TFNet2

TFNet 2018 updated to run with current tensorflow, uses Estimators, can run with MirroredStrategy on multiple GPUs

Known Issues

Training with current CUFFT/tf.signal.irfft may result in NaNs during training. Seems to be dataset dependent and will recover without problems if ran again (continue from last saved checkpoint). Current train script attempts to automatically detect this and restart training if NaNs are encountered.

Data Preparation

##VCTK-Corpus

1. Download and extract VCTK-Corpus
2. Run DATAROOT=/path/to/VCTK/wav48 TXT_ROOT=dataset_txt/vctk data_prepration/gen_all_rate.sh

Training

Baselines can be generated by running make tfnet_baselines.

Run make -n to see actual commands executed.