Deploy Your RF Classifier Without Fear — Introducing Checkpoint Mismatch Tolerance
By Ben Gilbert | November 9, 2025
You trained a perfect RF modulation classifier on [‘AM’, ‘FM’, ‘PSK’]. Now your customer wants [‘AM-DSB’, ‘FM’, ‘QPSK’]. Do you retrain? No. You just deploy — and it just works.
TA Drop-In Loader That Recovers
| Training | Deployment |
|---|---|
AM, FM, PSK | AM-DSB, FM, QPSK |
BPSK, QPSK | BPSK, 8PSK |
CW, FSK | GFSK |
model, M = load_from_checkpoint(
'model.pt',
runtime_classes=['AM-DSB', 'FM', 'QPSK'],
strategy='auto'
)What happens under the hood?
- Name Remap → AM → AM-DSB, FM → FM
- Hungarian Calibration → PSK → QPSK (using 50 labeled samples)
- Output Mapping → p_deployed = M @ p_checkpoint
No retraining. No downtime. < 10 ms.
How It Works: The Fallback Chain
In Other News:
NSIS Examples >> https://chatgpt.com/share/69110307-76f4-800d-89da-96f8d69ab4db