Spectrum Prediction

Contributions

1. Implemented multiple models for spectrum prediction ('TCN', 'informer', 'MLPMixer', 'rnn', 'lstm').
2. Calculated the params and latency of each model (params, latency).
3. Use automatic mixed precision training (torch.cuda.amp) to reduce the amount of model parameters and inference time while reduce part of the accuracy.
4. Simulated the attack behavior of the noise radio on the model prediction in the real complex environment (model attack).

Results of Expriments

in32out128, lr1e-2, bs=32, train_epochs=40, GTX3090

Methods	Structure	MSE	ocpy_acc	false_alarm	missing_alarm	avarage inference time	avarage training time per epoch	Training maximum memory usage	Params	avarage inference time with amp
MLPMixer	2 blocks + 1 fc	0.7032	0.6944	0.0782	0.2273	0.691ms	17.05s	2579MiB	65M	0.912ms
RNN	2 layers + 1 fc	0.7232	0.6876	0.0892	0.2232	2.424ms	18.73s	2529MiB	65M	2.117ms
LSTM	2 layers + 1 fc	0.7787	0.6826	0.0211	0.2963	6.566ms	19.92s	3595MiB	257M	4.425ms
Informer	2 en + 1 dec	0.7018	0.6938	0.1220	0.1841	5.624ms	31.26s	2905MiB	91M	6.071ms
TCN	2 layers + 1 fc	0.7098	0.6904	0.0630	0.2466	3.344ms	19.48s	3221MiB	209M	3.453ms

perturbation

Methods	Structure	ocpy_acc	false_alarm	missing_alarm
MLPMixer	2 blocks + 1 fc	0.6945	0.0770	0.2284
RNN	2 layers + 1 fc	0.6838	0.0000	0.3162
LSTM	2 layers + 1 fc	0.6826	0.0203	0.2972
Informer	2 en + 1 dec	0.6078	0.3045	0.0876
TCN	2 layers + 1 fc	0.6900	0.0679	0.2421

in32out128, lr1e-2, bs=32, train_epochs=40, GTX3090, use_amp

Methods	Structure	MSE	ocpy_acc	false_alarm	missing_alarm
MLPMixer	2 blocks + 1 fc	0.7033	0.6944	0.0784	0.2272
RNN	2 layers + 1 fc	0.7665	0.6721	0.1157	0.2121
LSTM	2 layers + 1 fc	0.8939	0.6838	0.0000	0.3162
Informer	2 en + 1 dec	0.6939	0.6984	0.1075	0.1941
TCN	2 layers + 1 fc	0.7075	0.6920	0.0636	0.2444