def ndlstm_base_dynamic(inputs,
noutput,
sequence_length,
scope=None,
reverse=False):
with tf.variable_scope(scope, "SeqLstm", [inputs]):
# TODO(tmb) make batch size, sequence_length dynamic
# example: sequence_length = tf.shape(inputs)[0]
_, batch_size, _ = tf.unstack(tf.shape(inputs)) #_shape(inputs)
lstm_cell = tf.contrib.rnn.BasicLSTMCell(noutput, state_is_tuple=True)
state = lstm_cell.zero_state(batch_size, dtype=tf.float32)
#state = array_ops.zeros([batch_size, lstm_cell.state_size])
#sequence_length = int(inputs.get_shape()[0])
sequence_lengths = tf.to_int64(tf.fill([batch_size], sequence_length))
if reverse:
inputs = tf.reverse_v2(inputs, [0])
outputs, _ = rnn.dynamic_rnn(lstm_cell,
inputs,
sequence_lengths,
state,
time_major=True)
#print(outputs)
if reverse:
outputs = tf.reverse_v2(outputs, [0])
return outputs
def ndlstm_base_dynamic(inputs, noutput, scope=None, reverse=False):
"""Run an LSTM, either forward or backward.
This is a 1D LSTM implementation using dynamic_rnn and
the TensorFlow LSTM op.
Args:
inputs: input sequence (length, batch_size, ninput)
noutput: depth of output
scope: optional scope name
reverse: run LSTM in reverse
Returns:
Output sequence (length, batch_size, noutput)
"""
with tf.variable_scope(scope, "SeqLstm", [inputs]):
# TODO(tmb) make batch size, sequence_length dynamic
# example: sequence_length = tf.shape(inputs)[0]
_, batch_size, _ = _shape(inputs)
lstm_cell = tf.nn.rnn_cell.BasicLSTMCell(noutput, state_is_tuple=False)
state = tf.zeros([batch_size, lstm_cell.state_size])
sequence_length = int(inputs.get_shape()[0])
sequence_lengths = tf.to_int64(tf.fill([batch_size], sequence_length))
if reverse:
inputs = tf.reverse_v2(inputs, [0])
outputs, _ = tf.nn.dynamic_rnn(lstm_cell,
inputs,
sequence_lengths,
state,
time_major=True)
if reverse:
outputs = tf.reverse_v2(outputs, [0])
return outputs
def ndlstm_base_dynamic(inputs, noutput, scope=None, reverse=False):
"""Run an LSTM, either forward or backward.
This is a 1D LSTM implementation using dynamic_rnn and
the TensorFlow LSTM op.
Args:
inputs: input sequence (length, batch_size, ninput)
noutput: depth of output
scope: optional scope name
reverse: run LSTM in reverse
Returns:
Output sequence (length, batch_size, noutput)
"""
with tf.variable_scope(scope, "SeqLstm", [inputs]):
# TODO(tmb) make batch size, sequence_length dynamic
# example: sequence_length = tf.shape(inputs)[0]
_, batch_size, _ = _shape(inputs)
lstm_cell = tf.nn.rnn_cell.BasicLSTMCell(noutput, state_is_tuple=False)
state = tf.zeros([batch_size, lstm_cell.state_size])
sequence_length = int(inputs.get_shape()[0])
sequence_lengths = tf.to_int64(tf.fill([batch_size], sequence_length))
if reverse:
inputs = tf.reverse_v2(inputs, [0])
outputs, _ = tf.nn.dynamic_rnn(lstm_cell, inputs, sequence_lengths, state, time_major=True)
if reverse:
outputs = tf.reverse_v2(outputs, [0])
return outputs
def predict_labels_multi_scale(images,
model_options,
eval_scales=(1.0, ),
add_flipped_images=False):
outputs_to_predictions = {
output: []
for output in model_options.outputs_to_num_classes
}
for i, image_scale in enumerate(eval_scales):
with tf.variable_scope(tf.get_variable_scope(),
reuse=True if i else None):
outputs_to_scales_to_logits = multi_scale_logits(
images,
model_options=model_options,
image_pyramid=[image_scale],
is_training=False,
fine_tune_batch_norm=False)
if add_flipped_images:
with tf.variable_scope(tf.get_variable_scope(), reuse=True):
outputs_to_scales_to_logits_reversed = multi_scale_logits(
tf.reverse_v2(images, [2]),
model_options=model_options,
image_pyramid=[image_scale],
is_training=False,
fine_tune_batch_norm=False)
for output in sorted(outputs_to_scales_to_logits):
scales_to_logits = outputs_to_scales_to_logits[output]
logits = tf.image.resize_bilinear(
scales_to_logits[MERGED_LOGITS_SCOPE],
tf.shape(images)[1:3],
align_corners=True)
outputs_to_predictions[output].append(
tf.expand_dims(tf.nn.softmax(logits), 4))
if add_flipped_images:
scales_to_logits_reversed = (
outputs_to_scales_to_logits_reversed[output])
logits_reversed = tf.image.resize_bilinear(
tf.reverse_v2(
scales_to_logits_reversed[MERGED_LOGITS_SCOPE], [2]),
tf.shape(images)[1:3],
align_corners=True)
outputs_to_predictions[output].append(
tf.expand_dims(tf.nn.softmax(logits_reversed), 4))
for output in sorted(outputs_to_predictions):
predictions = outputs_to_predictions[output]
