python - 如何将 int32 类型的 4D numpy 数组转换为 tfrecords？

Question

我有一个高光谱数据集，它是一个 numpy 数组，其尺寸为 (num_images, height=7, width=7, num_channels=144) 和数据类型 int32。

标签数组为(batch_size, num_classes=15)。我想将其转换为 tf.records 并正确读回。

到目前为止，我已经阅读了很多博客并尝试了很多不同的方法，但都失败了。这是我尝试过的？

问题是，当我用它训练模型时，代码不会抛出错误，但当我将它与使用 numpy 数组训练模型时的情况进行比较时，它的准确性结果没有任何意义。

问题是我在代码中哪里犯了错误？我在转换为 tfrecords 并读回时会犯任何错误吗？

def wrap_int64(value):
    return tf.train.Feature(int64_list=tf.train.Int64List(value=[value]))


def wrap_bytes(value):
    return tf.train.Feature(bytes_list=tf.train.BytesList(value=[value]))


def convert(images, labels, save_path, save_name):

"""

:param images: np.ndarray containing images with shape (num_images, 
 height, width, num_channels)
:param labels: np.ndarray containing labels with shape (num_labels,),
 i.e. one_hot=False
:param save_path: path in which we save the tfrecords
:return:
"""




out_path = os.path.join(save_path, save_name)
print("Converting: " + out_path)

assert images.dtype == np.int32

# Number of images
num_images = len(images)
print(num_images)
with tf.python_io.TFRecordWriter(out_path) as writer:
    for i in range(num_images):

        # Load a single image
        img = images[i]
        label = labels[i]

        # Convert the image to raw bytes.
        img_bytes = img.tostring()

        image_shape = np.array(np.shape(image)).astype(np.int32)

        # Convert the image to raw bytes.
        #########################################################
        # There is no need to flatten each image!!!
        ###########################################################
        img_bytes = image.tostring()
        img_shape_bytes = image_shape.tostring()
        # Create a dict with the data we want to save in the
        # TFRecords file. You can add more relevant data here.
        data = \
            {
                'image': wrap_bytes(tf.compat.as_bytes(img_bytes)),
                'image_shape': wrap_bytes(tf.compat.as_bytes(img_shape_bytes)),

                'label': wrap_int64(label)

            }

        # Wrap the data as TensorFlow Features.
        feature = tf.train.Features(feature=data)


        # Wrap again as a TensorFlow Example.
        example = tf.train.Example(features=feature)

        # Serialize the data.
        serialized = example.SerializeToString()

        # Write the serialized data to the TFRecords file.
        writer.write(serialized)

#

def parse(serialized, num_classes, normalization_factor):
    features = \
    {
        'image': tf.FixedLenFeature([], tf.string),
        'image_shape': tf.FixedLenFeature([], tf.string),
        'label': tf.FixedLenFeature([], tf.int64),
    }

    # Parse the serialized data so we get a dict with our data.
    parsed_example = \
    tf.parse_single_example(
        serialized=serialized,
        features=features)

    # Get the image, shape and label as raw bytes.
    image_raw = parsed_example['image']
    image_shape_raw = parsed_example['image_shape']
    label = parsed_example['label']

    # Decode the raw bytes so it becomes a tensor with type.

    # have to be converted to the exact same datatype as it was before 
      starting conversion to tfrecords

    image = tf.decode_raw(image_raw, tf.int32)
    image_shape = tf.decode_raw(image_shape_raw, tf.int32)

    # reshape the image back to its original shape
    image_reshaped = tf.reshape(image, image_shape)

    # let's cast the image to tf.float32 and normalize it. Let's 
    # change the label to one_hot as well.
    image_normed = normalization_factor * tf.cast(image_reshaped, tf.float32)
    label_one_hot = tf.one_hot(label, num_classes)

    # The image and label are now correct TensorFlow types.
    return image_normed, label_one_hot

#

def input_fn(filenames, num_classes, normalization_factor, train, batch_size=1024, prefetch_buffer_size=5):


    buffer_size = 10 * batch_size

    dataset = tf.data.TFRecordDataset(filenames=filenames)

    dataset = dataset.map(lambda x: parse(x, num_classes, normalization_factor))
    if train:
        dataset = dataset.shuffle(buffer_size=buffer_size)

    # Allow infinite reading of the data.
        num_repeat = None
    else:
        num_repeat = 1

    # Repeat the dataset the given number of times.
    dataset = dataset.repeat(num_repeat)

    # Get a batch of data with the given size.
    dataset = dataset.batch(batch_size)

    dataset = dataset.prefetch(buffer_size=prefetch_buffer_size)

    # Create an iterator for the dataset and the above modifications.
    iterator = dataset.make_one_shot_iterator()

    # Get the next batch of images and labels.
    batch_images_tf, batch_labels_tf = iterator.get_next()

    return batch_images_tf, batch_labels_tf

Answer 1

例如，您需要使用tf.train.Feature(假设你的标签是整数)

对于 int 值:

def _int64_feature(value):
  return tf.train.Feature(int64_list=tf.train.Int64List(value=[value]))

对于字节:

def _bytes_feature(value):
  return tf.train.Feature(bytes_list=tf.train.BytesList(value=[value]))

他们:

 data = {'label': _int64_feature(label),
         'image': _bytes_feature(tf.compat.as_bytes(img_bytes))}

应该可以完成工作