Importing Trained Model

[noinformationavailable]

I am sure I have missed something obvious with this.

I have trained a model both through colab and reconfigured a local build option using that as a guide.

I am not sure how to configure access to the local model once built in a custom python application (not home assistant)

from the base demo example I can use the pre trained without issue
I modified the parser
parser.add_argument(
"--model",
help="The model to use for openWakeWord, leave blank to use all available models",
type=str,
required=False,
default=""
)

or even
parser.add_argument(
"--model",
help="The model to use for openWakeWord, leave blank to use all available models",
type=str,
required=False,
default="alexa"
)

and these work as expected for all or for one

However I have been unable to modify it to find a custom model. Have tried putting the files in the root directory of the script and referring to them with and without extensions. Placing them in a custom folder etc. I am sure it is just something basic but I am missing it.

Edit specific error: UnboundLocalError: cannot access local variable 'model_path' where it is not associated with a value

[dscripka]

Are you using the latest version of openWakeWord (either v0.5.1, or the main branch of the repo)?

In recent versions of openWakeWord and the detect_from_microphone.py example, the argument name is --model_path and takes the full path to the trained model. The error you are getting seems to indicate that the variable is present in the script, but your arguments may not be providing it.

[noinformationavailable]

Thanks! - The detec from mic is working fine.

I was using the base script from capture_activations.py

As it provided a clear way of preventing the wake words from persisting beyond the initial activation. and worked well with the pretrained.

I will have a go at altering the script to match. An updated capture_activations.py would be great to use as a model but I understand how busy you would be. Thanks for pointing me in the right direction.

[noinformationavailable]

Seems to have been as simple as modifying the Loading of the openwakwords model from the detect script. Thanks again.

[dscripka]

Ah, you are correct, that script is out-of-date! I will try to update that to match the other examples scripts soon.

[noinformationavailable]

Is there a way or format to use to load more than one custom model to use?

[dscripka]

Yes, when you instantiate the openWakeword Model object, you can provide as many paths to custom models as you like:

# Instantiate the model(s)
model = Model(
    wakeword_models=["path/to/custom_model1.tflite", "path/to/custom_model2.tflite"],  
)

Calling predict will then return a dictionary with predictions from each loaded model.

[noinformationavailable]

Thanks alot! Its such a valuable tool and one I have been looking for for a while. Your work is really appreciated.