Hey guys, so I was thinking about the best way to add Tensorboard to the TalkNet training colab, and from looking through the code, it seems like it's installed as part of Tensorflow. Considering that in the pipeline training notebook, there's just a cell to launch Tensorboard, before you begin training, I feel like it would be very simple to implement this. Would someone be willing to try and implement this feature? I'm also willing to give it a shot, but I'm not 100% sure how to add cells to an already existing notebook.

unfortunately you also need to add code to the training process that will output the things to tensorboard i might give it a shot but no guarantees

@everyone can anyone solve it?

your audio is either in stereo or in 44khz

follow this guide

ok thanks

@hecko thanks it was in stereo you saved me a lot of time

You can't ping everyone in this server

ok sorry

Is it possible to train a HifiGAN with a pipeline model? I've noticed the pipeline checkpoints are

.pt

files but the standard tacotron one's aren't.

I think it should be fine You just need to direct the cell to the trained model Idk if it actually accepts them though, i haven't tried it

Keep running into this with pipeline models:

Traceback (most recent call last):
  File "GTA.py", line 177, in <module>
    GTA_Synthesis(args.output_directory, args.checkpoint_path, args.n_gpus, args.rank, args.group_name, hparams)
  File "GTA.py", line 95, in GTA_Synthesis
    model = warm_start_model(checkpoint_path, model)
  File "GTA.py", line 79, in warm_start_model
    model.load_state_dict(checkpoint_dict['state_dict'])
KeyError: 'state_dict

mega b says he has a script for training hifigan on pipeline but is yet to share it

Damn Should we... ping him

@mega b the people want high-fidelity generative adversarial network for pipeline

I'm at school

heck

Okay wait but I don't think it was answered, or I'm an idioot and cant read for shit: Did the pipline update, and is it now better than just the regualr arpabet train notebook?

nbdev supports ipa

(' - ') assuming both is yes

Keep getting this airer when training TalkNet, on step 5. ": [Errno 2] No such file or directory: 'wavs/100.wav'" There is no "100.wav" in the dataset or transcript.

I originally had 100.wav when I was transcribing with the auto transcription notebook, but ended up deleting that file from the data set, because the transcript didn't include it for some reason. I'm not sure what it's complaining about.

Oh NVM figured it out. 100 actually was in the transcript, but it wasn't in the order of the files in terms of numbering so I missed it the first time around. Sorry.

Ok now getting this on step 7. ": Cannot convert 'pathlib.PosixPath' to string: '/content/drive/My Drive/talknet/kary/TalkNetDurs/2022-05-20_00-01-25/checkpoints' full_key: checkpoint_callback_params.dirpath reference_type=Optional[CallbackParams] object_type=CallbackParams"

I got this error a few times before, so that step is broken for now.

Hopefully it will be fixed before my colab subscription runs out in a few weeks

Ok. Do I just move on to the next step?

Probably not a good idea.

I'm sure the duration predictor is important for something

Well, I guess it's back to tacotron for me then lol. Kind of a shame too, since TalkNet has been a lot easier to get good results from in my experience. Was literally working like three days ago, wonder what happened.