def read_and_decode(filename_queue, label_type, shape):
reader = tf.TFRecordReader()
_, serialized_example = reader.read(filename_queue)
features = tf.parse_single_example(
serialized_example,
features={
'image_raw': tf.FixedLenFeature([], tf.string),
'label_raw': tf.FixedLenFeature([], tf.string),
})
image = tf.decode_raw(features['image_raw'], tf.uint8)
image = tf.cast(image, tf.float32)
image = (image - 127.5) * (1. / 128.0)
image.set_shape([shape * shape * 3])
image = tf.reshape(image, [shape, shape, 3])
label = tf.decode_raw(features['label_raw'], tf.float32)
if label_type == 'cls':
image = tf.image.random_flip_left_right(image)
image = tf.image.random_flip_up_down(image)
label.set_shape([2])
elif label_type == 'bbx':
label.set_shape([4])
elif label_type == 'pts':
label.set_shape([10])
return image, label
def _binary_parse_function_example(serialized_example_protocol):
'''
DESCRIPTION:
This function will deserialize, decompress and then transform
the image and label in the appropriate shape based on the (new) merged
structure of the dataset.
'''
#Parsing the exampe from the binary format
features={
'image': tf.FixedLenFeature((),tf.string),
'label': tf.FixedLenFeature((),tf.string)
}
parsed_feature=tf.parse_single_example(serialized_example_protocol,
features)
#Now setting the appropriate tranformation (decoding and reshape)
height=514
width=513
depth=40
#Decoding the image from biary
image=tf.decode_raw(parsed_feature['image'],tf.float32)#BEWARE of dtype
image.set_shape([depth*height*width])
#Now reshape in usual way since reshape automatically read in c-order
image=tf.reshape(image,[height,width,depth])
#Now decoding the label
target_len=6
label=tf.decode_raw(parsed_feature['label'],tf.float32)
label.set_shape([target_len])
#Reshaping appropriately
label=tf.reshape(label,[target_len,])
#Returing the example tuple finally
return image,label
def read_and_decode(filename_queue):
# input: filename
# output: image, label pair
# setup a TF record reader
reader = tf.TFRecordReader()
_, serialized_example = reader.read(filename_queue)
# list the features we want to extract, i.e., the image and the label
features = tf.parse_single_example(
serialized_example,
features={
'img_raw': tf.FixedLenFeature([], tf.string),
'label_raw': tf.FixedLenFeature([], tf.string),
})
# Decode the training image
# Convert from a scalar string tensor (whose single string has
# length 256*256) to a float tensor
image = tf.decode_raw(features['img_raw'], tf.int64)
image.set_shape([65536])
image_re = tf.reshape(image, (256,256))
# Scale input pixels by 1024
image_re = tf.cast(image_re, tf.float32) * (1. / 1024)
# decode the label image, an image with all 0's except 1's where the left
# ventricle exists
label = tf.decode_raw(features['label_raw'], tf.uint8)
label.set_shape([65536])
label_re = tf.reshape(label, [256,256])
return image_re, label_re
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_raw': tf.FixedLenFeature([], tf.string),
'mask_raw': tf.FixedLenFeature([], tf.string),
}
)
# must be read back as uint8 here
image = tf.decode_raw(features['image_raw'], tf.uint8)
segmentation = tf.decode_raw(features['mask_raw'], tf.uint8)
image.set_shape([224*224*3])
segmentation.set_shape([224*224*1])
image = tf.reshape(image,[224,224,3])
segmentation = tf.reshape(segmentation,[224,224])
rgb = tf.cast(image, tf.float32)
rgb = rgb * (1./255)
rgb = tf.cast(image, tf.float32)
mask = tf.cast(segmentation, tf.float32)
mask = (mask / 255.) * 20
mask = tf.cast(mask, tf.int64)
return rgb, mask
def deserialize(examples_serialized):
"""Called by Dataset.map() to convert batches of records to tensors."""
features = tf.parse_single_example(examples_serialized, feature_map)
users = tf.reshape(tf.decode_raw(
features[movielens.USER_COLUMN], tf.int32), (batch_size,))
items = tf.reshape(tf.decode_raw(
features[movielens.ITEM_COLUMN], tf.uint16), (batch_size,))
if params["use_tpu"] or params["use_xla_for_gpu"]:
items = tf.cast(items, tf.int32) # TPU and XLA disallows uint16 infeed.
if not training:
dupe_mask = tf.reshape(tf.cast(tf.decode_raw(
features[rconst.DUPLICATE_MASK], tf.int8), tf.bool), (batch_size,))
return {
movielens.USER_COLUMN: users,
movielens.ITEM_COLUMN: items,
rconst.DUPLICATE_MASK: dupe_mask,
}
labels = tf.reshape(tf.cast(tf.decode_raw(
features["labels"], tf.int8), tf.bool), (batch_size,))
return {
movielens.USER_COLUMN: users,
movielens.ITEM_COLUMN: items,
}, labels
def read_single_example_and_decode(filename_queue):
tfrecord_options = tf.python_io.TFRecordOptions(tf.python_io.TFRecordCompressionType.ZLIB)
reader = tf.TFRecordReader(options=tfrecord_options)
_, serialized_example = reader.read(filename_queue)
features = tf.parse_single_example(
serialized=serialized_example,
features={
'img_name': tf.FixedLenFeature([], tf.string),
'img_height': tf.FixedLenFeature([], tf.int64),
'img_width': tf.FixedLenFeature([], tf.int64),
'img': tf.FixedLenFeature([], tf.string),
'gtboxes_and_label': tf.FixedLenFeature([], tf.string),
'num_objects': tf.FixedLenFeature([], tf.int64)
}
)
img_name = features['img_name']
img_height = tf.cast(features['img_height'], tf.int32)
img_width = tf.cast(features['img_width'], tf.int32)
img = tf.decode_raw(features['img'], tf.uint8)
img = tf.reshape(img, shape=[img_height, img_width, 3])
gtboxes_and_label = tf.decode_raw(features['gtboxes_and_label'], tf.int32)
gtboxes_and_label = tf.reshape(gtboxes_and_label, [-1, 9])
num_objects = tf.cast(features['num_objects'], tf.int32)
return img_name, img, gtboxes_and_label, num_objects
def build_next_batch_op(self):
reader = tf.TFRecordReader()
_, serialized_experience = reader.read(self.filename_queue)
features = tf.parse_single_example(serialized_experience, features={
'state': tf.FixedLenFeature([], tf.string),
'action': tf.FixedLenFeature([2], tf.float32),
'reward': tf.FixedLenFeature([], tf.float32),
'next_state': tf.FixedLenFeature([], tf.string),
'is_episode_finished': tf.FixedLenFeature([], tf.int64)})
state = tf.decode_raw(features['state'], tf.uint8)
state.set_shape([86*86*4])
action = features['action']
reward = features['reward']
next_state = tf.decode_raw(features['next_state'], tf.uint8)
next_state.set_shape([86*86*4])
is_episode_finished = features['is_episode_finished']
state = tf.reshape(state, [86, 86, 4])
next_state = tf.reshape(next_state, [86, 86, 4])
state_batch, action_batch, reward_batch, next_state_batch, is_episode_finished_batch = tf.train.shuffle_batch(
[state, action, reward, next_state, is_episode_finished], batch_size=self.batch_size, capacity=100,
min_after_dequeue=0)
return state_batch, action_batch, reward_batch, next_state_batch, is_episode_finished_batch
def parse_sequence_example(self, record_string):
features_dict = {
'images_raw': tf.FixedLenFeature([], tf.string),
'label': tf.FixedLenFeature([], tf.int64),
'width': tf.FixedLenFeature([], tf.int64),
'height': tf.FixedLenFeature([], tf.int64),
'depth': tf.FixedLenFeature([], tf.int64),
'sequence_length': tf.FixedLenFeature([], tf.int64)
}
if ADD_GEOLOCATIONS:
features_dict['geo'] = tf.FixedLenFeature([], tf.string)
features = tf.parse_single_example(record_string, features_dict)
images = tf.decode_raw(features['images_raw'], tf.float32)
width = tf.cast(features['width'], tf.int32)
height = tf.cast(features['height'], tf.int32)
depth = tf.cast(features['depth'], tf.int32)
label = tf.cast(features['label'], tf.int32)
sequence_length = tf.cast(features['sequence_length'], tf.int32)
images = tf.reshape(images, [sequence_length, height, width, depth])
if ADD_GEOLOCATIONS:
geo = tf.decode_raw(features['geo'], tf.float32)
geo = tf.reshape(geo, [2, ])
return images, label, geo
else:
return images, label
def read_and_decode(self, filename_queue):
"""
A definition of how TF should read the file record.
Slightly altered version from https://github.com/tensorflow/tensorflow/blob/r0.7/tensorflow/examples/how_tos/ \
reading_data/fully_connected_reader.py
:param filename_queue: The file name queue to be read.
:type filename_queue: tf.QueueBase
:return: The read file data including the image data and depth data.
:rtype: (tf.Tensor, tf.Tensor)
"""
reader = tf.TFRecordReader()
_, serialized_example = reader.read(filename_queue)
features = tf.parse_single_example(
serialized_example,
features={
'image_raw': tf.FixedLenFeature([], tf.string),
'depth_raw': tf.FixedLenFeature([], tf.string),
})
image = tf.decode_raw(features['image_raw'], tf.uint8)
image = tf.reshape(image, [self.height, self.width, self.channels])
image = tf.cast(image, tf.float32) * (1. / 255) - 0.5
depth = tf.decode_raw(features['depth_raw'], tf.float32)
depth = tf.reshape(depth, [self.height, self.width, 1])
return image, depth
def make_readers(file_prefix):
"""
Return states and qvals tensors
:param file_prefix:
:return:
"""
FLOAT_SIZE = 4
states_reader = tf.FixedLengthRecordReader(STATES_HISTORY * N_STATE * FLOAT_SIZE)
next_states_reader = tf.FixedLengthRecordReader(STATES_HISTORY * N_STATE * FLOAT_SIZE)
actions_reader = tf.FixedLengthRecordReader(1)
rewards_reader = tf.FixedLengthRecordReader(FLOAT_SIZE)
_, states = states_reader.read(tf.train.string_input_producer([file_prefix + ".states"]))
_, next_states = next_states_reader.read(tf.train.string_input_producer([file_prefix + ".next_states"]))
_, actions = actions_reader.read(tf.train.string_input_producer([file_prefix + ".actions"]))
_, rewards = rewards_reader.read(tf.train.string_input_producer([file_prefix + ".rewards"]))
states = tf.decode_raw(states, tf.float32, name="decode_states")
states = tf.reshape(states, (STATES_HISTORY * N_STATE, ), name="reshape_states")
next_states = tf.decode_raw(next_states, tf.float32, name="decode_next_states")
next_states = tf.reshape(next_states, (STATES_HISTORY * N_STATE, ), name="reshape_next_states")
actions = tf.decode_raw(actions, tf.int8, name="decode_actions")
actions = tf.reshape(actions, (1, ), name="reshape_actions")
actions = tf.to_int32(actions)
rewards = tf.decode_raw(rewards, tf.float32, name="decode_rewards")
rewards = tf.reshape(rewards, (1, ), name="reshape_qvals")
return states, actions, rewards, next_states
def parser(self, record):
keys_to_features = {
'labels': tf.FixedLenFeature([], tf.string),
'userIds': tf.VarLenFeature(tf.int64),
'itemIds': tf.VarLenFeature(tf.int64),
'user_profiles_indices': tf.FixedLenFeature([], tf.string),
'user_profiles_values': tf.VarLenFeature(tf.int64),
'user_profiles_weights': tf.VarLenFeature(tf.float32),
'user_profiles_shape': tf.FixedLenFeature([2], tf.int64),
'item_profiles_indices': tf.FixedLenFeature([], tf.string),
'item_profiles_values': tf.VarLenFeature(tf.int64),
'item_profiles_weights': tf.VarLenFeature(tf.float32),
'item_profiles_shape': tf.FixedLenFeature([2], tf.int64)
}
parsed = tf.parse_single_example(record, keys_to_features)
labels = tf.reshape(tf.decode_raw(parsed['labels'], tf.float32), [-1, 1])
userIds = tf.sparse_tensor_to_dense(parsed['userIds'])
itemIds = tf.sparse_tensor_to_dense(parsed['itemIds'])
user_profiles_indices = tf.reshape(tf.decode_raw(parsed['user_profiles_indices'], tf.int64), [-1, 2])
user_profiles_values = tf.sparse_tensor_to_dense(parsed['user_profiles_values'])
user_profiles_weights = tf.sparse_tensor_to_dense(parsed['user_profiles_weights'])
user_profiles_shape = parsed['user_profiles_shape']
item_profiles_indices = tf.reshape(tf.decode_raw(parsed['item_profiles_indices'], tf.int64), [-1, 2])
item_profiles_values = tf.sparse_tensor_to_dense(parsed['item_profiles_values'])
item_profiles_weights = tf.sparse_tensor_to_dense(parsed['item_profiles_weights'])
item_profiles_shape = parsed['item_profiles_shape']
return labels, userIds, itemIds, \
user_profiles_indices, user_profiles_values, user_profiles_weights, user_profiles_shape, \
item_profiles_indices, item_profiles_values, item_profiles_weights, item_profiles_shape
def buildSpImageConverter(channelOrder, img_dtype):
"""
Convert a imageIO byte encoded image into a image tensor suitable as input to ConvNets
The name of the input must be a subset of those specified in `image.imageIO.imageSchema`.
:param img_dtype: the type of data the underlying image bytes represent
"""
with IsolatedSession() as issn:
# Flat image data -> image dimensions
# This has to conform to `imageIO.imageSchema`
height = tf.placeholder(tf.int32, [], name="height")
width = tf.placeholder(tf.int32, [], name="width")
num_channels = tf.placeholder(tf.int32, [], name="nChannels")
image_buffer = tf.placeholder(tf.string, [], name="data")
# The image is packed into bytes with height as leading dimension
# This is the default behavior of Python Image Library
shape = tf.reshape(tf.stack([height, width, num_channels], axis=0),
shape=(3,), name='shape')
if img_dtype == 'uint8':
image_uint8 = tf.decode_raw(image_buffer, tf.uint8, name="decode_raw")
image_float = tf.to_float(image_uint8)
elif img_dtype == 'float32':
image_float = tf.decode_raw(image_buffer, tf.float32, name="decode_raw")
else:
raise ValueError('''unsupported image data type "%s", currently only know how to
handle uint8 and float32''' % img_dtype)
image_reshaped = tf.reshape(image_float, shape, name="reshaped")
image_reshaped = imageIO.fixColorChannelOrdering(channelOrder, image_reshaped)
image_input = tf.expand_dims(image_reshaped, 0, name="image_input")
gfn = issn.asGraphFunction([height, width, image_buffer, num_channels], [image_input])
return gfn
def read_to_numpy(self, file_name, data_type=None):
"""
Reads entire TFRecords file as NumPy.
:param file_name: The TFRecords file name to read.
:type file_name: str
:param data_type: Data type of data. Used if that data type doesn't include things like labels.
