Some questions about training model

[guoaoo]

I have trained a synthesizer used by default hparams from scratch in 14 days, and there are some plot in Tensorboard.

Real-Time-Voice-Cloning/synthesizer/hparams.py

Lines 52 to 57 in 7432046

	tts_schedule = [(2, 1e-3, 20_000, 12), # Progressive training schedule
	(2, 5e-4, 40_000, 12), # (r, lr, step, batch_size)
	(2, 2e-4, 80_000, 12), #
	(2, 1e-4, 160_000, 12), # r = reduction factor (# of mel frames
	(2, 3e-5, 320_000, 12), # synthesized for each decoder iteration)
	(2, 1e-5, 640_000, 12)], # lr = learning rate

Some questions to ask:

1. The pretrain model is trained to 295k, is that because it is good enough? The size of the .pt file does not increase with the number of training steps.
2. I used the default hparams during training and only used about 6G of GPU memory, there is still a lot left for 12G 2080ti and 32G V100, how do I adjust the hparams to speed up the training? If increase batch_size, how much should I increase the learning rate? Is reduction factor an adjustable number, and will it have a bad effect on the results?
3. After reading Training a new model based on LibriTTS #449 , do you recommend that I modify the hparams and train a synthesizer from scratch or just fine-tuning on pretained model?

[ghost]

1. No, the number of steps is arbitrary though I believe the training had converged earlier than that.
2. You need to experiment with batch_size and learning rate to find the best values. The reduction factor r is adjustable (integer >= 1). Training and inference are faster when it is set to a higher number. Quality is better when r is lower.
3. Train from scratch.