# Compute average prediction across different scales and flipped images.
predictions = tf.reduce_mean(tf.concat(predictions, 4), axis=4)
outputs_to_predictions[output] = tf.argmax(predictions, 3)
return outputs_to_predictions
def get_batch(dataset_dir,
num_readers=2,
batch_size=3,
net=None,
FLAGS=None,
file_pattern='*.tfrecord',
is_training=True,
shuffle=False):
filename = glob.glob(os.path.join(dataset_dir,
"*train*" + file_pattern))[0]
print(filename)
filename_queue = tf.train.string_input_producer([filename], num_epochs=50)
reader = tf.TFRecordReader()
_, serialized_example = reader.read(filename_queue)
features = tf.parse_single_example(
serialized_example,
features={
'image/encoded': tf.FixedLenFeature([],
tf.string,
default_value=''),
'label/encoded': tf.FixedLenFeature([],
tf.string,
default_value=''),
}) # return image and label
with tf.name_scope('cmc/label'):
label = tf.decode_raw(features['label/encoded'], tf.float32)
label = tf.transpose(tf.reshape(label, [3, 1, 240, 240]), (0, 2, 3, 1))
with tf.name_scope('cmc/image'):
image = tf.decode_raw(features['image/encoded'], tf.float32)
image = tf.transpose(tf.reshape(image, [3, 4, 240, 240]), (0, 2, 3, 1))
#image = tf.reshape(image, [32, 100, 3])
image = tf.cast(image, tf.float32)
label = tf.cast(label, tf.int32)
print("image", image.get_shape(), label.get_shape())
"""
if (np.random.choice([0, 1]) == 1):
image = tf.image.flip_left_right(image)
label = tf.image.flip_left_right(label)
"""
do_a_crop_flip = tf.random_uniform([], seed=None)
do_a_crop_flip = tf.greater(do_a_crop_flip, 0.5)
image = tf.cond(do_a_crop_flip, lambda: tf.reverse_v2(image, [2]),
lambda: image)
label = tf.cond(do_a_crop_flip, lambda: tf.reverse_v2(label, [2]),
lambda: label)
sh_images, sh_labels = tf.train.shuffle_batch([image, label],
batch_size=batch_size,
num_threads=1,
capacity=100 * batch_size,
min_after_dequeue=batch_size)
return sh_images, sh_labels
def _rotate_pp(data):
data["label"] = tf.constant([0, 1, 2, 3])
# We use our own instead of tf.image.rot90 because that one broke
# internally shortly before deadline...
data["image"] = tf.stack([
data["image"],
tf.transpose(tf.reverse_v2(data["image"], [1]), [1, 0, 2]),
tf.reverse_v2(data["image"], [0, 1]),
tf.reverse_v2(tf.transpose(data["image"], [1, 0, 2]), [1]),
])
return data
def predict_labels_multi_scale(images,
model_options,
eval_scales=(1.0,),
add_flipped_images=False):
"""预测分割标签
:param images: tensor,[batch, height, width, channels]
:param model_options: 卷积网络模型选择来配置模型
:param eval_scales: 用于调整图像大小以进行评估的比例
:param add_flipped_images:是否添加翻转图像用于评估
:return:具有指定output_type（例如，语义预测）的键的字典和存储表示预测的张量的值（通道上的argmax）.
