def inference(input_tensor, regularizer=None, return_feature=False):
# do_conv(name, input_tensor, out_channel, ksize, stride=[1, 1, 1, 1], is_pretrain=True, dropout=False, regularizer=None):
for key in Config.CONV_LAYERS_CONFIG:
layer_config = Config.CONV_LAYERS_CONFIG[key]
if layer_config['batch_norm']:
input_tensor = batch_norm(input_tensor)
conv_res = do_conv(key,
input_tensor,
layer_config['deep'],
[layer_config['size'], layer_config['size']],
dropout=layer_config['dropout'],
is_pretrain=layer_config['trainable'],
batch_normalization=layer_config['batch_norm'])
input_tensor = conv_res
if layer_config['pooling']['exists']:
pooling = pool(layer_config['pooling']['name'], input_tensor)
input_tensor = pooling
# FC_layer(layer_name, x, out_nodes, regularizer=None):
fc1 = FC_layer('fc1', input_tensor, Config.FC_SIZE, regularizer)
fc1 = batch_norm(fc1)
fc2 = FC_layer('fc2', fc1, Config.OUTPUT_NODE, regularizer)
if return_feature:
return fc2, fc1
return fc2
def inference(input_tensors, regularizer):
CONV_LAYER_OUT = None
for index, input_tensor in enumerate(input_tensors):
# 对于每个尺度的图像，进行下面的ＲＮＮ-Part
input_shape = input_tensor.get_shape().as_list()
state = tf.zeros([Config.BATCH_SIZE, Config.STATE_FEATURE_DIM])
with tf.variable_scope('rnn-part-' + str(index), reuse=False):
# 对不同phase之间参数的一样的
for i in range(input_shape[3]):
if i == 0:
reuse = False
else:
reuse = True
cur_input = input_tensor[:, :, :, i]
cur_input = tf.reshape(
cur_input,
[input_shape[0], input_shape[1], input_shape[2], 1])
for key in Config.CONV_LAYERS_CONFIG:
layer_config = Config.CONV_LAYERS_CONFIG[key]
conv_res = do_conv(
key,
cur_input,
layer_config['deep'],
[layer_config['size'], layer_config['size']],
dropout=layer_config['dropout'],
reuse=reuse)
cur_input = conv_res
if layer_config['pooling']['exists']:
pooling = pool(
layer_config['pooling']['name'],
cur_input,
is_max_pool=layer_config['pooling']['is_max'])
cur_input = pooling
# 完成了卷积操作
cur_input = convert_two_dim(cur_input) # 展成二维的变量
print cur_input
cur_input = tf.concat([cur_input, state], axis=1)
print cur_input
state = FC_layer('extract_state', cur_input,
Config.STATE_FEATURE_DIM, regularizer,
reuse) # 更新状态
if CONV_LAYER_OUT is None:
CONV_LAYER_OUT = state
else:
CONV_LAYER_OUT = tf.concat([CONV_LAYER_OUT, state], axis=1)
input_tensor = CONV_LAYER_OUT
for key in Config.FC_LAYERS_CONFIG:
layer_config = Config.FC_LAYERS_CONFIG[key]
if not layer_config['regularizer']:
cur_regularizer = None
else:
cur_regularizer = regularizer
input_tensor = FC_layer(key, input_tensor, layer_config['size'],
cur_regularizer)
if layer_config['batch_norm']:
input_tensor = batch_norm(input_tensor)
return input_tensor
def inference(input_tensor, bg_tensor, regularizer=None):
# do_conv(name, input_tensor, out_channel, ksize, stride=[1, 1, 1, 1], is_pretrain=True, dropout=False, regularizer=None):
for key in Config.CONV_LAYERS_CONFIG:
layer_config = Config.CONV_LAYERS_CONFIG[key]
conv_res = do_conv(
key,
input_tensor,
layer_config['deep'],
[layer_config['size'], layer_config['size']],
dropout=layer_config['dropout']
)
input_tensor = conv_res
if layer_config['pooling']['exists']:
pooling = pool(
layer_config['pooling']['name'],
input_tensor
)
input_tensor = pooling
for key in Config.CONV_LAYERS_CONFIG_BG:
layer_config = Config.CONV_LAYERS_CONFIG_BG[key]
conv_res = do_conv(
key,
bg_tensor,
layer_config['deep'],
[layer_config['size'], layer_config['size']],
dropout=layer_config['dropout']
)
bg_tensor = conv_res
if layer_config['pooling']['exists']:
pooling = pool(
layer_config['pooling']['name'],
bg_tensor,
is_max_pool=layer_config['pooling']['max_pooling']
)
bg_tensor = pooling
input_tensor = convert_two_dim(input_tensor)
bg_tensor = convert_two_dim(bg_tensor)
input_tensor = tf.concat([input_tensor, bg_tensor], axis=1)
print input_tensor
# FC_layer(layer_name, x, out_nodes, regularizer=None):
fc1 = FC_layer(
'fc1',
input_tensor,
Config.FC_SIZE,
regularizer
)
fc1 = batch_norm(fc1)
fc2 = FC_layer(
'fc2',
fc1,
Config.OUTPUT_NODE,
regularizer
)
return fc2
def fc_layers(self):
# def FC_layer(layer_name, x, out_nodes):
images = FC_layer('fc6',
self.convs_output,
4096,
regularizer=self.regularizer,
dropout=True)
images = batch_norm(images)
images = FC_layer('fc7',
images,
4096,
regularizer=self.regularizer,
dropout=True)
images = batch_norm(images)
self.feature = images
images = FC_layer('fc8',
images,
self.classesnumber,
regularizer=self.regularizer)
self.fcs_output = images
def inference(input_tensor, regularizer):
input_shape = input_tensor.get_shape().as_list()
CONV_OUT = None
with tf.variable_scope('rnn-part'):
for i in range(input_shape[3]):
if i == 0:
reuse = False
else:
reuse = True
cur_input = input_tensor[:, :, :, i]
cur_input = tf.reshape(
cur_input, [input_shape[0], input_shape[1], input_shape[2], 1])
layer_keys = list(Config.CONV_LAYERS_CONFIG)
layer_keys.sort()
for key in layer_keys:
layer_config = Config.CONV_LAYERS_CONFIG[key]
conv_res = do_conv(
key,
cur_input,
layer_config['deep'],
[layer_config['size'], layer_config['size']],
dropout=layer_config['dropout'],
reuse=reuse)
cur_input = conv_res
if layer_config['pooling']['exists']:
pooling = pool(layer_config['pooling']['name'], cur_input)
cur_input = pooling
# 完成了卷积操作
cur_input = convert_two_dim(cur_input) # 展成二维的变量
if CONV_OUT is None:
CONV_OUT = cur_input
else:
CONV_OUT = tf.concat([CONV_OUT, cur_input], axis=1)
input_tensor = CONV_OUT
layer_keys = list(Config.FC_LAYERS_CONFIG)
layer_keys.sort()
for key in layer_keys:
layer_config = Config.FC_LAYERS_CONFIG[key]
if not layer_config['regularizer']:
cur_regularizer = None
else:
cur_regularizer = regularizer
input_tensor = FC_layer(key, input_tensor, layer_config['size'],
cur_regularizer)
if layer_config['batch_norm']:
input_tensor = batch_norm(input_tensor)
return input_tensor
def inference(input_tensors,
phase_names,
regularizer=None,
return_feature=False):
print input_tensors.get_shape().as_list()
# input_tensor 包含了三个期项，ｓｈａｐｅ是batchsize, image_w, image_h, 3
# do_conv(name, input_tensor, out_channel, ksize, stride=[1, 1, 1, 1], is_pretrain=True, dropout=False, regularizer=None):
keys = list(Config.CONV_LAYERS_CONFIG.keys())
keys.sort()
CONV_OUTPUT = None
for phase_index, phase in enumerate(phase_names):
phase_config = Config.CONV_LAYERS_CONFIG[phase]
conv_names = list(phase_config.keys())
conv_names.sort()
input_tensor = input_tensors[:, :, :, phase_index]
for conv_name in conv_names:
conv_layer_config = phase_config[conv_name]
conv_res = do_conv(
conv_name,
input_tensor,
conv_layer_config['deep'],
[conv_layer_config['size'], conv_layer_config['size']],
dropout=conv_layer_config['dropout'],
is_pretrain=conv_layer_config['trainable'],
batch_normalization=conv_layer_config['batch_norm'])
input_tensor = conv_res
if conv_layer_config['pooling']['exists']:
pooling = pool(conv_layer_config['pooling']['name'],
input_tensor)
input_tensor = pooling
input_tensor = convert_two_dim(input_tensor)
if CONV_OUTPUT is not None:
CONV_OUTPUT = tf.concat([CONV_OUTPUT, input_tensor], axis=1)
else:
CONV_OUTPUT = input_tensor
# FC_layer(layer_name, x, out_nodes, regularizer=None):
fc1 = FC_layer('fc1', CONV_OUTPUT, Config.FC_SIZE, regularizer)
fc1 = batch_norm(fc1)
fc2 = FC_layer('fc2', fc1, Config.OUTPUT_NODE, regularizer)
if return_feature:
return fc2, fc1
return fc2
def inference(input_tensors, regularizer=None):
conv_laerys_output = None
for index, input_tensor in enumerate(input_tensors):
if index == 0:
reuse = None
else:
reuse = True
for key in Config.CONV_LAYERS_CONFIG:
layer_config = Config.CONV_LAYERS_CONFIG[key]
conv_res = do_conv(key,
input_tensor,
layer_config['deep'],
[layer_config['size'], layer_config['size']],
dropout=layer_config['dropout'],
reuse=reuse)
input_tensor = conv_res
if layer_config['pooling']['exists']:
pooling = pool(layer_config['pooling']['name'], input_tensor)
input_tensor = pooling
if conv_laerys_output is None:
conv_laerys_output = convert_two_dim(input_tensor)
else:
conv_laerys_output = tf.concat(
[conv_laerys_output,
convert_two_dim(input_tensor)], 1)
input_tensor = conv_laerys_output
for key in Config.FC_LAYERS_CONFIG:
layer_config = Config.FC_LAYERS_CONFIG[key]
if not layer_config['regularizer']:
cur_regularizer = None
else:
cur_regularizer = regularizer
input_tensor = FC_layer(key, input_tensor, layer_config['size'],
cur_regularizer)
if layer_config['batch_norm']:
input_tensor = batch_norm(input_tensor)
return input_tensor
def inference(input_tensors, bg_tensors, regularizer):
CONV_OUTPUT = None
for index, input_tensor in enumerate(input_tensors):
# 同一个尺度
bg_tensor = bg_tensors[index]
shape = input_tensor.get_shape().as_list()
shape_bg = bg_tensor.get_shape().as_list()
state = tf.zeros(
shape=[
shape[0],
Config.STATE_FEATURE_DIM
],
dtype=tf.float32
)
with tf.variable_scope('rnn-part-' + str(index)):
for i in range(shape[3]):
cur_input_tensor = input_tensor[:, :, :, i] # 获取某一层的数据
cur_input_tensor = tf.reshape(
cur_input_tensor,
shape=[
shape[0],
shape[1],
shape[2],
1
]
)
cur_bg_tensor = bg_tensor[:, :, :, i]
