def img_to_tfrecord(train_imgs, train_labels, output_dir, name):
"""
:param image_dir: image_dir just like "data/Challenge2_Training_Task12_Images/*.jpg"
:param text_dir: label file dir
:param text_name: label file name
:param output_dir: output dir
:param name: output file name
:return: NULL
"""
tf_filename = _get_output_filename(output_dir, name)
imgLists = train_imgs # return a list
labels = train_labels
labels_encord, lengths = encode_labels(labels, alphabet)
image_format = b'JPEG'
with tf.python_io.TFRecordWriter(tf_filename) as tfrecord_writer:
for i, filename in enumerate(imgLists):
sys.stdout.write('\r>> Converting image %d/%d' %
(i + 1, len(imgLists)))
sys.stdout.flush()
example = tf.train.Example(features=tf.train.Features(
feature={
"label/value": int64_feature(labels_encord[i]),
"label/length": int64_feature(lengths[i]),
"image": bytes_feature(filename)
}))
tfrecord_writer.write(example.SerializeToString())
print('\nFinished converting the dataset!')
def dict_to_coco_example(img_data):
"""Convert python dictionary formath data of one image to tf.Example proto.
Args:
img_data: infomation of one image, inclue bounding box, labels of bounding box,\
height, width, encoded pixel data.
Returns:
example: The converted tf.Example
"""
bboxes = img_data['bboxes']
xmin, xmax, ymin, ymax = [], [], [], []
for bbox in bboxes:
xmin.append(bbox[0])
xmax.append(bbox[0] + bbox[2])
ymin.append(bbox[1])
ymax.append(bbox[1] + bbox[3])
example = tf.train.Example(features=tf.train.Features(feature={
'image/height': utils.int64_feature(img_data['height']),
'image/width': utils.int64_feature(img_data['width']),
'image/object/bbox/xmin': utils.float_list_feature(xmin),
'image/object/bbox/xmax': utils.float_list_feature(xmax),
'image/object/bbox/ymin': utils.float_list_feature(ymin),
'image/object/bbox/ymax': utils.float_list_feature(ymax),
'image/object/class/label': utils.int64_list_feature(img_data['labels']),
'image/encoded': utils.bytes_feature(img_data['pixel_data']),
'image/format': utils.bytes_feature('jpeg'.encode('utf-8')),
}))
return example
def create_tf_example(filename, encoded_jpeg, annotations, resize=True):
"""
This function create a tf.train.Example from the Waymo frame.
args:
- filename [str]: name of the image
- encoded_jpeg [bytes]: jpeg encoded image
- annotations [protobuf object]: bboxes and classes
returns:
- tf_example [tf.Train.Example]: tf example in the objection detection api format.
"""
if not resize:
encoded_jpg_io = io.BytesIO(encoded_jpeg)
image = Image.open(encoded_jpg_io)
width, height = image.size
width_factor, height_factor = image.size
else:
image_tensor = tf.io.decode_jpeg(encoded_jpeg)
height_factor, width_factor, _ = image_tensor.shape
image_res = tf.cast(tf.image.resize(image_tensor, (640, 640)), tf.uint8)
encoded_jpeg = tf.io.encode_jpeg(image_res).numpy()
width, height = 640, 640
mapping = {1: 'vehicle', 2: 'pedestrian', 4: 'cyclist'}
image_format = b'jpg'
xmins = []
xmaxs = []
ymins = []
ymaxs = []
classes_text = []
classes = []
filename = filename.encode('utf8')
for ann in annotations:
xmin, ymin = ann.box.center_x - 0.5 * ann.box.length, ann.box.center_y - 0.5 * ann.box.width
xmax, ymax = ann.box.center_x + 0.5 * ann.box.length, ann.box.center_y + 0.5 * ann.box.width
xmins.append(xmin / width_factor)
xmaxs.append(xmax / width_factor)
ymins.append(ymin / height_factor)
ymaxs.append(ymax / height_factor)
classes.append(ann.type)