每个预测的大小[批次，高度，宽度]。
"""
outputs_to_predictions = {
output: []
for output in model_options.outputs_to_num_classes
}
for i, image_scale in enumerate(eval_scales):
with tf.variable_scope(tf.get_variable_scope(), reuse=True if i else None):
outputs_to_scales_to_logits = multi_scale_logits(images,
model_options=model_options,
image_pyramid=[image_scale],
is_training=False,
fine_tune_batch_norm=False)
if add_flipped_images:
with tf.variable_scope(tf.get_variable_scope(), reuse=True):
outputs_to_scales_to_logits_reversed = multi_scale_logits(
tf.reverse_v2(images, [2]),
model_options=model_options,
image_pyramid=[image_scale],
is_training=False,
fine_tune_batch_norm=False)
for output in sorted(outputs_to_scales_to_logits):
scales_to_logits = outputs_to_scales_to_logits[output]
logits = tf.image.resize_bilinear(scales_to_logits[MERGED_LOGITS_SCOPE],
tf.shape(images)[1:3],
align_corners=True)
outputs_to_predictions[output].append(
tf.expand_dims(tf.nn.softmax(logits), 4))
if add_flipped_images:
scales_to_logits_reversed = (outputs_to_scales_to_logits_reversed[output])
logits_reversed = tf.image.resize_bilinear(
tf.reverse_v2(scales_to_logits_reversed[MERGED_LOGITS_SCOPE], [2]),
tf.shape(images)[1:3],
align_corners=True)
outputs_to_predictions[output].append(
tf.expand_dims(tf.nn.softmax(logits_reversed), 4))
for output in sorted(outputs_to_predictions):
predictions = outputs_to_predictions[output]
#计算不同尺度和翻转图像的平均预测。
predictions = tf.reduce_mean(tf.concat(predictions, 4), axis=4)
outputs_to_predictions[output] = tf.argmax(predictions, 3)
return outputs_to_predictions
def random_flip_lr(image, bboxes, seed=None):
"""Random flip left-right of an image and its bounding boxes."""
def flip_bboxes(bboxes):
"""Flip bounding boxes coordinates.
Args:
bboxes: [xmin, ymin, xmax, ymax]
"""
bboxes = tf.stack([1 - bboxes[:, 2], bboxes[:, 1],
1 - bboxes[:, 0], bboxes[:, 3]], axis=-1)
return bboxes
# Random flip. Tensorflow implementation.
with tf.name_scope('random_flip_left_right'):
image = tf.convert_to_tensor(image, name='image')
uniform_random = tf.random_uniform([], 0, 1.0, seed=seed)
mirror_cond = tf.less(uniform_random, .5)
# Flip image.
result = tf.cond(mirror_cond,
lambda: tf.reverse_v2(image, axis=[1]),
lambda: image)
# Flip bboxes.
bboxes = tf.cond(mirror_cond,
lambda: flip_bboxes(bboxes),
lambda: bboxes)
return fix_image_flip_shape(image, result), bboxes
def flip():
flipped = []
for tensor in tensor_list:
if dim < 0 or dim >= len(tensor.get_shape().as_list()):
raise ValueError('dim must represent a valid dimension.')
flipped.append(tf.reverse_v2(tensor, [dim]))
return flipped
def init_U(d):
translations = tf.cast(tf.reverse_v2(d, [0]), dtype=tf.float32)
shifted_original2 = tf.contrib.image.translate(original2, translations)
m = tf.minimum(original1, shifted_original2) / tf.maximum(original1, shifted_original2)
# if w and v are zero, we get nan, but we should get 0.
m = tf.where(tf.is_nan(m), tf.zeros_like(m), m)
return tf.reshape(m, tf.shape(m)[1:])
def flip_dim(tensor, prob=0.5, dim=1):
random_value = tf.random_uniform([])
is_flipped = tf.less_equal(random_value, prob)
flipped = tf.cond(is_flipped, lambda: tf.reverse_v2(tensor, [dim]),
lambda: tensor)
return flipped
def propagate_pooled(self, equal1, equal2, S, d):
translations = tf.cast(tf.reverse_v2(d, [0]), dtype=tf.float32)
shifted_equal2 = tf.contrib.image.translate(equal2, translations)
m = equal1 * shifted_equal2
return S * m
def get_em_acc_op(self, bucket_id):
"""Create a em_acc_op."""
with tf.name_scope("EMAcc_%d" % bucket_id):
# [sequence_length, batch_size]
input_ph = tf.placeholder(tf.int64, shape=(None, None))
# [sequence_length, batch_size, vocab_size]
output_ph = tf.placeholder(tf.float32,
shape=(None, None,
self.target_vocab_size))
input_op = tf.reverse_v2(input_ph, axis=[0])
output_op = tf.argmax(output_ph, axis=2)
def replace_eos_with_pad(in_seq):
"""Replace all tokens after EOS in sequence with PAD."""
