training.md

📚 Guide to Training 📚

To train UTMOSv2 following the methods described in the paper or used in the competition, please refer to this document.

📩 Install Training Dependencies 📩

To install the dependencies required for training, run the following command:

pip install --upgrade pip  # enable PEP 660 support
pip install -e .[train,optional]

Note

If you are using zsh, make sure to escape the square brackets like this:

pip install -e '.[train,optional]'

🚀 Train UTMOSv2 Using Your Own Data 🚀

To train UTMOSv2 using your own data, you need to create a JSON file that contains the location and name of your data. Here is an example structure for the JSON file:

{
  "data": [
    {
      "name": "dataset1",
      "dir": "/path/to/your/dataset1",
      "mos_list": "/path/to/your/moslist1.txt"
    },
    {
      "name": "dataset2",
      "dir": "/path/to/your/dataset2",
      "mos_list": "/path/to/your/moslist2.txt"
    }
    // Add more data entries as needed
  ]
}

Here, name is used to identify the data-domain ID, and dir specifies the directory where the corresponding .wav files are located. Additionally, mos_list records the MOS values for the .wav files in the directory, in the following format:

sys64e2f-utt491a78a,2.375
sys64e2f-utt8485f83,3.625
sys7ab3c-utt1417b69,4.0
...

The file extension .wav is optional and can be included or omitted. The common files between those in the dir and those specified in the mos_list will be used.

Specify the name, dir, and mos_list set for each dataset-domain ID you want to train.

Save this JSON file with an appropriate name, for example, data_config.json and run the following command:

python train.py --config spec_only --data_config data_config.json

🧪 Fine-tuning from Pre-trained Weights 🧪

To continue training from existing weights, specify the --weight option and train as follows. This is useful when you want to perform additional training using weights learned in a previous stage or when fine-tuning.

python train.py --config spec_only --data_config data_config.json --weight /path/to/your/weights.pth

The --weight option can specify either the configuration file name or the path to the weight .pth file. If the configuration file name is specified, models/{config_name}/fold{now_fold}_s{seed}_best_model.pth is used.

🔬 Using Weights & Biases (wandb) for Experiment Tracking 🔬

To use Weights & Biases (wandb) for experiment tracking, specify the --wandb option. You will also need to set the WANDB_API_KEY in your .env file or environment variables, or follow the prompt during execution to input your API key directly in the command line.

python train.py --config spec_only --data_config data_config.json --wandb