[Bug] Inverted lower/upper frequency signs in the generated HDF5 when `Spectrogram` is used as a transform

[tapyu]

Version

2.1.1 (Default)

System Information

❯ fastfetch --logo none
tapyu@felix-Alienware-m16-R1
----------------------------
OS: Ubuntu 24.04.4 LTS (Noble Numbat) x86_64
Host: Alienware m16 R1
Kernel: Linux 6.17.0-35-generic
Uptime: 22 hours, 43 mins
Packages: 2604 (dpkg), 16 (snap)
Shell: zsh 5.9
Terminal: /dev/pts/2
CPU: 13th Gen Intel(R) Core(TM) i9-13900HX (32) @ 5.40 GHz
GPU 1: NVIDIA GeForce RTX 4070 Max-Q / Mobile [Discrete]
GPU 2: Intel Raptor Lake-S UHD Graphics @ 1.65 GHz [Integrated]
Memory: 5.27 GiB / 31.03 GiB (17%)
Swap: 0 B / 8.00 GiB (0%)
Disk (/): 104.52 GiB / 191.19 GiB (55%) - ext4
Disk (/home): 142.18 GiB / 287.36 GiB (49%) - ext4
Local IP (enp109s0): 161.24.19.210/24
Battery (DELL 70N2F3B): 100% [AC Connected]
Locale: en_US.UTF-8
❯ uv pip list
Package                  Version     Editable project location
------------------------ ----------- ---------------------------------
aiobotocore              2.26.0
aiohappyeyeballs         2.6.1
aiohttp                  3.13.0
aiohttp-retry            2.9.1
aioitertools             0.13.0
aiosignal                1.4.0
amqp                     5.3.1
annotated-types          0.7.0
antlr4-python3-runtime   4.9.3
appdirs                  1.4.4
astroid                  4.0.1
asyncssh                 2.21.1
atpublic                 6.0.2
attrs                    25.4.0
billiard                 4.2.2
botocore                 1.41.5
celery                   5.5.3
certifi                  2025.10.5
cffi                     2.0.0
charset-normalizer       3.4.4
click                    8.3.0
click-didyoumean         0.3.1
click-plugins            1.1.1.2
click-repl               0.3.0
code-anatel-axis4        0.1.0       /home/tapyu/git/code-anatel-axis4
colorama                 0.4.6
configobj                5.0.9
contourpy                1.3.3
coverage                 7.11.0
cryptography             46.0.3
cycler                   0.12.1
dictdiffer               0.9.0
dill                     0.4.0
diskcache                5.6.3
distro                   1.9.0
dpath                    2.2.0
dulwich                  0.24.5
dvc                      3.63.0
dvc-data                 3.16.12
dvc-http                 2.32.0
dvc-objects              5.1.2
dvc-render               1.0.2
dvc-s3                   3.3.0
dvc-studio-client        0.22.0
dvc-task                 0.40.2
engineering-notation     0.13.2
entrypoints              0.4
filelock                 3.20.0
flatten-dict             0.4.2
flufl-lock               8.2.0
fonttools                4.60.1
frozenlist               1.8.0
fsspec                   2025.9.0
funcy                    2.0
gitdb                    4.0.12
gitpython                3.1.45
grandalf                 0.8
gto                      1.9.0
h5py                     3.15.0
hf-xet                   1.1.10
huggingface-hub          0.35.3
hydra-core               1.3.2
idna                     3.11
iniconfig                2.1.0
isort                    7.0.0
iterative-telemetry      0.0.10
jinja2                   3.1.6
jmespath                 1.1.0
joblib                   1.5.2
kiwisolver               1.4.9
kombu                    5.5.4
lightning                2.6.5
lightning-utilities      0.15.2
loguru                   0.7.3
markdown-it-py           4.0.0
markupsafe               3.0.3
matplotlib               3.10.7
mccabe                   0.7.0
mdurl                    0.1.2
mpmath                   1.3.0
multidict                6.7.0
networkx                 3.5
numpy                    2.4.6
nvidia-cublas-cu12       12.8.4.1
nvidia-cuda-cupti-cu12   12.8.90
nvidia-cuda-nvrtc-cu12   12.8.93
