Train MTDNN using pretrained models in Huggingface?

[varadhbhatnagar]

I am trying to train MTDNN for sequence classification on a language other than English. What changes do I need to make in the model configuration object MTDNNConfig to account for pretrained model and its vocab size?

I tried making changed to MTDNNConfig.vocab_size and MTDNNConfig.init_checkpoint but it lead to assertion errors in MTDNNModel.supported_init_checkpoints(). What are these supported init checkpoints? @namisan