python - 从对象检测 API 中使用的 TFRecord 文件中读取数据

Question

我想读取存储在我用作火车记录的 TFRecord 文件中的数据 TF Object Detection API .

但是，我收到了 InvalidArgumentError: Input to reshape is a tensor with 91090 values, but the requested shape has 921600 .我不明白错误的根源是什么，即使差异似乎是 10 倍。

问题:
如何在没有此错误的情况下读取文件？

我不能排除错误来自创建记录，或者错误在于我如何读取它。因此，我已经包含了我的代码。

我可以使用数据运行 object_detection/train.py，并从训练好的模型生成卡住图。

来自 this answer (及其提到的 GitHub 问题)，我发现我必须将 PNG 图像转换为 JPG，因此 as_jpg -part(见下面我的代码)。

我使用了来自 this answer 的代码作为读取文件的起点。

我使用 Tensorflow 1.7.0，Python 3.5

只有一类:“人类”。
该记录有 1000 张图像；每个图像可以有一个或多个边界框。 (对应图像中的每个人一个。)

我如何阅读 TFRecord :
如上所述:我使用了来自 this answer 的代码作为读取文件的起点:

train_record = 'train.record'

def read_and_decode(filename_queue):
    reader = tf.TFRecordReader()
    _, serialized_example = reader.read(filename_queue)
    features = tf.parse_single_example(
        serialized_example,
        # Defaults are not specified since both keys are required.
        features={
            'image/height': tf.FixedLenFeature([], tf.int64),
            'image/width': tf.FixedLenFeature([], tf.int64),
            'image/source_id': tf.FixedLenFeature([], tf.string),
            'image/encoded': tf.FixedLenFeature([], tf.string),
            'image/format': tf.FixedLenFeature([], tf.string),
            'image/object/bbox/xmin': tf.VarLenFeature(tf.float32),
            'image/object/bbox/xmax': tf.VarLenFeature(tf.float32),
            'image/object/bbox/ymin': tf.VarLenFeature(tf.float32),
            'image/object/bbox/ymax': tf.VarLenFeature(tf.float32),
            'image/object/class/text': tf.VarLenFeature(tf.string),
            'image/object/class/label': tf.VarLenFeature(tf.int64)
        })
    image = tf.decode_raw(features['image/encoded'], tf.uint8)
    # label = tf.cast(features['image/object/class/label'], tf.int32)
    height = tf.cast(features['image/height'], tf.int32)
    width = tf.cast(features['image/width'], tf.int32)
    return image, height, width

def get_all_records(FILE):
    with tf.Session() as sess:
        filename_queue = tf.train.string_input_producer([ FILE ])
        image, height, width = read_and_decode(filename_queue)
        image = tf.reshape(image, tf.stack([height, width, 3]))
        image.set_shape([640,480,3])
        init_op = tf.initialize_all_variables()
        sess.run(init_op)
        coord = tf.train.Coordinator()
        threads = tf.train.start_queue_runners(coord=coord)
        for i in range(1):
            example, l = sess.run([image])
            img = Image.fromarray(example, 'RGB')
            img.save( "output/" + str(i) + '-train.png')

            print (example,l)
        coord.request_stop()
        coord.join(threads)


get_all_records(train_record)

创作 :

我做了一个类(class) Image对图像进行逻辑建模，还有一个类 Rect来表示边界框和标签。这不是很相关，但下面的代码在变量 img 时使用了它们。或 rect被看到。

相关部分可能是 get_bytes() -method，它更像是使用 PIL 的 Image.open(file_path) 的包装器:

class Image:

    # ... rest of class 


    def open_img(self):
        if self.file_path is not None:
            return Image.open(self.file_path)

    def get_bytes(self, as_jpg=False):
        if self.file_path is None:
             return None
        if as_jpg:
            # Convert to jpg:
            with BytesIO() as f:
                self.open_img().convert('RGB').save(f, format='JPEG', quality=95)
                return f.getvalue()
        else:  # Assume png
            return np.array(self.open_img().convert('RGB')).tobytes()

