tensorflow复现googlenet_v1，非高度集成模块

 

本次使用tensorflow的tf.nn函数，在尽可能少地定义函数的情况下复现googlenet-v1

这里demo在每一个inception是堆叠卷积后的结果，使用tf.concat函数可以。这里官方手册上给的通道在前，需要连接的多维矩阵在后，但是高版本的tensorflow是反过来的。

新版tensorflow：

layer_4a = tf.concat([layer_4a_1, layer_4a_2_2, layer_4a_3_2, layer_4a_4], 3)

旧版tensorflow：

layer_4a = tf.concat(3， [layer_4a_1, layer_4a_2_2, layer_4a_3_2, layer_4a_4])

最后附上完整的demo，未进行优化，demo较长，很笨重，但是较为方便理解。

import tensorflow as tf

# REGULARIZER = 0.01
BATCH_SIZE = 10

def weight_variable(shape, name=None):
    initial = tf.truncated_normal(shape, stddev=0.1)
    return tf.Variable(initial, name)

def bias_variable(shape,name=None):
    initial = tf.constant(0.1, shape=shape)
    return tf.Variable(initial, name)

def conv2d(input, filter, strides, padding="SAME", name=None):
    # filters with shape [filter_height * filter_width * in_channels, output_channels]
    # Must have strides[0] = strides[3] =1
    # For the most common case of the same horizontal and vertices strides, strides = [1, stride, stride, 1]
    '''
    Args:
        input: A Tensor. Must be one of the following types: float32, float64.
        filter: A Tensor. Must have the same type as input.
        strides: A list of ints. 1-D of length 4. The stride of the sliding window for each dimension of input.
        padding: A string from: "SAME", "VALID". The type of padding algorithm to use.
        use_cudnn_on_gpu: An optional bool. Defaults to True.
        name: A name for the operation (optional).
    '''
    return tf.nn.conv2d(input, filter, strides, padding="SAME", name=name)  # padding="SAME"用零填充边界

def GoogLeNet_v1(input, keep_prob = 0.4):
    '''参考论文: Going deeper with convolutions'''
    # input_size:(224,224,3)
    # Part1
    kernel_1 = weight_variable([7,7,3,64], name = 'kernel_1')
    bias_1 = bias_variable([64], name = 'bias_1')
    conv_layer_1 = conv2d(input, kernel_1, [1,2,2,1], name = 'conv_layer_1') + bias_1 # size = (112,112,64)
    layer_1 = tf.nn.relu(conv_layer_1, name = 'layer_1') # size = (112,112,64)
    maxpool_1 = tf.nn.max_pool(layer_1, ksize=[1,3,3,1], strides=[1,2,2,1], padding="SAME", name='maxpool_1') # size = (56,56,64)
    LocalRespNorm_1 = tf.nn.local_response_normalization(maxpool_1, name='LocalRespNorm_1') # size = (56,56,64)

    # Part2
    kernel_2_reduce = weight_variable([1,1,64,64], name = 'kernel_2_reduce')
    kernel_2 = weight_variable([3,3,64,192], name = 'kernel_2')
    bias_2_reduce = bias_variable([64], name = 'bias_2_reduce')
    bias_2 = bias_variable([192], name = 'bias_2')
    conv_layer_2_reduce = conv2d(LocalRespNorm_1, kernel_2_reduce, [1,1,1,1],padding='VALID', name = 'conv_layer_2_reduce') + bias_2_reduce # size = (56,56,64) 这里的padding不确定是不是SAME,论文上可能是VALID
    layer_2_reduce = tf.nn.relu(conv_layer_2_reduce, name = 'layer_2_reduce') # size = (56,56,64)
    conv_layer_2 = conv2d(layer_2_reduce, kernel_2, [1,1,1,1], name = 'conv_layer_2') + bias_2 # size = (56,56,192)
    layer_2 = tf.nn.relu(conv_layer_2, name = 'layer_2') # size = (56,56,192)
    LocalRespNorm_2 = tf.nn.local_response_normalization(layer_2, name='LocalRespNorm_2') # size = (56,56,192)
    max_pool_2 = tf.nn.max_pool(LocalRespNorm_2, ksize=[1,3,3,1], strides=[1,2,2,1], padding="SAME", name='maxpool_2') # size = (28,28,192)
    
