Function generateInputFromDef

e2e/benchmarks/benchmark_util.js:80–142

* Generates a random input for input definition. * * ```js * const input = generateInput(inputDefs); * * console.log(`Generated input: ${Object.values(input)}`); * console.log(`Prediction for the generated input: ${prediction}`); * ``` * * @param inputDefs The input definition that is used

(inputDefs, isForGraphModel = false)

 * @param isForGraphModel flag for whether to generate inputs for GraphModel
 */
function generateInputFromDef(inputDefs, isForGraphModel = false) {
  if (inputDefs == null) {
    throw new Error('The inputDef cannot be found.');
  }

  const tensorArray = [];
  try {
    inputDefs.forEach((inputDef, inputDefIndex) => {
      const inputShape = inputDef.shape;

      // Construct the input tensor.
      let inputTensor;
      if (inputDef.dtype === 'float32' || inputDef.dtype === 'int32') {
        // We assume a bell curve normal distribution. In this case,
        // we use below approximation:
        // mean ~= (min + max) / 2
        // std ~= (max - min) / 4
        // Note: for std, our approximation is based on the fact that
        // 95% of the data is within the range of 2 stds above and
        // below the mean. So 95% of the data falls in the range of
        // 4 stds.
        const min = inputDef.range[0];
        const max = inputDef.range[1];
        const mean = (min + max) / 2;
        const std = (max - min) / 4;
        generatedRaw = tf.randomNormal(inputShape, mean, std, inputDef.dtype);
        // We clip the value to be within [min, max], because 5% of
        // the data generated maybe outside of [min, max].
        inputTensor = tf.clipByValue(generatedRaw, min, max);
        generatedRaw.dispose();
      } else if (inputDef.dtype === 'string') {
        size = tf.util.sizeFromShape(inputDef.shape);
        data = [...Array(size)].map(
            () => Math.random().toString(36).substring(2, 7));
        inputTensor = tf.tensor(data, inputShape, inputDef.dtype);
      } else {
        throw new Error(
            `The ${inputDef.dtype} dtype of '${inputDef.name}' input ` +
            `at model.inputs[${inputDefIndex}] is not supported.`);
      }
      tensorArray.push(inputTensor);
    });

    // Return tensor map for tf.GraphModel.
    if (isForGraphModel) {
      const tensorMap = inputDefs.reduce((map, inputDef, i) => {
        map[inputDef.name] = tensorArray[i];
        return map;
      }, {});
      return tensorMap;
    }

    return tensorArray;
  } catch (e) {
    // Dispose all input tensors when the input construction is failed.
    tensorArray.forEach(tensor => {
      if (tensor instanceof tf.Tensor) {
        tensor.dispose();