Question

我正在尝试使用以下代码以TensorFlow的protobuf格式编写可变大小的图像：

img_feature = tf.train.Feature(
    bytes_list=tf.train.BytesList(value=[
        img.flatten().tostring()]))
# Define how the sequence length is stored
seq_len_feature = tf.train.Feature(
    int64_list=tf.train.Int64List(value=[seq_len]))
# Define how the label list is stored
label_list_feature = tf.train.Feature(
    int64_list=tf.train.Int64List(value=label_list))
# Define the feature dictionary that defines how the data is stored
feature = {
    IMG_FEATURE_NAME: img_feature,
    SEQ_LEN_FEATURE_NAME: seq_len_feature,
    LABEL_LIST_FEATURE_NAME: label_list_feature}
# Create an example object to store
example = tf.train.Example(
    features=tf.train.Features(feature=feature))

我保存的图像img具有固定高度但可变长度。

现在，如果我想使用以下代码解析此图像：

# Define how the features are read from the example
features_dict = {
  IMG_FEATURE_NAME: tf.FixedLenFeature([], tf.string),
  SEQ_LEN_FEATURE_NAME: tf.FixedLenFeature([1], tf.int64),
  LABEL_LIST_FEATURE_NAME: tf.VarLenFeature(tf.int64),
}
features = tf.parse_single_example(
    serialized_example,
    features=features_dict)
# Decode string to uint8 and reshape to image shape
img = tf.decode_raw(features[IMG_FEATURE_NAME], tf.uint8)
img = tf.reshape(img, (self.img_shape, -1))
seq_len = tf.cast(features[SEQ_LEN_FEATURE_NAME], tf.int32)
# Convert list of labels
label_list = tf.cast(features[LABEL_LIST_FEATURE_NAME], tf.int32)

我收到以下错误： ValueError: All shapes must be fully defined: [TensorShape([Dimension(28), Dimension(None)]), TensorShape([Dimension(1)]), TensorShape([Dimension(3)])]

有没有办法存储大小可变的图像（在我的情况下更具体地说是可变宽度）并用TFRecordReader读取它们？

Answer 1

首先，我无法重现错误。以下代码可以正常工作：

import tensorflow as tf
import numpy as np

image_height = 100
img = np.random.randint(low=0, high=255, size=(image_height,200), dtype='uint8')
IMG_FEATURE_NAME = 'image/raw'

with tf.Graph().as_default():
  img_feature = tf.train.Feature(
      bytes_list=tf.train.BytesList(value=[
          img.flatten().tostring()]))
  feature = {IMG_FEATURE_NAME: img_feature}

  example = tf.train.Example(features=tf.train.Features(feature=feature))
  serialized_example = example.SerializeToString()

  features_dict = {IMG_FEATURE_NAME: tf.FixedLenFeature([], tf.string)}
  features = tf.parse_single_example(serialized_example, features=features_dict)
  img_tf = tf.decode_raw(features[IMG_FEATURE_NAME], tf.uint8)
  img_tf = tf.reshape(img_tf, (image_height, -1))

  with tf.Session() as sess:
    img_np = sess.run(img_tf)

  print(img_np)

print('Images are identical: %s' % (img == img_np).all())

输出：

图片相同：正确

其次，我建议存储编码为PNG而不是RAW的图像，并使用tf.VarLenFeature + tf.image.decode_png读取它们。它将为您节省大量空间，并自然支持可变大小的图像。

Answer 2

我最终能够使用以下代码使其工作，以创建protobuf数据文件：

_, img_png = cv2.imencode('.png', img)
img_png = img_png.tostring()
label_list_feature = [
    tf.train.Feature(bytes_list=tf.train.BytesList(value=[label]))
    for label in label_list]
img_feature = tf.train.Feature(bytes_list=tf.train.BytesList(
        value=[img_png]))
# Define feature for sequence length
seq_len_feature = tf.train.Feature(
    int64_list=tf.train.Int64List(value=[seq_len]))
# Feature list that contains list of labels
feature_list = {
    LABEL_LIST_FEATURE_NAME: tf.train.FeatureList(
        feature=label_list_feature)
}
# Context that contains sequence lenght and image
context = tf.train.Features(feature={
    IMG_FEATURE_NAME: img_feature,
    SEQ_LEN_FEATURE_NAME: seq_len_feature
})
feature_lists = tf.train.FeatureLists(feature_list=feature_list)
# Add sequence length as context
example = tf.train.SequenceExample(
    feature_lists=feature_lists,
    context=context)

以下代码从protobuf中读取：

# Sequence length is a context feature
context_features = {
    IMG_FEATURE_NAME: tf.FixedLenFeature([], dtype=tf.string),
    SEQ_LEN_FEATURE_NAME: tf.FixedLenFeature([], dtype=tf.int64)
}
# Image and target word is a sequence feature
sequence_features = {
    LABEL_LIST_FEATURE_NAME: tf.FixedLenSequenceFeature(
        [], dtype=tf.string)
}
# Parse the example
context_parsed, sequence_parsed = tf.parse_single_sequence_example(
    serialized=serialized_example,
    context_features=context_features,
    sequence_features=sequence_features
)
seq_len = tf.cast(context_parsed[SEQ_LEN_FEATURE_NAME], tf.int32)
# Process the image
img = context_parsed[IMG_FEATURE_NAME]
img = tf.image.decode_png(img, dtype=tf.uint8, channels=nb_channels)
img = tf.reshape(img, (img_height, -1, nb_channels))
labels = sequence_parsed[LABEL_LIST_FEATURE_NAME]
return img, seq_len, labels

注意：在这个例子中，我将整数标签列表更改为字符串标签列表（在我的情况下更自然）。我还将图像存储为png字节字符串。

如何以TensorFlow的protobuf格式保存和读取可变大小的图像

2 个答案: