def extracted_features(input_tensor):
conv1 = layers.conv_layer('conv1', input_tensor, ksize=[5, 5], depth=128, kstep=[1, 1], padding='SAME')
pool2 = layers.pool_layer('pool2', conv1, ksize=[2, 2], kstep=[2, 2], padding='SAME', pool_fun=tf.nn.max_pool)
conv3 = layers.conv_layer('conv3', pool2, ksize=[5, 5], depth=256, kstep=[1, 1], padding='SAME')
pool4 = layers.pool_layer('pool4', conv3, ksize=[2, 2], kstep=[2, 2], padding='SAME', pool_fun=tf.nn.max_pool)
conv5 = layers.conv_layer('conv5', pool4, ksize=[5, 5], depth=128, kstep=[1, 1], padding='SAME')
pool6 = layers.pool_layer('pool6', conv5, ksize=[2, 2], kstep=[2, 2], padding='SAME', pool_fun=tf.nn.max_pool)
return pool6
def image_classify_cnn(input,classnum):
conv1 = layers.conv_layer('conv1',input,ksize=[5,5],depth=128,kstep=[1,1],padding='SAME')
pool2 = layers.pool_layer('pool2',conv1,ksize=[2,2],kstep=[2,2],padding='SAME',pool_fun=tf.nn.max_pool)
conv3 = layers.conv_layer('conv3',pool2,ksize=[5,5],depth=256,kstep=[1,1],padding='SAME')
pool4 = layers.pool_layer('pool4',conv3,ksize=[2,2],kstep=[2,2],padding='SAME',pool_fun=tf.nn.max_pool)
conv5 = layers.conv_layer('conv5',pool4,ksize=[5,5],depth=128,kstep=[1,1],padding='SAME')
pool6 = layers.pool_layer('pool6',conv5,ksize=[2,2],kstep=[2,2],padding='SAME',pool_fun=tf.nn.max_pool)
input = layers.convert_n_to_2('convert',pool6)
return image_classify(input,classnum)
def reduction_B_block(name, input):
'''
:param name:
:param input:
:return:
'''
with tf.name_scope(name) as scope:
pool_1b1 = layers.pool_layer(scope + 'pool_1b1',
input,
ksize=[3, 3],
kstep=[2, 2],
padding='VALID',
pool_fun=tf.nn.max_pool)
conv_1b2 = layers.conv_layer(scope + 'conv_1b2',
input,
ksize=[1, 1],
depth=192,
kstep=[1, 1],
padding='SAME')
conv_2b2 = layers.conv_layer(scope + 'conv_2b2',
conv_1b2,
ksize=[3, 3],
depth=192,
kstep=[2, 2],
padding='VALID')
conv_3b3 = layers.conv_layer(scope + 'conv_3b3',
input,
ksize=[1, 1],
depth=256,
kstep=[1, 1],
padding='SAME')
conv_4b3 = layers.conv_layer(scope + 'conv_4b3',
conv_3b3,
ksize=[1, 7],
depth=256,
kstep=[1, 1],
padding='SAME')
conv_5b3 = layers.conv_layer(scope + 'conv_5b3',
conv_4b3,
ksize=[7, 1],
depth=320,
kstep=[1, 1],
padding='SAME')
conv_6b3 = layers.conv_layer(scope + 'conv_6b3',
conv_5b3,
ksize=[3, 3],
depth=320,
kstep=[2, 2],
padding='VALID')
concat_0 = tf.concat(3, [pool_1b1, conv_2b2, conv_6b3],
name=scope + 'concat_0')
return concat_0
def stem_block(name, input):
'''
inception-v4神经网络的stem模块
:param name: 该模块的名称
:param input: 该模块的输入矩阵
:return:
'''
with tf.name_scope(name) as scope:
conv_1 = layers.conv_layer(scope + 'conv_1',
input,
ksize=[3, 3],
depth=32,
kstep=[2, 2],
padding='VALID')
conv_2 = layers.conv_layer(scope + 'conv_2',
conv_1,
ksize=[3, 3],
depth=32,
kstep=[1, 1],
padding='VALID')
conv_3 = layers.conv_layer(scope + 'conv_3',
conv_2,
ksize=[3, 3],
depth=64,
kstep=[1, 1],
padding='SAME')
pool_1b1 = layers.pool_layer(scope + 'pool_1b1',
conv_3,
ksize=[3, 3],
kstep=[2, 2],
padding='VALID',
pool_fun=tf.nn.max_pool)
conv_4b2 = layers.conv_layer(scope + 'conv_4b2',
conv_3,
ksize=[3, 3],
depth=96,
kstep=[2, 2],
padding='VALID')
concat_0 = tf.concat(3, [pool_1b1, conv_4b2], name=scope + 'concat_0')
conv_5b3 = layers.conv_layer(scope + 'conv_5b3',
concat_0,
ksize=[1, 1],
depth=64,
kstep=[1, 1],
padding='SAME')
conv_6b3 = layers.conv_layer(scope + 'conv_6b3',
conv_5b3,
ksize=[3, 3],
depth=96,
kstep=[1, 1],
padding='VALID')
conv_7b4 = layers.conv_layer(scope + 'conv_7b4',
