__init__() missing 2 required positional arguments: 'configs' and 'tokenizer'

[wl-junlin]

when run openspeech_cli.hydra_eval.py, load_from_checkpoint method report this error:

__init__() missing 2 required positional arguments: 'configs' and 'tokenizer'

I think it is caused by that the checkpoit doesn't pickle 'configs' and 'tokennizer' params in model class.

[sooftware]

@hasangchun Can you check?

[upskyy]

Thanks for letting me know. I'll fix it as soon as I have time.

[wl-junlin]

Thanks for letting me know. I'll fix it as soon as I have time.

Thanks, the tokennizer seems can't be pickled. when I tried to add self.save_hyperparameters() in OpenspeechModel.__init()__, it reported an error.

[yinruiqing]

I think you can override on_save_checkpoint or on_load_checkpoint like what I did to solve this bug. But it is not very elegant.

[wuxiuzhi738]

I have the same problem, and there are many problems need to modify when I use (configs=configs, ......,tokenizer=None) . Im not sure thats right, but no errors .

[OleguerCanal]

Hi, I'm trying to load an already-trained model doing something like this:

model = MODEL_REGISTRY[configs.eval.model_name]
model = model.load_from_checkpoint(configs.eval.checkpoint_path, configs=configs, tokenizer=tokenizer)

However, I get this error (for all the layers):

RuntimeError: Error(s) in loading state_dict for ConformerLSTMModel:
	Unexpected key(s) in state_dict: "encoder.conv_subsample.conv.sequential.0.weight ...

Have you guys managed to load a pre-trained model @wuxiuzhi738 @yinruiqing ? Am I doing something wrong??

Thank you a lot

[OleguerCanal]

Just realized that the model didn't have those layers because the build_model() method was not being called. I solved it calling it in the __init__ and now it seems to work well.

[panmareksadowski]

Hi,

I made fixes based on the above tips and evaluation works, but I'm getting much worse results on the validation data set than during training on the same validation dataset. Is it using somehow ground truth during training validation evaluation that causes better results during training validation?