def Inceptionv3_module3(self, net,name):
with tf.variable_scope(name) as scope:
with tf.variable_scope('branch1x1'):
branch1x1 = tl.conv2d(net, 320, (1, 1))
branch1x1 = tl.bn_new(branch1x1)
branch1x1 = tl.relu(branch1x1)
with tf.variable_scope('branch3x3'):
branch3x3 = tl.conv2d(net, 384, (1, 1),name='branch3x3_0')
branch3x3 = tl.bn_new(branch3x3)
branch3x3 = tl.relu(branch3x3)
branch3x3 = tf.concat([tl.conv2d(branch3x3, 384, (1, 3),name='branch3x3_1'),
tl.conv2d(branch3x3, 384, (3, 1),name='branch3x3_2')],3)
with tf.variable_scope('branch3x3dbl'):
branch3x3dbl = tl.conv2d(net, 448, (1, 1),name='branch3x3dbl_1')
branch3x3dbl = tl.bn_new(branch3x3dbl,name='branch3x3dbl_1')
branch3x3dbl = tl.relu(branch3x3dbl)
branch3x3dbl = tl.conv2d(branch3x3dbl, 384, (3, 3),name='branch3x3dbl_2')
branch3x3dbl = tl.bn_new(branch3x3dbl,name='branch3x3dbl_2')
branch3x3dbl = tl.relu(branch3x3dbl)
branch3x3dbl = tf.concat([tl.conv2d(branch3x3dbl, 384, (1, 3),name='branch3x3dbl_3'),
tl.conv2d(branch3x3dbl, 384, (3, 1),name='branch3x3dbl_4')],3)
with tf.variable_scope('branch_pool'):
branch_pool = tl.avg_pool2d(net, (3, 3))
branch_pool = tl.conv2d(branch_pool, 192, (1, 1))
branch_pool = tl.bn_new(branch_pool)
branch_pool = tl.relu(branch_pool)
net = tf.concat( [branch1x1, branch3x3, branch3x3dbl, branch_pool],3)
return net
def MobileNet_separable_2d(self, net, out_channel, stride = (1, 1), name='separable'):
width_multiplier=1
net = tl.depthwise_conv2d(net, width_multiplier, (3, 3), stride, name = name + 'dep')
net = tl.bn_new(net, name = name + 'dep')
net = tl.relu(net)
net = tl.conv2d(net, out_channel, (1,1), stride, name = name + 'conv')
net = tl.bn_new(net, name = name + 'conv')
net = tl.relu(net)
return net
def _conv_layer(self, input, out_channels, conv_ksize, conv_stride_size, pool_ksize, pool_stride_size, name):
with tf.variable_scope(name) as scope:
lconv = tl.conv2d(input, out_channels, conv_ksize, conv_stride_size, name='conv')
lpool = tl.max_pool2d(lconv, pool_ksize, pool_stride_size, name='pool')
lbn = tl.bn_new(lpool)
lrelu = tl.relu(lbn)
return lrelu
def inference_MobileNet(self, im):
n=6
num_classes=10
width_multiplier=1
net = tl.conv2d(im, round(32 * width_multiplier), (3, 3), (2, 2), padding='SAME', name='conv_1')
net = tl.bn_new(net, name='conv_1')
net = tl.relu(net)
net = self.MobileNet_separable_2d(net, 32, name ='sep1')
net = self.MobileNet_separable_2d(net, 64, (2, 2), name='sep2')
net = self.MobileNet_separable_2d(net, 128, name ='sep3')
net = self.MobileNet_separable_2d(net, 128, (2, 2), name='sep4')
net = self.MobileNet_separable_2d(net, 256, name ='sep5')
net = self.MobileNet_separable_2d(net, 256, (2, 2), name='sep6')
net = self.MobileNet_separable_2d(net, 512, name ='sep7')
for i in range(n):
net = self.MobileNet_separable_2d(net, 512, name ='sep' + str(i + 8))
net = self.MobileNet_separable_2d(net, 512, (2, 2), name ='sep' + str(n + 9))
net = self.MobileNet_separable_2d(net, 1024, name='sep' + str(n + 10))