out_seq = in_seq.copy()
for idx in range(in_seq.shape[-1]):
eos_list = in_seq[:, idx].reshape(in_seq.shape[0]).tolist()
eos_idx = eos_list.index(
EOS_ID) if EOS_ID in eos_list else -1
out_seq[eos_idx:, idx] = PAD_ID
return out_seq
eos_op = tf.py_func(replace_eos_with_pad, [output_op], tf.int64)
equal_op = tf.equal(
tf.reduce_sum(tf.abs(input_op - eos_op), axis=0), 0)
em_acc_op = tf.reduce_mean(tf.cast(equal_op, tf.float32), axis=0)
summary_op = tf.summary.scalar("EMAccSummary", em_acc_op)
return input_ph, output_ph, em_acc_op, summary_op
def sort_by_field(boxlist, field, order=SortOrder.descend, scope=None):
"""Sort boxes and associated fields according to a scalar value.
In-place operation.
`boxlist` must have that field which stores a 1-D numeric tensor.
Args:
boxlist: a BoxList holding `n` boxes.
field: string scalar, the field by which boxes are sorted. Must be 1-D.
order: int scalar, indicating descend or ascend. Defaults to descend.
scope: string scalar, name scope.
Returns:
sorted_boxlist: a BoxList holding `n` boxes where boxes are ordered by
the values in the field `field` in the specified order.
"""
with tf.name_scope(scope, 'SortByField'):
if order != SortOrder.descend and order != SortOrder.ascend:
raise ValueError('Invalid sort order.')
field_to_sort = boxlist.get_field(field)
num_boxes = boxlist.num_boxes()
_, sorted_indices = tf.nn.top_k(field_to_sort, num_boxes, sorted=True)
if order == SortOrder.ascend:
sorted_indices = tf.reverse_v2(sorted_indices, [0])
return gather(boxlist, sorted_indices)
def to_image_tf(tensor_in, colormap=None):
"""Reverse the second dimension and swap the second dimension and the third dimension"""
if colormap is None:
colormap = get_colormap()
shape = tf.stack([-1, tensor_in.get_shape()[1], tensor_in.get_shape()[2], 3])
recolored = tf.reshape(tf.matmul(tf.reshape(tensor_in, [-1, 5]), colormap), shape)
return tf.transpose(tf.reverse_v2(recolored, axis=[2]), [0, 2, 1, 3])
def flip():
flipped = []
for tensor in tensor_list:
if dim < 0 or dim >= len(tensor.get_shape().as_list()):
raise ValueError('dim must represent a valid dimension.')
flipped.append(tf.reverse_v2(tensor, [dim])) # 反转张量的特定维度dim
return flipped
def estimate_age_and_gender(aligned_images, model_path):
with tf.Graph().as_default():
session = tf.Session()
images_pl = tf.placeholder(tf.float32,
shape=[None, 160, 160, 3],
name='input_image')
images = tf.map_fn(lambda frame: tf.reverse_v2(frame, [-1]),
images_pl) #BGR TO RGB
images_norm = tf.map_fn(
lambda frame: tf.image.per_image_standardization(frame), images)
train_mode = tf.placeholder(tf.bool)
age_logits, gender_logits, _ = inception_resnet_v1.inference(
images_norm,
keep_probability=0.8,
phase_train=train_mode,
weight_decay=1e-5)
gender = tf.argmax(tf.nn.softmax(gender_logits), 1)
age_ = tf.cast(tf.constant([i for i in range(0, 101)]), tf.float32)
age = tf.reduce_sum(tf.multiply(tf.nn.softmax(age_logits), age_),
axis=1)
init_op = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
session.run(init_op)
saver = tf.train.Saver()
ckpt = tf.train.get_checkpoint_state(model_path)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(session, ckpt.model_checkpoint_path)
else:
pass
return session.run([age, gender],
feed_dict={
images_pl: aligned_images,
train_mode: False
})
def _encoder(self):
with tf.variable_scope("encoder"):
with tf.variable_scope("embedding-layer"):
if not self.pretrained_wordvec:
embedding_encoder = tf.get_variable(
name="decoder-embedding",
shape=[self.vocab_size_x, self.embed_size],
dtype=tf.float32)
else:
embedding_encoder = np.load("glove.npy")
tf.summary.histogram(name="embedding-encoder",
values=embedding_encoder)
embeded_x = tf.nn.embedding_lookup(
embedding_encoder,
self.input_x) # [None, sequence_len, embed_size]
with tf.variable_scope("backward-rnn"):
# backward rnn
embeded_x_back = tf.reverse_v2(embeded_x, axis=[1])
self.encoder_rnn_cell = self._get_cell(self.num_units,
self.rnn_type)
init_state = self.encoder_rnn_cell.zero_state(self.batch_size,
dtype=tf.float32)
encoder_output, encoder_state = tf.nn.dynamic_rnn(
cell=self.encoder_rnn_cell,
inputs=embeded_x_back,
initial_state=init_state)
return encoder_state # [None, sequence_len, num_units], [None, num_units]
def random_flip_left_right(image, bboxes, seed=None):
"""Random flip left-right of an image and its bounding boxes.
