wandb instrumentation

[syllogismos]

I did wandb instrumentation of this repo, to log histograms, plots, audios while training. You can see the resulting report here

Encoder Training results

Synthesizer Training results

Vocoder Training results

[daciansolgen]

how did you get this to work?? i cant even install the software :(

[syllogismos]

Hey, I trained p2.xlarge EC2 instance, used amazon deep learning AMI as base machine, I can give you the conda environment file if you needed.

Also I trained on part of the training dataset. And even with part of the training dataset, I needed a 1TB harddrive to save all the generated models, data of encoder, synthesizer and vocoder models.

[syllogismos]

Here is the conda environment file https://gist.github.com/syllogismos/0b3097c7714ce6119db500a3fcdcc8dd

you can install this environment using this command conda env create -f environment.yml

It has additional slack and twitter pip packages installed as I tried building slack and twitter mention bots that reply with an audio of what you typed automatically. Will post the code here.

[ghost]

This is cool, but I think this is more of an optional feature since not everyone will sign up for wandb to take advantage of the instrumentation. I am going to recommend that we close this issue without merging, and point people to your fork if anyone asks about wandb support.

[ghost]

@CorentinJ Requesting that this PR be closed without merging. Please see my review for rationale.

Another possibility is to start a wandb branch, but I don't think we should take on a feature if it's not going to be maintained.

[CorentinJ]

The PR is riddled with unnecessary code anyway.

[syllogismos]

Thank you for responding and comments @CorentinJ @blue-fish This is a really cool project.

The initial pull request only has the wandb instrumentation. After that I saw this repo is not supported anymore. I wasn't hoping that the slackbot and twitter bot to be merged into this. I can do a separate pull request with only wandb instrumentation if anyone is interested.

I did the twitterbot/slackbot as a fun experiment. I used the trained model you provided for the bots, but the results are okay. Does the voice have to be in the training dataset? I wanted to generate people's mentions on twitter into the voice of Stephen Fry. And I used 20 minutes of his voice. I don't know, the cloned voice seemed really monotonous, and isn't capturing the gaps between the words, it seems like the gaps also has some information regarding a persons voice? I don't know. I didn't really check out other voice cloning projects. Do you have any suggestions on getting better results?

Thank you.