gentlely way to read tfrecords data into batches

Question

all the follow question are based on tensorflow 1.0 API

I am now be able to write images under directory which named by class name, and this is I generate tfrecords codes:

def _convert_to_example(filename, image_buffer, label, text, height, width):
    colorspace = 'RGB'
    channels = 3
    image_format = 'JPEG'

    example = tf.train.Example(features=tf.train.Features(feature={
      'image/height': _int64_feature(height),
      'image/width': _int64_feature(width),
      'image/colorspace': _bytes_feature(tf.compat.as_bytes(colorspace)),
      'image/channels': _int64_feature(channels),
      'image/class/label': _int64_feature(label),
      'image/class/text': _bytes_feature(tf.compat.as_bytes(text)),
      'image/format': _bytes_feature(tf.compat.as_bytes(image_format)),
      'image/filename': _bytes_feature(tf.compat.as_bytes(os.path.basename(filename))),
      'image/encoded': _bytes_feature(tf.compat.as_bytes(image_buffer))}))
    return example

this is the main method, so here I stored height, widht, channels(this value con't readout) etc.

And I am able read tfrecords out, this is my code:

def read_tfrecords():
    print('reading from tfrecords file {}'.format(FLAGS.record_file))
    record_iterator = tf.python_io.tf_record_iterator(path=FLAGS.record_file)

    with tf.Session() as sess:
        for string_record in record_iterator:
            example = tf.train.Example()
            example.ParseFromString(string_record)

            height_ = int(example.features.feature['image/height'].int64_list.value[0])
            width_ = int(example.features.feature['image/width'].int64_list.value[0])
            channels_ = int(example.features.feature['image/channels'].int64_list.value[0])

            image_bytes_ = example.features.feature['image/encoded'].bytes_list.value[0]
            label_ = int(example.features.feature['image/class/label'].int64_list.value[0])
            text_bytes_ = example.features.feature['image/class/text'].bytes_list.value[0]

            # image_array_ = np.fromstring(image_bytes_, dtype=np.uint8).reshape((height_, width_, 3))
            image_ = tf.image.decode_jpeg(image_bytes_)
            image_ = sess.run(image_)
            text_ = text_bytes_.decode('utf-8')

            print('tfrecords height {0}, width {1}, channels {2}: '.format(height_, width_, channels_))
            print('decode image shape: ', image_.shape)
            print('label text: ', text_)
            print('label: ', label_)
            # io.imshow(image_)
            # plt.show()

All goes fair, however, the problem occurs when I try load tfrecords data into batches and feed it into network

Here is all the code I load batches:

tf.app.flags.DEFINE_integer('target_image_height', 150, 'train input image height')
tf.app.flags.DEFINE_integer('target_image_width', 200, 'train input image width')

tf.app.flags.DEFINE_integer('batch_size', 12, 'batch size of training.')
tf.app.flags.DEFINE_integer('num_epochs', 100, 'epochs of training.')
tf.app.flags.DEFINE_float('learning_rate', 0.01, 'learning rate of training.')

FLAGS = tf.app.flags.FLAGS


def read_and_decode(filename_queue):
    reader = tf.TFRecordReader()
    _, serialized_example = reader.read(filename_queue)
    features = tf.parse_single_example(
        serialized=serialized_example,
        features={
            'image/height': tf.FixedLenFeature([], tf.int64),
            'image/width': tf.FixedLenFeature([], tf.int64),
            'image/channels': tf.FixedLenFeature([], tf.int64),
            'image/encoded': tf.FixedLenFeature([], tf.string),
            'image/class/label': tf.FixedLenFeature([], tf.int64),
        })

    image = tf.decode_raw(features['image/encoded'], out_type=tf.uint8)
    height = tf.cast(features['image/height'], dtype=tf.int32)
    width = tf.cast(features['image/width'], dtype=tf.int32)
    channels = tf.cast(features['image/channels'], dtype=tf.int32)
    label = tf.cast(features['image/class/label'], dtype=tf.int32)

