Not able to adjust the NeuralNetwork parameters

Summary

I am using the “Extracting the configuration into a file” approach for training and running the code, and I have not been able to adjust the neural network options (such as errorThresh) in the settings of the conf.json file

Simplest Example to Reproduce

Here is the code I am running (from quickstart)

const { dockStart } = require('@nlpjs/basic');

(async () => {
  const dock = await dockStart();
  const nlp = dock.get('nlp');
  await nlp.train();
  const response = await nlp.process('en', 'some query');
...

and I am using this settings in the conf.json file…

{
  "settings": {
    "nlp": {
          "corpora": ["./model/corpus.json"],
	  "threshold": 0.8,
	  "nlu" : {  "log": true,
		  "errorThresh" : 0.0000005,
		  "deltaErrorThresh" : 0.00000001
	  }
	}
  },
  "use": ["Basic", "LangEn"]
}

I look throughout the documentation and the code but there was not definite example on how to structure the input. However, there was one hint that the way shown above might be the way. But no matter what values I change the errorThresh or deltaErrorThresh parameters to, the training log does not change. I also tried several other ways to input the settings for these parameters in conf.json with no effect on the training log. For all the combinations I tried, the training results were identical and no errors were indicated.

How should the conf.json be structured to change these parameters? Thanks in advance!

Current Behavior

I am not able to affect the training history with the inputs I have tried.

Context

My model is training very quickly, but I don’t know if the training is stopping prematurely. That is why I would like to adjust these training parameters.

[aigloss]

HI @sjscotti @Apollon77 ,

the parameters errorThresh and deltaErrorThresh are actually used by the NeuralNetwork when training the corpus:

https://github.com/axa-group/nlp.js/blob/66d556cca6b766568c9f87c3e19defac6ddd0dc2/packages/neural/src/neural-network.js#L172-L173.

To let those parameters reach NeuralNetwork you’ll have to add them in the ‘nlu-<language>’ section of your settings, what would make your configuration look like this

{
  "settings": {
    "nlp": {
      "corpora": ["./corpus-en.json"]
    },
    "nlu-??": {
      "log": true,
      "errorThresh": 0.0000005,
      "deltaErrorThresh": 0.00000001
    }
  },
  "use": ["Basic", "ConsoleConnector"]
}

Note the use of -??. That is a common pattern used in this project to act as a wildcard. So, in the example above, those settings would be applied for any NeuralNetwork. If, on the other hand, you’d like to target the NLU for a specific language, you could just something like nlu-en.

Be careful when tweaking those values though, as lower values might lead to a much higher training time, with a not so improved accuracy.

[Apollon77]

Looking forward to hear from you 😃