How to export model.nemo from ckpt?

[Minxiangliu]

Hi,
I want to export the model from ckpt for testing during model training.
Is there any script to export model.nemo from ckpt?

Many thank!

[VahidooX]

You need to load the model with method load_from_checkpoint and then save it as nemo file. For instance if it is an ASR model, you may do it like this:

    model = nemo.collections.asr.models.EncDecCTCModelBPE.load_from_checkpoint(model_ckpt)
    model.save_to(output_path)

You may also set the following in your config, so checkoints are stored as nemo file instead of PTL checkpoints:

exp_manager.checkpoint_callback_params.always_save_nemo = true

[Arkymeng]

hello, why there occurs ' ValueError: cfg must have tokenizer config to create a tokenizer !' when I load model from ckpt?

[Akshaya-Devadiga]

code:
import nemo.collections.nlp as nemo_nlp
model = nemo.collections.nlp.models.distilbert-base-uncased.load_from_checkpoint('pre_trained_bert/checkpoints/SequenceClassifier-STEP-7210.pt')

ERROR:

AttributeError Traceback (most recent call last)
in
----> 1 model = nemo.collections.nlp.models.distilbert-base-uncased.load_from_checkpoint('pre_trained_bert/checkpoints/SequenceClassifier-STEP-7210.pt')
2 #model.save_to(output_path)

AttributeError: module 'nemo.collections.nlp' has no attribute 'models'

@VahidooX could u plz help me solve this error

[VahidooX]

You need to use the name of the model class instead of distilbert-base-uncased. Would you please try the following:

model = nemo.collections.nlp.models.TokenClassificationModel.load_from_checkpoint('pre_trained_bert/checkpoints/SequenceClassifier-STEP-7210.pt')