def attention_gate(hi_input, lo_input, filters, keep_prob_):
hi_channel_att = global_avg_pool(hi_input, name='hi_channel_att')
lo_channel_att = global_avg_pool(lo_input, name='lo_channel_att')
hi_channel_att = tf.layers.conv2d(hi_channel_att,
filters,
1,
padding='same',
kernel_initializer='he_normal')
hi_channel_att = tf.nn.relu(hi_channel_att)
lo_channel_att = tf.layers.conv2d(lo_channel_att,
filters,
1,
padding='same',
kernel_initializer='he_normal')
lo_channel_att = tf.nn.relu(lo_channel_att)
att = tf.add(hi_channel_att, lo_channel_att)
att = tf.layers.conv2d(att,
filters,
1,
padding='same',
kernel_initializer='he_normal')
att = tf.nn.softmax(att)
output = att * lo_input
output = tf.concat([hi_input, output], 3)
return output
def selective_kernel_layer(sk_conv1, sk_conv2, sk_conv3, middle, out_dim):
sum_u = sk_conv1 + sk_conv2 + sk_conv3
squeeze = global_avg_pool(sum_u)
squeeze = tf.reshape(squeeze, [-1, 1, 1, out_dim])
z = tf.layers.dense(squeeze, use_bias=True, units=middle)
z = tf.nn.relu(z)
a1 = tf.layers.dense(z, use_bias=True, units=out_dim)
a2 = tf.layers.dense(z, use_bias=True, units=out_dim)
a3 = tf.layers.dense(z, use_bias=True, units=out_dim)
before_softmax = tf.concat([a1, a2, a3], 1)
after_softmax = tf.nn.softmax(before_softmax, dim=1)
a1 = after_softmax[:, 0, :, :]
a1 = tf.reshape(a1, [-1, 1, 1, out_dim])
a2 = after_softmax[:, 1, :, :]
a2 = tf.reshape(a2, [-1, 1, 1, out_dim])
a3 = after_softmax[:, 2, :, :]
a3 = tf.reshape(a3, [-1, 1, 1, out_dim])
select_1 = sk_conv1 * a1
select_2 = sk_conv2 * a2
select_3 = sk_conv3 * a3
out = select_1 + select_2 + select_3
return out
def se_net_res(input_x, ratio, layer_name, is_training, bn_decay):
with tf.name_scope(layer_name):
out_dim = input_x.get_shape()[-1].value
squeeze = global_avg_pool(input_x)
squeeze = tf.reshape(squeeze, [-1, 1, out_dim])
excitation = tf_util.conv1d(squeeze,
out_dim / ratio,
1,
padding='SAME',
bn=True,
is_training=is_training,
scope=layer_name + 'fc1',
bn_decay=bn_decay)
excitation = tf.nn.relu(excitation)
excitation = tf_util.conv1d(excitation,
out_dim,
1,
padding='SAME',
bn=True,
is_training=is_training,
scope=layer_name + 'fc2',
bn_decay=bn_decay)
excitation = tf.nn.sigmoid(excitation)
excitation = tf.reshape(excitation, [-1, 1, 1, out_dim])
scale = input_x * excitation
scale = scale + input_x
return scale
def Global_Average_Pooling(x, stride=1):
# width = np.shape(x)[1]
# height = np.shape(x)[2]
# pool_size = [width, height]
# return tf.layers.average_pooling2d(inputs=x, pool_size=pool_size, strides=stride)
return global_avg_pool(x, name='Global_avg_pooling')
def mobile_net_v2(input_shape, n_classes, img_prep=None, img_aug=None): """MobileNetv2 This function defines a MobileNetv2 architectures. Parameters ---------- input_shape: An integer or tuple/list of 3 integers, shape of input tensor. n_classes: Number of classes. img_prep: Function handle for image pre-processing img_aug: Function handle for image augmentation # Returns MobileNetv2 model. """ inputs = input_data(shape=input_shape, data_preprocessing=img_prep, data_augmentation=img_aug) x = reshape(inputs, [-1, input_shape[0], input_shape[1], 1]) x = _conv_block(x, 32, (3, 3), strides=(2, 2)) x = _inverted_residual_block(x, 16, (3, 3), t=1, strides=1, n=1) x = _inverted_residual_block(x, 24, (3, 3), t=6, strides=2, n=2) x = _inverted_residual_block(x, 32, (3, 3), t=6, strides=2, n=3) x = _inverted_residual_block(x, 64, (3, 3), t=6, strides=2, n=4) x = _inverted_residual_block(x, 96, (3, 3), t=6, strides=1, n=3) x = _inverted_residual_block(x, 160, (3, 3), t=6, strides=2, n=3) x = _inverted_residual_block(x, 320, (3, 3), t=6, strides=1, n=1) x = _conv_block(x, 1280, (1, 1), strides=(1, 1)) x = global_avg_pool(x) x = reshape(x, [-1, 1, 1, 1280]) x = dropout(x, 0.3, name='Dropout') x = conv_2d(x, n_classes, (1, 1), padding='same', activation='softmax', weights_init='xavier') output = reshape(x, [-1, n_classes]) return output
def build_SEnet(self, input_x):
input_x = self.conv_bn_layer(
input_x,
filters=self.out_dims[0],
filter_size=3,
stride=1,
scope="first_layer")
print(input_x)
for i, out_dim in enumerate(self.out_dims[1:]):
x = self.residual_layer(
(x if i else input_x),
out_dim=out_dim,
num_block=self.num_block,
depth=self.depth,
cardinality=self.cardinality,
reduction_ratio=self.reduction_ratio,
layer_num=str(i+1))
print(x)
x = global_avg_pool(x)
print(x)
x = flatten(x)
print(x)
return tf.layers.dense(inputs=x, use_bias=False, units=9)
def alchNet19(img_prep, img_aug, learning_rate):
network = input_data(shape=[None, 64, 64, 3],
data_preprocessing=img_prep,
data_augmentation=img_aug)
network = conv_2d(network, 64, 3, activation='relu')
network = batch_normalization(network)
network = resLayer(network, 64)
network = resLayer(network, 64)
network = resLayer(network, 128, stride = 2)
network = resLayer(network, 128)
network = resLayer(network, 256, stride = 2)
network = resLayer(network, 256)
network = resLayer(network, 512, stride = 2)
network = resLayer(network, 512)
network = global_avg_pool(network)
network = fully_connected(network, 1024, activation='relu')
network = batch_normalization(network, stddev=0.002, trainable=True, restore=True, reuse=False)
network = dropout(network, 0.5)
network = fully_connected(network, 200, activation='softmax')
network = regression(network, optimizer='momentum',
loss='categorical_crossentropy',
learning_rate=learning_rate)
return network
def squeeze_excitation_layer(input_x, out_dim, middle):
squeeze = global_avg_pool(input_x)
excitation = tf.layers.dense(squeeze, use_bias=True, units=middle)
excitation = tf.nn.relu(excitation)
excitation = tf.layers.dense(excitation, use_bias=True, units=out_dim)
excitation = tf.nn.sigmoid(excitation)
excitation = tf.reshape(excitation, [-1, 1, 1, out_dim])
scale = input_x * excitation
return scale
def squeeze_and_excitation(input_x,ratio=16):
x=global_avg_pool(input_x)
x=tf.layers.dense(x,units=input_x.shape[3]//ratio)
x=tf.nn.relu(x)
x=tf.layers.dense(x,units=input_x.shape[3])
x=tf.nn.sigmoid(x)
x=tf.reshape(x,[-1,1,1,input_x.shape[3]])
scale=x*input_x
return scale
def create(self, is_training=False):
with tf.variable_scope(self.scope, reuse=self.reuse):
with slim.arg_scope([slim.fully_connected],
activation_fn=tf.nn.relu):
with slim.arg_scope([slim.conv2d],
activation_fn=tf.nn.relu,
padding='VALID'):
net = self.inputs
net = slim.conv2d(net, 64, 5, scope='conv1')
self.conv1 = net
net = slim.max_pool2d(net, 2, stride=2, scope='pool1')
self.pool1 = net
net = slim.conv2d(net, 128, 5, scope='conv2')
