nextBatch method

Map<String, Tensor> nextBatch(int batchSize)

Loads a batch of batchSize triplets directly into GPU memory

Implementation

Map<String, Tensor> nextBatch(
  int batchSize, //, List<Tensor> tracker
) {
  List<double> anchors = [];
  List<double> positives = [];
  List<double> negatives = [];

  for (int i = 0; i < batchSize; i++) {
    // 1. Pick a random person for Anchor & Positive
    final people = _identityMap.keys.toList();
    final personA = people[_random.nextInt(people.length)];
    final imagesA = _identityMap[personA]!;

    final imgIdx1 = _random.nextInt(imagesA.length);
    int imgIdx2;
    do {
      imgIdx2 = _random.nextInt(imagesA.length);
    } while (imgIdx1 == imgIdx2 && imagesA.length > 1);

    // 2. Pick a different person for Negative
    String personB;
    do {
      personB = people[_random.nextInt(people.length)];
    } while (personA == personB);
    final imagesB = _identityMap[personB]!;

    // 3. Process and Flatten
    anchors.addAll(_processImage(imagesA[imgIdx1]));
    positives.addAll(_processImage(imagesA[imgIdx2]));
    negatives.addAll(_processImage(imagesB[_random.nextInt(imagesB.length)]));
  }

  final anchorTensor = Tensor.fromList([
    batchSize,
    imageSize * imageSize * 3,
  ], anchors);
  final positiveTensor = Tensor.fromList([
    batchSize,
    imageSize * imageSize * 3,
  ], positives);
  final negativeTensor = Tensor.fromList([
    batchSize,
    imageSize * imageSize * 3,
  ], negatives);

  // tracker.addAll([anchorTensor, positiveTensor, negativeTensor]);

  return {
    'anchor': anchorTensor,
    'positive': positiveTensor,
    'negative': negativeTensor,
  };
}