concat_0,
ksize=[1, 1],
depth=64,
kstep=[1, 1],
padding='SAME')
conv_8b4 = layers.conv_layer(scope + 'conv_8b4',
conv_7b4,
ksize=[7, 1],
depth=64,
kstep=[1, 1],
padding='SAME')
conv_9b4 = layers.conv_layer(scope + 'conv_9b4',
conv_8b4,
ksize=[1, 7],
depth=64,
kstep=[1, 1],
padding='SAME')
conv_10b4 = layers.conv_layer(scope + 'conv_10b4',
conv_9b4,
ksize=[3, 3],
depth=96,
kstep=[1, 1],
padding='VALID')
concat_1 = tf.concat(3, [conv_6b3, conv_10b4], name=scope + 'concat_1')
conv_11b5 = layers.conv_layer(scope + 'conv_11b5',
concat_1,
ksize=[3, 3],
depth=192,
kstep=[1.1],
padding='VALID')
pool_2b6 = layers.pool_layer(scope + 'pool_2b6',
concat_1,
ksize=[1, 1],
kstep=[2, 2],
padding='VALID',
pool_fun=tf.nn.max_pool)
concat_2 = tf.concat(3, [conv_11b5, pool_2b6], name=scope + 'concat_2')
return concat_2
def inception_C_block(name, input):
'''
:param name:
:param input:
:return:
'''
with tf.name_scope(name) as scope:
pool_1b1 = layers.pool_layer(scope + 'pool_1b1',
input,
ksize=[1, 1],
kstep=[1, 1],
padding='SAME',
pool_fun=tf.nn.avg_pool)
conv_1b1 = layers.conv_layer(scope + 'conv_1b1',
pool_1b1,
ksize=[1, 1],
depth=256,
kstep=[1, 1],
padding='SAME')
conv_2b2 = layers.conv_layer(scope + 'conv_2b2',
input,
ksize=[1, 1],
depth=256,
kstep=[1, 1],
padding='SAME')
conv_3b3 = layers.conv_layer(scope + 'conv_3b3',
input,
ksize=[1, 1],
depth=384,
kstep=[1, 1],
padding='SAME')
conv_4b3_1 = layers.conv_layer(scope + 'conv_4b3_1',
conv_3b3,
ksize=[1, 3],
depth=256,
kstep=[1, 1],
padding='SAME')
conv_5b3_2 = layers.conv_layer(scope + 'conv_5b3_2',
conv_3b3,
ksize=[3, 1],
depth=256,
kstep=[1, 1],
padding='SAME')
conv_6b4 = layers.conv_layer(scope + 'conv_6b4',
input,
ksize=[1, 1],
depth=384,
kstep=[1, 1],
padding='SAME')
conv_7b4 = layers.conv_layer(scope + 'conv_7b4',
conv_6b4,
ksize=[1, 3],
depth=448,
kstep=[1, 1],
padding='SAME')
conv_8b4 = layers.conv_layer(scope + 'conv_8b4',
conv_7b4,
ksize=[3, 1],
depth=512,
kstep=[1, 1],
padding='SAME')
conv_9b4_1 = layers.conv_layer(scope + 'conv_9b4_1',
conv_8b4,
ksize=[3, 1],
depth=256,
kstep=[1, 1],
padding='SAME')
conv_10b4_2 = layers.conv_layer(scope + 'conv_10b4_2',
conv_8b4,
ksize=[1, 3],
depth=256,
kstep=[1, 1],
padding='SAME')
concat_0 = tf.concat(3, [
conv_1b1, conv_2b2, conv_4b3_1, conv_5b3_2, conv_9b4_1, conv_10b4_2
],
name=scope + 'concat_0')
return concat_0
def inception_A_block(name, input):
'''
:param name:
:param input:
:return:
'''
with tf.name_scope(name) as scope:
pool_1b1 = layers.pool_layer(scope + 'pool_1b1',
input,
ksize=[1, 1],
kstep=[1, 1],
padding='SAME',
pool_fun=tf.nn.avg_pool)
conv_1b1 = layers.conv_layer(scope + 'conv_1b1',
pool_1b1,
ksize=[1, 1],
depth=96,
kstep=[1, 1],
padding='SAME')
conv_2b2 = layers.conv_layer(scope + 'conv_2b2',
input,
ksize=[1, 1],
depth=96,
kstep=[1, 1],
padding='SAME')
conv_3b3 = layers.conv_layer(scope + 'conv_3b3',
input,
ksize=[1, 1],
depth=64,
kstep=[1, 1],
padding='SAME')
conv_4b3 = layers.conv_layer(scope + 'conv_4b3',
conv_3b3,
ksize=[3, 3],
depth=96,
kstep=[1, 1],
padding='SAME')
conv_5b4 = layers.conv_layer(scope + 'conv_5b4',
input,
ksize=[1, 1],
depth=64,
kstep=[1, 1],
padding='SAME')
conv_6b4 = layers.conv_layer(scope + "conv_6b4",
conv_5b4,
ksize=[3, 3],
depth=96,
kstep=[1, 1],
padding='SAME')
conv_7b4 = layers.conv_layer(scope + 'conv_7b4',
conv_6b4,
ksize=[3, 3],
depth=96,
kstep=[1, 1],
padding='SAME')
concat_0 = tf.concat(3, [conv_1b1, conv_2b2, conv_4b3, conv_7b4],
name=scope + 'concat_0')
return concat_0