[Feature] Add “family-uniform” class_distribution for data generation

[romanziske]

Is there an existing feature already?

• Yes, I have checked the existing features.

Description

When training family-level classifiers, the current configuration only supports per-class (signal-level) distributions. This leads to heavy oversampling of large families (e.g. FSK, OFDM) and undersampling of small ones (e.g. OOK, FM), skewing both training and evaluation. It would be very useful to introduce a “family_uniform” mode that automatically assigns equal total probability to each family, regardless of how many individual signals it contains.

category	#instances	category	#instances	category	#instances
am	2892	ask	3420	chirp	2180
fm	790	fsk	10944	ofdm	8468
ook	748	psk	4036	qam	5718
tone	742
total	39938

I currently use follwoing work around:

metadata["overrides"]["class_list"] = [
        "ook",
        "4ask", "8ask", "16ask", "32ask", "64ask",
        "2fsk", "2gfsk", "2msk", "2gmsk",
        "4fsk", "4gfsk", "4msk", "4gmsk",
        "8fsk", "8gfsk", "8msk", "8gmsk",
        "16fsk", "16gfsk", "16msk", "16gmsk",
        "bpsk", "qpsk", "8psk", "16psk", "32psk", "64psk",
        "16qam", "32qam", "32qam_cross", "64qam", "128qam_cross",
        "256qam", "512qam_cross", "1024qam",
        "ofdm-64", "ofdm-72", "ofdm-128", "ofdm-180",
        "ofdm-256", "ofdm-300", "ofdm-512", "ofdm-600",
        "ofdm-900", "ofdm-1024", "ofdm-1200", "ofdm-2048",
        "fm",
        "am-dsb-sc", "am-dsb", "am-lsb", "am-usb",
        "lfm_data", "lfm_radar", "chirpss",
        "tone",
    ]

    metadata["overrides"]["class_distribution"] = [
        # ook (1 signal ⇒ 0.1)
        0.1,
        # ask (5 signals ⇒ each 0.02)
        0.02, 0.02, 0.02, 0.02, 0.02,
        # fsk (16 signals ⇒ each 1/(10*16)=0.00625)
        *[0.00625]*16,
        # psk (6 signals ⇒ each 1/(10*6)=0.016666666666666666)
        *[0.016666666666666666]*6,
        # qam (8 signals ⇒ each 1/(10*8)=0.0125)
        *[0.0125]*8,
        # ofdm (12 signals ⇒ each 1/(10*12)=0.008333333333333333)
        *[0.008333333333333333]*12,
        # fm (1 ⇒ 0.1)
        0.1,
        # am (4 ⇒ each 1/(10*4)=0.025)
        *[0.025]*4,
        # chirp (3 ⇒ each 1/(10*3)=0.03333333333333333)
        *[0.03333333333333333]*3,
        # tone (1 ⇒ 0.1)
        0.1,
    ]

Now each family is equally sampled:

category	#instances	category	#instances	category	#instances
am	3796	ask	4098	chirp	4248
fm	4026	fsk	4140	ofdm	3744
ook	3864	psk	4080	qam	4216
tone	4026
total	40238

[MattCarrickPL]

This is a good idea, and I appreciate you posting your yaml for the work around.

My thought this is best implemented in some pre-made selection in a GUI such that the user doesn't have to hand code the distribution into the yaml. I'm tagging this with the GUI issue (#283) such that they can be completed together

[ereoh]

Update: we are restructuring our SignalBuilder for v2.1.0 and planning for this feature be available upon release.