Train

Use the teklia-dan train command to train a new DAN model. It is able to train a DAN model at line or document-level.

To train DAN on your dataset:

1. Create a training JSON configuration file. Refer to the dedicated page for a description of parameters.
2. Run teklia-dan train --config path/to/your/config.json.
3. (Optional) Train a language model. Refer to the dedicated page.
4. Look into the training results in the output folder indicated in your configuration:
   • checkpoints contains model weights for the last trained epoch and for the epoch giving the best valid CER.
   • results contains the tensorboard log file and the parameters file.

Additional pages

• Jean Zay tutorial
• Data augmentation