We use deterministic random number generation to ensure replicability, even of noise oscillators. Nonetheless, there are small numeric differences between the CPU and GPU. The mean-average-error between audio samples generated on CPU and GPU are < 1e-2.
Reproducibility is currently guaranteed when
batch_size is multiple
of 32 and you use the default
If a train/test split is desired, 10% of the samples are marked as test. Because researchers with larger GPUs seek higher-throughput with batchsize 1024, \(9 \cdot 1024\) samples are designated as train, the next 1024 samples as test, etc.