#shape = net.get_shape()
#net = tl.avg_pool2d(net, shape[1:3], name='avg_pool')
net = tf.reshape(net, [self.batch_size, -1])
logits = tl.fc(net, num_classes, name='fc')
return logits
def residual_block(self,x, n_out, subsample, name):
with tf.variable_scope(name+'conv_1')as scope:
if subsample:
y = tl.conv2d(x, n_out, (3,3), (2,2), padding='SAME',name=name+'conv_1')
shortcut = tl.conv2d(x, n_out, (3,3), (2,2), padding='SAME',name=name +'shortcut')
else:
y = tl.conv2d(x, n_out, (3,3), (1,1), padding='SAME', name=name+'conv_1')
shortcut = tl.identity(x)
y = tl.bn_new(y)
y = tl.relu(y)
with tf.variable_scope(name+'conv_2')as scope:
y = tl.conv2d(y, n_out, (3,3), (1,1), padding='SAME', name=name+'conv_2')
y = tl.bn_new(y)
y = y + shortcut
y = tf.nn.relu(y, name='relu_2')
return y
def Inceptionv3_module1(self, net, kernel_size,name):
with tf.variable_scope(name):
with tf.variable_scope('branch1x1') as scope:
branch1x1 = tl.conv2d(net, 64, (1, 1))
branch1x1 = tl.bn_new(branch1x1)
branch1x1 = tl.relu(branch1x1)
with tf.variable_scope('branch5x5') as scope:
branch5x5 = tl.conv2d(net, 48, (1, 1),name='branch5x5_1')
branch5x5 = tl.bn_new(branch5x5,name='branch5x5_1')
branch5x5 = tl.relu(branch5x5)
branch5x5 = tl.conv2d(branch5x5, 64, kernel_size,name='branch5x5_2')
branch5x5 = tl.bn_new(branch5x5,name='branch5x5_2')
branch5x5 = tl.relu(branch5x5)
with tf.variable_scope('branch3x3dbl') as scope:
branch3x3dbl = tl.conv2d(net, 64, (1, 1),name='branch3x3dbl_1')
branch3x3dbl = tl.bn_new(branch3x3dbl,name='branch3x3dbl_1')
branch3x3dbl = tl.relu(branch3x3dbl)
branch3x3dbl = tl.conv2d(branch3x3dbl, 96, (3, 3),name='branch3x3dbl_2')
branch3x3dbl = tl.bn_new(branch3x3dbl,name='branch3x3dbl_2')
branch3x3dbl = tl.relu(branch3x3dbl)
branch3x3dbl = tl.conv2d(branch3x3dbl, 96, (3, 3),name='branch3x3dbl_3')
branch3x3dbl = tl.bn_new(branch3x3dbl,name='branch3x3dbl_3')
branch3x3dbl = tl.relu(branch3x3dbl)
with tf.variable_scope('branch_pool') as scope:
branch_pool = tl.avg_pool2d(net, (3, 3))
branch_pool = tl.conv2d(branch_pool, 32, (1, 1))
branch_pool = tl.bn_new(branch_pool)
branch_pool = tl.relu(branch_pool)
net = tf.concat([branch1x1, branch5x5, branch3x3dbl, branch_pool],3)
return net
def Inceptionv3_module2(self, net,name):
with tf.variable_scope(name) as scope:
with tf.variable_scope('branch1x1'):
branch1x1 = tl.conv2d(net, 192, (1, 1))
branch1x1 = tl.bn_new(branch1x1)
branch1x1 = tl.relu(branch1x1)
with tf.variable_scope('branch7x7'):
branch7x7 = tl.conv2d(net, 192, (1, 1),name='branch7x7_1')
branch7x7 = tl.bn_new(branch7x7,name='branch7x7_1')
branch7x7 = tl.relu(branch7x7)
branch7x7 = tl.conv2d(branch7x7, 192, (1, 7),name='branch7x7_2')
branch7x7 = tl.bn_new(branch7x7,name='branch7x7_2')
branch7x7 = tl.relu(branch7x7)
branch7x7 = tl.conv2d(branch7x7, 192, (7, 1),name='branch7x7_3')
branch7x7 = tl.bn_new(branch7x7,name='branch7x7_3')
branch7x7 = tl.relu(branch7x7)
with tf.variable_scope('branch7x7dbl'):