self.conv2 = net
net = slim.max_pool2d(net, 2, stride=2, scope='pool2')
self.pool2 = net
self.f = tf.contrib.layers.flatten(net)
###############################################################################
self.avg = global_avg_pool(net, name="Global_avg_pooling")
excitation = slim.fully_connected(self.avg,
128 / 16,
activation_fn=tf.nn.relu,
scope='atten_fc1')
excitation = slim.fully_connected(
excitation,
128,
activation_fn=tf.nn.sigmoid,
scope='atten_fc2')
self.atten_weight = tf.reshape(excitation, [-1, 1, 1, 128])
net = net * self.atten_weight
self.cmp = net
self.att_flat = tf.contrib.layers.flatten(self.cmp)
###############################################################################
net = tf.contrib.layers.flatten(self.pool2)
net = slim.fully_connected(net,
1024,
activation_fn=tf.nn.relu,
scope='fc3')
self.fc3 = net
net = slim.dropout(net,
0.5,
is_training=self.training_flag)
net = slim.fully_connected(net,
64,
activation_fn=tf.tanh,
scope='fc4')
self.fc4 = net
net = slim.fully_connected(net,
10,
activation_fn=None,
scope='fc5')
self.fc5 = net
self.softmax_output = slim.softmax(net, scope='prediction')
def Global_Average_Pooling(x,stride=1):
"""
width = np.shape(x)[1]
height = np.shape(x)[2]
pool_size = [width,height]
return tf.layer.average_pooling2d(inputs=x,pool_size=pool_size,strides=stride)
# The strdie value does not matter.It is global average pooling without tflearn
"""
return global_avg_pool(x,name='Global_avg_pooling')
def build_network(self, loadModel=False):
"""
构建模型
"""
# Smaller 'AlexNet'
# https://github.com/tflearn/tflearn/blob/master/examples/images/alexnet.py
print('[+] Building CNN')
img_aug = ImageAugmentation()
img_aug.add_random_flip_leftright()
# img_aug.add_random_rotation(max_angle=25.)
img_aug.add_random_blur(sigma_max=0.3)
# 输入数据 http://tflearn.org/layers/core/#input-data
self.network = input_data(shape=[None, SIZE_FACE, SIZE_FACE, 1], data_augmentation=img_aug)
# self.network = input_data(shape=[None, SIZE_FACE, SIZE_FACE, 1])
# 卷积层 http://tflearn.org/layers/conv/#convolution-2d
# 激活函数 http://tflearn.org/activations/
self.network = conv_2d(self.network, 64, 3, activation='relu')
# self.gap1 = global_avg_pool(self.network)
# 池化层 http://tflearn.org/layers/conv/#max-pooling-2d
self.network = max_pool_2d(self.network, 2, strides=2)
# 卷积层
self.network = conv_2d(self.network, 96, 3, activation='relu')
# self.gap2 = global_avg_pool(self.network)
# 池化层
self.network = max_pool_2d(self.network, 2, strides=2)
# 卷积层
self.network = conv_2d(self.network, 128, 3, activation='relu')
self.network = global_avg_pool(self.network)
# 全连接层 http://tflearn.org/layers/core/#fully-connected
self.network = fully_connected(self.network, 2048, activation='relu',
weight_decay=0.001)
# dropout随机将部分输出改为0，避免过拟合 http://tflearn.org/layers/core/#dropout
self.network = dropout(self.network, 0.8)
# 全连接层：softmax分类
# self.network = merge([self.gap1, self.gap2, self.gap3], mode="concat", name="concat")
self.network = fully_connected(self.network, len(EMOTIONS), activation='softmax')
# 定义损失函数和优化器 http://tflearn.org/layers/estimator/#regression
self.network = regression(self.network,
# http://tflearn.org/optimizers/
optimizer='Adam',
# optimizer='SGD',
# http://tflearn.org/objectives/
loss='categorical_crossentropy',
