Transfer Learning for RF Domain Adaptation – Synthetic Dataset

Citation Author(s):
Submitted by:
Lauren Wong
Last updated:
Thu, 03/24/2022 - 18:21
0 ratings - Please login to submit your rating.


A synthetic dataset designed to evaluate transfer learning performance for RF domain adaptation in the publication Assessing the Value of Transfer Learning Metrics for RF Domain Adaptation. The dataset contains a total of 13.8 million examples, with 600k examples each of 22 modulation schemes (given below) and AWGN noise (200k each for training, validation, and testing); 512 raw IQ samples per example. For each example, the signal-to-noise ratio is uniformly selected from the range [-10dB, 20dB], and the frequency offset is uniformly selected from the range [-10%, 10%] of sample rate. Further details can be found in the publication.


Modulation schemes included:

·      PSK of order 2, 4, 8, and 16

·      DPSK of order 4

·      QAM of order 16, 32, and 64

·      APSK of order 16 and 32

·      FSK with 5k and 75k carrier spacing

·      GFSK with 5k and 75k carrier spacing

·      MSK

·      GMSK

·      Narrowband and wideband FM

·      Double sideband, double sideband suppressed carrier, lower sideband, and upper sideband AM



This dataset was generated using Python wrappers around liquid-dsp (, and is saved in SigMF format such that each example is saved in an individual ‘.sigmf-data’ file with an associated ‘.sigmf-meta’ file  of the same name. The ‘.sigmf-data’ file contains the interleaved raw IQ samples in binary format and can be read using the numpy.load() function. The ‘.sigmf-meta’ file contains all metadata parameters used to generate the example including the number of samples, modulation type, signal-to-noise ratio, frequency offset, and filtering parameters, is in json format ,and can be read using json.load(). Further details and code examples for loading the dataset can be found at