我如何创建示例 :

use_jpg = True

def create_tf_example(img):
    image_format= b'jpg' if use_jpg else b'png'
    encoded_image_data = img.get_bytes(as_jpg=use_jpg) # Encoded image bytes

    relative_path = img.get_file_path()
    if relative_path is None or not img.has_person():
        return None  # Ignore images without humans or image data
    else:
        filename = str(Path(relative_path).resolve()) # Absolute filename of the image. Empty if image is not from file

    xmins = []  # List of normalized left x coordinates in bounding box (1 per box)
    xmaxs = []  # List of normalized right x coordinates in bounding box (1 per box)
    ymins = []  # List of normalized top y coordinates in bounding box (1 per box)
    ymaxs = []  # List of normalized bottom y coordinates in bounding box (1 per box)
    classes_text = []  # List of string class name of bounding box (1 per box)
    classes = []  # List of integer class id of bounding box (1 per box)

    for rect in img.rects:
        if not rect.is_person:
            continue  # For now, ignore negative samples as TF does this by default
        else:
            xmin, xmax, ymin, ymax = rect.get_normalized_xy_min_max()
            xmins.append(xmin)
            xmaxs.append(xmax)
            ymins.append(ymin)
            ymaxs.append(ymax)
            # Human class:
            classes.append(1)
            classes_text.append('Human'.encode())

    return tf.train.Example(features=tf.train.Features(feature={
        'image/height': dataset_util.int64_feature(height),
        'image/width': dataset_util.int64_feature(width),
        #  'image/filename': dataset_util.bytes_feature(filename.encode()),
        'image/source_id': dataset_util.bytes_feature(filename.encode()),
        'image/encoded': dataset_util.bytes_feature(encoded_image_data),
        'image/format': dataset_util.bytes_feature(image_format),
        'image/object/bbox/xmin': dataset_util.float_list_feature(xmins),
        'image/object/bbox/xmax': dataset_util.float_list_feature(xmaxs),
        'image/object/bbox/ymin': dataset_util.float_list_feature(ymins),
        'image/object/bbox/ymax': dataset_util.float_list_feature(ymaxs),
        'image/object/class/text': dataset_util.bytes_list_feature(classes_text),
        'image/object/class/label': dataset_util.int64_list_feature(classes),
    }))

我如何创建 TFRecord :

def convert_to_tfrecord(imgs, output_file_path):
    with tf.python_io.TFRecordWriter(output_file_path) as writer:
        for img in imgs:
            tf_example = create_tf_example(img)
            if tf_example is not None:
                writer.write(tf_example.SerializeToString())


convert_to_tfrecord(train_imgs, 'train.record')
convert_to_tfrecord(validation_imgs, 'validate.record')
convert_to_tfrecord(test_imgs, 'test.record')

来自 dataset_util模块:

def int64_feature(value):
    return tf.train.Feature(int64_list=tf.train.Int64List(value=[value]))


def int64_list_feature(value):
    return tf.train.Feature(int64_list=tf.train.Int64List(value=value))


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


def bytes_list_feature(value):
    return tf.train.Feature(bytes_list=tf.train.BytesList(value=value))


def float_list_feature(value):
    return tf.train.Feature(float_list=tf.train.FloatList(value=value))

Answer 1

我通过使用 tf.image.decode_jpeg 将数据解码为 jpeg 解决了这个问题.

代替:

def read_and_decode(filename_queue):
    # ...

    image = tf.decode_raw(features['image/encoded'], tf.uint8)

    # ...

我做了:

def read_and_decode(filename_queue):
    # ...

    image = tf.image.decode_jpeg(features['image/encoded'])

    # ...

这解释了为什么预期大小和给定大小之间的差异如此之大的原因:给定(读取)字节是“仅”压缩的 JPEG 数据，
而不是全尺寸的“完整”位图图像。