:type data_type: str
:return: The images and labels NumPy
:rtype: (np.ndarray, np.ndarray)
"""
feature_types = self.attain_feature_types(data_type)
images = []
labels = []
for tfrecord in tf.python_io.tf_record_iterator(file_name):
with tf.Graph().as_default() as graph: # Create a separate as this runs slow when on one graph.
features = tf.parse_single_example(tfrecord, features=feature_types)
image_shape, label_shape = self.extract_shapes_from_tfrecords_features(features, data_type)
flat_image = tf.decode_raw(features['image_raw'], tf.uint8)
image_tensor = tf.reshape(flat_image, image_shape)
image_tensor = tf.squeeze(image_tensor)
if data_type != 'deploy':
flat_label = tf.decode_raw(features['label_raw'], tf.float32)
label_tensor = tf.reshape(flat_label, label_shape)
label_tensor = tf.squeeze(label_tensor)
else:
label_tensor = tf.constant(-1.0, dtype=tf.float32, shape=[1, 1, 1])
with tf.Session(graph=graph) as session:
initialize_op = tf.global_variables_initializer()
session.run(initialize_op)
image, label = session.run([image_tensor, label_tensor])
images.append(image)
labels.append(label)
return np.stack(images), np.stack(labels)
def create_image_and_label_inputs_from_file_name_queue(self, file_name_queue, data_type=None):
"""
Creates the inputs for the image and label for a given file name queue.
:param file_name_queue: The file name queue to be used.
:type file_name_queue: tf.Queue
:param data_type: The type of data (train, validation, test, deploy, etc) to determine how to process.
:type data_type: str
:return: The image and label inputs.
:rtype: (tf.Tensor, tf.Tensor)
"""
reader = tf.TFRecordReader()
_, serialized_example = reader.read(file_name_queue)
feature_types = self.attain_feature_types(data_type)
features = tf.parse_single_example(serialized_example, features=feature_types)
image_shape, label_shape = self.extract_shapes_from_tfrecords_features(features, data_type)
flat_image = tf.decode_raw(features['image_raw'], tf.uint8)
image = tf.reshape(flat_image, image_shape)
if data_type != 'deploy':
flat_label = tf.decode_raw(features['label_raw'], tf.float32)
label = tf.reshape(flat_label, label_shape)
else:
# Makes a fake label tensor for preprocessing to work on.
label = tf.constant(-1.0, dtype=tf.float32, shape=[1, 1, 1])
return image, label
def read_decode_tfrecord_list(file_list, do_augment = False):
''''Read TFRecord content'''
reader = tf.TFRecordReader()
_, serialized_example = reader.read(file_list)
features = tf.parse_single_example(
serialized_example,
# Defaults are not specified since both keys are required.
features={
'image': tf.FixedLenFeature([], tf.string),
'shape': tf.FixedLenFeature([], tf.string),
'label': tf.FixedLenFeature([], tf.float32),
})
shape = tf.decode_raw(features['shape'], tf.uint8)
#print('Shape (shape) is:', shape.shape)
image = tf.decode_raw(features['image'], tf.uint8)
#print('Shape (image) is:', image.shape)
label = tf.cast(features['label'], tf.float32)
# TODO: Infer from shape field from TFRecord
image.set_shape([256* 256* 3])
image = tf.reshape(image, [256, 256, 3])
image, label = process_features(image, label, do_augment)
return image, label
def batch_parse_tf_example(batch_size, example_batch):
'''
Args:
example_batch: a batch of tf.Example
Returns:
A dict of batched tensors
'''
features = {
'x': tf.FixedLenFeature([], tf.string),
'pi': tf.FixedLenFeature([], tf.string),
'outcome': tf.FixedLenFeature([], tf.float32),
}
parsed = tf.parse_example(example_batch, features)
x = tf.decode_raw(parsed['x'], tf.uint8)
x = tf.cast(x, tf.float32)
x = tf.reshape(x, [batch_size, go.N, go.N,
features_lib.NEW_FEATURES_PLANES])
pi = tf.decode_raw(parsed['pi'], tf.float32)
pi = tf.reshape(pi, [batch_size, go.N * go.N + 1])
outcome = parsed['outcome']
outcome.set_shape([batch_size])
return {
'pos_tensor': x,
'pi_tensor': pi,
'value_tensor': outcome,
}
def tfrecord_to_graph_ops(filenames, num_epochs):
file_queue = tf.train.string_input_producer(
filenames, name='file_queue', num_epochs=num_epochs
)
reader = tf.TFRecordReader(
options=tf.python_io.TFRecordOptions(
compression_type=tf.python_io.TFRecordCompressionType.GZIP
)
)
_, tfrecord = reader.read(file_queue)
tfrecord_features = tf.parse_single_example(
tfrecord,
features={
'images': tf.FixedLenFeature([], tf.string),
'labels': tf.FixedLenFeature([], tf.string),
},
name='data'
)
tfeat = tf.decode_raw(tfrecord_features['images'], tf.uint8)
# note, 'NCHW' is only supported on GPUs, so use 'NHWC'...
tfeat = tf.reshape(tfeat, [-1, 28, 28, 1])
ttarg = tf.decode_raw(tfrecord_features['labels'], tf.uint8)
ttarg = tf.one_hot(indices=ttarg, depth=10, on_value=1, off_value=0)
return tfeat, ttarg
def read_data(filename_queue, shape):
""" reads data from tfrecord files.
Args:
filename_queue: A que of strings with filenames
shape: image shape
Returns:
frames: the frame data in size (batch_size, image height, image width, frames)
"""
reader = tf.TFRecordReader()
key, serialized_example = reader.read(filename_queue)
features = tf.parse_single_example(
serialized_example,
features={
'image':tf.FixedLenFeature([],tf.string),
'mask':tf.FixedLenFeature([],tf.string)
})
image = tf.decode_raw(features['image'], tf.uint8)
mask = tf.decode_raw(features['mask'], tf.uint8)
image = tf.reshape(image, [shape[0], shape[1], 1])
mask = tf.reshape(mask, [shape[0], shape[1], 1])
image = tf.to_float(image)
mask = tf.to_float(mask)
image_mean = tf.reduce_mean(image)
image = image - image_mean
#image = image / 255.0
mask = mask / 255.0
return image, mask
def read_image(file_queue):
reader = tf.TFRecordReader()
# key, value = reader.read(file_queue)
_, serialized_example = reader.read(file_queue)
features = tf.parse_single_example(
serialized_example,
features={
'label': tf.FixedLenFeature([], tf.string),
'image_raw': tf.FixedLenFeature([], tf.string)
})
image = tf.decode_raw(features['image_raw'], tf.uint8)
# print('image ' + str(image))
image = tf.reshape(image, [INPUT_IMG_WIDE, INPUT_IMG_HEIGHT, INPUT_IMG_CHANNEL])
# image = tf.image.convert_image_dtype(image, dtype=tf.float32)
# image = tf.image.resize_images(image, (IMG_HEIGHT, IMG_WIDE))
# image = tf.cast(image, tf.float32) * (1. / 255) - 0.5
label = tf.decode_raw(features['label'], tf.uint8)
# label = tf.cast(label, tf.int64)
label = tf.reshape(label, [OUTPUT_IMG_WIDE, OUTPUT_IMG_HEIGHT])
# label = tf.decode_raw(features['image_raw'], tf.uint8)
# print(label)
# label = tf.reshape(label, shape=[1, 4])
return image, label
def read_raw_images(data_set):
dirs = './data/'+data_set+'/'
filename = list_binary_files(dirs)
print filename
filename_queue = tf.train.string_input_producer(filename)
if data_set is 'train':
image_bytes = FLAGS.height * FLAGS.width * FLAGS.depth
record_bytes = image_bytes + 1
reader = tf.FixedLengthRecordReader(record_bytes=record_bytes)
key, value = reader.read(filename_queue)
record_bytes = tf.decode_raw(value, tf.uint8)
label = tf.cast(tf.slice(record_bytes, [0], [1]), tf.int32)
depth_major = tf.reshape(tf.slice(record_bytes, [1], [image_bytes]),[FLAGS.depth, FLAGS.height, FLAGS.width])
uint8image = tf.transpose(depth_major, [1, 2, 0])
return label, uint8image
elif data_set is 'test':
image_bytes = FLAGS.height * FLAGS.width * FLAGS.depth
record_bytes = image_bytes + 1
reader = tf.FixedLengthRecordReader(record_bytes=record_bytes)
key, value = reader.read(filename_queue)
record_bytes = tf.decode_raw(value, tf.uint8)
depth_major = tf.reshape(tf.slice(record_bytes, [0], [image_bytes]),
[FLAGS.depth, FLAGS.height, FLAGS.width])
uint8image = tf.transpose(depth_major, [1, 2, 0])
return uint8image
def get_batch():
'''Makes batch queues from the training data.
Returns:
A Tuple of x (Tensor), y (Tensor).
x and y have the shape [batch_size, maxlen].
'''
import tensorflow as tf
# Load data
X, Y = load_train_data()
# Create Queues
x, y = tf.train.slice_input_producer([tf.convert_to_tensor(X),
tf.convert_to_tensor(Y)])
x = tf.decode_raw(x, tf.int32)
y = tf.decode_raw(y, tf.int32)
x, y = tf.train.batch([x, y],
shapes=[(None,), (None,)],
num_threads=8,
batch_size=hp.batch_size,
capacity=hp.batch_size * 64,
allow_smaller_final_batch=False,
dynamic_pad=True)
num_batch = len(X) // hp.batch_size
return x, y, num_batch # (N, None) int32, (N, None) int32, ()
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={
'height': tf.FixedLenFeature([], tf.int64),
'width': tf.FixedLenFeature([], tf.int64),
'image_raw': tf.FixedLenFeature([], tf.string),
'mask_raw': tf.FixedLenFeature([], tf.string)
})
# Convert from a scalar string tensor (whose single string has
# length mnist.IMAGE_PIXELS) to a uint8 tensor with shape
# [mnist.IMAGE_PIXELS].
image = tf.decode_raw(features['image_raw'], tf.uint8)
annotation = tf.decode_raw(features['mask_raw'], tf.uint8)
height = tf.cast(features['height'], tf.int32)
width = tf.cast(features['width'], tf.int32)
image_shape = tf.stack([height, width, 3])
annotation_shape = tf.stack([height, width, 3])
image = tf.reshape(image, image_shape)
annotation = tf.reshape(annotation, annotation_shape)
image_size_const = tf.constant(
(IMAGE_HEIGHT, IMAGE_WIDTH, 3), dtype=tf.int32)
annotation_size_const = tf.constant(
(IMAGE_HEIGHT, IMAGE_WIDTH, 3), dtype=tf.int32)
# Random transformations can be put here: right before you crop images
# to predefined size. To get more information look at the stackoverflow
# question linked above.
resized_image = tf.image.resize_image_with_crop_or_pad(image=image,
target_height=IMAGE_HEIGHT,
target_width=IMAGE_WIDTH)
resized_annotation = tf.image.resize_image_with_crop_or_pad(image=annotation,
target_height=IMAGE_HEIGHT,
target_width=IMAGE_WIDTH)
images, annotations = tf.train.shuffle_batch([resized_image, resized_annotation],
batch_size=2,
capacity=30,
num_threads=2,
min_after_dequeue=10)
return images, annotations
def parser(self, record):
keys_to_features = {
'attention_news_indices': tf.FixedLenFeature([], tf.string),
'attention_news_values': tf.VarLenFeature(tf.float32),
'attention_news_shape': tf.FixedLenFeature([2], tf.int64),
'attention_user_indices': tf.FixedLenFeature([], tf.string),
'attention_user_values': tf.VarLenFeature(tf.int64),
'attention_user_weights': tf.VarLenFeature(tf.float32),
'attention_user_shape': tf.FixedLenFeature([2], tf.int64),
'fm_feat_indices': tf.FixedLenFeature([], tf.string),
'fm_feat_val': tf.VarLenFeature(tf.float32),
'fm_feat_shape': tf.FixedLenFeature([2], tf.int64),
'labels': tf.FixedLenFeature([], tf.string),
'dnn_feat_indices': tf.FixedLenFeature([], tf.string),
'dnn_feat_values': tf.VarLenFeature(tf.int64),
'dnn_feat_weight': tf.VarLenFeature(tf.float32),
'dnn_feat_shape': tf.FixedLenFeature([2], tf.int64),
}
parsed = tf.parse_single_example(record, keys_to_features)
attention_news_indices = tf.reshape(tf.decode_raw(parsed['attention_news_indices'], \
tf.int64), [-1, 2])
attention_news_values = tf.sparse_tensor_to_dense(parsed['attention_news_values'])
attention_news_shape = parsed['attention_news_shape']
attention_user_indices = tf.reshape(tf.decode_raw(parsed['attention_user_indices'], \
tf.int64), [-1, 2])
attention_user_values = tf.sparse_tensor_to_dense(parsed['attention_user_values'])
attention_user_weights = tf.sparse_tensor_to_dense(parsed['attention_user_weights'])
attention_user_shape = parsed['attention_user_shape']
fm_feat_indices = tf.reshape(tf.decode_raw(parsed['fm_feat_indices'], \
tf.int64), [-1, 2])
fm_feat_val = tf.sparse_tensor_to_dense(parsed['fm_feat_val'])
fm_feat_shape = parsed['fm_feat_shape']
labels = tf.reshape(tf.decode_raw(parsed['labels'], tf.float32), [-1, 1])
dnn_feat_indices = tf.reshape(tf.decode_raw(parsed['dnn_feat_indices'], \
tf.int64), [-1, 2])
dnn_feat_values = tf.sparse_tensor_to_dense(parsed['dnn_feat_values'])
dnn_feat_weight = tf.sparse_tensor_to_dense(parsed['dnn_feat_weight'])
dnn_feat_shape = parsed['dnn_feat_shape']
return (attention_news_indices, attention_news_values, attention_news_shape, \
attention_user_indices, attention_user_values, attention_user_weights, \
attention_user_shape, fm_feat_indices, fm_feat_val, \
fm_feat_shape, labels, dnn_feat_indices, dnn_feat_values, \
dnn_feat_weight, dnn_feat_shape)
def read_tfrecord_and_decode_into_image_annotation_pair_tensors(tfrecord_filenames_queue):
"""Return image/annotation tensors that are created by reading tfrecord file.
The function accepts tfrecord filenames queue as an input which is usually
can be created using tf.train.string_input_producer() where filename
is specified with desired number of epochs. This function takes queue
produced by aforemention tf.train.string_input_producer() and defines
tensors converted from raw binary representations into
reshaped image/annotation tensors.
Parameters
----------
tfrecord_filenames_queue : tfrecord filename queue
String queue object from tf.train.string_input_producer()
Returns
-------
image, annotation : tuple of tf.int32 (image, annotation)
Tuple of image/annotation tensors
"""
reader = tf.TFRecordReader()
_, serialized_example = reader.read(tfrecord_filenames_queue)
features = tf.parse_single_example(
serialized_example,
features={
'height': tf.FixedLenFeature([], tf.int64),
'width': tf.FixedLenFeature([], tf.int64),
'image_raw': tf.FixedLenFeature([], tf.string),
'mask_raw': tf.FixedLenFeature([], tf.string)
})
image = tf.decode_raw(features['image_raw'], tf.uint8)
annotation = tf.decode_raw(features['mask_raw'], tf.uint8)
height = tf.cast(features['height'], tf.int32)
width = tf.cast(features['width'], tf.int32)
image_shape = tf.stack([height, width, 3])
# The last dimension was added because
# the tf.resize_image_with_crop_or_pad() accepts tensors
# that have depth. We need resize and crop later.
# TODO: See if it is necessary and probably remove third
# dimension
annotation_shape = tf.stack([height, width, 1])
image = tf.reshape(image, image_shape)
annotation = tf.reshape(annotation, annotation_shape)
return image, annotation
def parse_example_proto(example_serialized):
"""Parses an Example proto containing a training example of an image.
The output of the build_image_data.py image preprocessing script is a dataset
containing serialized Example protocol buffers. Each Example proto contains
the following fields:
Args:
example_serialized: scalar Tensor tf.string containing a serialized
Example protocol buffer.
Returns:
image_buffer: Tensor tf.string containing the contents of a JPEG file.
label: Tensor tf.int32 containing the label.
bbox: 3-D float Tensor of bounding boxes arranged [1, num_boxes, coords]
where each coordinate is [0, 1) and the coordinates are arranged as
[ymin, xmin, ymax, xmax].
text: Tensor tf.string containing the human-readable label.