classes_text.append(mapping[ann.type].encode('utf8'))
tf_example = tf.train.Example(features=tf.train.Features(feature={
'image/height': int64_feature(height),
'image/width': int64_feature(width),
'image/filename': bytes_feature(filename),
'image/source_id': bytes_feature(filename),
'image/encoded': bytes_feature(encoded_jpeg),
'image/format': bytes_feature(image_format),
'image/object/bbox/xmin': float_list_feature(xmins),
'image/object/bbox/xmax': float_list_feature(xmaxs),
'image/object/bbox/ymin': float_list_feature(ymins),
'image/object/bbox/ymax': float_list_feature(ymaxs),
'image/object/class/text': bytes_list_feature(classes_text),
'image/object/class/label': int64_list_feature(classes),
}))
return tf_example
def create_tf_example(filename, encoded_jpeg, annotations):
"""
convert to tensorflow object detection API format
args:
- filename [str]: name of the image
- encoded_jpeg [bytes-likes]: encoded image
- annotations [list]: bboxes and classes
returns:
- tf_example [tf.Example]
"""
# TO BE IMPLEMENTED
encoded_jpg_io = io.BytesIO(encoded_jpeg)
image = Image.open(encoded_jpg_io)
width, height = image.size
mapping = {1: 'vehicle', 2: 'pedestrian', 4: 'cyclist'}
image_format = b'jpg'
xmins = []
xmaxs = []
ymins = []
ymaxs = []
classes_text = []
classes = []
filename = filename.encode('utf8')
for ann in annotations:
xmin, ymin = ann.box.center_x - 0.5 * ann.box.length, ann.box.center_y - 0.5 * ann.box.width
xmax, ymax = ann.box.center_x + 0.5 * ann.box.length, ann.box.center_y + 0.5 * ann.box.width
xmins.append(xmin / width)
xmaxs.append(xmax / width)
ymins.append(ymin / height)
ymaxs.append(ymax / height)
classes.append(ann.type)
classes_text.append(mapping[ann.type].encode('utf8'))
tf_example = tf.train.Example(features=tf.train.Features(
feature={
'image/height': int64_feature(height),
'image/width': int64_feature(width),
'image/filename': bytes_feature(filename),
'image/source_id': bytes_feature(filename),
'image/encoded': bytes_feature(encoded_jpeg),
'image/format': bytes_feature(image_format),
'image/object/bbox/xmin': float_list_feature(xmins),
'image/object/bbox/xmax': float_list_feature(xmaxs),
'image/object/bbox/ymin': float_list_feature(ymins),
'image/object/bbox/ymax': float_list_feature(ymaxs),
'image/object/class/text': bytes_list_feature(classes_text),
'image/object/class/label': int64_list_feature(classes)
}))
return tf_example
def create_tf_example(image_buffer, label, width, height):
image_format = b'png'
return tf.train.Example(features=tf.train.Features(
feature={
'image/height': utils.int64_feature(height),
'image/width': utils.int64_feature(width),
'image/encoded': utils.bytes_feature(image_buffer),
'image/format': utils.bytes_feature(image_format),
'image/object/class/label': utils.int64_feature(label)
}))
def _convert_to_example(image, label):
example = tf.train.Example(features=tf.train.Features(
feature={
'image/height': int64_feature(image.shape[1]),
'image/width': int64_feature(image.shape[0]),
'image/ids': int64_feature(label),
# '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/array': bytes_feature(image.tobytes(order='C'))
}))
return example
def save_tfrecord(data):
coder = utils.ImageCoder()
tf.gfile.MakeDirs(FLAGS.output_dir)
for subset in data:
if subset == 'train' or subset == 'extra':
num_shards = 128
else:
num_shards = 64
shard_spacing = np.linspace(0, data[subset]['images'].shape[0],
num_shards + 1).astype(np.int)
cnt = 0
if subset == 'extra':
