Automatic Mixed Precision

Training with automatic mixed precision is available when training STT on an GPU.

Mixed precision training makes use of both FP32 and FP16 precisions where appropriate. FP16 operations can leverage the Tensor cores on NVIDIA GPUs (Volta, Turing or newer architectures) for improved throughput. Mixed precision training often allows larger batch sizes. Automatic mixed precision training can be enabled by including the flag --automatic_mixed_precision true at training time:

$ python -m coqui_stt_training.train \
    --train_files train.csv \
    --dev_files dev.csv \
    --test_files test.csv \
    --automatic_mixed_precision true

On a Volta generation V100 GPU, automatic mixed precision can speed up 🐸STT training and evaluation by approximately 30% to 40%.