cur_bg_tensor = tf.reshape(
cur_bg_tensor,
shape=[
shape_bg[0],
shape_bg[1],
shape_bg[2],
1
]
)
if i == 0:
reuse = False
else:
reuse = True
# 针对该ｐｈａｓｅ 计算roi的特征
for key in Config.CONV_LAYERS_CONFIG:
layer_config = Config.CONV_LAYERS_CONFIG[key]
# def do_conv(name, input_tensor, out_channel, ksize, stride=[1, 1, 1, 1], is_pretrain=True, dropout=False, regularizer=None, reuse=False):
conv_res = do_conv(
key,
cur_input_tensor,
layer_config['deep'],
[layer_config['size'], layer_config['size']],
dropout=layer_config['dropout'],
reuse=reuse
)
cur_input_tensor = conv_res
if layer_config['pooling']['exists']:
pooling = pool(
layer_config['pooling']['name'],
cur_input_tensor
)
cur_input_tensor = pooling
# 针对该ｐｈａｓｅ计算ｂｇ的特征
for key in Config.CONV_LAYERS_CONFIG_BG:
layer_config = Config.CONV_LAYERS_CONFIG_BG[key]
conv_res = do_conv(
key,
cur_bg_tensor,
layer_config['deep'],
[layer_config['size'], layer_config['size']],
dropout=layer_config['dropout'],
reuse=reuse
)
cur_bg_tensor = conv_res
if layer_config['pooling']['exists']:
pooling = pool(
layer_config['pooling']['name'],
cur_bg_tensor
)
cur_bg_tensor = pooling
cur_input_tensor = convert_two_dim(cur_input_tensor)
cur_bg_tensor = convert_two_dim(cur_bg_tensor)
fc_input = tf.concat([cur_input_tensor, cur_bg_tensor, state], axis=1)
state = FC_layer('extract_state', fc_input, Config.STATE_FEATURE_DIM, regularizer, reuse) # 更新状态
if CONV_OUTPUT is None:
CONV_OUTPUT = state
else:
CONV_OUTPUT = tf.concat([CONV_OUTPUT, state], axis=1)
input_tensor = CONV_OUTPUT
for key in Config.FC_LAYERS_CONFIG:
layer_config = Config.FC_LAYERS_CONFIG[key]
if not layer_config['regularizer']:
cur_regularizer = None
else:
cur_regularizer = regularizer
input_tensor = FC_layer(
key,
input_tensor,
layer_config['size'],
cur_regularizer
)
if layer_config['batch_norm']:
input_tensor = batch_norm(input_tensor)
return input_tensor
def inference(input_tensors, regularizer):
CONV_LAYER_OUT = None
params_shared = False
for index, input_tensor in enumerate(input_tensors):
# 对于每个尺度的图像，进行下面的ＲＮＮ-Part
input_shape = input_tensor.get_shape().as_list()
with tf.variable_scope('rnn-part-' + str(index), reuse=False):
# 对不同phase之间参数的一样的
ONE_SIZE_FEATURE = None
for i in range(input_shape[3]):
if i == 0:
reuse = False
else:
reuse = True
cur_input = input_tensor[:, :, :, i]
cur_input = tf.reshape(
cur_input,
[input_shape[0], input_shape[1], input_shape[2], 1])
for key in Config.CONV_LAYERS_CONFIG:
layer_config = Config.CONV_LAYERS_CONFIG[key]
cur_input = batch_norm(cur_input)
conv_res = do_conv(
key,
cur_input,
layer_config['deep'],
[layer_config['size'], layer_config['size']],
dropout=layer_config['dropout'],
reuse=reuse)
cur_input = conv_res
if layer_config['pooling']['exists']:
pooling = pool(
layer_config['pooling']['name'],
cur_input,
is_max_pool=layer_config['pooling']['is_max'])
cur_input = pooling
# 完成了卷积操作
cur_input = convert_two_dim(cur_input) # 展成二维的变量
print cur_input
if ONE_SIZE_FEATURE is None:
ONE_SIZE_FEATURE = cur_input
else:
