"""Provides data for the ucf11 dataset.
The dataset scripts used to create the dataset can be found at: convert_ucf11.py
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import os
import tensorflow as tf
from datasets import dataset_utils
from collections import namedtuple
_FILE_PATTERN = 'ucf11_%s_*.tfrecord'
SPLITS_TO_SIZES = {'train': 1147, 'test': 449}
_NUM_CLASSES = 11
_ITEMS_TO_DESCRIPTIONS = {
'video': 'A sequent of color image of varying size.',
'label': 'A single integer between 0 and 10',
}
_ALPHA = 2
_BETA = 11
KEY_TO_FEATURES = {
'video/encoded': tf.FixedLenFeature((), tf.string, default_value=''),
# 'video/format': tf.FixedLenFeature((), tf.string, default_value='mpg'),
'video/class/label': tf.FixedLenFeature(
[], tf.int64, default_value=tf.zeros([], dtype=tf.int64)),
'video/frame_count': tf.FixedLenFeature(
[], tf.int64, default_value=tf.zeros([], dtype=tf.int64)),
'video/height': tf.FixedLenFeature(
[], tf.int64, default_value=tf.zeros([], dtype=tf.int64)),
'video/width': tf.FixedLenFeature(
[], tf.int64, default_value=tf.zeros([], dtype=tf.int64)),
'video/channel': tf.FixedLenFeature(
[], tf.int64, default_value=[3]),
}
Dataset = namedtuple('Dataset', ['filenames', 'num_samples', 'num_classes',
'labels_to_names', 'items_to_descriptions'])
def get_split(split_name, dataset_dir, file_pattern=None):
"""Gets a dataset tuple with instructions for reading flowers.
Args:
split_name: A train/validation split name.
dataset_dir: The base directory of the dataset sources.
file_pattern: The file pattern to use when matching the dataset sources.
It is assumed that the pattern contains a '%s' string so that the split
name can be inserted.
reader: The TensorFlow reader type.
Returns:
A `Dataset` namedtuple.
Raises:
ValueError: if `split_name` is not a valid train/validation split.
"""
if split_name not in SPLITS_TO_SIZES:
raise ValueError('split name %s was not recognized.' % split_name)
if not file_pattern:
file_pattern = _FILE_PATTERN
file_pattern = os.path.join(dataset_dir, file_pattern % split_name)
# Allowing None in the signature so that dataset_factory can use the default.
# if reader is None:
# reader = tf.TFRecordReader
labels_to_names = None
if dataset_utils.has_labels(dataset_dir):
labels_to_names = dataset_utils.read_label_file(dataset_dir)
filenames = tf.gfile.Glob(file_pattern)
return Dataset(
filenames=filenames,
num_samples=SPLITS_TO_SIZES[split_name],
num_classes=_NUM_CLASSES,
labels_to_names=labels_to_names,
items_to_descriptions=_ITEMS_TO_DESCRIPTIONS
)
def read_and_decode(filename_queue):
reader = tf.TFRecordReader()
_, serialized_example = reader.read(filename_queue)
features = tf.parse_single_example(serialized_example, KEY_TO_FEATURES)
encoded = tf.decode_raw(features['video/encoded'], tf.uint8)
frame_count = tf.cast(features['video/frame_count'], tf.int32)
label = tf.cast(features['video/class/label'], tf.int32)
height = tf.cast(features['video/height'], tf.int32)
width = tf.cast(features['video/width'], tf.int32)
channel = tf.cast(features['video/channel'], tf.int32)
encoded = tf.reshape(encoded, [frame_count, height, width, channel])
return encoded, label, [frame_count, height, width, channel]
def build_data(dataset, is_training=True):
filename_queue = tf.train.string_input_producer(dataset.filenames)
# Read one video
encoded, label, shape = read_and_decode(filename_queue)
# c = lambda encoded, label, shape: tf.less(shape[0], 25)
# b = lambda encoded, label, shape: read_and_decode(filename_queue)
# encoded, label, shape = tf.while_loop(c, b, (encoded, label, shape))
label = tf.reshape(label, [])
frame_count = shape[0]
delta = frame_count // _BETA
# delta = tf.cond(tf.greater(delta, 0), lambda: delta, lambda: 1)
p_max = frame_count - 1 - (_ALPHA - 1) * delta
if is_training:
p = tf.cast(tf.random_uniform([], minval=0, maxval=tf.cast(p_max, tf.float32)), tf.int32)
index = tf.range(p, p + (_ALPHA - 1) * delta + 1, delta)
exampled_frames = tf.gather(encoded, index)
exampled_frames.set_shape([_ALPHA, 240, 240, 3])
else:
p = 0
index = tf.range(p, p + (_BETA - 1) * delta + 1, delta)
exampled_frames = tf.gather(encoded, index)
exampled_frames.set_shape([_BETA, 240, 240, 3])
return exampled_frames, label
# Make queue
# if mode == 'train':
# example_queue = tf.RandomShuffleQueue(
# capacity=16 * batch_size,
# min_after_dequeue=8 * batch_size,
# dtypes=[tf.float32, tf.int32],
# shapes=[out_shapes, [1]]
# )
# num_threads = 16
# elif mode == 'validate':
# example_queue = tf.FIFOQueue(
# 3 * batch_size,
# dtypes=[tf.float32, tf.int64],
# shapes=[out_shapes, [1]]
# )
# num_threads = 1
# else:
# example_queue = tf.FIFOQueue(
# 3 * batch_size,
# dtypes=[tf.float32, tf.int64],
# shapes=[out_shapes, [1]]
# )
# num_threads = 1
#
# example_queue_op = example_queue.enqueue([extract_frames, label])
# tf.train.add_queue_runner(tf.train.queue_runner.QueueRunner(
# example_queue, enqueue_ops=[example_queue_op] * num_threads))
#
# # Read batch frames, label
# inputs, labels = example_queue.dequeue_many(batch_size)
# labels = tf.reshape(labels, [batch_size])
# labels = tf.one_hot(labels, dataset.num_classes)
#
# assert len(inputs.shape) == 5
# assert inputs.shape[0] == batch_size
# assert inputs.shape[-1] == out_shapes[-1]
# assert inputs.shape[1] == out_shapes[0]
#
# return inputs, labels