output_dir = os.path.join(FLAGS.output_dir, 'unlabeled')
tf.gfile.MakeDirs(output_dir)
else:
output_dir = FLAGS.output_dir
for i in range(len(shard_spacing) - 1):
if FLAGS.full_train_data:
output_filename = '%s_full-%.5d-of-%.5d' % (subset, i,
num_shards)
else:
output_filename = '%s-%.5d-of-%.5d' % (subset, i, num_shards)
if i % 10 == 0:
tf.logging.info('saving {}'.format(output_filename))
writer = tf.python_io.TFRecordWriter(
os.path.join(output_dir, output_filename))
for j in range(shard_spacing[i], shard_spacing[i + 1]):
cnt += 1
image = data[subset]['images'][j]
label = data[subset]['labels'][j] + 1
image_data = coder.encode_jpeg(image)
features = {
'image/encoded': utils.bytes_feature(image_data),
'image/class/label': utils.int64_feature(label),
}
if subset == 'extra':
features['probabilities'] = utils.float_feature([0] * 10)
features['label'] = utils.int64_feature(0)
features['prob'] = utils.float_feature(0)
example = tf.train.Example(features=tf.train.Features(
feature=features))
writer.write(example.SerializeToString())
writer.close()
assert cnt == data[subset]['images'].shape[0]
def write_tfexamples(data, tfexample_path):
colorspace = "RGB"
channels = 1
image_format = "JPEG"
height = FLAGS.canvas_size
width = FLAGS.canvas_size
image_coder = utils.ImageCoder(FLAGS.canvas_size)
builder = tf.python_io.TFRecordWriter(tfexample_path)
tf.logging.info("Writing %s TF Example.", tfexample_path)
for index, datum in enumerate(data):
# Convert the contents of the datum into a tf example.
image_encoded = image_coder.encode_image(datum.images)
latents = [float(x) for x in datum.true_latents]
labels = datum.attribute_labels
example = tf.train.Example(features=tf.train.Features(
feature={
"image/height": utils.int64_feature(height),
"image/width": utils.int64_feature(width),
"image/colorspace": utils.string_feature(colorspace.encode()),
"image/channels": utils.int64_feature(channels),
"latents": utils.float_feature(latents),
"labels": utils.int64_feature(labels),
"image/format": utils.string_feature(image_format.encode()),
"image/encoded": utils.bytes_feature(image_encoded)
}))
builder.write(example.SerializeToString())
if index % 1000 == 0:
tf.logging.info("Wrote %d datapoints to the tfexample.", index)
builder.close()
def get_image_list(features):
dump_features = {}
prob = keep_idx[worker_id][cnt][1]
label = keep_idx[worker_id][cnt][0]
include_copy = keep_idx[worker_id][cnt][2]
image_bytes = features['image/encoded'][0]
dump_features['image/encoded'] = utils.bytes_feature(image_bytes)
dump_features['prob'] = utils.float_feature(prob)
dump_features['probabilities'] = utils.float_feature(
keep_idx[worker_id][cnt][3])
dump_features['label'] = utils.int64_feature(label)
example = tf.train.Example(features=tf.train.Features(
feature=dump_features))
cur_image_list = []
for j in range(include_copy):
image_info = {
'example': example,
'label': label,
'prob': prob,
'image_bytes': image_bytes,
'cnt': cnt,
}
cur_image_list += [image_info]
return cur_image_list
def run_prediction(estimator):
global shard_id
shard_id_list = FLAGS.shard_id.split(',')
for cur_shard_id in shard_id_list:
shard_id = int(cur_shard_id)
worker_image_num = get_num_image()
cnt, predict_result_list = predict_on_dataset(estimator,
worker_image_num)
tf.logging.info('predicted on %d images', cnt)
assert cnt == worker_image_num, (cnt, worker_image_num)
tf.gfile.MakeDirs(FLAGS.output_dir)
if FLAGS.reassign_label:
sample_dir = os.path.join(FLAGS.output_dir, 'samples')
