Availability of the Colab notebook

[34j]

Many issues have been opened, but I think the notebook is just barely working, although it is rubbish.

Describe the bug

1. f-strings in commands no longer working
2. chore(notebook): copy configs files from Google Drive #1064 might be reducing the the simplicity unfortunately
3. Logs are not displayed
4. Annoying popup: WARNING: Upgrading ipython, ipykernel, tornado, prompt-toolkit, pyzmq can cause your runtime to repeatedly crash. https://colab.research.google.com/ #1163

To Reproduce

Run notebook

Additional context

No response

Version

2024/05/10

Platform

Google Colab

Code of Conduct

• I agree to follow this project's Code of Conduct.

No Duplicate

• I have checked existing issues to avoid duplicates.

[xD0135]

I had the same experience trying to run the provided Colab notebook and decided on a simpler approach, as shown below. Each code block here represents a code block that you add in Colab, and I'm focusing only on training and not including inference, as the original Colab notebook has. Also, make sure that you select a GPU instance, otherwise this won't work.

#@title Mount Google Drive
from google.colab import drive
drive.mount('/content/drive')

#@title Install Dependencies
!python -m pip install so-vits-svc-fork

#@title Verify Dependencies
!svc --help

For this next step, I created a folder MyDrive/TTS/sovits in my Google Drive, hence it listed in the path. Inside the folder I then added my training data in the expected location dataset_raw/{speaker_id}/**/{wav_file}.{any_format} before running the following:

#@title Generate Config
!cd /content/drive/MyDrive/TTS/sovits && svc pre-resample && svc pre-config

This will create a few folders, and because I'm running on the free T4 Colab instance, I optimized my config file located in my Google Drive under MyDrive/TTS/sovits/configs/44k/config.json with the following (only showing the modified lines)

{
  "train": {
    "epochs": 201,
    "batch_size": 16
  }
}

If you made changes to the config file, save/upload them before proceeding. It's worth noting that using 16 as the batch_size keeps the VRAM stable at around 11.9 / 15.0 GB. Anything over will cause it to run OOM.

#@title Generate Hubert
!cd /content/drive/MyDrive/TTS/sovits && svc pre-hubert

#@title Start Training
!cd /content/drive/MyDrive/TTS/sovits && svc train

And that's it! For the training step, if you want to load the TensorBoard so that you have a visual representation of the training, you can use the following instead of the previous:

#@title Start Training
%load_ext tensorboard
%tensorboard --logdir /content/drive/MyDrive/TTS/sovits/logs/44k
!cd /content/drive/MyDrive/TTS/sovits && svc train

For my dataset with about 13 minutes of training audio, it took just over an hour to complete. And now that I understand how long it takes to run, I can increase the epochs parameter to match my desired tradeoff of training data quality. Hope it helps!

[Gambozza]

I'm experiencing an issue while trying to train a model using so-vits-svc-fork on Google Colab. I followed the provided notebook instructions and structured my files in Google Drive as follows:

📂 MyDrive/TTS/sovits/
├── 📂 dataset_raw/{speaker_id}/**/*.wav
├── 📂 logs/44k/
├── 📂 configs/44k/config.json

After mounting Google Drive and installing dependencies, I executed the preprocessing steps successfully:

Mount Google Drive

from google.colab import drive
drive.mount('/content/drive')

Install dependencies

!python -m pip install so-vits-svc-fork

Generate config and preprocess data

!cd /content/drive/MyDrive/TTS/sovits && svc pre-resample && svc pre-config

Generate Hubert features

!cd /content/drive/MyDrive/TTS/sovits && svc pre-hubert

Start training

!cd /content/drive/MyDrive/TTS/sovits && svc train

The training command executes but doesn’t seem to progress. The output only shows:
[INFO] NumExpr defaulting to 2 threads.
and nothing else happens.

I checked TensorBoard logs at /content/drive/MyDrive/TTS/sovits/logs/44k/lightning_logs/version_0/, but no relevant data appears.
The dataset is correctly structured, and running ls dataset_raw/{speaker_id}/ confirms that all .wav files are present.

Additionally, GPU availability is confirmed (nvidia-smi shows Tesla T4 with CUDA 12.4), but it's not being utilized (GPU-Util: 0%).

I have tried:

Restarting the training process multiple times.
Checking logs in logs/44k/, but they don’t seem to update.
Running on CPU (CUDA_VISIBLE_DEVICES=-1), but the issue persists.
Ensuring TensorFlow and dependencies are up to date (pip show tensorflow shows version 2.15.0).
Attempting to run svc train --debug, but the flag isn’t recognized.

At this point, I’m unsure what is preventing the training from progressing. Has anyone encountered a similar issue on Colab? Is there a way to enable detailed logs or debug mode to identify the problem?