"""
# Dense features in Example proto.
feature_map = {
'height': tf.FixedLenFeature((), tf.int64),
'width': tf.FixedLenFeature((), tf.int64),
'channel': tf.FixedLenFeature((), tf.int64),
'label': tf.FixedLenFeature((), tf.int64),
'label_depth': tf.FixedLenFeature((), tf.int64),
'label_one_hot_raw': tf.FixedLenFeature((), tf.string),
'image_raw': tf.FixedLenFeature((), tf.string),
'location_raw': tf.FixedLenFeature((), tf.string)}
#sparse_float32 = tf.VarLenFeature(dtype=tf.float32)
# Sparse features in Example proto.
features = tf.parse_single_example(example_serialized, feature_map)
image_raw = tf.decode_raw(features["image_raw"], tf.uint8)
image = tf.reshape(image_raw, [64, 64, 3])
label = tf.cast(features['label'], dtype=tf.int32)
label_one_hot = tf.decode_raw(features['label_one_hot_raw'], tf.float64)
location = tf.decode_raw(features['location_raw'], tf.int64)
# Note that we impose an ordering of (y, x) just to make life difficult.
#bbox = tf.concat(axis=0, values=[ymin, xmin, ymax, xmax])
# Force the variable number of bounding boxes into the shape
# [1, num_boxes, coords].
#bbox = tf.expand_dims(bbox, 0)
#bbox = tf.transpose(bbox, [0, 2, 1])
return image, location, label_one_hot
def decode(self,batched_serialized_tensors,batch_size):
"""Decodes the input from batch of serialized tensors
Formats and reshapes image
Args:
batched_serialized_tensors: tensor output from Batcher containing read in
serialized tensors
Returns:
batched_decoded_tensors: dict of batches of decoded TFRecords of batch_size
"""
#faster to decode tensors as a batch
batched_decoded_tensors = tf.parse_example(batched_serialized_tensors[fields.InputDataFields.serialized],
self._keys_to_features)
#Decode and cast tensors if needed
for label in self._multi_task_labels:
tensor = batched_decoded_tensors[label.name]
#only strings need t obe decoded
if label.dtype == "string":
if label.decodetype:
tensor = tf.decode_raw(tensor, TYPE_MAP[label.decodetype])
else:
raise ValueError("string type must have a type to be decoded to.")
if label.casttype:
tensor = tf.cast(tensor, TYPE_MAP[label.casttype])
if label.shape:
tensor = tf.reshape(tensor, [batch_size,*label.shape])
tensor.set_shape([batch_size, *label.shape])
batched_decoded_tensors[label.name] = tensor
#input is handlded separately
image_float = tf.cast(
tf.decode_raw(batched_decoded_tensors['input'],
tf.uint8),
tf.float32)
image_float = tf.reshape(image_float,[batch_size,
self._image_height,
self._image_width,
self._channels])
image_float.set_shape([batch_size,
self._image_height,
self._image_width,
self._channels])
batched_decoded_tensors['input'] = image_float
return batched_decoded_tensors
def read_and_decode(filename, is_train):
filename_queue = tf.train.string_input_producer([filename])
reader = tf.TFRecordReader()
_,serialized_example = reader.read(filename_queue)
if is_train == True:
features = tf.parse_single_example(serialized_example,
features={
"hat_label": tf.FixedLenFeature([], tf.int64),
"hair_label": tf.FixedLenFeature([], tf.int64),
"gender_label": tf.FixedLenFeature([], tf.int64),
"top_label": tf.FixedLenFeature([], tf.int64),
"down_label": tf.FixedLenFeature([], tf.int64),
"shoes_label": tf.FixedLenFeature([], tf.int64),
"bag_label": tf.FixedLenFeature([], tf.int64),
"img_raw": tf.FixedLenFeature([], tf.string),
})
img = tf.decode_raw(features['img_raw'], tf.uint8)
img = tf.reshape(img, [128, 256, 3])
#image = Image.frombytes('RGB', (224, 224), img[0])
img = tf.cast(img, tf.float32) * (1. / 255) - 0.5
#print(type(img))
#img = np.asarray(img, dtype=np.uint8)
#print(type(img))
#tl.visualize.frame(I=img, second=5, saveable=False, name='frame', fig_idx=12836)
hat_label = tf.cast(features['hat_label'], tf.int32)
hair_label = tf.cast(features['hair_label'], tf.int32)
gender_label = tf.cast(features['gender_label'], tf.int32)
top_label = tf.cast(features['top_label'], tf.int32)
down_label = tf.cast(features['down_label'], tf.int32)
shoes_label = tf.cast(features['shoes_label'], tf.int32)
bag_label = tf.cast(features['bag_label'], tf.int32)
labels = {"hat":hat_label, "hair":hair_label, "gender":gender_label,
"top":top_label, "down":down_label, "shoes":shoes_label,
"bag":bag_label}
return img, labels
else:
features = tf.parse_single_example(serialized_example,
features={
"img_raw": tf.FixedLenFeature([], tf.string),
})
img = tf.decode_raw(features['img_raw'], tf.uint8)
img = tf.reshape(img, [128, 256, 3])
img = tf.cast(img, tf.float32) * (1. / 255) - 0.5
#tl.visualize.frame(I=img, second=5, saveable=False, name='frame', fig_idx=12833)
return img
def parse_function(planes, probs, winner):
"""
Convert unpacked record batches to tensors for tensorflow training
"""
planes = tf.decode_raw(planes, tf.uint8)
probs = tf.decode_raw(probs, tf.float32)
winner = tf.decode_raw(winner, tf.float32)
planes = tf.to_float(planes)
planes = tf.reshape(planes, (ChunkParser.BATCH_SIZE, 112, 8*8))
probs = tf.reshape(probs, (ChunkParser.BATCH_SIZE, 1858))
winner = tf.reshape(winner, (ChunkParser.BATCH_SIZE, 1))
return (planes, probs, winner)
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={
'vector': tf.FixedLenFeature([], tf.string),
'label': tf.FixedLenFeature([], tf.int64),
})
# features = tf.parse_single_example(serialized_example, dense_keys=['vector', 'label'], dense_types=[tf.string, tf.int64])
# Convert from a scalar string tensor (whose single string has
# length tf_model.IMAGE_PIXELS) to a uint8 tensor with shape
# [tf_model.IMAGE_PIXELS].
image = tf.decode_raw(features['vector'], tf.float32)
image.set_shape([FEATURE_DIMENSIONALITY])
if FLAGS.transpose_input:
image = tf.reshape(image, FEATURE_INPUT_SHAPE)
image = tf.transpose(image, perm=[0,2,1])
image = tf.reshape(image, [-1])
# print("Image shape is %s" %(image.shape))
# OPTIONAL: Could reshape into a 28x28 image and apply distortions
# here. Since we are not applying any distortions in this
# example, and the next step expects the image to be flattened
# into a vector, we don't bother.
# Convert from [0, 255] -> [-0.5, 0.5] floats.
# image = tf.cast(image, tf.float32) * (1. / 255) - 0.5
# Convert label from a scalar uint8 tensor to an int32 scalar.
label = tf.cast(features['label'], tf.int32)
return image, label
def read_and_decode(filename_queue):
reader = tf.TFRecordReader()
_, serialized_example = reader.read(filename_queue)
features = tf.parse_single_example(
serialized_example,
dense_keys=['image_raw', 'label'],
# Defaults are not specified since both keys are required.
dense_types=[tf.string, tf.int64])
# Convert from a scalar string tensor (whose single string has
# length mnist.IMAGE_PIXELS) to a uint8 tensor with shape
# [mnist.IMAGE_PIXELS].
image = tf.decode_raw(features['image_raw'], tf.uint8)
image.set_shape([mnist.IMAGE_PIXELS])
# OPTIONAL: Could reshape into a 28x28 image and apply distortions
# here. Since we are not applying any distortions in this
# example, and the next step expects the image to be flattened
# into a vector, we don't bother.
# Convert from [0, 255] -> [-0.5, 0.5] floats.
image = tf.cast(image, tf.float32) * (1. / 255) - 0.5
# Convert label from a scalar uint8 tensor to an int32 scalar.
label = tf.cast(features['label'], tf.int32)
return image, label
def DecodeLabelAndImage(r):
r = tf.decode_raw(r, tf.uint8)
return tf.to_float(
tf.transpose(tf.reshape(r[1:], [3, 32, 32]),
[1, 2, 0])) / 255.0, tf.to_int32(r[0])
def extract_frame_level_features_per_tf_record(frame_file_path,
maximum_iter=False,
stop_at_iter=10):
'''
Extraction of Youtube tfrecords frame file features.
Args:
path to each tf_record (note: developed with assumption of storing on s3 bucket and assessing with glob)
maximum_iter - flag- if True, will limit number of videos extracted from each TF record
stop_at_iter - number of videos to extract
num_tf_records - number of records to extract - WARNING!!! this is VERY slow, if bigger than 1
Assumes each video in the tfrecord has following features:
'id' : bytes_list
'labels' : int64_list
'audio': float arr, each frame 128
'rgb', float arr, each frame 1024
returns:
numpy arrays of frame ids, frame multi-labels, frame audio, frame rgb
'''
frame_ids = []
frame_labels = []
feat_rgb = []
feat_audio = []
# ATTENTION: only use one TF record for debugging.
print(
f'There is {sum(1 for _ in tf.python_io.tf_record_iterator(frame_file_path))} videos in this TF record.'
)
iter_ = 0
for example in tf.python_io.tf_record_iterator(frame_file_path):
if maximum_iter and iter_ == stop_at_iter:
break
tf_example = tf.train.Example.FromString(example)
frame_ids.append(
tf_example.features.feature['id'].bytes_list.value[0].decode(
encoding='UTF-8'))
frame_labels.append(
tf_example.features.feature['labels'].int64_list.value)
tf_seq_example = tf.train.SequenceExample.FromString(example)
n_frames = len(
tf_seq_example.feature_lists.feature_list['audio'].feature)
rgb_frame = []
audio_frame = []
# iterate through frames
sys.stdout.flush()
for i in range(n_frames):
sess = tf.InteractiveSession()
sys.stdout.write('\r' + 'iterating video: ' + str(iter_) +
' ,frames: ' + str(i) + '/' + str(n_frames))
sys.stdout.flush()
rgb_frame.append(
tf.cast(
tf.decode_raw(
tf_seq_example.feature_lists.feature_list['rgb'].
feature[i].bytes_list.value[0], tf.uint8),
tf.float32).eval())
audio_frame.append(
tf.cast(
tf.decode_raw(
tf_seq_example.feature_lists.feature_list['audio'].
feature[i].bytes_list.value[0], tf.uint8),
tf.float32).eval())
tf.reset_default_graph()
sess.close()
feat_rgb.append(rgb_frame)
feat_audio.append(audio_frame)
iter_ += 1
return frame_ids, frame_labels, feat_rgb, feat_audio
def decode_label(label):
label = tf.decode_raw(label, tf.uint8) # tf.string -> [tf.uint8]
label = tf.reshape(label, []) # label is a scalar
return tf.to_int32(label)
def _gt_boxes_decoder(keys_to_tensors): bboxes = tf.decode_raw(keys_to_tensors['label/gt_boxes'], tf.float32) instances = tf.cast(keys_to_tensors['label/num_instances'], tf.int32) bboxes_shape = tf.stack([instances, 5]) return tf.reshape(bboxes, bboxes_shape)
def video_parse_function(example_proto):
"""Parses and preprocesses the features from a video tfrecord."""
features = {
"video":
tf.VarLenFeature(dtype=tf.string),
"video_length":
tf.FixedLenFeature(dtype=tf.int64, shape=[], default_value=10),
"video_height":
tf.FixedLenFeature(dtype=tf.int64, shape=[], default_value=128),
"video_width":
tf.FixedLenFeature(dtype=tf.int64, shape=[], default_value=128),
"video_channels":
tf.FixedLenFeature(dtype=tf.int64, shape=[], default_value=1),
}
parsed_features = tf.parse_single_example(example_proto, features)
video = tf.sparse_tensor_to_dense(parsed_features["video"],
default_value="")
video = tf.cast(tf.decode_raw(video, tf.uint8), tf.float32)
# Rescale video from [0, 255] to [-1, 1]
video = tf.reshape(
video,
tf.stack([
parsed_features["video_length"],
parsed_features["video_height"],
parsed_features["video_width"],
parsed_features["video_channels"]
]))
video = video * (2 / 255) - 1
start_index = tf.random_uniform(
1,
minval=0,
maxval=parsed_features["video_length"] - config.num_frames + 1,
dtype=tf.int64)
import pdb
pdb.set_trace() # Make sure OK
video = tf.gather(video, start_index, config.num_frames)
import pdb
pdb.set_trace() # Make sure this works
# start_index = 0
# video = video[start_index:start_index + config.num_frames, ...]
# TODO(drewjaegle): do we need these config fields?
# Otherwise, we need to ensure the tensor values match the config values.
video.set_shape([
config.num_frames, config.im_height, config.im_width,
config.im_channels
])
# Reshape to NCHW from NHWC
video = tf.transpose(video, [0, 3, 1, 2])
# TODO(drewjaegle): Allow resampling of images here
# Split images to input (10 frames) and predict (10 frames)
input_sequence = video[:config.input_seq_len, ...]
predict_sequence = video[config.input_seq_len:, ...]