uid_list = utils.get_uid_list()
for uid in uid_list:
tf.gfile.MakeDirs(os.path.join(sample_dir, uid))
image_bytes_placeholder = tf.placeholder(dtype=tf.string)
decoded_image = utils.decode_raw_image(image_bytes_placeholder)
raw_dst = get_input_fn({'batch_size': 1}, raw_data=True)
raw_iter = raw_dst.make_initializable_iterator()
raw_elem = raw_iter.get_next()
filename = utils.get_reassign_filename(FLAGS.label_data_dir,
FLAGS.file_prefix, shard_id,
FLAGS.num_shards,
FLAGS.worker_id)
record_writer = tf.python_io.TFRecordWriter(
os.path.join(FLAGS.output_dir, os.path.basename(filename)))
sample_prob = 30000. / (worker_image_num * FLAGS.num_shards)
with tf.Session() as sess:
sess.run(raw_iter.initializer)
for i in range(worker_image_num):
features = sess.run(raw_elem)
encoded_image = features['image/encoded']
features = {}
label = predict_result_list[i]['label']
prob = predict_result_list[i]['prob']
features['image/encoded'] = utils.bytes_feature(
encoded_image)
features['prob'] = utils.float_feature(prob)
features['label'] = utils.int64_feature(label)
features['probabilities'] = utils.float_feature(
predict_result_list[i]['probabilities'])
example = tf.train.Example(features=tf.train.Features(
feature=features))
record_writer.write(example.SerializeToString())
if np.random.random() < sample_prob:
uid = uid_list[label]
filename = os.path.join(
sample_dir, uid,
'image_{:d}_{:d}_{:.2f}.jpeg'.format(
shard_id, i, prob))
tf.logging.info('saving {:s}'.format(filename))
image = sess.run(
decoded_image,
feed_dict={image_bytes_placeholder: encoded_image})
utils.save_pic(image, filename)
record_writer.close()
else:
filename = 'train-info-%.5d-of-%.5d-%.5d' % (
shard_id, FLAGS.num_shards, FLAGS.worker_id)
writer = tf.python_io.TFRecordWriter(
os.path.join(FLAGS.output_dir, filename))
for result in predict_result_list:
features = {}
features['probabilities'] = utils.float_feature(
result['probabilities'])
features['classes'] = utils.int64_feature(result['label'])
example = tf.train.Example(features=tf.train.Features(
feature=features))
writer.write(example.SerializeToString())
writer.close()
def img_to_tfrecord(train_imgs,train_labels,output_dir,name):
"""
:param image_dir: image_dir just like "data/Challenge2_Training_Task12_Images/*.jpg"
:param text_dir: label file dir
:param text_name: label file name
:param output_dir: output dir
:param name: output file name
:return: NULL
"""
tf_filename = _get_output_filename(output_dir, name)
imgLists = train_imgs # return a list
labels = train_labels
labels_encord,lengths = encode_labels(labels)
image_format = b'JPEG'
with tf.python_io.TFRecordWriter(tf_filename) as tfrecord_writer:
for i, filename in enumerate(imgLists):
sys.stdout.write('\r>> Converting image %d/%d' % (i + 1, len(imgLists)))
sys.stdout.flush()
#image_data = tf.gfile.FastGFile(filename, 'rb').read()
image_data = load_image(filename)
#with tf.gfile.GFile(filename, 'rb') as fid:
# image_data = fid.read()
# with tf.Session() as sess:
# image = tf.image.decode_jpeg(image_data)
# image = sess.run(image)
# print(image.shape)#(32, 100, 3)
example = tf.train.Example(features=tf.train.Features(feature={"label/value": int64_feature(labels_encord[i]),
"label/length": int64_feature(lengths[i]),
"image/encoded": bytes_feature(image_data),
'image/format': bytes_feature(image_format)}))
tfrecord_writer.write(example.SerializeToString())
print('\nFinished converting the dataset!')