RML2022 Dataset

IQ modulation classification dataset with SNR in the range -18 to 20 dB.

Official links

• Dataset: https://github.com/venkateshsathya/RML22
• Paper: https://doi.org/10.1109/TWC.2023.3254490

Tasks supported

• Modulation classification (primary benchmark metric: accuracy)
• Optional: accuracy at different SNR buckets

Class labels

• 8PSK
• AM-DSB
• AM-SSB
• BPSK
• CPFSK
• GFSK
• PAM4
• QAM16
• QAM64
• QPSK
• WBFM

Split

• 80/20 random split

Preprocessing

1. Load the RML22.01A numpy archive containing (modulation, snr) keyed samples.
2. Cast IQ arrays to float32.
3. Normalize I and Q channels using dataset-wide mean/std for the 2022 release.
4. Expand to (2, 1, T) and store with modulation label index and SNR.
5. Compute class weights and write the HDF5 cache.

Script: preprocess_rml.py

python preprocessing/preprocess_rml.py \
  --data-path <RML_ROOT> \
  --version 2022 \
  --output data/rml2022.h5

Metric

• Primary: top-1 accuracy on the test split (overall).
• Optional: per-SNR accuracy (not used for ranking by default).

Citation

@article{10.1109/TWC.2023.3254490,
  author = {Sathyanarayanan, Venkatesh and Gerstoft, Peter and Gamal, Aly El},
  title = {RML22: Realistic Dataset Generation for Wireless Modulation Classification},
  journal = {Trans. Wireless. Comm.},
  year = {2023},
  volume = {22},
  number = {11},
  pages = {7663--7675},
  doi = {10.1109/TWC.2023.3254490}
}