When has the definition of num_classes been changed?

[whrichd]

I recently updated to a later version (1.6rc) of NeMo for building an RNNT model, but then I realized I can’t load the older conformer CTC models I had built with 1.3. The error is from nemo.collections.asr.modules.conv_asr.ConvASRDecoder saying len(vocabulary) is different from num_classes. In my case, the models I’ve tried, len(vocabulary) == num_classes +2, I think it’s because previously <unk> and _ are not counted in num_classes? I could be wrong. But since when has this changed and why? I went to see the def of this class, the checking of num_classes == len(vocabulary) has always been there since version 1.3, so I haven’t been able to figure out what actually caused the incompatibility.

[titu1994]

We haven't changed the support of num_classes. Are you using a Nemo file ? Could you print out the stack trace ?

[whrichd]

Traceback (most recent call last):
  File "speech_to_text_buffered_infer.py", line 174, in <module>
    main()  # noqa pylint: disable=no-value-for-parameter
  File "speech_to_text_buffered_infer.py", line 96, in main
    asr_model = nemo_asr.models.EncDecCTCModelBPE.restore_from(restore_path=args.asr_model)
  File "/home/——/NeMo/nemo/core/classes/modelPT.py", line 288, in restore_from
    cls, restore_path, override_config_path, map_location, strict, return_config, trainer
  File "/home//NeMo/nemo/core/connectors/save_restore_connector.py", line 142, in restore_from
    instance = calling_cls.from_config_dict(config=conf, trainer=trainer)
  File "/home//NeMo/nemo/core/classes/common.py", line 494, in from_config_dict
    instance = cls(cfg=config, trainer=trainer)
  File "/home//NeMo/nemo/collections/asr/models/ctc_bpe_models.py", line 207, in __init__
    super().__init__(cfg=cfg, trainer=trainer)
  File "/home//NeMo/nemo/collections/asr/models/ctc_models.py", line 174, in __init__
    self.decoder = EncDecCTCModel.from_config_dict(self._cfg.decoder)
  File "/home//NeMo/nemo/core/classes/common.py", line 452, in from_config_dict
    instance = hydra.utils.instantiate(config=config)
  File "/home//miniconda3/envs/mlflow/lib/python3.7/site-packages/hydra/_internal/instantiate/_instantiate2.py", line 180, in instantiate
    return instantiate_node(config, *args, recursive=_recursive_, convert=_convert_)
  File "/home//miniconda3/envs/mlflow/lib/python3.7/site-packages/hydra/_internal/instantiate/_instantiate2.py", line 249, in instantiate_node
    return _call_target(_target_, *args, **kwargs)
  File "/home//miniconda3/envs/mlflow/lib/python3.7/site-packages/hydra/_internal/instantiate/_instantiate2.py", line 66, in _call_target
    ).with_traceback(sys.exc_info()[2])
  File "/home//miniconda3/envs/mlflow/lib/python3.7/site-packages/hydra/_internal/instantiate/_instantiate2.py", line 62, in _call_target
    return _target_(*args, **kwargs)
  File "/home//NeMo/nemo/collections/asr/modules/conv_asr.py", line 404, in __init__
    f"If vocabulary is specified, it's length should be equal to the num_classes. Instead got: num_classes={num_classes} and len(vocabulary)={len(vocabulary)}"
ValueError: Error instantiating 'nemo.collections.asr.modules.conv_asr.ConvASRDecoder' : If vocabulary is specified, it's length should be equal to the num_classes. Instead got: num_classes=198 and len(vocabulary)=200

[whrichd]

here for example, i’m loading a model trained with 1.3 NeMo, but with 1.6 NeMo. Model is trained with 200 units vocab, but num_classes is 198 for some reason. Yeah I feel like it’s not this part of code itself changed, but something else.