随机翻转图像及其边界框的左右两侧。
"""
def flip_bboxes(bboxes):
"""Flip bounding boxes coordinates.
"""
bboxes = tf.stack(
[bboxes[:, 0], 1 - bboxes[:, 3], bboxes[:, 2], 1 - bboxes[:, 1]],
axis=-1)
return bboxes
# Random flip. Tensorflow implementation.
with tf.name_scope('random_flip_left_right'):
image = tf.convert_to_tensor(image, name='image')
_check3dimage(image, require_static=False)
uniform_random = tf.random_uniform([], 0, 1.0, seed=seed)
mirror_cond = tf.less(uniform_random, .5)
# Flip image.
result = tf.cond(mirror_cond, lambda: tf.reverse_v2(image, [1]),
lambda: image)
# Flip bboxes.
bboxes = tf.cond(mirror_cond, lambda: flip_bboxes(bboxes),
lambda: bboxes)
image_shape = image.get_shape()
result.set_shape(image_shape)
return result, bboxes
def image_mirroring(img, label):
"""
Randomly mirrors the images.
Args:
img: Training image to mirror.
label: Segmentation mask to mirror.
"""
distort_left_right_random = tf.random_uniform([1],
0,
1.0,
dtype=tf.float32)[0]
mirror = tf.less(tf.stack([1.0, distort_left_right_random, 1.0]), 0.5)
img = tf.reverse_v2(img, mirror)
label = tf.reverse_v2(label, mirror)
return img, label
def to_image_tf(tensor_in, colormap=None):
"""Reverse the second dimension and swap the second dimension and the third dimension"""
if colormap is None:
colormap = get_colormap()
shape = tf.stack([-1, tensor_in.get_shape()[1], tensor_in.get_shape()[2], 3])
recolored = tf.reshape(tf.matmul(tf.reshape(tensor_in, [-1, 2]), colormap), shape)
##recolored = tf.reshape(tf.matmul(tf.reshape(tensor_in, [-1, 6]), colormap), shape)
return tf.transpose(tf.reverse_v2(recolored, axis=[2]), [0, 2, 1, 3])
def propagate_convolved(self, original1, original2, S, d):
translations = tf.cast(tf.reverse_v2(d, [0]), dtype=tf.float32)
shifted_original2 = tf.contrib.image.translate(original2, translations)
m = tf.minimum(original1, shifted_original2) / tf.maximum(original1, shifted_original2)
# if w and v are zero, we get nan, but we should get 0.
m = tf.where(tf.is_nan(m), tf.zeros_like(m), m)
return S * m
def flip():
image_tensor_reversed = tf.reverse_v2(image_tensor, [1])
bbox_x_coords = 1.0 - bbox_tensor[:, 0]
bbox_y_coords = bbox_tensor[:, 1]
pts_x_coords = 1.0 - points_tensor[:, 0]
pts_y_coords = points_tensor[:, 1]
bbox_ret = tf.stack([bbox_x_coords, bbox_y_coords], axis=1)
pts_ret = tf.stack([pts_x_coords, pts_y_coords], axis=1)
return image_tensor_reversed, bbox_ret, pts_ret
def flip():
image_tensor_reversed = tf.reverse_v2(image_tensor, [0])
x_coords = points_tensor[:, 0]
y_coords = image_height - points_tensor[:, 1]
tmp = tf.stack([x_coords, y_coords], axis=1)
p1 = tmp[0, :]
p2 = tmp[1, :]
p3 = tmp[2, :]
p4 = tmp[3, :]
points_reversed = tf.stack([p3, p4, p1, p2])
return image_tensor_reversed, points_reversed
def gru_backward(self, input_x, hidden_size, name_variable):
"""
GRU backward
:param input_x:shape:[batch_size*num_sentence, sequence_length, embedding_size]
:param hidden_size: gru output hidden size
:param name_variable: name of gru variable
:return:GRU backward outputs and every time step state
"""
with tf.variable_scope(name_variable):
input_x = tf.reverse_v2(input_x, axis=[1])
gru_cell = self.create_gru_unit(hidden_size, 'gru_backward')
# init unit state
zero_state_length = tf.shape(input_x)[0]
gru_init_state = gru_cell.zero_state(zero_state_length,
dtype=tf.float32)
# run GRU backward
outputs, state = tf.nn.dynamic_rnn(gru_cell,
inputs=input_x,
initial_state=gru_init_state)
outputs = tf.reverse_v2(outputs, axis=[1])
return outputs, state
def load_session(model_path=MODEL_PATH):
global SESSION
global AGE
global GENDER
global IMAGES_PL
global TRAIN_MODE
global tf_config
graph = tf.Graph().as_default()
sess = tf.Session(config=tf_config)
images_pl = tf.placeholder(tf.float32,
shape=[None, 160, 160, 3],
name='input_image')