# TODO(drewjaegle): Do any preprocessing needed (i.e. downsample to 64)
return input_sequence, predict_sequence
def read_TFRecord(data_dir, batch_size, shuffle, in_classes):
# 分类数目
num_classes = in_classes
# 获取record文件
data_files = tf.gfile.Glob(data_dir)
# 读取文件。
filename_queue = tf.train.string_input_producer(data_files, shuffle=True)
reader = tf.TFRecordReader()
_, serialized_example = reader.read(filename_queue)
# 解析读取的样例。
features = tf.parse_single_example(serialized_example,
features={
'label':
tf.FixedLenFeature([], tf.int64),
'img_raw':
tf.FixedLenFeature([], tf.string),
'img_width':
tf.FixedLenFeature([], tf.int64),
'img_height':
tf.FixedLenFeature([], tf.int64),
}) # 取出包含image和label的feature对象
# tf.decode_raw可以将字符串解析成图像对应的像素数组
# 解析图片数据 string--unit8
image = tf.decode_raw(features['img_raw'], tf.uint8)
height = tf.cast(features['img_height'], tf.int32)
width = tf.cast(features['img_width'], tf.int32)
label = tf.cast(features['label'], tf.int32)
channel = 3
image = tf.reshape(image, [height, width, channel])
# reshape 向量---三维矩阵
# image = tf.reshape(image, [height, width, channel])
# 图像的缩放处理
image = tf.image.resize_image_with_crop_or_pad(image, 100, 100)
# image = tf.image.resize_images(image, [240,240], method=0)
image = tf.image.per_image_standardization(image)
# unit8 -- float32
# image = tf.cast(image, tf.float32) * (1. / 255) - 0.5
image = tf.cast(image, tf.float32)
#组合batch
min_after_dequeue = 1000
capacity = min_after_dequeue + 3 * batch_size
if shuffle:
image_batch, label_batch = tf.train.shuffle_batch(
[image, label],
batch_size=batch_size,
num_threads=64,
capacity=capacity,
min_after_dequeue=min_after_dequeue)
else:
image_batch, label_batch = tf.train.batch([image, label],
batch_size=batch_size,
num_threads=64,
capacity=capacity)
## ONE-HOT
label_batch = tf.reshape(label_batch, [batch_size, 1])
indices = tf.reshape(tf.range(0, batch_size, 1), [batch_size, 1])
label_batch = tf.sparse_to_dense(
tf.concat(values=[indices, label_batch], axis=1),
[batch_size, num_classes], 1.0, 0.0)
print(image_batch)
print(label_batch)
#n_classes = 10
#label_batch = tf.one_hot(label_batch, depth= n_classes)
#label_batch = tf.cast(label_batch, dtype=tf.int32)
#label_batch = tf.reshape(label_batch, [batch_size, n_classes])
return image_batch, label_batch
def read_and_decode(filename_queue=None, img_dims=[256,256,3], resize_to=[256,256], model_dims=[224,224,3], size_of_batch=32,\
labels=True, augmentations_dic=None, num_of_threads=1, shuffle=True):
"""
Reads in tf records and decodes the features of the image
Input: filename_queue - A node in a TensorFlow Graph used for asynchronous computations
img_dims - Dimensions of the tensor image stored as a tfrecord, example: [256, 256, 3]
model_dims - Dimensions of the tensor image that the model accepts, example: [224, 224, 3]
resize_to - Size to resize tf record to before training if resize_to is the same a img_dims no resizing will take place
size_of_batch - Size of the batch that will be fed into the model, example: 32
labels - Option for if the images stored in tfrecords have labels associated with them
augmentations_dic - Dictionary of augmentations that an image can have for training and validation. Augmentations
are chosen in the config
num_threads - Number of threads that execute a training op that dequeues mini-batches from the queue
shuffle - Boolean if batches fed into graph should be shuffled or not
Outputs: Tensor image, label of the image and filepath to the image. If labels is False only tensor image and filepath will be returned
"""
reader = tf.TFRecordReader()
_, serialized_example = reader.read(filename_queue)
if not labels:
features = tf.parse_single_example(
serialized_example,
features={
'image_raw': tf.FixedLenFeature([], tf.string),
'file_path': tf.FixedLenFeature([], tf.string),
})
image = tf.decode_raw(features['image_raw'], tf.uint8)
file_path = tf.cast(features['file_path'], tf.string)
image = tf.reshape(image, img_dims)
image = tf.cast(image, tf.float32)
image = tf.image.resize_images(image, resize_to)
image = tf.to_float(image)
image = image/255
if augmentations_dic and augmentations_dic['scale_jitter']:
random_size = randint(256,512)
image = tf.image.resize_images(image, [random_size, random_size])
else:
image = tf.image.resize_images(image,resize_to)
if augmentations_dic and augmentations_dic['rand_crop']:
image = tf.random_crop(image, model_dims)
else:
image = tf.image.resize_image_with_crop_or_pad(image, model_dims[0],\
model_dims[1])
if augmentations_dic and augmentations_dic['rand_color']:
random_color_ordering = randint(0,3)
image = distort_color(image,random_color_ordering)
if augmentations_dic and augmentations_dic['rand_flip_left_right']:
image = tf.image.random_flip_left_right(image)
if augmentations_dic and augmentations_dic['rand_flip_top_bottom']:
image = tf.image.random_flip_up_down(image)
if augmentations_dic and augmentations_dic['rand_rotate']:
random_angle = randint(0,359)
image = tf.contrib.image.rotate(image, random_angle)
if shuffle:
img, f = tf.train.shuffle_batch([image, file_path],
batch_size=size_of_batch,
capacity=1000 + 3 * size_of_batch,
min_after_dequeue=1000,
num_threads=num_of_threads)
else:
img, f = tf.train.batch([image, file_path],
batch_size=size_of_batch,
capacity=100000,
allow_smaller_final_batch=True,
num_threads=num_of_threads)
return img, f
else:
features = tf.parse_single_example(
serialized_example,
features={
'image_raw': tf.FixedLenFeature([], tf.string),
'file_path': tf.FixedLenFeature([], tf.string),
'label': tf.FixedLenFeature([], tf.int64)
})
image = tf.decode_raw(features['image_raw'], tf.uint8)
label = tf.cast(features['label'], tf.int32)
file_path = tf.cast(features['file_path'], tf.string)
image = tf.reshape(image, img_dims)
image = tf.cast(image, tf.float32)
image = tf.image.resize_images(image, resize_to)
image = tf.to_float(image)
image = image/255
if augmentations_dic and augmentations_dic['scale_jitter']:
random_size = randint(256,512)
image = tf.image.resize_images(image, [random_size, random_size])
else:
image = tf.image.resize_images(image,resize_to)
if augmentations_dic and augmentations_dic['rand_crop']:
image = tf.random_crop(image, model_dims)
else:
image = tf.image.resize_image_with_crop_or_pad(image, model_dims[0],\
model_dims[1])
if augmentations_dic and augmentations_dic['rand_color']:
random_color_ordering = randint(0,3)
image = distort_color(image,random_color_ordering)
if augmentations_dic and augmentations_dic['rand_flip_left_right']:
image = tf.image.random_flip_left_right(image)
if augmentations_dic and augmentations_dic['rand_flip_top_bottom']:
image = tf.image.random_flip_up_down(image)
if augmentations_dic and augmentations_dic['rand_rotate']:
random_angle = randint(0,359)
image = tf.contrib.image.rotate(image, random_angle)
if shuffle:
img, l, f = tf.train.shuffle_batch([image, label, file_path],
batch_size=size_of_batch,
capacity=1000 + 3 * size_of_batch,
min_after_dequeue=1000,
num_threads=num_of_threads)
else:
img, l, f = tf.train.batch([image, label, file_path],
batch_size=size_of_batch,
capacity=100000,
allow_smaller_final_batch=True,
num_threads=num_of_threads)
return img, l, f
def _decode_and_augment_image(example_proto):
keys_to_features = {
'label': tf.FixedLenFeature([], tf.int64),
'shape': tf.FixedLenFeature([], tf.string),
'image': tf.FixedLenFeature([], tf.string),
}
tfrecord_features = tf.parse_single_example(
example_proto, keys_to_features)
image = tf.decode_raw(tfrecord_features['image'], tf.uint8)
shape = tf.decode_raw(tfrecord_features['shape'], tf.int64)
if input_type == ".jpeg":
image = tf.reshape(image, target_size + [3])
else:
image = tf.reshape(image, target_size)
label = tfrecord_features['label']
if augment:
image = tf.image.random_flip_left_right(image)
image = tf.image.random_flip_up_down(image)
degrees = tf.random_uniform((), minval=-180, maxval=180)
image = tf.contrib.image.rotate(image, degrees)
width_shift = tf.random_uniform((), minval=0, maxval=0.05)
height_shift = tf.random_uniform((), minval=0, maxval=0.05)
horizontal_pad = tf.cast(tf.ceil(width_shift * target_size[0]),
tf.int32)
vertical_pad = tf.cast(tf.ceil(height_shift * target_size[1]),
tf.int32)
padding = tf.stack([
horizontal_pad, horizontal_pad, vertical_pad, vertical_pad,
tf.constant(0),
tf.constant(0)
])
padding = tf.reshape(padding, (3, 2))
image = tf.pad(image, padding)
image = tf.random_crop(image, target_size + [3])
zoom = tf.random_uniform((), minval=-0.1, maxval=0.1)
new_dim = tf.cast(tf.ceil((1 - zoom) * target_size[0]),
dtype=tf.int32)
image = tf.image.resize_image_with_crop_or_pad(
image, new_dim, new_dim)
image = tf.image.resize_images(
image, target_size, method=tf.image.ResizeMethod.BILINEAR)
if normalize:
std = tf.constant(np.array(
[70.53946096, 51.71475228, 43.03428563]),
dtype=tf.float32)
std = tf.expand_dims(tf.expand_dims(std, axis=0), axis=0)
mean = tf.constant(np.array(
[108.64628601, 75.86886597, 54.34005736]),
dtype=tf.float32)
mean = tf.expand_dims(tf.expand_dims(mean, axis=0), axis=0)
image = (tf.cast(image, dtype=tf.float32) - mean) / std
label = tf.reshape(label, [1])
if input_type == ".jpeg":
image = tf.reshape(image, target_size + [3])
else:
image = tf.reshape(image, target_size)
return {'shape': shape, 'image': image}, label
def read(bytes, dtype, shapex, shapey): inp = tf.decode_raw(bytes, dtype) retx = tf.reshape(inp[:np.prod(shapex)], shapex) rety = tf.reshape(inp[np.prod(shapex):], shapey) return (retx, rety)
batchsize = 2
batch = tf.train.shuffle_batch([ex],
batchsize,
capacity=batchsize * 10,
min_after_dequeue=batchsize * 5)
# 反序列化数据
example = tf.parse_example(batch, features=feature)
image = example['image']
label = example['label']
# 对byte数据解码成uint8类型的数据
image = tf.decode_raw(image, tf.uint8)
# 需要reshape，否则是一个向量
image = tf.reshape(image, [-1, 32, 32, 3])
with tf.Session() as sess:
# 线程的协调器
coord = tf.train.Coordinator()
sess.run(tf.local_variables_initializer())
threads = tf.train.start_queue_runners(sess, coord)
for i in range(1):
# image_bth, _ = sess.run([image, label])
# import cv2
# cv2.imshow("image", image_bth[0, ...])
# cv2.waitKey(0)
'lsize_dim0': tf.FixedLenFeature([], tf.int64),
'lsize_dim1': tf.FixedLenFeature([], tf.int64),
'lsize_dim2': tf.FixedLenFeature([], tf.int64),
'data_vol': tf.FixedLenFeature([], tf.string),
'label_vol': tf.FixedLenFeature([], tf.string)
}
with tf.Session() as sess:
queue = tf.train.string_input_producer(val_list,
num_epochs=None,
shuffle=False)
reader = tf.TFRecordReader()
fid, serialized_example = reader.read(queue)
parser = tf.parse_single_example(serialized_example,
features=decomp_feature)
data_vol = tf.decode_raw(parser['data_vol'], tf.float32)
label_vol = tf.decode_raw(parser['label_vol'], tf.float32)
data_vol = tf.reshape(data_vol, raw_size)
label_vol = tf.reshape(label_vol, raw_size)
data_vol = tf.slice(data_vol, [0, 0, 0], volume_size)
label_vol = tf.slice(label_vol, [0, 0, 1], label_size)
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(len(val_list)):
example, l = sess.run([data_vol, label_vol])
np.save(os.path.join(saveDir, str(i) + '.npy'), example)
np.save(os.path.join(saveDir_, str(i) + '.npy'), l)
import tensorflow as tf
reader = tf.TFRecordReader()
filename_queue = tf.train.string_input_producer(
["../image_data/output.tfrecords"])
_, serialized_example = reader.read(filename_queue)
features = tf.parse_single_example(serialized_example,
features={
'image_raw':
tf.FixedLenFeature([], tf.string),
'pixels':
tf.FixedLenFeature([], tf.int64),
'label':
tf.FixedLenFeature([], tf.int64),
})
image = tf.decode_raw(features['image_raw'], tf.uint8)
label = tf.cast(features['label'], tf.int32)
pixels = tf.cast(features['pixels'], tf.int32)
sess = tf.Session()
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
for i in range(10):
print sess.run([image, label, pixels])
# import tensorflow
import tensorflow as tf
# total bytes per image
TOTAL_BYTES = 3073
# create a list of filenames
filenames = ["dataset/data_batch_%d.bin" % i for i in range(1, 6)]
# create a queue of filenames
filename_queue = tf.train.string_input_producer(filenames, shuffle=False)
# initialize a reader to read TOTAL_BYTES bytes
reader = tf.FixedLengthRecordReader(TOTAL_BYTES)
# read TOTAL_BYTES bytes from the files
key, value = reader.read(filename_queue)
# decode read bytes to perceivable datatype
vector_bytes = tf.decode_raw(value, tf.uint8)
#create a session object
sess = tf.InteractiveSession()
#start queue runners
tf.train.start_queue_runners()
#print uint8 value of data read
print("vector_bytes ", sess.run(vector_bytes))
#print the number of elements in list vector_bytes
print("number of elements in vector_bytes ", len(sess.run(vector_bytes)))
label = tf.cast(tf.slice(vector_bytes, [0], [1]), tf.int32)
depth_major = tf.reshape(tf.slice(vector_bytes, [1], [3072]), [3, 32, 32])
uint8image = tf.transpose(depth_major, [1, 2, 0])
#print(sess.run(uint8image))
A = tf.constant([[[1, 2], [3, 4]], [[5, 6], [7, 8]], [[9, 10], [11, 12]]])
print(A.get_shape())
def load_train_batch(self):
"""Load a batch of training instances.
"""
opt = self.opt
# Load the list of training files into queues
#TODO
if opt.train_lite:
file_list = self.format_file_list(opt.dataset_dir,
opt.filelist_dir, 'train_lite')
else:
file_list = self.format_file_list(opt.dataset_dir,
opt.filelist_dir, 'train')
image_paths_queue = tf.train.string_input_producer(
file_list['image_file_list'], shuffle=False)
cam_paths_queue = tf.train.string_input_producer(
file_list['cam_file_list'], shuffle=False)
# Load camera intrinsics
cam_reader = tf.TextLineReader()
_, raw_cam_contents = cam_reader.read(cam_paths_queue)
rec_def = []
for i in range(9):
rec_def.append([1.])
