Is it possible to train the model in chunks?

[el-esnawy]

I was wondering if you code train a model and then save the JSON, and after some time load that same JSON and train it more?

[voidvoxel]

Not sure to be honest, but I had the same question myself last night. I'll check in my local instance and report back in a minute 😊

[voidvoxel]

I'm happy to inform you that it does work! 🎉

[el-esnawy]

Perfect 😄 Thanks

[voidvoxel]

Here's the code I used to test it on my local machine in case you'd like to try it too! 💕

// Define some inputs.
const inputs = [
  [0, 0], [0, 1], [1, 0], [1, 1]
];

// Define some outputs.
const outputs = [
  0, 1, 1, 0
];

// Map the inputs to the outputs.
const trainingData = inputs.map(
  (value, index) => ({input: value, output: [ outputs[index] ] })
);

// Create a new neural network.
let xornet = new brain.NeuralNetworkGPU(
  {
    inputSize: 2,
    hiddenLayers: [3],
    outputSize: 1
  }
);

// Train the network.
let results = xornet.train(
  trainingData, {
    iterations: 25000,
    errorThresh: 0.011
  }
);

// Log the results.
console.log(results);

// Save it to JSON and then immediately load it back to simulate saving to a JSON file and loading back from the file system.
xornet = xornet.fromJSON(xornet.toJSON());

// Just a helper function to run the inputs through the trained network and log the output in a pretty format <3
function run(...input) {
  console.log(`${input[0]} ^ ${input[1]} = ${Math.round(xornet.run(input))}`);
}

// Run each possible XOR combination.
run(0, 0);
run(0, 1);
run(1, 0);
run(1, 1);

// Train it some more! This should really be a smaller `errorThresh` than used below.
// I just used a higher value so that it would finish training immediately for the sake of the demo
// In a real world situation, set this number lower than previous training sessions and it will work towards
// the goal that you specified in the `errorThresh` option.
results = xornet.train(
  trainingData, {
    iterations: 25000,
    errorThresh: 0.22
  }
);

console.log(results);

// Train it yet again! You'll notice the error is still going down each time.
// This is because it was trained some more.
// Otherwise, with a higher `errorThresh` of 0.22, much higher than 0.011,
// the network's error would be closer to 0.22 instead of being below 0.011.
// This is because it trains an additional 100 iterations before callbacks are invoked,
// thus signaling the network to terminate the training session.
results = xornet.train(
  trainingData, {
    iterations: 25000,
    errorThresh: 0.22
  }
);

console.log(results);

[voidvoxel]

Perfect 😄 Thanks

Of course! Anytime 😊