learning_rate=0.001)
# 定义模型 http://tflearn.org/models/dnn/#deep-neural-network-model
self.model = tflearn.DNN(
self.network,
checkpoint_path=SAVE_DIRECTORY + '/emotion_recognition',
tensorboard_dir='c:\\tmp\\tflearn_logs',
max_checkpoints=1,
tensorboard_verbose=2
)
if loadModel:
self.load_model()
def Global_Average_Pooling(x, stride=1):
"""
width = np.shape(x)[1]
height = np.shape(x)[2]
pool_size = [width, height]
return tf.layers.average_pooling2d(inputs=x, pool_size=pool_size, strides=stride) # The stride value does not matter
It is global average pooling without tflearn
"""
return global_avg_pool(x, name='Global_avg_pooling')
def global_average_pool(x, name='global_avg_pooling'):
"""
width = np.shape(x)[1]
height = np.shape(x)[2]
pool_size = [width, height]
return tf.layers.average_pooling2d(inputs=x, pool_size=pool_size, strides=stride)
"""
net = global_avg_pool(x, name=name)
print('{}: {}'.format(name, net.get_shape()))
return net
def construct_inceptionv4onfire(x,y, training=False, enable_batch_norm=True):
network = input_data(shape=[None, y, x, 3])
#stem of inceptionV4
conv1_3_3 = conv_2d(network,32,3,strides=2,activation='relu',name='conv1_3_3_s2',padding='valid')
conv2_3_3 = conv_2d(conv1_3_3,32,3,activation='relu',name='conv2_3_3')
conv3_3_3 = conv_2d(conv2_3_3,64,3,activation='relu',name='conv3_3_3')
b_conv_1_pool = max_pool_2d(conv3_3_3,kernel_size=3,strides=2,padding='valid',name='b_conv_1_pool')
if enable_batch_norm:
b_conv_1_pool = batch_normalization(b_conv_1_pool)
b_conv_1_conv = conv_2d(conv3_3_3,96,3,strides=2,padding='valid',activation='relu',name='b_conv_1_conv')
b_conv_1 = merge([b_conv_1_conv,b_conv_1_pool],mode='concat',axis=3)
b_conv4_1_1 = conv_2d(b_conv_1,64,1,activation='relu',name='conv4_3_3')
b_conv4_3_3 = conv_2d(b_conv4_1_1,96,3,padding='valid',activation='relu',name='conv5_3_3')
b_conv4_1_1_reduce = conv_2d(b_conv_1,64,1,activation='relu',name='b_conv4_1_1_reduce')
b_conv4_1_7 = conv_2d(b_conv4_1_1_reduce,64,[1,7],activation='relu',name='b_conv4_1_7')
b_conv4_7_1 = conv_2d(b_conv4_1_7,64,[7,1],activation='relu',name='b_conv4_7_1')
b_conv4_3_3_v = conv_2d(b_conv4_7_1,96,3,padding='valid',name='b_conv4_3_3_v')
b_conv_4 = merge([b_conv4_3_3_v, b_conv4_3_3],mode='concat',axis=3)
b_conv5_3_3 = conv_2d(b_conv_4,192,3,padding='valid',activation='relu',name='b_conv5_3_3',strides=2)
b_pool5_3_3 = max_pool_2d(b_conv_4,kernel_size=3,padding='valid',strides=2,name='b_pool5_3_3')
if enable_batch_norm:
b_pool5_3_3 = batch_normalization(b_pool5_3_3)
b_conv_5 = merge([b_conv5_3_3,b_pool5_3_3],mode='concat',axis=3)
net = b_conv_5
# inceptionV4 modules
net=inception_block_a(net)
net=inception_block_b(net)
net=inception_block_c(net)
pool5_7_7=global_avg_pool(net)
if(training):
pool5_7_7=dropout(pool5_7_7,0.4)
loss = fully_connected(pool5_7_7, 2,activation='softmax')
if(training):
network = regression(loss, optimizer='rmsprop',
loss='categorical_crossentropy',
learning_rate=0.001)
else:
network=loss
model = tflearn.DNN(network, checkpoint_path='inceptionv4onfire',
max_checkpoints=1, tensorboard_verbose=0)
return model
def SE_layer(x, channel, reduction=4):
with tf.name_scope('SElayer'):
avg_pool = global_avg_pool(x, name='Global_avg_pool')
fc1 = tf.layers.dense(inputs=avg_pool,
