def _get_data_tensor(self, dataset, batch_size, data_format):
# 数据
provider = slim.dataset_data_provider.DatasetDataProvider(
dataset,
common_queue_capacity=20 * batch_size,
common_queue_min=10 * batch_size,
shuffle=True)
# 提取数据
[image, labels,
bboxes] = provider.get(['image', 'object/label', 'object/bbox'])
# 数据预处理
image, labels, bboxes = ssd_vgg_preprocessing.preprocess_for_train(
image, labels, bboxes, self.img_shape, data_format)
# 编码label和boxes：Encode ground-truth labels and bboxes.
classes, localisations, scores = self.ssd_net.bboxes_encode(
labels, bboxes, self.ssd_anchors)
# reshape_list：拉直
batch_tensors = self._reshape_list(
[image, classes, localisations, scores])
r = tf.train.batch(batch_tensors,
batch_size=batch_size,
capacity=5 * batch_size)
# reshape_list：变成原来的形状
return self._reshape_list(r, shape=[1] + [len(self.ssd_anchors)] * 3)
def __preprocess_data(self, image, labels, bboxes):
out_shape = g_ssd_model.img_shape
if self.is_training_data:
image, labels, bboxes = preprocess_for_train(image, labels, bboxes, out_shape = out_shape)
else:
image, labels, bboxes, _ = preprocess_for_eval(image, labels, bboxes, out_shape = out_shape)
return image, labels, bboxes