images = tf.map_fn(lambda frame: tf.reverse_v2(frame, [-1]),
images_pl) #BGR TO RGB
images_norm = tf.map_fn(
lambda frame: tf.image.per_image_standardization(frame), images)
train_mode = tf.placeholder(tf.bool)
age_logits, gender_logits, _ = inception_resnet_v1.inference(
images_norm,
keep_probability=1.0, #0.8
phase_train=train_mode,
weight_decay=1e-5)
gender = tf.argmax(tf.nn.softmax(gender_logits), 1)
age_ = tf.cast(tf.constant([i for i in range(0, 101)]), tf.float32)
age = tf.reduce_sum(tf.multiply(tf.nn.softmax(age_logits), age_), axis=1)
GENDER = gender
AGE = age
IMAGES_PL = images_pl
TRAIN_MODE = train_mode
init_op = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
sess.run(init_op)
saver = tf.train.Saver()
ckpt = tf.train.get_checkpoint_state(model_path)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
print("restore and continue training! : {}".format(
ckpt.model_checkpoint_path))
else:
sys.exit("Age-Gender Model not found")
SESSION = sess
def sort_by_field(boxlist, field, order=SortOrder.descend, scope=None):
"""Sort boxes and associated fields according to a scalar field.
A common use case is reordering the boxes according to descending scores.
Args:
boxlist: BoxList holding N boxes.
field: A BoxList field for sorting and reordering the BoxList.
order: (Optional) descend or ascend. Default is descend.
scope: name scope.
Returns:
sorted_boxlist: A sorted BoxList with the field in the specified order.
Raises:
ValueError: if specified field does not exist
ValueError: if the order is not either descend or ascend
"""
with tf.name_scope(scope, 'SortByField'):
if order != SortOrder.descend and order != SortOrder.ascend:
raise ValueError('Invalid sort order')
field_to_sort = boxlist.get_field(field)
if len(field_to_sort.shape.as_list()) != 1:
raise ValueError('Field should have rank 1')
num_boxes = boxlist.num_boxes()
num_entries = tf.size(field_to_sort)
length_assert = tf.Assert(tf.equal(num_boxes, num_entries), [
'Incorrect field size: actual vs expected.', num_entries, num_boxes
])
with tf.control_dependencies([length_assert]):
# TODO(derekjchow): Remove with tf.device when top_k operation runs
# correctly on GPU.
with tf.device('/cpu:0'):
_, sorted_indices = tf.nn.top_k(field_to_sort,
num_boxes,
sorted=True)
if order == SortOrder.ascend:
sorted_indices = tf.reverse_v2(sorted_indices, [0])
return gather(boxlist, sorted_indices)
def _flip_patches(self, tensor, name):
with tf.variable_scope('flip-' + name):
batch_size, seq_len, out_size = tensor.get_shape().as_list()
out_h, out_w = self._output_shape(out_size, self.input_factor)
output_number = self._output_number(seq_len)
# prepare for flipping
tmp = tf.reshape(tensor,
[batch_size, output_number, -1, out_h, out_w])
tmp = tf.transpose(tmp, [0, 1, 3, 2, 4])
tmp = tf.reshape(
tmp,
[batch_size, out_h * output_number, out_w * output_number])
# flip
tmp = tf.reverse_v2(tmp, [1])
# transform back
tmp = tf.reshape(tmp,
[batch_size, output_number, out_h, -1, out_w])
tmp = tf.transpose(tmp, [0, 1, 3, 2, 4])
return tf.reshape(tmp, [batch_size, seq_len, out_size])
def polyconv(self, M):
"""
compute the coefficients of the product of polynomials with individual coefs as the rows of M. Fixed width
based on DXmax
"""
w = 4 * DXmax + 1
N = tf.concat(
[tf.zeros([self.latent_input_dim, 4]),
tf.cast(M, tf.float32)],
axis=1)
P = tf.foldl(
lambda a, x: tf.reshape(
tf.nn.convolution(
tf.reshape(a, [1, w, 1]),
tf.reshape(tf.reverse_v2(x, axis=[0]), [w, 1, 1]), 'SAME'),
[w]), N)
Q = tf.slice(P, [4], [-1])
# Q=tf.Print(Q,[M,Q,tf.shape(Q)],summarize=100,message='q')
return tf.cast(Q, tf.float64)
def __init__(self):
self.model_path = "./models"
self.shape_detector = "shape_predictor_68_face_landmarks.dat"
cuda = True
font_scale = 1
thickness = 1
#log('1')
print('1')
with tf.Graph().as_default():
config = tf.ConfigProto()