use_bias=True,
units=channel // reduction)
ReLU = tf.nn.relu(fc1)
fc2 = tf.layers.dense(inputs=ReLU, use_bias=True, units=channel)
Sigmoid = tf.nn.sigmoid(fc2)
excitation = tf.reshape(Sigmoid, [-1, 1, 1, channel])
return x * excitation
def Global_Average_Pooling(x, stride=1):
"""
图像的行和列数据集合
width=np.shape(x)[1]
height=np.shape(x)[2]
:param x:
:param stride:
:return:
下面使用h5对其进行相应的存储
"""
return global_avg_pool(x,name='Global_Average_Pooling')
def channel_attention(input_x, out_dim, ratio, layer_name):
with tf.name_scope(layer_name):
squeeze = global_avg_pool(input_x, name='Global_avg_pooling')
excitation = tf.layers.dense(
inputs=squeeze, use_bias=False, units=out_dim / ratio, name=layer_name+'_'+'dense1')
excitation = tf.nn.relu(excitation)
excitation = tf.layers.dense(
inputs=excitation, use_bias=False, units=out_dim, name=layer_name+'_'+'dense2')
excitation = tf.nn.sigmoid(excitation)
excitation = tf.reshape(excitation, [-1, 1, 1, out_dim])
scale = input_x * excitation
return scale
class GoogLeNet:
network = input_data(shape=[None, 1024, 1024, 1])
network = conv_2d(network, 64, 9, strides=4, activation='relu', bias=False)
'''Bottleneck'''
network = tflearn.residual_bottleneck(network,
nb_blocks=3,
bottleneck_size=16,
out_channels=64)
network = tflearn.residual_bottleneck(network,
nb_blocks=1,
bottleneck_size=32,
out_channels=128,
downsample=True)
network = tflearn.residual_bottleneck(network,
nb_blocks=2,
bottleneck_size=32,
out_channels=128)
network = tflearn.residual_bottleneck(network,
nb_blocks=1,
bottleneck_size=64,
out_channels=256,
downsample=True)
network = tflearn.residual_bottleneck(network,
nb_blocks=2,
bottleneck_size=64,
out_channels=256)
network = batch_normalization(network)
network = tflearn.activation(network, 'relu')
network = global_avg_pool(network)
'''Output layer'''
output = fully_connected(network, 15, activation='sigmoid')
network = regression(output,
optimizer='momentum',
loss='binary_crossentropy',
learning_rate=0.01)
'''Set model + Save parameters + Tensorboard'''
model = tflearn.DNN(network,
checkpoint_path='params_resnet_cxr',
max_checkpoints=1,
tensorboard_verbose=0)
'''Feed the oxflowers17 dataset to the model'''
model.fit(train_x,
train_t,
n_epoch=10,
validation_set=(test_x, test_t),
show_metric=True,
batch_size=16,
snapshot_epoch=False,
snapshot_step=100,
run_id='resnet_cxr')
def ch_attn(x):
batch_size, height, width, ch = x.get_shape().as_list()
squeeze = global_avg_pool(x, name='Global_avg_pooling')
excitation = Fully_connected(squeeze, units=ch // 16, layer_name='_fully_connected1')
excitation = lrelu(excitation, 0.2)
excitation = Fully_connected(excitation, units=ch, layer_name='_fully_connected2')
excitation = tf.nn.sigmoid(excitation)
excitation = tf.reshape(excitation, [-1,1,1,ch])
scale = x * excitation
return scale
def exit_flow(input_data, num_classes):
net = slim.separable_conv2d(input_data, 728, [3, 3], depth_multiplier=1)
net = slim.separable_conv2d(net, 1024, [3, 3], depth_multiplier=1)
net = slim.conv2d(net, 1024, [2, 2], 2)
net1 = slim.conv2d(input_data, 1024, [1, 1], 2)
net = tf.add(net, net1)
net = slim.separable_conv2d(net, 1536, [3, 3], depth_multiplier=1)
net = slim.separable_conv2d(net, 2048, [3, 3], depth_multiplier=1)
net = global_avg_pool(net)