ONE_SIZE_FEATURE = tf.concat([ONE_SIZE_FEATURE, cur_input],
axis=1)
FC_OUT = FC_layer('CONVERT-' + str(index), ONE_SIZE_FEATURE,
Config.SIZE_FEATURE_DIM, regularizer)
FC_OUT = batch_norm(FC_OUT)
print 'FC OUT is ', FC_OUT
if CONV_LAYER_OUT is None:
CONV_LAYER_OUT = FC_OUT
else:
CONV_LAYER_OUT = tf.concat([CONV_LAYER_OUT, FC_OUT], axis=1)
print 'CONVLAYEROUT is ', CONV_LAYER_OUT
print 'CONVLAYEROUT is ', CONV_LAYER_OUT
input_tensor = CONV_LAYER_OUT
for key in Config.FC_LAYERS_CONFIG:
layer_config = Config.FC_LAYERS_CONFIG[key]
if not layer_config['regularizer']:
cur_regularizer = None
else:
cur_regularizer = regularizer
input_tensor = FC_layer(key, input_tensor, layer_config['size'],
cur_regularizer)
if layer_config['batch_norm']:
input_tensor = batch_norm(input_tensor)
return input_tensor
def convlayers(self):
self.parameters = []
# zero-mean input
# with tf.name_scope('preprocess') as scope:
# mean = tf.constant([123.68, 116.779, 103.939], dtype=tf.float32, shape=[1, 1, 1, 3], name='img_mean')
# images = self.imgs-mean
# 对输入数据进行归一化
images = batch_norm(self.imgs)
# images = self.imgs
# images = self.imgs
# conv1_1
# do_conv(name, input_tensor, out_channel, ksize, stride=[1, 1, 1, 1], is_pretrain=True):
self.conv1_1 = do_conv('conv1_1',
images,
64, [3, 3],
is_pretrain=self.trainable['conv1_1'])
self.conv1_2 = do_conv('conv1_2',
self.conv1_1,
64, [3, 3],
is_pretrain=self.trainable['conv1_2'],
batch_normalization=False)
self.pooling1 = pool('pooling1', self.conv1_2, is_max_pool=True)
self.conv2_1 = do_conv('conv2_1',
self.pooling1,
128, [3, 3],
is_pretrain=self.trainable['conv2_1'])
self.conv2_2 = do_conv('conv2_2',
self.conv2_1,
128, [3, 3],
is_pretrain=self.trainable['conv2_2'],
batch_normalization=False)
self.pooling2 = pool('pooling2', self.conv2_2, is_max_pool=True)
self.conv3_1 = do_conv('conv3_1',
self.pooling2,
256, [3, 3],
is_pretrain=self.trainable['conv3_1'])
self.conv3_2 = do_conv('conv3_2',
self.conv3_1,
256, [3, 3],
is_pretrain=self.trainable['conv3_2'])
self.conv3_3 = do_conv('conv3_3',
self.conv3_2,
256, [3, 3],
is_pretrain=self.trainable['conv3_3'],
batch_normalization=False)
self.pooling3 = pool('pooing3', self.conv3_3, is_max_pool=True)
self.conv4_1 = do_conv('conv4_1',
self.pooling3,
512, [3, 3],
is_pretrain=self.trainable['conv4_1'])
self.conv4_2 = do_conv('conv4_2',
self.conv4_1,
512, [3, 3],
is_pretrain=self.trainable['conv4_2'])
self.conv4_3 = do_conv('conv4_3',
self.conv4_2,
512, [3, 3],
is_pretrain=self.trainable['conv4_3'],
batch_normalization=False)
self.pooling4 = pool('pooling4', self.conv4_3, is_max_pool=True)
self.conv5_1 = do_conv('conv5_1',
self.pooling4,
512, [3, 3],
is_pretrain=self.trainable['conv5_1'])
self.conv5_2 = do_conv('conv5_2',
self.conv5_1,
512, [3, 3],
is_pretrain=self.trainable['conv5_2'])
self.conv5_3 = do_conv('conv5_3',
self.conv5_2,
512, [3, 3],
is_pretrain=self.trainable['conv5_3'],
batch_normalization=False)
self.pooling5 = pool('pooling5', self.conv5_3, is_max_pool=True)
self.convs_output = self.pooling5