nvidia-cuda-runtime-cu12 12.8.90
nvidia-cudnn-cu12        9.10.2.21
nvidia-cufft-cu12        11.3.3.83
nvidia-cufile-cu12       1.13.1.3
nvidia-curand-cu12       10.3.9.90
nvidia-cusolver-cu12     11.7.3.90
nvidia-cusparse-cu12     12.5.8.93
nvidia-cusparselt-cu12   0.7.1
nvidia-nccl-cu12         2.27.5
nvidia-nvjitlink-cu12    12.8.93
nvidia-nvshmem-cu12      3.3.20
nvidia-nvtx-cu12         12.8.90
omegaconf                2.3.0
opencv-python            4.13.0.92
orjson                   3.11.3
packaging                25.0
pathspec                 0.12.1
pillow                   12.0.0
platformdirs             4.5.0
pluggy                   1.6.0
prompt-toolkit           3.0.52
propcache                0.4.1
psutil                   7.1.0
pycparser                2.23
pydantic                 2.12.3
pydantic-core            2.41.4
pydantic-settings        2.11.0
pydot                    4.0.1
pygit2                   1.18.2
pygments                 2.19.2
pygtrie                  2.5.0
pylint                   4.0.1
pyparsing                3.2.5
pytest                   8.4.2
pytest-cov               7.0.0
python-dateutil          2.9.0.post0
python-dotenv            1.1.1
pytorch-lightning        2.5.5
pyyaml                   6.0.3
requests                 2.32.5
rich                     14.2.0
ruamel-yaml              0.18.15
ruamel-yaml-clib         0.2.14
ruff                     0.15.17
s3fs                     2025.9.0
safetensors              0.6.2
scikit-learn             1.7.2
scipy                    1.16.2
scmrepo                  3.5.2
semver                   3.0.4
setuptools               80.9.0
shellingham              1.5.4
shortuuid                1.0.13
shtab                    1.7.2
six                      1.17.0
smmap                    5.0.2
sqltrie                  0.11.2
sympy                    1.14.0
tabulate                 0.9.0
threadpoolctl            3.6.0
timm                     1.0.20
tomlkit                  0.13.3
torch                    2.9.0
torchaudio               2.9.0
torchmetrics             1.8.2
torchsig                 2.1.1
torchvision              0.24.0
tqdm                     4.67.1
triton                   3.5.0
typer                    0.19.2
typing-extensions        4.15.0
typing-inspection        0.4.2
tzdata                   2025.2
urllib3                  2.5.0
vine                     5.1.0
voluptuous               0.15.2
wcwidth                  0.2.14
wrapt                    1.17.3
yarl                     1.22.0
zc-lockfile              4.0

Description

From the Torchsig-generated HDF5 file:

Some instancies in the generated HDF5 dataset have their lower and upper frequencies with flipped sign.

How to Reproduce the Bug

Any command that uses a TorchSigIterableDataset object, e.g.:

iter_dataset = TorchSigIterableDataset(
    signal_generators=signal_generators,
    metadata=dataset_metadata,
    transforms=[impairments.dataset_transforms, Spectrogram(fft_size=cfg.data.ex.trace.n_samples)],
    component_transforms=[impairments.signal_transforms],
    target_labels=cfg.data.target,
    seed=cfg.data.seed,
)

dataloader = WorkerSeedingDataLoader(
    iter_dataset,
    batch_size=cfg.process.batch_size,
    num_workers=cfg.process.n_workers,
    collate_fn=default_collate_fn,
)

dataset_creator = DatasetCreator(
    dataset_length=cfg.data.ex.n,
    dataloader=dataloader,
    root=data_prefix,
    overwrite=cfg.process.is_overwrite,
    multithreading=cfg.process.is_multithreading,
)
dataset_creator.create()

An MWE can be generated and one can see that happening for some spectrograms

[tapyu]

It follows a self-contained MWE

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()

You can see that my # HACK code block fixes the TorchSig's bug