net = slim.fully_connected(net, num_classes)
net = slim.softmax(net)
return net
def ANN(WIDTH, HEIGHT, CHANNELS, LABELS):
dropout_value = 0.35
# Real-time data preprocessing
img_prep = ImagePreprocessing()
img_prep.add_featurewise_zero_center()
img_prep.add_featurewise_stdnorm()
# Building the network
network = input_data(shape=[None, WIDTH, HEIGHT, CHANNELS],
data_preprocessing=img_prep,
name='input')
network = conv_2d(network, 64, 3, activation='relu', bias=False)
# Residual blocks'
network = residual_bottleneck(network, 3, 16, 64)
network = residual_bottleneck(network, 1, 32, 128, downsample=True)
network = residual_bottleneck(network, 2, 32, 128)
network = residual_bottleneck(network, 1, 64, 256, downsample=True)
network = residual_bottleneck(network, 2, 64, 256)
network = residual_bottleneck(network, 1, 128, 512, downsample=True)
network = residual_bottleneck(network, 2, 128, 512)
network = batch_normalization(network)
network = activation(network, 'relu')
network = global_avg_pool(network)
# Output layer
network = fully_connected(network, LABELS, activation='softmax')
network = regression(network,
optimizer='adam',
learning_rate=0.01,
loss='categorical_crossentropy',
name='target')
'''
# Regression
network = regression(network, optimizer = 'momentum',
loss = 'categorical_crossentropy',
learning_rate = 0.1)
'''
model = tflearn.DNN(network,
tensorboard_verbose=0,
tensorboard_dir='./logs',
best_checkpoint_path='./checkpoints/best/best_val',
max_checkpoints=1)
return model
def squeeze_excitation_layer(self, input_x, out_dim, reduction_ratio,
layer_name):
with tf.name_scope(layer_name):
pool = global_avg_pool(input_x)
squeeze = tf.layers.dense(
pool,
use_bias=False,
units=out_dim/reduction_ratio,
)
squeeze = tf.nn.relu(squeeze)
excitation = tf.layers.dense(
squeeze, units=out_dim, use_bias=False)
excitation = tf.nn.sigmoid(excitation)
excitation = tf.reshape(excitation, [-1, 1, 1, out_dim])
return input_x*excitation
def network(self):
in_layer = input_data([None, 1, self.str_len * 2 + 2, 1])
indices = in_layer[:, 0, :2, 0]
if self.emb > 1:
lstm1 = lstm(embedding(in_layer[:, 0, 2:, 0], 26, self.emb),
300,
return_seq=True)
else:
lstm1 = lstm(in_layer[:, 0, 2:, :], 300, return_seq=True)
# lstm branch
lstm2 = lstm(lstm1, 300, return_seq=True)
lstm3 = lstm(lstm2, 300, return_seq=True)
lstm4 = lstm(lstm3, 300)
# cnn branch
in_layer = bn(in_layer)
conv1 = conv_2d(in_layer, 64, [1, 7], 1)
norm1 = relu(bn(conv1))
block1 = self.residual_block(norm1, 128, [1, 3], 2, stride=2)
block2 = self.residual_block(block1, 256, [1, 3], 2, stride=2)
block3 = self.residual_block(block2, 512, [1, 3], 2)
block4 = self.residual_block(block3, 1024, [1, 3], 2)
n_out_filters = block4.get_shape().as_list()[-1]
gap = tf.reshape(global_avg_pool(block4), [-1, n_out_filters])
# fully-connected branch
fc_ind = fc(indices, 100, activation='tanh')
fc_ind2 = fc(fc_ind, 100, activation='tanh')
# merge lstm, conv, and fc layers
merged = tf.concat([lstm4, gap, fc_ind2], 1)
out = fc(merged, self.num_classes,
activation='softmax') # output layer
# describe optimization
net = regression(out,
optimizer='adam',
loss='categorical_crossentropy',
learning_rate=self.lr)
# build model
model = tflearn.DNN(net, tensorboard_verbose=2, tensorboard_dir='.')