config.gpu_options.per_process_gpu_memory_fraction = 0.4
# session = tf.Session(config=config, ...)
self.sess = tf.Session(config=config)
self.images_pl = tf.placeholder(tf.float32,
shape=[None, 160, 160, 3],
name='input_image')
images = tf.map_fn(lambda frame: tf.reverse_v2(frame, [-1]),
self.images_pl) #BGR TO RGB
images_norm = tf.map_fn(
lambda frame: tf.image.per_image_standardization(frame),
images)
self.train_mode = tf.placeholder(tf.bool)
age_logits, gender_logits, _ = inception_resnet_v1.inference(
images_norm,
keep_probability=0.8,
phase_train=self.train_mode,
weight_decay=1e-5)
self.gender = tf.argmax(tf.nn.softmax(gender_logits), 1)
age_ = tf.cast(tf.constant([i for i in range(0, 101)]), tf.float32)
self.age = tf.reduce_sum(tf.multiply(tf.nn.softmax(age_logits),
age_),
axis=1)
init_op = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
self.sess.run(init_op)
saver = tf.train.Saver()
ckpt = tf.train.get_checkpoint_state(self.model_path)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(self.sess, ckpt.model_checkpoint_path)
print("restore and continue training!")
else:
pass
def sort_by_field(boxlist, field, order=SortOrder.descend, scope=None):
"""Sort boxes and associated fields according to a scalar field.
A common use case is reordering the boxes according to descending scores.
Args:
boxlist: BoxList holding N boxes.
field: A BoxList field for sorting and reordering the BoxList.
order: (Optional) descend or ascend. Default is descend.
scope: name scope.
Returns:
sorted_boxlist: A sorted BoxList with the field in the specified order.
Raises:
ValueError: if specified field does not exist
ValueError: if the order is not either descend or ascend
"""
with tf.name_scope(scope, 'SortByField'):
if order != SortOrder.descend and order != SortOrder.ascend:
raise ValueError('Invalid sort order')
field_to_sort = boxlist.get_field(field)
if len(field_to_sort.shape.as_list()) != 1:
raise ValueError('Field should have rank 1')
num_boxes = boxlist.num_boxes()
num_entries = tf.size(field_to_sort)
length_assert = tf.Assert(
tf.equal(num_boxes, num_entries),
['Incorrect field size: actual vs expected.', num_entries, num_boxes])
with tf.control_dependencies([length_assert]):
# TODO: Remove with tf.device when top_k operation runs
# correctly on GPU.
with tf.device('/cpu:0'):
_, sorted_indices = tf.nn.top_k(field_to_sort, num_boxes, sorted=True)
if order == SortOrder.ascend:
sorted_indices = tf.reverse_v2(sorted_indices, [0])
return gather(boxlist, sorted_indices)
def _create_session(self):
logger.debug("Creating session")
with tf.Graph().as_default():
self.session = tf.Session()
images_pl = tf.placeholder(tf.float32,
shape=[None, 160, 160, 3],
name='input_image')
images = tf.map_fn(lambda frame: tf.reverse_v2(frame, [-1]),
images_pl) #BGR TO RGB
images_norm = tf.map_fn(
lambda frame: tf.image.per_image_standardization(frame),
images)
train_mode = tf.placeholder(tf.bool)
age_logits, gender_logits, _ = resnet.inference(
images_norm,
keep_probability=0.8,
phase_train=train_mode,
weight_decay=1e-5)
gender = tf.argmax(tf.nn.softmax(gender_logits), 1)
age_ = tf.cast(tf.constant([i for i in range(0, 101)]), tf.float32)
age = tf.reduce_sum(tf.multiply(tf.nn.softmax(age_logits), age_),
axis=1)
init_op = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
self.session.run(init_op)
saver = tf.train.Saver()
ckpt = tf.train.get_checkpoint_state(self.model_path)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(self.session, ckpt.model_checkpoint_path)
self._age = age
self._gender = gender
self._images_pl = images_pl
self._train_mode = train_mode
logger.debug("Session restorted")
else:
logger.debug("Could not create session")
profiler.tick("Created session")
def __init__(self,
k=50,
filter_width=5,
pooling_size=4,
axis_num=3,
thread_num=4,
param_scope_name='MotionBasisLearner',
save_params_path='./save_variables/params',
summary_dir='./summaries/',
accumulated_steps=0):
self.k = k
self.filter_width = filter_width
self.pooling_size = pooling_size
self.axis_num = axis_num
self.save_params_path = save_params_path
self.summary_dir = summary_dir
self.accumulated_steps = accumulated_steps
self.param_scope_name = param_scope_name
assert thread_num >= 1
self.thread_num = thread_num # number of thread to run on.
