RangeError: Source is too large

What is wrong?

RangeError: Source is too large

Where does it happen?

In the a Neural Network when trying to train a net in GPU mode

How do we replicate the issue?

• Step 1

Create a dummy tanning data set that total around 32k records like below,

    {
        "input": [
            "boston apartments",
            "condos for sale in boston",
            "marblehead real estate",
            "massachusetts real estate"
        ],
        "output": "Condos"
    }

• Step 2

Now, Import it and run it like below,

const brain = require('brain.js');
const natural = require('natural')
const jsonfile = require('jsonfile');
const trainingDataSet = require('./sample.json');

const net = new brain.NeuralNetworkGPU();


net.train({trainingDataSet}{ log: (stats) => console.log(stats), iterations: 20000 });

jsonfile.writeFile("./trainedNet.json", net.toJSON());

Expected behavior (i.e. solution)

Should have run without this error

Version information

Nodejs:

v12.16.1

Browser:

Brain.js:

2.0.0-beta.2

How important is this (1-5)?

5

Other Comments

[gerald-dotcom]

neuralNetworkGPU: net Doesn’t work. I am getting an error which ask to provide valid neural network.

Other thing is that without using GPU but using LSTM, I’m not getting the state info but it create a network.

    const trainStream = new brain.TrainStream({
        neuralNetwork: net,
        floodCallback: function () {
            encodedData.forEach(function (data) {
                trainStream.write(data);
            })
            trainStream.endInputs();
        },
        doneTrainingCallback: function (stats) {
            console.log(
                `trained in ${stats.iterations} iterations with error: ${stats.error}`
              );
        },
    });

[robertleeplummerjr]

The original questions mentions a data shape of the following:

"input": [
            "boston apartments",
            "condos for sale in boston",
            "marblehead real estate",
            "massachusetts real estate"
        ],
        "output": "Condos"

This must be broken up, and used with brain.recurrent.LSTM for it to work. brain.NeuralNetwork isn’t compatible with strings.

Example of how data should look:

[
  { "input": "boston apartments", "output": "Condos" },
  { "input": "condos for sale in boston", "output": "Condos" },
  { "input": "marblehead real estate", "output": "Condos" },
  { "input": "massachusetts real estate", "output": "Condos" },
]