raw_cam_vec = tf.decode_csv(raw_cam_contents, record_defaults=rec_def)
raw_cam_vec = tf.stack(raw_cam_vec)
intrinsics = tf.reshape(raw_cam_vec, [3, 3])
# Load images
img_reader = tf.WholeFileReader()
_, image_contents = img_reader.read(image_paths_queue)
image_seq = tf.image.decode_jpeg(image_contents)
tgt_image, src_image_stack = \
self.unpack_image_sequence(
image_seq, opt.img_height, opt.img_width, opt.num_source)
#TODO Load Semantics
# See cityscape label defs in https://github.com/mcordts/cityscapesScripts/blob/master/cityscapesscripts/helpers/labels.py#L62
# Also notice that deeplabv3+ uses `train_id` https://github.com/tensorflow/models/blob/69b016449ffc797421bf003d8b7fd8545db866d7/research/deeplab/datasets/build_cityscapes_data.py#L46
# Color maps are in https://github.com/tensorflow/models/blob/69b016449ffc797421bf003d8b7fd8545db866d7/research/deeplab/utils/get_dataset_colormap.py#L207
if opt.sem_assist:
sem_paths_queue = tf.train.string_input_producer(
file_list['sem_image_file_list'], shuffle=False)
sem_reader = tf.WholeFileReader()
sem_keys, sem_contents = sem_reader.read(sem_paths_queue)
if opt.load_from_raw:
sem_seq = tf.reshape(
tf.decode_raw(sem_contents, tf.uint8),
[1, opt.img_height, (opt.num_source + 1) * opt.img_width])
else:
sem_seq = tf.py_func(read_npy_file, [sem_keys], [
tf.uint8,
])
#TODO Load Instances: we use COCO
# Two channels: class and id level. For id level we only use the edge
if opt.ins_assist:
ins_paths_queue = tf.train.string_input_producer(
file_list['ins_image_file_list'], shuffle=False)
ins_reader = tf.WholeFileReader()
ins_keys, ins_contents = ins_reader.read(ins_paths_queue)
if opt.load_from_raw:
ins_seq = tf.reshape(tf.decode_raw(ins_contents, tf.uint8), [
1, opt.img_height, (opt.num_source + 1) * opt.img_width, 2
])
else:
ins_seq = tf.py_func(read_npy_file, [ins_keys], [
tf.uint8,
])
#TODO 1. SHUFFLE BATCH
# Form training batches
seed = random.randint(0, 2**31 - 1)
min_after_dequeue = 2048
capacity = min_after_dequeue + opt.num_threads * opt.batch_size
if opt.sem_assist and opt.ins_assist:
src_image_stack, tgt_image, intrinsics, sem_seq, ins_seq = tf.train.shuffle_batch(
[
src_image_stack, tgt_image, intrinsics, sem_seq[0],
ins_seq[0]
], opt.batch_size, capacity, min_after_dequeue,
opt.num_threads, seed)
elif opt.sem_assist:
src_image_stack, tgt_image, intrinsics, sem_seq = tf.train.shuffle_batch(
[src_image_stack, tgt_image, intrinsics, sem_seq[0]],
opt.batch_size, capacity, min_after_dequeue, opt.num_threads,
seed)
elif opt.ins_assist:
src_image_stack, tgt_image, intrinsics, ins_seq = tf.train.shuffle_batch(
[src_image_stack, tgt_image, intrinsics, ins_seq[0]],
opt.batch_size, capacity, min_after_dequeue, opt.num_threads,
seed)
else:
src_image_stack, tgt_image, intrinsics = tf.train.shuffle_batch(
[src_image_stack, tgt_image, intrinsics], opt.batch_size,
capacity, min_after_dequeue, opt.num_threads, seed)
# semantic segmentation
tgt_sem = None
tgt_sem_map = None
tgt_sem_mask = None
tgt_sem_edge = None
src_sem_stack = None
src_sem_map_stack = None
src_sem_mask_stack = None
src_sem_edge_stack = None
# ins0 ~ instance level, but still class segmentation
tgt_ins0 = None
tgt_ins0_map = None
tgt_ins0_edge = None
src_ins0_stack = None
src_ins0_map_stack = None
src_ins0_edge_stack = None
# ins1 ~ instance level, but this is id segmentation
tgt_ins1_edge = None
src_ins1_edge_stack = None
#TODO 2. TRAMSFORMATION AND UNPACKING
if opt.sem_assist:
#TODO get one-hot encoded sem_oh_seq (4,128,1248,19)X{0,1}
sem_oh_seq = tf.cast(
tf.one_hot(sem_seq, on_value=1, depth=opt.sem_num_class),
tf.uint8)
#TODO decouple tgt_sem (4,128,1248,19)X{0,1} src_sem_stack (4,128,1248,2*19)X{0,1}
tgt_sem, src_sem_stack = self.unpack_sem_sequence_batch_atom(
sem_oh_seq, opt.sem_num_class)
#TODO get densemap sem_map_seq (4,128,1248,1)X{0,1,...,18}
sem_map_seq = tf.expand_dims(sem_seq, -1)
#TODO decouple tgt_sem_map (4,128,1248,1)X{0,1,...,18} src_sem_map_stack (4,128,1248,2*1)X{0,1,...,18}
tgt_sem_map, src_sem_map_stack = self.unpack_sem_sequence_batch_atom(
sem_map_seq, 1)
if opt.sem_mask_explore:
#TODO get sem mask sem_mask_seq (4,128,1248,c) here we assume c=1
sem_mask_seq = self.get_sem_mask_batch(sem_seq)
#TODO decouple tgt_sem_mask (4,128,1248,c) src_sem_mask_stack (4,128,1248,2*c)
tgt_sem_mask, src_sem_mask_stack = self.unpack_sem_sequence_batch_atom(
sem_mask_seq, 1)
if opt.sem_edge_explore:
#TODO get sem edge sem_edge_seq (4,128,1248,c) here we assume c=1
sem_edge_seq = self.get_sem_edge_batch(sem_seq)
#TODO decouple tgt_sem_edge (4,128,1248,c) src_sem_edge_stack (4,128,1248,2*c)
tgt_sem_edge, src_sem_edge_stack = self.unpack_sem_sequence_batch_atom(
sem_edge_seq, 1)
if opt.ins_assist:
ins0_seq = ins_seq[:, :, :, 0]
ins1_seq = ins_seq[:, :, :, 1]
#TODO get one-hot ins0_oh_seq (4,128,1248,81)X{0,1}
ins0_oh_seq = tf.cast(
tf.one_hot(ins0_seq, on_value=1, depth=opt.ins_num_class),
tf.uint8)
#ins1_oh_seq = tf.cast(tf.one_hot(ins1_seq, on_value=1, depth = 255), tf.uint8)
#TODO decouple tgt_ins0 (4,128,1248,81)X{0,1} src_ins0_stack (4,128,1248,2*81)X{0,1}
tgt_ins0, src_ins0_stack = self.unpack_sem_sequence_batch_atom(
ins0_oh_seq, opt.ins_num_class)
#tgt_ins1, src_ins1_stack = self.unpack_sem_sequence_batch_atom(ins1_oh_seq, opt.ins_num_class)
#TODO get densemap sem_ins0_seq (4,128,1248,1)X{0,1,...,80}
ins0_map_seq = ins_seq[:, :, :, :1]
ins1_map_seq = ins_seq[:, :, :, 1:]
#TODO decouple tgt_ins0_map (4,128,1248,1)X{0,1,...,80} src_ins0_map_stack (4,128,1248,2*1)X{0,1,...,80}
tgt_ins0_map, src_ins0_map_stack = self.unpack_sem_sequence_batch_atom(
ins0_map_seq, 1)
tgt_ins1_map, src_ins1_map_stack = self.unpack_sem_sequence_batch_atom(
ins1_map_seq, 1)
if opt.ins0_edge_explore:
#TODO get edge ins0_edge_seq (4,128,1248,c) here we assume c=1
ins0_edge_seq = self.get_sem_edge_batch(ins0_seq)
#TODO decouple tgt_ins0_edge (4,128,1248,c) src_ins0_edge_stack (4,128,1248,2*c)
tgt_ins0_edge, src_ins0_edge_stack = self.unpack_sem_sequence_batch_atom(
ins0_edge_seq, 1)
if opt.ins1_edge_explore:
#TODO get edge ins1_edge_seq (4,128,1248,c) here we assume c=1
ins1_edge_seq = self.get_sem_edge_batch(ins1_seq)
#TODO decouple tgt_ins1_edge (4,128,1248,c) src_ins1_edge_stack (4,128,1248,2*c)
tgt_ins1_edge, src_ins1_edge_stack = self.unpack_sem_sequence_batch_atom(
ins1_edge_seq, 1)
#TODO 3. DATA AUGMENTATION
image_all = tf.concat([tgt_image, src_image_stack], axis=3)
image_all, intrinsics, aug_params = self.data_augmentation(
image_all, intrinsics, opt.img_height,
opt.img_width) #TODO changed API
if opt.sem_assist:
##TODO Do the same data augmentation for semantic segmentations
tgt_sem, src_sem_stack = self.data_aug(tgt_sem, src_sem_stack,
aug_params, "bilinear")
tgt_sem_map, src_sem_map_stack = self.data_aug(
tgt_sem_map, src_sem_map_stack, aug_params, "neighbor")
if self.opt.sem_mask_explore:
tgt_sem_mask, src_sem_mask_stack = \
self.data_aug(tgt_sem_mask, src_sem_mask_stack, aug_params, "bilinear")
if self.opt.sem_edge_explore:
tgt_sem_edge, src_sem_edge_stack = \
self.data_aug(tgt_sem_edge, src_sem_edge_stack, aug_params, "bilinear")
#TODO maybe transfer needs this settings self.data_aug(tgt_sem_edge, src_sem_edge_stack, aug_params, "neighbor")
if opt.ins_assist:
##TODO Do the same data augmentation for instance segmentations
tgt_ins0, src_ins0_stack = self.data_aug(tgt_ins0, src_ins0_stack,
aug_params, "bilinear")
#tgt_ins1, src_ins1_stack = self.data_aug(tgt_ins1, src_ins1_stack, aug_params, "bilinear")
tgt_ins0_map, src_ins0_map_stack = self.data_aug(
tgt_ins0_map, src_ins0_map_stack, aug_params, "neighbor")
#tgt_ins1_map, src_ins1_map_stack = self.data_aug(tgt_ins1_map, src_ins1_map_stack, aug_params, "neighbor")
if self.opt.ins0_edge_explore:
tgt_ins0_edge, src_ins0_edge_stack = \
self.data_aug(tgt_ins0_edge, src_ins0_edge_stack, aug_params, "bilinear")
#TODO maybe transfer needs this settings self.data_aug(tgt_ins0_edge, src_ins0_edge_stack, aug_params, "neighbor")
if self.opt.ins1_edge_explore:
tgt_ins1_edge, src_ins1_edge_stack = \
self.data_aug(tgt_ins1_edge, src_ins1_edge_stack, aug_params, "bilinear")
#TODO maybe transfer needs this settings self.data_aug(tgt_ins1_edge, src_ins1_edge_stack, aug_params, "neighbor")
# 4. RETURN
# image_channels=3*opt.seq_length
tgt_image = image_all[:, :, :, :3]
src_image_stack = image_all[:, :, :, 3:] #3:image_channels]
intrinsics = self.get_multi_scale_intrinsics(intrinsics,
opt.num_scales)
# if opt.sem_assist and opt.ins_assist:
return tgt_image, src_image_stack, intrinsics, \
[tgt_sem, tgt_sem_map, tgt_sem_mask, tgt_sem_edge], \
[src_sem_stack, src_sem_map_stack, src_sem_mask_stack, src_sem_edge_stack], \
[tgt_ins0, tgt_ins0_map, tgt_ins0_edge, tgt_ins1_edge], \
[src_ins0_stack, src_ins0_map_stack, src_ins0_edge_stack, src_ins1_edge_stack]
num_reader_threads=
1 # number of threads for prefetching SequenceExample protos.
)
serialized_sequence_example = input_queue.dequeue()
# serialized_sequence_example = tf.train.string_input_producer(["train.cat_caption"]) # don't work
context, sequence = tf.parse_single_sequence_example(
serialized=serialized_sequence_example,
context_features={
"image/img_raw": tf.FixedLenFeature([], dtype=tf.string)
},
sequence_features={
"image/caption": tf.FixedLenSequenceFeature([], dtype=tf.string),
"image/caption_ids": tf.FixedLenSequenceFeature([], dtype=tf.int64),
})
img = tf.decode_raw(context["image/img_raw"], tf.uint8)
img = tf.reshape(img, [height, width, 3])
img = tf.image.convert_image_dtype(img, dtype=tf.float32)
try:
# for TensorFlow 0.11
img = tf.image.resize_images(img,
size=(resize_height, resize_width),
method=tf.image.ResizeMethod.BILINEAR)
except Exception:
# for TensorFlow 0.10
img = tf.image.resize_images(img,
new_height=resize_height,
new_width=resize_width,
method=tf.image.ResizeMethod.BILINEAR)
# Crop to final dimensions.
def read_and_decode(tfrecords_filename, batch_size):
filename_queue = tf.train.string_input_producer(tfrecords_filename,
num_epochs=100)
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={
'label': tf.FixedLenFeature([], tf.int64),
'seq_len': tf.FixedLenFeature([], tf.int64),
'read_length': tf.FixedLenFeature([], tf.string),
'read_addr': tf.FixedLenFeature([], tf.string),
'ip': tf.FixedLenFeature([], tf.string),
'sp': tf.FixedLenFeature([], tf.string),
'write_length': tf.FixedLenFeature([], tf.string),
'write_addr': tf.FixedLenFeature([], tf.string),
'instruction_id': tf.FixedLenFeature([], tf.string),
'read2_length': tf.FixedLenFeature([], tf.string),
'read2_addr': tf.FixedLenFeature([], tf.string),
'bp': tf.FixedLenFeature([], tf.string)
})
label = tf.cast(features['label'], tf.int32)
#seq_len = tf.cast(features['seq_len'], tf.int32)
seq_shape = [SEQ_LEN, 1]
list_vars = []
for field in [
'read_length', 'read_addr', 'ip', 'sp', 'write_length',
'write_addr', 'read2_addr', 'read2_length', 'bp'
]:
new_field = tf.reshape(tf.decode_raw(features[field], tf.float64),
seq_shape)
relative_field = tf.cast(cos_relative_positions(new_field), tf.float32)
tf.summary.histogram('cos_relative_positions_{}'.format(field),
relative_field)
list_vars.append(relative_field)
print(relative_field)
print(tf.concat(list_vars, axis=-1))
instruction_id = tf.cast(
tf.reshape(tf.decode_raw(features['instruction_id'], tf.float64),
seq_shape), tf.int32)
X, instruction_ids, Y = tf.train.shuffle_batch(
[tf.concat(list_vars, axis=-1), instruction_id, label],
batch_size=batch_size,
capacity=1000,
num_threads=16,
min_after_dequeue=2)
tf.summary.histogram('All X', X)
return X, instruction_ids, Y
def read_cifar10(filename_queue):
"""Reads and parses examples from CIFAR10 data files.
Recommendation: if you want N-way read parallelism, call this function
N times. This will give you N independent Readers reading different
files & positions within those files, which will give better mixing of
examples.
Args:
filename_queue: A queue of strings with the filenames to read from.
Returns:
An object representing a single example, with the following fields:
height: number of rows in the result (32)
width: number of columns in the result (32)
depth: number of color channels in the result (3)
key: a scalar string Tensor describing the filename & record number
for this example.
label: an int32 Tensor with the label in the range 0..9.
uint8image: a [height, width, depth] uint8 Tensor with the image data
"""
class CIFAR10Record(object):
pass
result = CIFAR10Record()
# Dimensions of the images in the CIFAR-10 dataset.
# See http://www.cs.toronto.edu/~kriz/cifar.html for a description of the
# input format.
label_bytes = 1 # 2 for CIFAR-100
result.height = 32
result.width = 32
result.depth = 3
image_bytes = result.height * result.width * result.depth
# Every record consists of a label followed by the image, with a
# fixed number of bytes for each.
record_bytes = label_bytes + image_bytes
# Read a record, getting filenames from the filename_queue. No
# header or footer in the CIFAR-10 format, so we leave header_bytes
# and footer_bytes at their default of 0.
reader = tf.FixedLengthRecordReader(record_bytes=record_bytes)
result.key, value = reader.read(filename_queue)
# Convert from a string to a vector of uint8 that is record_bytes long.
record_bytes = tf.decode_raw(value, tf.uint8)
# The first bytes represent the label, which we convert from uint8->int32.
result.label = tf.cast(tf.strided_slice(record_bytes, [0], [label_bytes]),
tf.int32)
# The remaining bytes after the label represent the image, which we reshape
# from [depth * height * width] to [depth, height, width].
depth_major = tf.reshape(
tf.strided_slice(record_bytes, [label_bytes],
[label_bytes + image_bytes]),
[result.depth, result.height, result.width])
# Convert from [depth, height, width] to [height, width, depth].
result.uint8image = tf.transpose(depth_major, [1, 2, 0])
return result
features = tf.parse_single_example(serialized_example,
features = {
'image': tf.FixedLenFeature([], tf.string),
'label': tf.FixedLenFeature([], tf.int64),
'height': tf.FixedLenFeature([], tf.int64),
'weight': tf.FixedLenFeature([], tf.int64),
'channels': tf.FixedLenFeature([], tf.int64),
}
)
image = features['image']
label = features['label']
height = features['height']
weight = features['weight']
channels = features['channels']
decode_image = tf.decode_raw(image, tf.uint8)#解码tensor
decode_image.set_shape([height,weight,channels])
#前面的预处理图片函数
image_size = 299
distort_image = preprocess_for_train(decode_image, image_size, image_size, None)
#整理成输入batch队列
min_after_dequeue = 10000
batch_size = 100
capacity = min_after_dequeue + 3 * batch_size
image_batch, label_batch = tf.train.shuffle_batch(
[distorted_image, label], batch_size=batch_size,
capacity=capacity, min_after_dequeue=min_after_dequeue
)
logit = inference(image_batch)
def get_batch(paths, options):
"""Returns a data split of the RECOLA dataset, which was saved in tfrecords format.
Args:
paths: list with paths to data files
options: dict with data settings
Returns:
The raw audio examples and the corresponding arousal/valence
labels.