    # cast image int64 to float32 [0, 255] -> [-0.5, 0.5]
    image = tf.cast(image, tf.float32) * (1. / 255) - 0.5
    image_shape = tf.stack([height, width, 3])
    image = tf.reshape(image, image_shape)

    return image, label


def inputs(train, batch_size, num_epochs):
    if not num_epochs:
        num_epochs = None
    filenames = ['./data/tiny_5_tfrecords/train-00000-of-00002',
                 './data/tiny_5_tfrecords/train-00001-of-00002']
    print(filenames)
    with tf.name_scope('input'):
        filename_queue = tf.train.string_input_producer(filenames, num_epochs=num_epochs)
    print(filename_queue)
    image, label = read_and_decode(filename_queue)
    images, sparse_labels = tf.train.shuffle_batch(
        [image, label],
        batch_size=batch_size,
        num_threads=2,
        capacity=1000 + 3 * batch_size,
        min_after_dequeue=1000)

    return images, sparse_labels


def run_training():
    images, labels = inputs(train=True, batch_size=FLAGS.batch_size,
                            num_epochs=FLAGS.num_epochs)

    images = tf.Print(images, [images], message='this is images:')
    images.eval()
    predictions = inference.lenet(images=images, num_classes=5, activation_fn='relu')
    slim.losses.softmax_cross_entropy(predictions, labels)

    total_loss = slim.losses.get_total_loss()
    tf.summary.scalar('loss', total_loss)

    optimizer = tf.train.RMSPropOptimizer(0.001, 0.9)

    train_op = slim.learning.create_train_op(total_loss=total_loss,
                                             optimizer=optimizer,
                                             summarize_gradients=True)
    slim.learning.train(train_op=train_op, save_summaries_secs=20)


def main(_):
    run_training()


if __name__ == '__main__':
    tf.app.run()

I run this program, got this error:

raceback (most recent call last):
  File "train_tiny5_tensorflow.py", line 111, in <module>
    tf.app.run()
  File "/usr/local/lib/python3.6/site-packages/tensorflow/python/platform/app.py", line 44, in run
    _sys.exit(main(_sys.argv[:1] + flags_passthrough))
  File "train_tiny5_tensorflow.py", line 107, in main
    run_training()
  File "train_tiny5_tensorflow.py", line 88, in run_training
    num_epochs=FLAGS.num_epochs)
  File "train_tiny5_tensorflow.py", line 81, in inputs
    min_after_dequeue=1000)
  File "/usr/local/lib/python3.6/site-packages/tensorflow/python/training/input.py", line 1165, in shuffle_batch
    name=name)
  File "/usr/local/lib/python3.6/site-packages/tensorflow/python/training/input.py", line 724, in _shuffle_batch
    dtypes=types, shapes=shapes, shared_name=shared_name)
  File "/usr/local/lib/python3.6/site-packages/tensorflow/python/ops/data_flow_ops.py", line 624, in __init__
    shapes = _as_shape_list(shapes, dtypes)
  File "/usr/local/lib/python3.6/site-packages/tensorflow/python/ops/data_flow_ops.py", line 77, in _as_shape_list
    raise ValueError("All shapes must be fully defined: %s" % shapes)
ValueError: All shapes must be fully defined: [TensorShape([Dimension(None), Dimension(None), Dimension(3)]), TensorShape([])]

apparently, program didn't got tfrecords file at all.

I have tried this: 1. I thought it maybe filenames not right, I change it into both relative path and absolute path, either works; 2. I places tfrecords file just beside script, and write tfrecords file name directly didn't work.

So, basicly, I got this problem:

1. What's the officially and reasonable way to write a as short as possiable program load tfrecords file into batches and feed into network

2. BTW, what's the simplest and elegantest way to write tensorflow layer? slim is a good choice, original way is ugly and complicated!

Answer 1

I don't know if your own answer implies what I'm writing here since I don't fully understand your own answer. However, the issue causing the ValueError is that shuffle_batch needs the image sizes to all be the same and known. Thus what would fix the issue is in the function read_and_decode, simply make image_shape have the same values for all images, for example:
image_shape = tf.stack([FLAGS.target_image_height, FLAGS.target_image_width, 3])
and then using:
image = tf.reshape(image, image_shape)
to make all images have the same size.

Answer 2

For anyone maybe occured same questions, I made some mistakes in above code. Simply do not using decode_raw, using tf.image.decode_jpeg instead, and my code function

def inputs(train, batch_size, num_epochs):
    if not num_epochs:
        num_epochs = None
    filenames = ['./data/tiny_5_tfrecords/train-00000-of-00002',
                 './data/tiny_5_tfrecords/train-00001-of-00002']
    print(filenames)
    with tf.name_scope('input'):
        filename_queue = tf.train.string_input_producer(filenames, num_epochs=num_epochs)
    print(filename_queue)
    image, label = read_and_decode(filename_queue)
    images, sparse_labels = tf.train.shuffle_batch(
        [image, label],
        batch_size=batch_size,
        num_threads=2,
        capacity=1000 + 3 * batch_size,
        min_after_dequeue=1000)

    return images, sparse_labels

I missed a tab for the last 2 lines.