如何在 TF 2 中通过自定义函数使用 tf.data.Dataset.interleave()？

Question

我正在使用 TF 2.2，并尝试使用 tf.data 创建管道。

以下工作正常：

def load_image(filePath, label):

    print('Loading File: {}' + filePath)
    raw_bytes = tf.io.read_file(filePath)
    image = tf.io.decode_image(raw_bytes, expand_animations = False)

    return image, label

# TrainDS Pipeline
trainDS = getDataset()
trainDS = trainDS.shuffle(size['train'])
trainDS = trainDS.map(load_image, num_parallel_calls=AUTOTUNE)

for d in trainDS:
    print('Image: {} - Label: {}'.format(d[0], d[1]))

我想将load_image()与 一起使用Dataset.interleave()。然后我尝试：

# TrainDS Pipeline
trainDS = getDataset()
trainDS = trainDS.shuffle(size['train'])
trainDS = trainDS.interleave(lambda x, y: load_image_with_label(x, y), cycle_length=4)

for d in trainDS:
    print('Image: {} - Label: {}'.format(d[0], d[1]))

但我收到以下错误：

Exception has occurred: TypeError
`map_func` must return a `Dataset` object. Got <class 'tuple'>
  File "/data/dev/train_daninhas.py", line 44, in <module>
    trainDS = trainDS.interleave(lambda x, y: load_image_with_label(x, y), cycle_length=4)

如何调整我的代码以并行Dataset.interleave()读取load_image()图像？

Answer 1

正如错误所示，您需要修改load_image以便它返回一个Dataset对象，我已经展示了一个带有两个图像的示例，说明如何执行此操作tensorflow 2.2.0：

import tensorflow as tf
filenames = ["./img1.jpg", "./img2.jpg"]
labels = ["A", "B"]

def load_image(filePath, label):
    print('Loading File: {}' + filePath)
    raw_bytes = tf.io.read_file(filePath)
    image = tf.io.decode_image(raw_bytes, expand_animations = False)
    return tf.data.Dataset.from_tensors((image, label))

dataset = tf.data.Dataset.from_tensor_slices((filenames, labels))
dataset = dataset.interleave(lambda x, y: load_image(x, y), cycle_length=4)

for i in dataset.as_numpy_iterator():
    image = i[0]
    label = i[1]
    print(image.shape)
    print(label.decode())

# (275, 183, 3)
# A
# (275, 183, 3)
# B