"""
batch_size = options['batch_size']
frame_size = options['frame_size']
num_channels = options['num_channels']
num_classes = options['num_classes']
crop_size = options['crop_size']
# max_in_len = options['max_in_len']
# max_out_len = options['max_out_len']
time_window_len = options['time_window_len']
if options['shuffle']:
shuffle = options['shuffle']
else:
shuffle = False
if options['horizontal_flip']:
horizontal_flip = options['horizontal_flip']
else:
horizontal_flip = False
if options['random_crop']:
random_crop = options['random_crop']
else:
random_crop = False
# root_path = Path(dataset_dir) / split_name
# paths = [str(x) for x in root_path.glob('*.tfrecords')]
filename_queue = tf.train.string_input_producer(paths, shuffle=shuffle)
reader = tf.TFRecordReader()
_, serialized_example = reader.read(filename_queue)
features = tf.parse_single_example(
serialized_example,
features={
'video': tf.FixedLenFeature([], tf.string),
'label': tf.FixedLenFeature([], tf.string),
'in_seq_len': tf.FixedLenFeature([], tf.int64),
'out_seq_len': tf.FixedLenFeature([], tf.int64)
}
)
video = tf.cast(tf.decode_raw(features['video'], tf.uint8), tf.float32) # / 255.
label = tf.cast(tf.decode_raw(features['label'], tf.uint8), tf.int32)
in_seq_len = tf.cast(features['in_seq_len'], tf.int32)
out_seq_len = tf.cast(features['out_seq_len'], tf.int32)
# perform bucketing with input_length being the single filter (need to add out_length buckets)
# Number of threads should always be one, in order to load samples sequentially.
_seq_lens, [encoder_inputs, target_labels, encoder_inputs_lengths, target_labels_lengths] = \
tf.contrib.training.bucket_by_sequence_length(in_seq_len,
[video, label, in_seq_len, out_seq_len], batch_size,
[20, 30, 50, 60, 88, 400],
num_threads=1, capacity=500, dynamic_pad=True,
allow_smaller_final_batch=True)
# encoder_inputs, target_labels, encoder_inputs_lengths, target_labels_lengths = \
# tf.train.batch([video, label, in_seq_len, out_seq_len], batch_size,
# num_threads=1, capacity=500, dynamic_pad=True,
# allow_smaller_final_batch=True)
encoder_inputs = tf.reshape(encoder_inputs, (batch_size,
tf.reduce_max(encoder_inputs_lengths),
frame_size,
frame_size,
num_channels))
target_labels = tf.reshape(target_labels, (batch_size, -1))
# create decoder_inputs
# add <sos> token
# decoder_inputs = tf.identity(target_labels)
sos_slice = tf.constant(options['num_classes'] - 2, dtype=tf.int32, shape=[options['batch_size'], 1])
decoder_inputs = tf.concat([sos_slice, target_labels], axis=1)
decoder_inputs = tf.one_hot(decoder_inputs, num_classes)
if crop_size is not None and random_crop:
encoder_inputs = tf.random_crop(encoder_inputs, [batch_size,
tf.reduce_max(encoder_inputs_lengths),
crop_size,
crop_size,
num_channels])
elif crop_size:
start_xy = int((frame_size - crop_size) / 2)
encoder_inputs = tf.slice(encoder_inputs,
[0, 0, start_xy, start_xy, 0],
[batch_size, tf.reduce_max(encoder_inputs_lengths),
crop_size, crop_size, num_channels])
encoder_inputs = tf.reshape(encoder_inputs, [batch_size, -1, crop_size, crop_size, 1])
# random left right flip
if horizontal_flip:
sample = tf.random_uniform(shape=[], minval=0, maxval=1, dtype=tf.float32)
option = tf.less(sample, 0.5)
encoder_inputs = tf.cond(option,
lambda: tf.map_fn(video_left_right_flip, encoder_inputs),
lambda: tf.map_fn(tf.identity, encoder_inputs))
encoder_inputs = normalize(encoder_inputs)
# slicw video to time_window_len consecutive frames with stride 1
if time_window_len != 1:
# pad encoder_inputs s.t. each frame is in the same number of slices
ei_paddings = [[0, 0], [time_window_len-1, time_window_len-1], [0, 0], [0, 0], [0, 0]]
padded_encoder_inputs = tf.pad(encoder_inputs, ei_paddings, 'CONSTANT', constant_values=0)
encoder_inputs = slice_video(padded_encoder_inputs,
dims=[batch_size, tf.reduce_max(encoder_inputs_lengths) + 2*(time_window_len - 1),
crop_size, crop_size, num_channels],
time_window=time_window_len)
encoder_inputs = tf.reshape(encoder_inputs, [batch_size, -1, crop_size, crop_size, time_window_len])
encoder_inputs_lengths = encoder_inputs_lengths + time_window_len - 1
return encoder_inputs, target_labels, decoder_inputs, encoder_inputs_lengths, target_labels_lengths
def crnn_fn(features, labels, mode, params):
"""
:param features: dict {
'image'
'images_width'
'corpora'
}
:param labels: labels. flattend (1D) array with encoded label (one code per character)
:param mode:
:param params: dict {
'Params'
}
:return:
"""
parameters = params.get('Params')
assert isinstance(parameters, Params)
# Load pre-trained cnn model
if parameters.cnn_pretained_ckpt_path:
exclude = ['deep_bidirectional_lstm']
variables_to_restore = tf.contrib.slim.get_variables_to_restore(
exclude=exclude)
tf.train.init_from_checkpoint(
parameters.cnn_pretained_ckpt_path,
{v.name.split(':')[0]: v
for v in variables_to_restore})
if mode != tf.estimator.ModeKeys.TRAIN:
parameters.keep_prob_dropout = 1.0
conv = deep_cnn(features['image'], (mode == tf.estimator.ModeKeys.TRAIN),
parameters.cnn_model,
summaries=False)
logprob, raw_pred = deep_bidirectional_lstm(conv,
features['corpus'],
params=parameters,
summaries=False)
# Compute seq_len from image width
n_pools = parameters.width_down_sampling
seq_len_inputs = tf.divide(
features['image_width'], n_pools, name='seq_len_input_op') - 1
predictions_dict = {'prob': logprob, 'raw_predictions': raw_pred}
if not mode == tf.estimator.ModeKeys.PREDICT:
# Alphabet and codes
keys = [c for c in parameters.alphabet.encode('latin1')]
values = parameters.alphabet_codes
# Convert string label to code label
with tf.name_scope('str2code_conversion'):
table_str2int = tf.contrib.lookup.HashTable(
tf.contrib.lookup.KeyValueTensorInitializer(
keys, values, key_dtype=tf.int64, value_dtype=tf.int64),
-1)
splitted = tf.string_split(labels, delimiter='')
values_int = tf.cast(
tf.squeeze(tf.decode_raw(splitted.values, tf.uint8)), tf.int64)
codes = table_str2int.lookup(values_int)
codes = tf.cast(codes, tf.int32)
sparse_code_target = tf.SparseTensor(splitted.indices, codes,
splitted.dense_shape)
seq_lengths_labels = tf.bincount(
tf.cast(sparse_code_target.indices[:, 0],
tf.int32), #array of labels length
minlength=tf.shape(predictions_dict['prob'])[1])
# Loss
# ----
# >>> Cannot have longer labels than predictions -> error
with tf.control_dependencies([
tf.less_equal(sparse_code_target.dense_shape[1],
tf.reduce_max(tf.cast(seq_len_inputs, tf.int64)))
]):
loss_ctc = tf.nn.ctc_loss(
labels=sparse_code_target,
inputs=predictions_dict['prob'],
sequence_length=tf.cast(seq_len_inputs, tf.int32),
preprocess_collapse_repeated=False,
ctc_merge_repeated=True,
ignore_longer_outputs_than_inputs=
True, # returns zero gradient in case it happens -> ema loss = NaN
time_major=True)
loss_ctc = tf.reduce_mean(loss_ctc)
loss_ctc = tf.Print(loss_ctc, [loss_ctc], message='* Loss : ')
global_step = tf.train.get_or_create_global_step()
# # Create an ExponentialMovingAverage object
ema = tf.train.ExponentialMovingAverage(decay=0.99,
num_updates=global_step,
zero_debias=True)
# Create the shadow variables, and add op to maintain moving averages
maintain_averages_op = ema.apply([loss_ctc])
loss_ema = ema.average(loss_ctc)
# Train op
# --------
if parameters.learning_rate_decay:
learning_rate = tf.train.exponential_decay(
parameters.learning_rate,
global_step,
parameters.learning_rate_steps,
parameters.learning_rate_decay,
staircase=True)
else:
learning_rate = tf.constant(parameters.learning_rate)
if parameters.optimizer == 'ada':
optimizer = tf.train.AdadeltaOptimizer(learning_rate)
elif parameters.optimizer == 'momentum':
optimizer = tf.train.MomentumOptimizer(learning_rate, momentum=0.9)
elif parameters.optimizer == 'adam':
optimizer = tf.train.AdamOptimizer(
learning_rate, beta1=0.5,
epsilon=1e-07) # at 1e-08 sometimes exploding gradient
elif parameters.optimizer == 'rms':
optimizer = tf.train.RMSPropOptimizer(learning_rate)
if not parameters.train_cnn:
trainable = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
'deep_bidirectional_lstm')
print('Training LSTM only')
else:
trainable = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
opt_op = optimizer.minimize(loss_ctc,
global_step=global_step,
var_list=trainable)
with tf.control_dependencies(update_ops + [opt_op]):
train_op = tf.group(maintain_averages_op)
# Summaries
# ---------
tf.summary.scalar('learning_rate', learning_rate)
tf.summary.scalar('losses/ctc_loss', loss_ctc)
else:
loss_ctc, train_op = None, None
if mode in [
tf.estimator.ModeKeys.EVAL, tf.estimator.ModeKeys.PREDICT,
tf.estimator.ModeKeys.TRAIN
]:
with tf.name_scope('code2str_conversion'):
keys = tf.cast(parameters.alphabet_decoding_codes, tf.int64)
values = [c for c in parameters.alphabet_decoding]
table_int2str = tf.contrib.lookup.HashTable(
tf.contrib.lookup.KeyValueTensorInitializer(keys, values), '?')
sparse_code_pred, log_probability = tf.nn.ctc_beam_search_decoder(
predictions_dict['prob'],
sequence_length=tf.cast(seq_len_inputs, tf.int32),
merge_repeated=False,
beam_width=100,
top_paths=parameters.nb_logprob)
# likelihoood. For future rename it as confidence and take softmax of log_probability
predictions_dict['score'] = log_probability
sequence_lengths_pred = [
tf.bincount(tf.cast(sparse_code_pred[i].indices[:, 0],
tf.int32),
minlength=tf.shape(predictions_dict['prob'])[1])
for i in range(parameters.top_paths)
]
pred_chars = [
table_int2str.lookup(sparse_code_pred[i])
for i in range(parameters.top_paths)
]
list_preds = [
get_words_from_chars(pred_chars[i].values,
sequence_lengths=sequence_lengths_pred[i])
for i in range(parameters.top_paths)
]
predictions_dict['words'] = tf.stack(list_preds)
tf.summary.text('predicted_words',
predictions_dict['words'][0][:10])
# Evaluation ops
# --------------
if mode == tf.estimator.ModeKeys.EVAL:
with tf.name_scope('evaluation'):
CER = tf.metrics.mean(tf.edit_distance(
sparse_code_pred[0], tf.cast(sparse_code_target,
dtype=tf.int64)),
name='CER')
# Convert label codes to decoding alphabet to compare predicted and groundtrouth words
target_chars = table_int2str.lookup(
tf.cast(sparse_code_target, tf.int64))
target_words = get_words_from_chars(target_chars.values,
seq_lengths_labels)
accuracy = tf.metrics.accuracy(target_words,
predictions_dict['words'][0],
name='accuracy')
eval_metric_ops = {
'eval/accuracy': accuracy,
'eval/CER': CER,
}
CER = tf.Print(CER, [CER], message='-- CER : ')
accuracy = tf.Print(accuracy, [accuracy], message='-- Accuracy : ')
else:
eval_metric_ops = None
export_outputs = {
'predictions': tf.estimator.export.PredictOutput(predictions_dict)
}
return tf.estimator.EstimatorSpec(mode=mode,
predictions=predictions_dict,
loss=loss_ctc,
train_op=train_op,
eval_metric_ops=eval_metric_ops,
export_outputs=export_outputs,
scaffold=tf.train.Scaffold())
def read_and_decode(filename_queue):#输入文件名队列
reader = tf.TFRecorder()
_, serialized_example = reader.read(filename_queue)
features = tf.parse_sing_example(#解析example
serialized_example,
#必须写明features里面的key的名称
features={
'image_raw': tf.FixedLenFeature([], tf.string), #图片是string类型
'label': tf.FixedLenFeature([], tf.int64), #标记是int64类型
})
#对于BytesList,要重新进行解码，把string类型的0维Tensor变成uint8类型的一维Tensor
image = tf.decode_raw(features['image_raw'], tf.uint8)
mnist = input_data.read_data_sets('data/',
dtype=tf.uint8, # 注意这里编码是uint8
reshape=False,
# validation_size=FLAGS.validation_size
)
image.set_shape([mnist.IMAGE_PIXELS])
#Tensor("input/DecodeRaw:0", shape=(784,), dtype=uint8)
#image张量的形状为：Tensor("input/sub:0", shape=(784,), dtype=float32)
image = tf.cast(image, tf.float32) * (1. / 255) - 0.5
#把标记从unint8类型转换为int32类型
#label张量的形状为Tensor("input/Cast_1；0“， shape=(), dtype=int32)
label = tf.cast(features['label'], tf.int32)
return image, label
#接下来使用tf.train.shuffle_batch将前面生成的样本随机化，获得一个最小批次的张量
def inputs(train, batch_size, num_epochs):
#输入参数
#batch_size：训练的每一批有多少个样本
#num_epochs:过几遍数据，设置为O/None表示永远训练下去
'''
返回结果:A tuple (images, labels)
*images:类型float,形状[batch_size, mnist.IMAGE_PIXELS], 范围[-0.5, 0.5].