return model
def Dense_net(X, num_filter, nb_blocks):
conv = tf.layers.conv2d(
inputs=X, filters=2*num_filter, kernel_size=[3, 3], padding='SAME',
strides=2, kernel_initializer=tf.contrib.layers.xavier_initializer())
net = tf.layers.max_pooling2d(inputs=X, strides=2, pool_size=[3, 3], padding='VALID')
for idx, i in enumerate(nb_blocks):
net = dense_block(X=net, num_filter=num_filter, nb_layers=i, layer_name='dense_'+str(idx))
net = transition_layer(net, scope='trans_'+str(idx))
net = tf.layers.batch_normalization(inputs=net)
net = tf.nn.relu(net)
net = global_avg_pool(net, name='Global_avg_pooling')
net = tf.contrib.layers.flatten(net)
net = tf.layers.dense(inputs=net, units=OUTPUT_SIZE, name='linear')
return net
def MY_ResneXt(Img, ImageSize, MiniBatchSize):
num_classes = 10
depth = 8
cardinality = 32
blocks = 3
net = Img
net = conv_layer_with_relu(net=net,
num_filters=32,
kernel_size=3,
strides=1)
net = single_resneXt_block(net=net,
strides=1,
kernel_size=3,
depth=depth,
cardinality=cardinality)
net = conv_layer_with_relu(net=net,
num_filters=128,
kernel_size=1,
strides=2)
cardinality = 16
net = single_resneXt_block(net=net,
strides=1,
kernel_size=3,
depth=depth,
cardinality=cardinality)
net = global_avg_pool(net, name='Global_avg_pooling')
net = tf.contrib.layers.flatten(net)
# net = tf.layers.dense(inputs=net, name='layer_fc1',
# units=128, activation=tf.nn.relu)
#Contruct Fully-connected layer
net = tf.layers.dense(inputs=net,
name='layer_fc_out',
units=num_classes,
activation=None)
prLogits = net
prSoftMax = tf.nn.softmax(net)
return prLogits, prSoftMax
def Squeeze_excitation_layer(self, input_x, out_dim, ratio, layer_name):
with tf.name_scope(layer_name):
squeeze = global_avg_pool(input_x, name='Global_avg_pooling')
excitation = Fully_connected(squeeze,
units=out_dim / ratio,
layer_name=layer_name +
'_fully_connected1')
excitation = tf.nn.relu(excitation)
excitation = Fully_connected(excitation,
units=out_dim,
layer_name=layer_name +
'_fully_connected2')
excitation = tf.nn.sigmoid(excitation)
excitation = tf.reshape(excitation, [-1, 1, 1, out_dim])
scale = input_x * excitation
return scale
def ch_attn(input, channels, scope):
with tf.variable_scope(scope):
squeeze = global_avg_pool(input)
squeeze = tf.reshape(squeeze, [-1, 1, 1, channels])
squeeze = conv(squeeze,
channels // 8,
kernel=1,
stride=1,
scope="squeeze_conv")
excite = relu(squeeze)
excite = conv(excite,
channels,
kernel=1,
stride=1,
scope="excite_conv")
attn = tf.nn.sigmoid(excite)
output = attn * input
return output
def Squeeze_Excitation_Block(input_x, output_chn, ratio=16):
batch, in_depth, in_height, in_width, in_channels = [
int(d) for d in input_x.get_shape()
] # 取出各维度大小
Squeeze = tf.reshape(input_x,
(batch, in_depth, in_height * in_width, in_channels))
