My trained model is 4GB, how that? about waveglow HOT 11 CLOSED

because the default config is huage. 12*8(512channel)

from waveglow.

We haven't done a lot of ablative analysis yet to see how few channels we could get away with or how few layers. A lot of architecture decisions were made based on the early parts of the training curves which seem to favor bigger models. But if smaller models were trained for 500k iterations they might sound essentially as good.

from waveglow.

@hcwu1993 the trained model or the checkpoint file that is saved during training and includes includes the optimizer states?

from waveglow.

it should save model parameters and structure information according to the pytorch doc.

from waveglow.

By the way, the given model is 2GB. So the config is different from the paper? And i got a unusual result using this model. The F0 of generation wav is lower than the natural, it sounds like male voice.

from waveglow.

@hcwu1993 Unlike the checkpoints saved during draining that include optimizer states, the checkpoint we shared with the pretrained model only contains the model. Hence the difference in size.

from waveglow.

thank you. And i have another question. I use this command to synthesis wav, the default sampling rate is 22050Hz. it sounds like male voice. is there any problem in my command?

from waveglow.

Your mel-spectrograms must have the same parameters (sampling_rate, filter_lenght, hop_length, win_length, mel_fmin, mel_fmax) as your model.

The pretrained model we share was trained with "mel_fmax": 8000.0.
We eventually updated the config.json file to have "mel_fmax": 8000.0. as the default. https://github.com/NVIDIA/waveglow/blob/master/config.json#L20

If you trained your model before this update, it is possible that your model was trained with librosa's default: "mel_fmax": sampling_rate/2.

from waveglow.

@ Did your batch_size=24, train with fp16？

from waveglow.

No, we trained with FP32.

from waveglow.

Closing issue. Please re-open if needed.

from waveglow.