*labels:类型int32,形状[bathch_size],范围[0, mnist.NUM_CLASSES]
注意tf.train.QueueRunner必须用tf.train.start_queue_runners()来启动线程
'''
if not num_epochs:num_epochs = None
#获取文件路径，即/tmp/data/train.tfrecords, /tmp/data/validation.records
filename = os.path.join('/home/niangu/桌面/TensorFlow/test.tfrecords', )
with tf.name_scope('input'):
#tf.train.string_input_producer返回一个QueueRunner, 里面有一个FIFOQQueue
filename_queue = tf.train.string_input_producer(
[filename], num_epochs=num_epochs)#如果样本量很大，可以分成若干文件，把文件名列表传入
image, label = read_and_decode(filename_queue)
#随机化example，并把它们规整成batch_size大小
#tf.train.shuffle_batch生成了RandomShuffleQueue，并开启俩个线程
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)#留下一部分队列，来保存每次有足够的数据做随机打乱
return images, sparse_labels
#最后我们把生成的batch张量作为网络的输入，进行训练
def run_training():
with tf.Graph().as_default():
#输入images和labels
images, labels = inputs(train=True, batch_size=FLAGS.bathch_size, num_epochs=FLAGS.num_epochs)
#构建一个从推理模型来预测数据的图
logits = mnist.inference(images, FLAGS.hidden1, FLAGS.hidden2)
loss = mnist.loss(logits, labels) #定义损失函数
#Add to the Graph operations that train the model
train_op = mnist.training(loss, FLAGS.learning_rate)
#初始化参数，特别注意:string_input_producer内部创建了一个epoch计数变量
#归入tf.GraphKeys.LOCAL_VARIABLES集和中，必须单独用initialize_local_variables()初始化
init_op = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
sess = tf.Session()
sess.run(init_op)
#Start input enqueue threads
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
try:
step = 0
while not coord.should_stop():#进入永久循环
start_time = time.time()
_, loss_value = sess.run([train_op, loss])
duration = time.time() - start_time
#每100次训练输出一次结果
if step % 100 == 0:
print('Step %d:loss = %.2f (%.3f sec)' % (step, loss_value, duration))
step += 1
except tf.errors.OutOfRangeError:
print('Done training for %d epochs, %d steps.' % (FLAGS.num_epochs, step))
finally:
coord.request_stop() #通知其他线程关闭
coord.join(threads)
sess.close()
def read_data(filename_queue, is_train):
reader = tf.TFRecordReader()
_, serialized_example = reader.read(filename_queue)
features = tf.parse_single_example(serialized_example,
features={
'name':
tf.FixedLenFeature([], tf.string),
'm':
tf.FixedLenFeature([], tf.int64),
'n':
tf.FixedLenFeature([], tf.int64),
'query':
tf.FixedLenFeature([], tf.string),
'align':
tf.FixedLenFeature([], tf.string),
'y':
tf.FixedLenFeature([], tf.string),
'mask':
tf.FixedLenFeature([], tf.string),
'gap':
tf.VarLenFeature(tf.float32),
'identity':
tf.VarLenFeature(tf.float32),
'identity_cons':
tf.VarLenFeature(tf.float32),
'ss_dssp':
tf.FixedLenFeature([], tf.string),
'asa_num':
tf.VarLenFeature(tf.int64),
})
name = features["name"]
m = tf.cast(features["m"], tf.int32)
n = tf.cast(features["n"], tf.int32)
align = tf.reshape(tf.decode_raw(features["align"], tf.uint8),
tf.stack([m, n]))
query = tf.decode_raw(features["query"], tf.uint8)
y = tf.reshape(tf.decode_raw(features["y"], tf.uint8), tf.stack([n, n]))
mask = tf.reshape(tf.decode_raw(features["mask"], tf.uint8),
tf.stack([n, n]))
gap = features["gap"].values
identity = features["identity"].values
identity_cons = features["identity_cons"].values
ss_dssp = tf.decode_raw(features["ss_dssp"], tf.uint8)
asa_num = tf.cast(features["asa_num"].values, tf.int32)
gap = features["gap"].values
identity = features["identity"].values
identity_cons = features["identity_cons"].values
#clip
def clipping(align, query, ss_dssp, asa_num, y, mask):
begin = tf.random_uniform([],
maxval=tf.shape(align)[1] - n_clip,
dtype=tf.int32)
align = align[:, begin:begin + n_clip]
query = query[begin:begin + n_clip]
ss_dssp = ss_dssp[begin:begin + n_clip]
asa_num = asa_num[begin:begin + n_clip]
y = y[begin:begin + n_clip, begin:begin + n_clip]
mask = mask[begin:begin + n_clip, begin:begin + n_clip]
return align, query, ss_dssp, asa_num, y, mask
align, query, ss_dssp, asa_num, y, mask = tf.cond(
(n > n_clip) & (is_train),
lambda: clipping(align, query, ss_dssp, asa_num, y, mask), lambda:
(align, query, ss_dssp, asa_num, y, mask))
#sampling
def sampling(align, gap, identity, identity_cons):
idx = tf.random_uniform([n_alignment], maxval=m, dtype=tf.int32)
align = tf.gather_nd(align, tf.expand_dims(idx, 1))
gap = tf.gather_nd(gap, tf.expand_dims(idx, 1))
identity = tf.gather_nd(identity, tf.expand_dims(idx, 1))
identity_cons = tf.gather_nd(identity_cons, tf.expand_dims(idx, 1))
return align, gap, identity, identity_cons
align, gap, identity, identity_cons = tf.cond(
(m > n_alignment) & (is_train),
lambda: sampling(align, gap, identity, identity_cons), lambda:
(align, gap, identity, identity_cons))
return name, align, query, y, mask, gap, identity, identity_cons, ss_dssp, asa_num
def get(self):
""" Provides input data to the graph. """
# calculate size of each record (this lists what is contained in the db and how many bytes are occupied)
record_bytes = 0
encoding_bytes = 4
kp_xyz_entries = 3 * self.num_kp
record_bytes += encoding_bytes * kp_xyz_entries
encoding_bytes = 4
kp_uv_entries = 2 * self.num_kp
record_bytes += encoding_bytes * kp_uv_entries
kp_vis_entries = self.num_kp
record_bytes += encoding_bytes * kp_vis_entries
image_bytes = self.image_size[0] * self.image_size[1] * 3
record_bytes += image_bytes
""" READ DATA ITEMS"""
# Start reader
reader = tf.FixedLengthRecordReader(header_bytes=0,
record_bytes=record_bytes)
_, value = reader.read(
tf.train.string_input_producer([self.path_to_db]))
# decode to floats
bytes_read = 0
data_dict = dict()
record_bytes_float32 = tf.decode_raw(value, tf.float32)
# 1. Read keypoint xyz
keypoint_xyz21 = tf.reshape(
tf.slice(record_bytes_float32, [bytes_read // 4],
[kp_xyz_entries]), [self.num_kp, 3])
bytes_read += encoding_bytes * kp_xyz_entries
keypoint_xyz21 /= 1000.0 # scale to meters
keypoint_xyz21 = self.convert_kp(keypoint_xyz21)
# calculate wrist coord
if self.use_wrist_coord:
wrist_xyz = keypoint_xyz21[16, :] + 2.0 * (keypoint_xyz21[0, :] -
keypoint_xyz21[16, :])
keypoint_xyz21 = tf.concat(
[tf.expand_dims(wrist_xyz, 0), keypoint_xyz21[1:, :]], 0)
data_dict['keypoint_xyz21'] = keypoint_xyz21
# 2. Read keypoint uv AND VIS
keypoint_uv_vis21 = tf.reshape(
tf.slice(record_bytes_float32, [bytes_read // 4],
[kp_uv_entries + kp_vis_entries]), [self.num_kp, 3])
bytes_read += encoding_bytes * (kp_uv_entries + kp_vis_entries)
keypoint_uv_vis21 = self.convert_kp(keypoint_uv_vis21)
keypoint_uv21 = keypoint_uv_vis21[:, :2]
keypoint_vis21 = tf.equal(keypoint_uv_vis21[:, 2], 1.0)
# calculate wrist vis
if self.use_wrist_coord:
wrist_vis = tf.logical_or(keypoint_vis21[16], keypoint_vis21[0])
keypoint_vis21 = tf.concat(
[tf.expand_dims(wrist_vis, 0), keypoint_vis21[1:]], 0)
wrist_uv = keypoint_uv21[16, :] + 2.0 * (keypoint_uv21[0, :] -
keypoint_uv21[16, :])
keypoint_uv21 = tf.concat(
[tf.expand_dims(wrist_uv, 0), keypoint_uv21[1:, :]], 0)
data_dict['keypoint_vis21'] = keypoint_vis21
if self.coord_uv_noise:
noise = tf.truncated_normal([42, 2],
mean=0.0,
stddev=self.coord_uv_noise_sigma)
keypoint_uv21 += noise
data_dict['keypoint_uv21'] = keypoint_uv21
# decode to uint8
record_bytes_uint8 = tf.decode_raw(value, tf.uint8)
# 4. Read image
image = tf.reshape(
tf.slice(record_bytes_uint8, [bytes_read], [image_bytes]),
[self.image_size[0], self.image_size[1], 3])
image = tf.cast(image, tf.float32)
bytes_read += image_bytes
# subtract mean
image = image / 255.0 - 0.5
if self.hue_aug:
image = tf.image.random_hue(image, self.hue_aug_max)
data_dict['image'] = image
""" CONSTANTS """
# Camera intrinsics
sx = 822.79041
sy = 822.79041
tx = 318.47345
ty = 250.31296
data_dict['cam_mat'] = tf.constant([[sx, 0.0, tx], [0.0, sy, ty],
[0.0, 0.0, 1.0]])
# Hand side: this dataset only contains left hands
data_dict['hand_side'] = tf.one_hot(tf.constant(0, dtype=tf.int32),
depth=2,
on_value=1.0,
off_value=0.0,
dtype=tf.float32)
assert bytes_read == record_bytes, "Doesnt add up."
""" DEPENDENT DATA ITEMS: XYZ represenations. """
# make coords relative to root joint
kp_coord_xyz_root = keypoint_xyz21[0, :] # this is the palm coord
kp_coord_xyz21_rel = keypoint_xyz21 - kp_coord_xyz_root # relative coords in metric coords
index_root_bone_length = tf.sqrt(
tf.reduce_sum(
tf.square(kp_coord_xyz21_rel[12, :] -
kp_coord_xyz21_rel[11, :])))
data_dict['keypoint_scale'] = index_root_bone_length
data_dict[
'keypoint_xyz21_normed'] = kp_coord_xyz21_rel / index_root_bone_length # normalized by length of 12->11
# calculate local coordinates
kp_coord_xyz21_local = bone_rel_trafo(
data_dict['keypoint_xyz21_normed'])
kp_coord_xyz21_local = tf.squeeze(kp_coord_xyz21_local)
data_dict['keypoint_xyz21_local'] = kp_coord_xyz21_local
# calculate viewpoint and coords in canonical coordinates
kp_coord_xyz21_rel_can, rot_mat = canonical_trafo(
data_dict['keypoint_xyz21_normed'])
kp_coord_xyz21_rel_can, rot_mat = tf.squeeze(
kp_coord_xyz21_rel_can), tf.squeeze(rot_mat)
data_dict['keypoint_xyz21_can'] = kp_coord_xyz21_rel_can
data_dict['rot_mat'] = tf.matrix_inverse(rot_mat)
""" DEPENDENT DATA ITEMS: HAND CROP """
if self.hand_crop:
crop_center = keypoint_uv21[12, ::-1]
# catch problem, when no valid kp available (happens almost never)
crop_center = tf.cond(tf.reduce_all(tf.is_finite(crop_center)),
lambda: crop_center,
lambda: tf.constant([0.0, 0.0]))
crop_center.set_shape([
2,
])
if self.crop_center_noise:
noise = tf.truncated_normal(
[2], mean=0.0, stddev=self.crop_center_noise_sigma)
crop_center += noise
crop_scale_noise = tf.constant(1.0)
if self.crop_scale_noise:
crop_scale_noise = tf.squeeze(
tf.random_uniform([1], minval=1.0, maxval=1.2))
if not self.use_wrist_coord:
wrist_uv = keypoint_uv21[16, :] + 2.0 * (keypoint_uv21[0, :] -
keypoint_uv21[16, :])
keypoint_uv21 = tf.concat(
[tf.expand_dims(wrist_uv, 0), keypoint_uv21[1:, :]], 0)
# select visible coords only
kp_coord_h = tf.boolean_mask(keypoint_uv21[:, 1], keypoint_vis21)
kp_coord_w = tf.boolean_mask(keypoint_uv21[:, 0], keypoint_vis21)
kp_coord_hw = tf.stack([kp_coord_h, kp_coord_w], 1)
# determine size of crop (measure spatial extend of hw coords first)
min_coord = tf.maximum(tf.reduce_min(kp_coord_hw, 0), 0.0)
max_coord = tf.minimum(tf.reduce_max(kp_coord_hw, 0),
self.image_size)
# find out larger distance wrt the center of crop
crop_size_best = 2 * tf.maximum(max_coord - crop_center,
crop_center - min_coord)
crop_size_best = tf.reduce_max(crop_size_best)
crop_size_best = tf.minimum(tf.maximum(crop_size_best, 50.0),
500.0)
# catch problem, when no valid kp available
crop_size_best = tf.cond(
tf.reduce_all(tf.is_finite(crop_size_best)),
lambda: crop_size_best, lambda: tf.constant(200.0))
crop_size_best.set_shape([])
# calculate necessary scaling
scale = tf.cast(self.crop_size, tf.float32) / crop_size_best
scale = tf.minimum(tf.maximum(scale, 1.0), 10.0)
scale *= crop_scale_noise
data_dict['crop_scale'] = scale
if self.crop_offset_noise:
noise = tf.truncated_normal(
[2], mean=0.0, stddev=self.crop_offset_noise_sigma)
crop_center += noise
# Crop image
img_crop = crop_image_from_xy(tf.expand_dims(image, 0),
crop_center, self.crop_size, scale)
data_dict['image_crop'] = tf.squeeze(img_crop)
# Modify uv21 coordinates
crop_center_float = tf.cast(crop_center, tf.float32)
keypoint_uv21_u = (
data_dict['keypoint_uv21'][:, 0] -
crop_center_float[1]) * scale + self.crop_size // 2
keypoint_uv21_v = (
data_dict['keypoint_uv21'][:, 1] -
crop_center_float[0]) * scale + self.crop_size // 2
keypoint_uv21 = tf.stack([keypoint_uv21_u, keypoint_uv21_v], 1)
data_dict['keypoint_uv21'] = keypoint_uv21
# Modify camera intrinsics
scale = tf.reshape(scale, [
1,
])
scale_matrix = tf.dynamic_stitch([
[0], [1], [2], [3], [4], [5], [6], [7], [8]
], [scale, [0.0], [0.0], [0.0], scale, [0.0], [0.0], [0.0], [1.0]])
scale_matrix = tf.reshape(scale_matrix, [3, 3])
crop_center_float = tf.cast(crop_center, tf.float32)
trans1 = crop_center_float[0] * scale - self.crop_size // 2
trans2 = crop_center_float[1] * scale - self.crop_size // 2
trans1 = tf.reshape(trans1, [
1,
])
trans2 = tf.reshape(trans2, [
1,
])
trans_matrix = tf.dynamic_stitch(
[[0], [1], [2], [3], [4], [5], [6], [7], [8]],
[[1.0], [0.0], -trans2, [0.0], [1.0], -trans1, [0.0], [0.0],
[1.0]])
trans_matrix = tf.reshape(trans_matrix, [3, 3])
data_dict['cam_mat'] = tf.matmul(
trans_matrix, tf.matmul(scale_matrix, data_dict['cam_mat']))
""" DEPENDENT DATA ITEMS: Scoremap from the SUBSET of 21 keypoints"""
# create scoremaps from the subset of 2D annoataion
keypoint_hw21 = tf.stack([keypoint_uv21[:, 1], keypoint_uv21[:, 0]],
-1)
scoremap_size = self.image_size
if self.hand_crop:
scoremap_size = (self.crop_size, self.crop_size)
scoremap = self.create_multiple_gaussian_map(keypoint_hw21,
scoremap_size,
self.sigma,
valid_vec=keypoint_vis21)
if self.scoremap_dropout:
scoremap = tf.nn.dropout(scoremap,
self.scoremap_dropout_prob,
noise_shape=[1, 1, 21])
scoremap *= self.scoremap_dropout_prob
data_dict['scoremap'] = scoremap
if self.random_crop_to_size:
tensor_stack = tf.concat([
data_dict['image'],
tf.expand_dims(tf.cast(data_dict['hand_parts'], tf.float32),
-1),
tf.cast(data_dict['hand_mask'], tf.float32)
], 2)
s = tensor_stack.get_shape().as_list()
tensor_stack_cropped = tf.random_crop(
tensor_stack,
[self.random_crop_size, self.random_crop_size, s[2]])
data_dict = dict(
) # delete everything else because the random cropping makes the data invalid anyway
data_dict['image'], data_dict['hand_parts'], data_dict['hand_mask'] = tensor_stack_cropped[:, :, :3],\
tf.cast(tensor_stack_cropped[:, :, 3], tf.int32),\
tf.cast(tensor_stack_cropped[:, :, 4:], tf.int32)
names, tensors = zip(*data_dict.items())
if self.shuffle:
tensors = tf.train.shuffle_batch_join([tensors],
batch_size=self.batch_size,
capacity=100,
min_after_dequeue=50,
enqueue_many=False)
else:
tensors = tf.train.batch_join([tensors],
batch_size=self.batch_size,
capacity=100,
enqueue_many=False)
return dict(zip(names, tensors))
def parse_tfrecord_tf(record):
features = tf.parse_single_example(record, features={
'shape': tf.FixedLenFeature([3], tf.int64),
'data': tf.FixedLenFeature([], tf.string)})
data = tf.decode_raw(features['data'], tf.uint8)
return tf.reshape(data, features['shape'])
def decode_image(image):
# Normalize from [0, 255] to [0.0, 1.0]
image = tf.decode_raw(image, tf.uint8)
image = tf.cast(image, tf.float32)
image = tf.reshape(image, [784])
return image / 255.0
features={
'image/height' : tf.FixedLenFeature([], tf.int64 ),
'image/width' : tf.FixedLenFeature([], tf.int64 ),
'image/colorspace' : tf.FixedLenFeature([], tf.string),
'image/channels' : tf.FixedLenFeature([], tf.int64 ),
'image/class/label': tf.FixedLenFeature([], tf.int64 ),
'image/class/text' : tf.FixedLenFeature([], tf.string),
'image/format' : tf.FixedLenFeature([], tf.string),
'image/filename' : tf.FixedLenFeature([], tf.string),
'image/encoded' : tf.FixedLenFeature([], tf.string)
},
name='features'
)
