prediction script

[wuxiuzhi738]

❓ Questions & Help

Details

do you have prediction script?

[upskyy]

Prediction is not supported yet. We will add more later.

[wuxiuzhi738]

In hydra_eval.py, line 47, "model = MODEL_REGISTRY[configs.eval.model_name].load_from_checkpoint(configs.eval.checkpoint_path)",
An error occurs that config and tokenizer parameters are missing. I use "model = MODEL_REGISTRY['joint_ctc_conformer_lstm'].load_from_checkpoint(configs=configs, checkpoint_path=configs.eval.checkpoint_path, tokenizer=None)" and do some other modifications, program runs without errors , but the predicted results looks like not right that there are a lot of repetitive words.

[wuxiuzhi738]

As #86 saied.

[sooftware]

Can you show me the result?

[wuxiuzhi738]

and now ,the train_cer of training is already less than 0.5.

[sooftware]

I'm sure it's weird. Can you organize and write both train and evaluation scripts about which model you used?

[wuxiuzhi738]

model: joint_ctc_conformer_lstm, The training script is basically unchanged Major modifications in hydra_eval.py script:

[wuxiuzhi738]

I'm sure it's weird. Can you organize and write both train and evaluation scripts about which model you used?

I wonder if can load the model with 'torch. load',or only use 'load_from_checkpoint' in pytorch-lightninig.