Method toFunction

src/neural-network.js:936–993

* * @returns {Function}

()

   * @returns {Function}
   */
  toFunction() {
    const activation = this.activation;
    const leakyReluAlpha = this.leakyReluAlpha;
    let needsVar = false;
    function nodeHandle(layers, layerNumber, nodeKey) {
      if (layerNumber === 0) {
        return (typeof nodeKey === 'string'
          ? `input['${nodeKey}']`
          : `input[${nodeKey}]`);
      }

      const layer = layers[layerNumber];
      const node = layer[nodeKey];
      let result = ['(' , node.bias];
      for (let w in node.weights) {
        if (node.weights[w] < 0) {
          result.push(`${node.weights[w]}*${nodeHandle(layers, layerNumber - 1, w)}`);
        } else {
          result.push(`+${node.weights[w]}*${nodeHandle(layers, layerNumber - 1, w)}`);
        }
      }
      result.push(')');

      switch (activation) {
        case 'sigmoid':
          return `1/(1+1/Math.exp(${result.join('')}))`;
        case 'relu': {
          needsVar = true;
          return `((v=${result.join('')})<0?0:v)`;
        }
        case 'leaky-relu': {
          needsVar = true;
          return `((v=${result.join('')})<0?0:${leakyReluAlpha}*v)`;
        }
        case 'tanh':
          return `Math.tanh(${result.join('')})`;
        default:
          throw new Error('unknown activation type ' + activation);
      }
    }

    const layers = this.toJSON().layers;
    const layersAsMath = [];
    let result;
    for (let i in layers[layers.length - 1]) {
      layersAsMath.push(nodeHandle(layers, layers.length - 1, i));
    }
    if (this.outputLookup) {
      result = `{${
        Object.keys(this.outputLookup)
          .map((key, i) => `'${key}':${layersAsMath[i]}`)
        }}`;
    } else {
      result = `[${layersAsMath.join(',')}]`;
    }

    return new Function('input', `${ needsVar ? 'var v;' : '' }return ${result};`);
  }