import os
import unittest
import tensorflow as tf
import numpy as np
from video2tfrecord import convert_videos_to_tfrecord
from tensorflow.python.platform import gfile
height = 720
width = 1280
num_depth = 4
in_path = "./example/input"
out_path = "./example/output"
n_videos_per_record = 1
class Testvideo2tfrecord(unittest.TestCase):
def test_example1(self):
n_frames = 5
convert_videos_to_tfrecord(source_path=in_path, destination_path=out_path,
n_videos_in_record=n_videos_per_record,
n_frames_per_video=n_frames,
dense_optical_flow=True,
file_suffix="*.mp4")
filenames = gfile.Glob(os.path.join(out_path, "*.tfrecords"))
n_files = len(filenames)
self.assertTrue(filenames)
self.assertEqual(n_files * n_videos_per_record,
get_number_of_records(filenames, n_frames))
" travis ressource exhaust, passes locally for 3.6 and 3.4"
# def test_example2(self):
# n_frames = 'all'
# convert_videos_to_tfrecord(source_path=in_path, destination_path=out_path,
# n_videos_in_record=n_videos_per_record,
# n_frames_per_video=n_frames,
# n_channels=num_depth, dense_optical_flow=False,
# file_suffix="*.mp4")
#
# filenames = gfile.Glob(os.path.join(out_path, "*.tfrecords"))
# n_files = len(filenames)
#
# self.assertTrue(filenames)
# self.assertEqual(n_files * n_videos_per_record,
# get_number_of_records(filenames, n_frames))
def read_and_decode(filename_queue, n_frames):
"""Creates one image sequence"""
reader = tf.TFRecordReader()
_, serialized_example = reader.read(filename_queue)
image_seq = []
if n_frames == 'all':
n_frames = 354 # travis kills due to too large tfrecord
for image_count in range(n_frames):
path = 'blob' + '/' + str(image_count)
feature_dict = {path: tf.FixedLenFeature([], tf.string),
'height': tf.FixedLenFeature([], tf.int64),
'width': tf.FixedLenFeature([], tf.int64),
'depth': tf.FixedLenFeature([], tf.int64)}
features = tf.parse_single_example(serialized_example,
features=feature_dict)
image_buffer = tf.reshape(features[path], shape=[])
image = tf.decode_raw(image_buffer, tf.uint8)
image = tf.reshape(image, tf.stack([height, width, num_depth]))
image = tf.reshape(image, [1, height, width, num_depth])
image_seq.append(image)
image_seq = tf.concat(image_seq, 0)
return image_seq
def get_number_of_records(filenames, n_frames):
"""
this function determines the number of videos available in all tfrecord files. It also checks on the correct shape of the single examples in the tfrecord
files.
:param filenames: a list, each entry containign a (relative) path to one tfrecord file
:return: the number of overall videos provided in the filenames list
"""
num_examples = 0
if n_frames == 'all':
n_frames_in_test_video = 354
else:
n_frames_in_test_video = n_frames
# create new session to determine batch_size for validation/test data
with tf.Session() as sess_valid:
filename_queue_val = tf.train.string_input_producer(filenames, num_epochs=1)
image_seq_tensor_val = read_and_decode(filename_queue_val, n_frames)
init_op = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
sess_valid.run(init_op)
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(coord=coord)
try:
while True:
video = sess_valid.run([image_seq_tensor_val])
assert np.shape(video) == (1, n_frames_in_test_video, height, width,
num_depth), "shape in the data differs from the expected shape"
num_examples += 1
except tf.errors.OutOfRangeError as e:
coord.request_stop(e)
finally:
coord.request_stop()
coord.join(threads)
return num_examples
if __name__ == '__main__':
unittest.main()