Squeeze = global_avg_pool(Squeeze)
#print("squeeze",Squeeze.shape)#1+通道数量
Excitation = tf.layers.dense(Squeeze,
units=output_chn / ratio,
use_bias=False) #进行压缩，学习注意参数
Excitation = tf.nn.relu(Excitation)
Excitation = tf.layers.dense(Excitation, units=output_chn, use_bias=False)
Excitation = tf.sigmoid(Excitation)
Excitation = tf.reshape(Excitation, [-1, 1, 1, 1, output_chn]) #生成了缩放尺寸
# print("ex:",Excitation.shape)
# print("input_x:", input_x.shape)
scale = input_x * Excitation #对输入进行缩放，美滋滋
return scale
def NCAM_Module(in_dim):
""" Position attention module"""
# Ref from SAGAN
chanel_in = in_dim
#gamma = tf.Variable(tf.zeros([1]),name='gamma')
m_batchsize, height, width, C = combined_static_and_dynamic_shape(
chanel_in)
globel_avg = global_avg_pool(chanel_in)
channel_avg_weights = tf.reshape(globel_avg, [1, C, -1])
globel_max = global_max_pool(chanel_in)
channel_max_weights = tf.reshape(globel_max, [1, -1, C])
energy = tf.matmul(channel_avg_weights, channel_max_weights) # 矩阵乘法
attention = tf.nn.softmax(energy, dim=-1) # 添加非线性函数
proj_value_CAM = tf.reshape(chanel_in, [m_batchsize, C, -1])
out = tf.matmul(attention, proj_value_CAM)
out = tf.reshape(out, [m_batchsize, height, width, C]) # reshape到原图
# out =gamma * out
out = PAM_Module(out)
out = out + chanel_in
return out
def Global_Average_Pooling(x, voxel, stride=1):
"""
width = np.shape(x)[1]
height = np.shape(x)[2]
pool_size = [width, height]
return tf.layers.average_pooling2d(inputs=x, pool_size=pool_size, strides=stride) # The stride value does not matter
It is global average pooling without tflearn
"""
start = 0
for i in range(len(voxel)):
end = start + voxel[i]
# 各个维度相加取平均
x_tmp = np.sum(x[start:end, :, :, :], axis=0) / (end - start)
if i == 0:
x_ave = x_tmp
else:
x_ave = np.vstack((x_ave, x_tmp))
start = end
return global_avg_pool(x_ave, name='Global_avg_pooling')
img_aug.add_random_flip_leftright()
img_aug.add_random_rotation(max_angle=25.)
# Convolutional network building
net = input_data(shape=[None, 32, 32, 3],
data_preprocessing=img_prep,
data_augmentation=img_aug)
filters = [64,128,256,512]
for f in filters:
net = fractal_conv2d(net, 4, f, 3,
normalizer_fn=batch_normalization)
net = slim.max_pool2d(net,2, 2)
net = fractal_conv2d(net, 4, 512, 2,
normalizer_fn=batch_normalization)
net = conv_2d(net, 10, 1)
net = global_avg_pool(net)
net = softmax(net)
net = regression(net, optimizer='adam',
loss='categorical_crossentropy',
learning_rate=.002)
# Train using classifier
model = tflearn.DNN(net, tensorboard_verbose=0)
model.fit(X, Y, n_epoch=400, shuffle=True, validation_set=(X_test, Y_test),
show_metric=True, batch_size=32, run_id='cifar10_cnn')
def Global_Average_Pooling(x):
return global_avg_pool(x, name='Global_avg_pooling')