with tf.variable_scope(param_scope_name) as crbm_scope:
self.w = tf.get_variable(
'weights',
shape=(axis_num, filter_width, 1, k),
dtype=tf.float32,
initializer=tf.random_normal_initializer())
self.w_r = tf.reverse_v2(self.w, [0, 1])
self.hb = tf.get_variable(
'hidden_biase',
shape=(k, ),
dtype=tf.float32,
initializer=tf.random_normal_initializer())
self.vb = tf.get_variable(
'visible_biase',
shape=(1, ),
dtype=tf.float32,
initializer=tf.random_normal_initializer())
self.sess = tf.Session()
# initialize parameters
self.sess.run(tf.global_variables_initializer())
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),
'age': tf.FixedLenFeature([], tf.int64),
'gender': tf.FixedLenFeature([], tf.int64),
'file_name': 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.image.decode_jpeg(features['image_raw'], channels=3)
# image = tf.image.resize_images(image, [64, 64])
# image = tf.cast(image, 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 = image * (1. / 255) - 0.5
image = tf.decode_raw(features['image_raw'], tf.uint8)
image.set_shape([160 * 160 * 3])
image = tf.reshape(image, [160, 160, 3])
image = tf.reverse_v2(image, [-1])
image = tf.image.per_image_standardization(image)
# image = tf.cast(image,tf.float32) * (1. / 255) - 0.5
# Convert label from a scalar uint8 tensor to an int32 scalar.
age = features['age']
gender = features['gender']
file_path = features['file_name']
return image, age, gender, file_path
def predict_labels_multi_scale(images,
model_options,
eval_scales=(1.0,),
add_flipped_images=False):
"""Predicts segmentation labels.
Args:
images: A tensor of size [batch, height, width, channels].
model_options: A ModelOptions instance to configure models.
eval_scales: The scales to resize images for evaluation.
add_flipped_images: Add flipped images for evaluation or not.
Returns:
A dictionary with keys specifying the output_type (e.g., semantic
prediction) and values storing Tensors representing predictions (argmax
over channels). Each prediction has size [batch, height, width].
"""
outputs_to_predictions = {
output: []
for output in model_options.outputs_to_num_classes
}
for i, image_scale in enumerate(eval_scales):
with tf.variable_scope(tf.get_variable_scope(), reuse=True if i else None):
outputs_to_scales_to_logits = multi_scale_logits(
images,
model_options=model_options,
image_pyramid=[image_scale],
is_training=False,
fine_tune_batch_norm=False)
if add_flipped_images:
with tf.variable_scope(tf.get_variable_scope(), reuse=True):
outputs_to_scales_to_logits_reversed = multi_scale_logits(
tf.reverse_v2(images, [2]),
model_options=model_options,
image_pyramid=[image_scale],
is_training=False,
fine_tune_batch_norm=False)
for output in sorted(outputs_to_scales_to_logits):
scales_to_logits = outputs_to_scales_to_logits[output]
logits = tf.image.resize_bilinear(
scales_to_logits[MERGED_LOGITS_SCOPE],
tf.shape(images)[1:3],
align_corners=True)
outputs_to_predictions[output].append(
tf.expand_dims(tf.nn.softmax(logits), 4))
if add_flipped_images:
scales_to_logits_reversed = (
outputs_to_scales_to_logits_reversed[output])
logits_reversed = tf.image.resize_bilinear(
tf.reverse_v2(scales_to_logits_reversed[MERGED_LOGITS_SCOPE], [2]),
tf.shape(images)[1:3],
align_corners=True)
outputs_to_predictions[output].append(
tf.expand_dims(tf.nn.softmax(logits_reversed), 4))
for output in sorted(outputs_to_predictions):
predictions = outputs_to_predictions[output]
# Compute average prediction across different scales and flipped images.
predictions = tf.reduce_mean(tf.concat(predictions, 4), axis=4)
outputs_to_predictions[output] = tf.argmax(predictions, 3)
return outputs_to_predictions
