        target_labels=['class_name', 'start_in_samples', 'duration_in_samples', 'lower_freq', 'upper_freq', 'snr_db'],
        seed=14,
    )

    dataloader = WorkerSeedingDataLoader(
        iter_dataset,
        batch_size=8,
        num_workers=10,
        collate_fn=default_collate_fn,
    )

    dataset_creator = DatasetCreator(
        dataset_length=200,
        dataloader=dataloader,
        root='dataset',
        overwrite=True,
        multithreading=True,
    )
    dataset_creator.create()
    print(f"Dataset created at {dataset_creator.root}")

def plot_spec_bboxes():
    
    with h5py.File("dataset/data.h5", "r") as f_raw:

        example_idxs_raw = sorted(f_raw['data'].keys(), key=lambda value: int(value))
        
        # Plot and save the spectrogram with physical axes (time, frequency)
        for i in range(0, len(f_raw['data'].keys()), 2):
            spectrogram = f_raw['data'][str(i)][()]
            spectrogram = np.flip(spectrogram, axis=0) # flip spectrogram vertically to match the RF axis orientation (lowest matrix rows should correspond to the highest RF frequencies)
            emitter = f_raw["metadata"][str(i+1)]
            plt.figure(figsize=(10, 4))
            plt.imshow(
                spectrogram,
                aspect='auto',
                origin='lower',
                extent=[0, ex_duration, rf_min, rf_max],
            )
            # add metadata
            ax = plt.gca()
            mod_name = emitter['class_name'][()].decode("utf-8").lower()
            snr_db = emitter['snr_db'][()]
            
            t_start_samp = emitter['start_in_samples'][()]
            t_duration_samp = emitter['duration_in_samples'][()]

            t_start_s = t_start_samp * samp_period
            t_duration_s = t_duration_samp * samp_period
            
            bb_start_freq_hz = emitter['_lower_frequency'][()]
            bb_end_freq_hz = emitter['_upper_frequency'][()]
            bw_hz = emitter['bandwidth'][()]

            center_freq_hz = emitter['center_freq'][()]

            # HACK: this solves a bug in TorchSig where the sign of the baseband frequencies is flipped for some examples for unknown reasons
            if (
                center_freq_hz < 0 and (bb_start_freq_hz > 0 or bb_end_freq_hz > 0)
                or
                center_freq_hz > 0 and (bb_start_freq_hz < 0 or bb_end_freq_hz < 0)
            ):
                bb_start_freq_hz, bb_end_freq_hz = -1*bb_end_freq_hz, -1*bb_start_freq_hz


            # Draw bounding box (true)
            rf_start_hz = fc + bb_start_freq_hz
            rect = Rectangle(
                (t_start_s, rf_start_hz), t_duration_s, bw_hz,
                linewidth=2, edgecolor='red', facecolor='none'
            )
            ax.add_patch(rect)
            
            # format axes
            plt.colorbar(label='Intensity [dB]')
            ax.xaxis.set_major_formatter(EngFormatter(unit='s', sep=' '))
            ax.yaxis.set_major_locator(MultipleLocator(10e6))
            ax.yaxis.set_major_formatter(EngFormatter(unit='Hz', sep=' '))
            plt.title('Spectrogram Example')
            plt.xlabel('Time')
            plt.ylabel('Frequency')
            # Save the figure
            plt.savefig(f"spectrogram_{i}.png", bbox_inches='tight')


def main():
    get_dataset()
    plot_spec_bboxes()


if __name__ == "__main__":
    main()

This a generated spectrogram.

We have neither information about the x/y axes nor the support vector of them. Is x axis in time? frequency? What are their values? The dataset metadata on the axis are not self-contained, we need the source code to recover the support vector of the spectrogram. It should be included in the generated HDF5 file.