Can we fine tune a fine tuned model for quartznet architecture?

[wiamfa]

Hello , is there any pre-trained models for the french language? or Is there any possible way to fine-tune a fine-tuned model the second time?

thanks,

[alancucki]

Hi @wiamfa

there is a pre-trained model for French from the NeMo project.

You can use it in NeMo, or follow these steps to load it in DeepLearningExamples:

1. Change the extension from .nemo to .tar.gz.
2. After unpacking you'll find two files: model_weights.ckpt and model_config.yaml.
3. Make a copy of configs/quartznet15x5_speedp-online-1.15_speca.yaml and replace labels with those from model_config.yaml.
4. Rename the keys in model_weights.ckpt:

   import torch
   remap = lambda k: (k.replace('encoder.encoder', 'encoder.layers')
                  .replace('decoder.decoder_layers', 'decoder.layers')
                  .replace('conv.weight', 'weight'))
   ckpt = torch.load('model_weights.ckpt')
   torch.save({'state_dict': {remap(k): v for k, v in ckpt.items() if 'preproc' not in k}},
          'qn_15x5_french.pt')

   Voilà!

Is there any possible way to fine-tune a fine-tuned model the second time?

This should work fine too.