Tensorflow 无法解码 tfrecords 中的 jpeg 字节

Question

I tried to write some images to a tfrecord file, but I found it is too large. Then I tried to write the origin jpeg bytes into The tfrecord file. but when I try to read it, there is exception: ValueError: Shape must be rank 0 but is rank 1 for 'DecodeJpeg' (op: 'DecodeJpeg') with input shapes: [32].

Followed is my code

import tensorflow as tf
import os


def write_features(example_image_paths, tf_records_path):
    with tf.python_io.TFRecordWriter(tf_records_path) as writer:
        for image_path in example_image_paths:
            with open(image_path, 'rb') as f:
                image_bytes = f.read()
            feautres = tf.train.Features(
                feautres={
                    'images':
                    tf.train.Feature(bytes_list=tf.train.BytesList(
                        value=image_bytes))
                })
            example = tf.train.Example(feautres)
            writer.write(example.SerializeToString())


def extract_features_batch(serialized_batch):
    """

    :param serialized_batch:
    :return:
    """
    features = tf.parse_example(
        serialized_batch,
        features={'images': tf.FixedLenFeature([], tf.string)})
    bs = features['images'].shape[0]
    images = tf.image.decode_image(features['images'], 3)
    w, h = (280, 32)
    images = tf.cast(x=images, dtype=tf.float32)
    images = tf.reshape(images, [bs, h, w, 3])

    return images


def inputs(tfrecords_path, batch_size, num_epochs, num_threads=4):
    """

    :param tfrecords_path:
    :param batch_size:
    :param num_epochs:
    :param num_threads:
    :return: input_images, input_labels, input_image_names
    """

    if not num_epochs:
        num_epochs = None

    dataset = tf.data.TFRecordDataset(tfrecords_path)

    dataset = dataset.batch(batch_size, drop_remainder=True)

    # The map transformation takes a function and applies it to every element
    # of the dataset.
    dataset = dataset.map(map_func=extract_features_batch,
                          num_parallel_calls=num_threads)
    dataset = dataset.shuffle(buffer_size=1000)
    dataset = dataset.repeat()

    iterator = dataset.make_one_shot_iterator()

    return iterator.get_next(name='IteratorGetNext')


if __name__ == '__main__':
    pass
    # img_names = os.listdir('./images')
    # img_paths = []
    # for img_name in img_paths:
    #     img_paths.append(os.path.join('./images', img_name))
    # write_features(img_paths, 'test.tfrecords')

    images = inputs('./test.tfrecords', 32, None)

How can I read and decode the jpeg bytes properly? Thanks!

Answer 1

You need to decode images before batching the dataset. In other words, in your inputs() function the 'correct' order would be:

dataset = dataset.map(map_func=extract_features_batch,
                      num_parallel_calls=num_threads) 

dataset = dataset.batch(batch_size, drop_remainder=True)

The documentation says ( https://www.tensorflow.org/api_docs/python/tf/io/decode_image ) that tf.io.decode_image expects an image in a form of a scalar or 0-dimensional string (0-D string is considered a scalar) while if you batch the dataset object first the tf.io.decode_image receives a list (or a batch) of images (represented as a list of batch_size times 0 dimensional strings). It then complains that it expected 0-dimensional array while received an array with the shape of [32] (which is the batch size in your case).

I have no idea about how we could optimize input pipeline for batch-processing other than inefficiently do batching after processing. As usual, there is nothing about it in docs on tf 2.0.