branch7x7dbl = tl.conv2d(net, 192, (1, 1),name='branch7x7dbl_1')
branch7x7dbl = tl.bn_new(branch7x7dbl,name='branch7x7dbl_1')
branch7x7dbl = tl.relu(branch7x7dbl)
branch7x7dbl = tl.conv2d(branch7x7dbl, 192, (7, 1),name='branch7x7dbl_2')
branch7x7dbl = tl.bn_new(branch7x7dbl,name='branch7x7dbl_2')
branch7x7dbl = tl.relu(branch7x7dbl)
branch7x7dbl = tl.conv2d(branch7x7dbl, 192, (1, 7),name='branch7x7dbl_3')
branch7x7dbl = tl.bn_new(branch7x7dbl,name='branch7x7dbl_3')
branch7x7dbl = tl.relu(branch7x7dbl)
branch7x7dbl = tl.conv2d(branch7x7dbl, 192, (7, 1),name='branch7x7dbl_4')
branch7x7dbl = tl.bn_new(branch7x7dbl,name='branch7x7dbl_4')
branch7x7dbl = tl.relu(branch7x7dbl)
branch7x7dbl = tl.conv2d(branch7x7dbl, 192, (1, 7),name='branch7x7dbl_5')
branch7x7dbl = tl.bn_new(branch7x7dbl,name='branch7x7dbl_5')
branch7x7dbl = tl.relu(branch7x7dbl)
with tf.variable_scope('branch_pool'):
branch_pool = tl.avg_pool2d(net, (3, 3))
branch_pool = tl.conv2d(branch_pool, 192, (1, 1))
branch_pool = tl.bn_new(branch_pool)
branch_pool = tl.relu(branch_pool)
net = tf.concat([branch1x1, branch7x7, branch7x7dbl, branch_pool],3)
return net
def inference_ResNet(self, im):
n=5
n_class=10
with tf.variable_scope('conv_0') as scope:
y = tl.conv2d(im, 16, (3,3), (1,1), padding='SAME', name='conv_init')
y = tl.bn_new(y)
y = tl.relu(y)
y = self.residual_group(y, 16, n, False, name='group_1')
y = self.residual_group(y, 32, n, True, name='group_2')
y = self.residual_group(y, 64, n, True, name='group_3')
#print(y.get_shape())
#shape = y.get_shape()
#y = tl.avg_pool2d(y, shape[1:3], (1, 1), padding='VALID', name='avg_pool')
y = tf.reshape(y, [self.batch_size, -1])
y = tl.fc(y, n_class, name='fc')
return y
def inference_Inceptionv3(self,im):
n=3
# n_class=10
n_class=101
with tf.variable_scope('Inceptionv3') as scope:
y = tl.conv2d(im, 32, (3, 3), (2, 2), name='conv0')
y = tl.bn_new(y,name='conv1')
y = tl.relu(y)
y = tl.conv2d(y, 32, (3, 3), (1, 1), name='conv1')
y = tl.bn_new(y,name='conv2')
y = tl.relu(y)
y = tl.conv2d(y, 64, (3, 3), (1, 1), padding='SAME', name='conv2')
y = tl.bn_new(y,name='conv3')
y = tl.relu(y)
y = tl.max_pool2d(y, (3, 3), (2, 2), name='pool1')
y = tl.bn_new(y,name='conv4')
y = tl.relu(y)
# 73 x 73 x 64
y = tl.conv2d(y, 80, (3, 3), (1, 1), name='conv3')
y = tl.bn_new(y,name='conv5')
y = tl.relu(y)
# 73 x 73 x 80.
y = tl.conv2d(y, 192, (3, 3), (2, 2), name='conv4')
y = tl.bn_new(y,name='conv6')
y = tl.relu(y)
# 71 x 71 x 192.
y = tl.max_pool2d(y, (3, 3), (2, 2), name='pool2')
# 35 x 35 x 192.
for i in range(n - 1):
y = self.Inceptionv3_module1(y, (5, 5),name='mixed_35x35x256a'+str(i))
for i in range(n - 1):
y = self.Inceptionv3_module2(y,name='mixed_17x17x768e'+str(i))
for i in range(n - 1):
y = self.Inceptionv3_module3(y,name='mixed_8x8x2048a'+str(i))
shape = y.get_shape()
y = tl.avg_pool2d(y, shape[1:3], padding='VALID', name='pool')
# 1 x 1 x 2048
y = tl.dropout(y, self.keep_prob)
y = tl.flatten(y)
# 2048
logits = tl.fc(y, n_class, name='logits')
return logits