# image was saved as uint8, so we have to decode tf.string as uint8.
imageT = tf.decode_raw(tfrecord['image/encoded'], tf.uint8)
# since exported as tf.int64 there is no need for tf.decode_raw
heightT = tfrecord['image/height']
widthT = tfrecord['image/width' ]
# remember, it's all just ops, have to run to get result
with tf.Session() as sess:
# init vars
sess.run(tf.global_variables_initializer())
sess.run(tf. local_variables_initializer())
# init summary file writer
sfw = tf.summary.FileWriter(os.getcwd(),graph=sess.graph)
def read_cifar10(data_dir, is_train, batch_size, shuffle):
"""Read CIFAR10
Args:
data_dir: the directory of CIFAR10
is_train: boolen
batch_size:
shuffle:
Returns:
label: 1D tensor, tf.int32
image: 4D tensor, [batch_size, height, width, 3], tf.float32
"""
img_width = 32
img_height = 32
img_depth = 3
label_bytes = 1
image_bytes = img_width * img_height * img_depth
with tf.name_scope('input'):
if is_train:
filenames = [
os.path.join(data_dir, 'data_batch_%d.bin' % ii)
for ii in np.arange(1, 6)
]
else:
filenames = [os.path.join(data_dir, 'test_batch.bin')]
filename_queue = tf.train.string_input_producer(filenames)
reader = tf.FixedLengthRecordReader(label_bytes + image_bytes)
key, value = reader.read(filename_queue)
record_bytes = tf.decode_raw(value, tf.uint8)
label = tf.slice(record_bytes, [0], [label_bytes])
label = tf.cast(label, tf.int32)
image_raw = tf.slice(record_bytes, [label_bytes], [image_bytes])
image_raw = tf.reshape(image_raw, [img_depth, img_height, img_width])
image = tf.transpose(image_raw,
(1, 2, 0)) # convert from D/H/W to H/W/D
image = tf.cast(image, tf.float32)
image = tf.image.per_image_standardization(
image) #substract off the mean and divide by the variance
if shuffle:
images, label_batch = tf.train.shuffle_batch(
[image, label],
batch_size=batch_size,
num_threads=64,
capacity=20000,
min_after_dequeue=3000)
else:
images, label_batch = tf.train.batch([image, label],
batch_size=batch_size,
num_threads=64,
capacity=2000)
# ONE-HOT
n_classes = 10
label_batch = tf.one_hot(label_batch, depth=n_classes)
label_batch = tf.cast(label_batch, dtype=tf.int32)
label_batch = tf.reshape(label_batch, [batch_size, n_classes])
return images, label_batch
def mnist_tfrecord_input(data_dir,
training=True,
sequence_length=20,
img_size=None,
batch_size=1,
seed=None):
"""Create input tfrecord tensors and queues.
TFRecord:
TFRecord(s) are assumed to be placed at `data_dir`.
Each sample contains raw uint8 image sequences with key `'img_i'`.
Training and validation set are pre-splitted and their corresponding
record file have suffix `_trn.tfrecords` or `_val_tfrecords`.
Preprocessing:
Crop each image to a square one with size `min(ORIGINAL_HEIGHT, ORIGINAL_WIDTH)`
and resize (bicubic) to `(IMG_WIDTH, IMG_HEIGHT)`. Normalize pixel value from
[0, 255] to [0, 1]
Args:
data_dir: directory holding TFRecord(s).
training: whether to use training or validation data.
sequence_length: length of the video sequence.
img_size: the (hight, width) of processed img input, if None use original size.
batch_size: size of data mimi-batches.
seed: random seed for `shuffle_batch` generator.
Returns:
list of tensors corresponding to images. The images
tensor is 5D, batch x time x height x width x 1.
Raises:
RuntimeError: if no files found.
"""
file_suffix = '*_trn.tfrecords' if training else '*_val.tfrecords'
filenames = gfile.Glob(os.path.join(data_dir, file_suffix))
if not filenames:
raise RuntimeError('No data files found.')
filename_queue = tf.train.string_input_producer(filenames, shuffle=True)
reader = tf.TFRecordReader()
_, serialized_example = reader.read(filename_queue)
image_seq = []
for i in range(sequence_length):
# extract image tensor
image_name = 'img_{}'.format(i)
features = tf.parse_single_example(
serialized_example,
features={image_name: tf.FixedLenFeature([], tf.string)}
)
image = tf.decode_raw(features[image_name], tf.uint8)
image = tf.reshape(image, shape=[ORIGINAL_HEIGHT, ORIGINAL_WIDTH, COLOR_CHAN])
# preprocessing
crop_size = min(ORIGINAL_HEIGHT, ORIGINAL_WIDTH)
image = tf.image.resize_image_with_crop_or_pad(image, crop_size, crop_size)
image = tf.reshape(image, [1, crop_size, crop_size, COLOR_CHAN])
if img_size is None:
img_size = (ORIGINAL_HEIGHT, ORIGINAL_WIDTH)
if img_size[0] != img_size[1]:
raise ValueError('Unequal height and width unsupported')
image = tf.image.resize_bicubic(image, img_size)
image = tf.cast(image, tf.float32) / 255.0
image_seq.append(image)
image_seq = tf.concat(axis=0, values=image_seq)
image_batch = tf.train.shuffle_batch(
tensors=[image_seq],
batch_size=batch_size,
capacity=100 * batch_size,
min_after_dequeue=50 * batch_size,
num_threads=batch_size,
seed=seed
)
return image_batch
def read_cifar100(filename_queue, coarse_or_fine='fine'):
"""Reads and parses examples from CIFAR100 data files.
Recommendation: if you want N-way read parallelism, call this function
N times. This will give you N independent Readers reading different
files & positions within those files, which will give better mixing of
examples.
Args:
filename_queue: A queue of strings with the filenames to read from.
Returns:
An object representing a single example, with the following fields:
height: number of rows in the result (32)
width: number of columns in the result (32)
depth: number of color channels in the result (3)
key: a scalar string Tensor describing the filename & record number
for this example.
label: an int32 Tensor with the label in the range 0..9.
uint8image: a [height, width, depth] uint8 Tensor with the image data
"""
class CIFAR100Record(object):
pass
result = CIFAR100Record()
coarse_label_bytes = 1
fine_label_bytes = 1
result.height = 32
result.width = 32
result.depth = 3
image_bytes = result.height * result.width * result.depth
record_bytes = coarse_label_bytes + fine_label_bytes + image_bytes
reader = tf.FixedLengthRecordReader(record_bytes=record_bytes,
header_bytes=0,
footer_bytes=0)
result.key, value = reader.read(filename_queue)
record_bytes = tf.decode_raw(value, tf.uint8)
coarse_label = tf.cast(
tf.strided_slice(record_bytes, [0], [coarse_label_bytes]), tf.int32)
fine_label = tf.cast(
tf.strided_slice(record_bytes, [coarse_label_bytes],
[coarse_label_bytes + fine_label_bytes]), tf.int32)
if coarse_or_fine == 'fine':
result.label = fine_label #
else:
result.label = coarse_label #
depth_major = tf.reshape(
tf.strided_slice(
record_bytes, [coarse_label_bytes + fine_label_bytes],
[coarse_label_bytes + fine_label_bytes + image_bytes]),
[result.depth, result.height, result.width])
result.uint8image = tf.transpose(depth_major, [1, 2, 0])
return result
def cifar10(path=pathcifar,
activation="sigmoid",
conv_channels=(16, 16, 16),
linear_layers=32,
batch_size=128,
num_threads=4,
min_queue_examples=1000,
mode="train"):
"""Cifar10 classification with a convolutional network."""
# Data.
_open_cifar10(path)
if activation == "sigmoid":
activation_op = tf.sigmoid
elif activation == "relu":
activation_op = tf.nn.relu
else:
raise ValueError("{} activation not supported".format(activation))
# Read images and labels from disk.
if mode == "train":
filenames = [
os.path.join(path, CIFAR10_FOLDER, "data_batch_{}.bin".format(i))
for i in range(1, 6)
]
elif mode == "test":
filenames = [os.path.join(path, "test_batch.bin")]
else:
raise ValueError("Mode {} not recognised".format(mode))
depth = 3
height = 32
width = 32
label_bytes = 1
image_bytes = depth * height * width
record_bytes = label_bytes + image_bytes
reader = tf.FixedLengthRecordReader(record_bytes=record_bytes)
_, record = reader.read(tf.train.string_input_producer(filenames))
record_bytes = tf.decode_raw(record, tf.uint8)
label = tf.cast(tf.slice(record_bytes, [0], [label_bytes]), tf.int32)
raw_image = tf.slice(record_bytes, [label_bytes], [image_bytes])
image = tf.cast(tf.reshape(raw_image, [depth, height, width]), tf.float32)
# height x width x depth.
image = tf.transpose(image, [1, 2, 0])
image = tf.div(image, 255)
queue = tf.RandomShuffleQueue(
capacity=min_queue_examples + 3 * batch_size,
min_after_dequeue=min_queue_examples,
dtypes=[tf.float32, tf.int32],
shapes=[image.get_shape(), label.get_shape()])
enqueue_ops = [queue.enqueue([image, label]) for _ in range(num_threads)]
tf.train.add_queue_runner(tf.train.QueueRunner(queue, enqueue_ops))
with tf.name_scope('Optimizee_loss'):
def compute_loss():
image_batch, label_batch = queue.dequeue_many(batch_size)
label_batch = tf.reshape(label_batch, [batch_size])
output = image_batch
with tf.variable_scope('ConvMLP', reuse=tf.AUTO_REUSE):
conv1_w = tf.get_variable(
"conv1_w",
shape=[5, 5, depth, conv_channels[0]],
dtype=tf.float32,
initializer=tf.random_normal_initializer(stddev=0.01))
conv1_b = tf.get_variable(
"conv1_b",
shape=[
conv_channels[0],
],
dtype=tf.float32,
initializer=tf.random_normal_initializer(stddev=0.01))
conv1_beta = tf.get_variable(
"conv1_beta",
shape=[1, 1, 1, conv_channels[0]],
dtype=tf.float32,
initializer=tf.random_normal_initializer(stddev=0.01))
conv2_w = tf.get_variable(
"conv2_w",
shape=[5, 5, conv_channels[0], conv_channels[1]],
dtype=tf.float32,
initializer=tf.random_normal_initializer(stddev=0.01))
conv2_b = tf.get_variable(
"conv2_b",
shape=[
conv_channels[1],
],
dtype=tf.float32,
initializer=tf.random_normal_initializer(stddev=0.01))
conv2_beta = tf.get_variable(
"conv2_beta",
shape=[1, 1, 1, conv_channels[1]],
dtype=tf.float32,
initializer=tf.random_normal_initializer(stddev=0.01))
conv3_w = tf.get_variable(
"conv3_w",
shape=[5, 5, conv_channels[1], conv_channels[2]],
dtype=tf.float32,
initializer=tf.random_normal_initializer(stddev=0.01))
conv3_b = tf.get_variable(
"conv3_b",
shape=[
conv_channels[2],
],
dtype=tf.float32,
initializer=tf.random_normal_initializer(stddev=0.01))
conv3_beta = tf.get_variable(
"conv3_beta",
shape=[1, 1, 1, conv_channels[2]],
dtype=tf.float32,
initializer=tf.random_normal_initializer(stddev=0.01))
output = tf.nn.convolution(output,
conv1_w,
padding='SAME',
strides=[1, 1])
output = tf.nn.relu(tf.nn.bias_add(output, conv1_b))
output = tf.nn.max_pool(output,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME')
b_m_1, b_v_1 = tf.nn.moments(output, axes=[0, 1, 2])
output = tf.nn.batch_normalization(output,
b_m_1,
b_v_1,
conv1_beta,
scale=None,
variance_epsilon=1e-8)
output = tf.nn.convolution(output,
conv2_w,
padding='SAME',
strides=[1, 1])
output = tf.nn.relu(tf.nn.bias_add(output, conv2_b))
output = tf.nn.max_pool(output,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME')
b_m_2, b_v_2 = tf.nn.moments(output, [0, 1, 2])
output = tf.nn.batch_normalization(output,
b_m_2,
b_v_2,
conv2_beta,
scale=None,
variance_epsilon=1e-8)
output = tf.nn.convolution(output,
conv3_w,
padding='SAME',
strides=[1, 1])
output = tf.nn.relu(tf.nn.bias_add(output, conv3_b))
output = tf.nn.max_pool(output,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME')
b_m_3, b_v_3 = tf.nn.moments(output, [0, 1, 2])
output = tf.nn.batch_normalization(output,
b_m_3,
b_v_3,
conv3_beta,
scale=None,
variance_epsilon=1e-8)
output = tf.layers.flatten(output)
W_in = tf.get_variable(
"W_in",
shape=[output.shape[1], linear_layers],
dtype=tf.float32,
initializer=tf.random_normal_initializer(stddev=0.01))
b_in = tf.get_variable(
"b_in",
shape=[
linear_layers,
],
dtype=tf.float32,
initializer=tf.random_normal_initializer(stddev=0.01))
W_out = tf.get_variable(
"W_out",
shape=[linear_layers, 10],
dtype=tf.float32,
initializer=tf.random_normal_initializer(stddev=0.01))
b_out = tf.get_variable(
"b_out",
shape=[
10,
],
dtype=tf.float32,
initializer=tf.random_normal_initializer(stddev=0.01))
layer_out = activation_op(tf.add(tf.matmul(output, W_in), b_in))
output = tf.add(tf.matmul(layer_out, W_out), b_out)
loss = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=output, labels=label_batch)
return tf.reduce_mean(loss)
with tf.name_scope('Convex_loss'):
def convex_loss():
with tf.variable_scope('conv_var', reuse=tf.AUTO_REUSE):
v = tf.get_variable(
"v",
shape=[1, 10],
dtype=tf.float32,
initializer=tf.random_normal_initializer(stddev=0.01))
# Non-trainable variables.
target = tf.get_variable(
"target",
shape=[1, 10],
dtype=tf.float32,
initializer=tf.random_uniform_initializer(),
trainable=False)
return tf.reduce_mean(
tf.clip_by_value(tf.square(v - target), 0, 10))
return collections.OrderedDict([('Opt_loss', compute_loss),
('Aux_loss', convex_loss)])