    # Part3a
    kernel_3a_1 = weight_variable([1,1,192,64], name = 'kernel_3a_1')
    kernel_3a_2_1 = weight_variable([1,1,192,96], name = 'kernel_3a_2_1')
    kernel_3a_2_2 = weight_variable([3,3,96,128], name = 'kernel_3a_2_2')
    kernel_3a_3_1 = weight_variable([1,1,192,16], name = 'kernel_3a_3_1')
    kernel_3a_3_2 = weight_variable([5,5,16,32], name = 'kernel_3a_3_2')
    kernel_3a_4 = weight_variable([1,1,192,32], name = 'kernel_3a_4')
    bias_3a_1 = bias_variable([64], name = 'bias_3a_1')
    bias_3a_2_1 = bias_variable([96], name = 'bias_3a_2_1')
    bias_3a_2_2 = bias_variable([128], name = 'bias_3a_2_2')
    bias_3a_3_1 = bias_variable([16], name = 'bias_3a_3_1')
    bias_3a_3_2 = bias_variable([32], name = 'bias_3a_3_2')
    bias_3a_4 = bias_variable([32], name = 'bias_3a_4')
    layer_3a_1 = tf.nn.relu(conv2d(max_pool_2, kernel_3a_1, [1,1,1,1], name = 'conv_layer_3a_1') + bias_3a_1, name = 'layer_3a_1') # size = (28,28,64)
    layer_3a_2_1 = tf.nn.relu(conv2d(max_pool_2, kernel_3a_2_1, [1,1,1,1], name = 'conv_layer_3a_2_1') + bias_3a_2_1, name = 'layer_3a_2_1') # size = (28,28,96)
    layer_3a_2_2 = tf.nn.relu(conv2d(layer_3a_2_1, kernel_3a_2_2, [1,1,1,1], name = 'conv_layer_3a_2_2') + bias_3a_2_2, name = 'layer_3a_2_2') # size = (28,28,128)
    layer_3a_3_1 = tf.nn.relu(conv2d(max_pool_2, kernel_3a_3_1, [1,1,1,1], name = 'conv_layer_3a_3_1') + bias_3a_3_1, name = 'layer_3a_3_1') # size = (28,28,16)
    layer_3a_3_2 = tf.nn.relu(conv2d(layer_3a_3_1, kernel_3a_3_2, [1,1,1,1], name = 'conv_layer_3a_3_2') + bias_3a_3_2, name = 'layer_3a_3_2') # size = (28,28,32)
    max_pool_3a_1 = tf.nn.max_pool(max_pool_2, ksize=[1,3,3,1], strides=[1,1,1,1], padding="SAME", name='maxpool_3a_1') # size = (28,28,192)
    layer_3a_4 = tf.nn.relu(conv2d(max_pool_3a_1, kernel_3a_4, [1,1,1,1], name = 'conv_layer_3a_4') + bias_3a_4, name = 'layer_3a_4') # size = (28,28,32)
    layer_3a = tf.concat([layer_3a_1, layer_3a_2_2, layer_3a_3_2, layer_3a_4], 3) # size = (28,28,256)

    # Part3b
    kernel_3b_1 = weight_variable([1,1,256,128], name = 'kernel_3b_1')
    kernel_3b_2_1 = weight_variable([1,1,256,128], name = 'kernel_3b_2_1')
    kernel_3b_2_2 = weight_variable([3,3,128,192], name = 'kernel_3b_2_2')
    kernel_3b_3_1 = weight_variable([1,1,256,32], name = 'kernel_3b_3_1')
    kernel_3b_3_2 = weight_variable([5,5,32,96], name = 'kernel_3b_3_2')
    kernel_3b_4 = weight_variable([1,1,256,64], name = 'kernel_3b_4')
    bias_3b_1 = bias_variable([128], name = 'bias_3b_1')
    bias_3b_2_1 = bias_variable([128], name = 'bias_3b_2_1')
    bias_3b_2_2 = bias_variable([192], name = 'bias_3b_2_2')
    bias_3b_3_1 = bias_variable([32], name = 'bias_3b_3_1')
    bias_3b_3_2 = bias_variable([96], name = 'bias_3b_3_2')
    bias_3b_4 = bias_variable([64], name = 'bias_3b_4')
    layer_3b_1 = tf.nn.relu(conv2d(layer_3a, kernel_3b_1, [1,1,1,1], name = 'conv_layer_3b_1') + bias_3b_1, name = 'layer_3b_1') # size = (28,28,128)
    layer_3b_2_1 = tf.nn.relu(conv2d(layer_3a, kernel_3b_2_1, [1,1,1,1], name = 'conv_layer_3b_2_1') + bias_3b_2_1, name = 'layer_3b_2_1') # size = (28,28,128)
    layer_3b_2_2 = tf.nn.relu(conv2d(layer_3b_2_1, kernel_3b_2_2, [1,1,1,1], name = 'conv_layer_3b_2_2') + bias_3b_2_2, name = 'layer_3b_2_2') # size = (28,28,192)
    layer_3b_3_1 = tf.nn.relu(conv2d(layer_3a, kernel_3b_3_1, [1,1,1,1], name = 'conv_layer_3b_3_1') + bias_3b_3_1, name = 'layer_3b_3_1') # size = (28,28,32)
    layer_3b_3_2 = tf.nn.relu(conv2d(layer_3b_3_1, kernel_3b_3_2, [1,1,1,1], name = 'conv_layer_3b_3_2') + bias_3b_3_2, name = 'layer_3b_3_2') # size = (28,28,96)
    max_pool_3b_1 = tf.nn.max_pool(layer_3a, ksize=[1,3,3,1], strides=[1,1,1,1], padding="SAME", name='maxpool_3b_1') # size = (28,28,256)
    layer_3b_4 = tf.nn.relu(conv2d(max_pool_3b_1, kernel_3b_4, [1,1,1,1], name = 'conv_layer_3b_4') + bias_3b_4, name = 'layer_3b_4') # size = (28,28,64)
    layer_3b = tf.concat([layer_3b_1, layer_3b_2_2, layer_3b_3_2, layer_3b_4], 3) # size = (28,28,480)

    # Part4
    MaxPool_4 = tf.nn.max_pool(layer_3b, ksize=[1,3,3,1], strides=[1,2,2,1], padding="SAME", name='MaxPool_5') # size = (14,14,480)

    # Part4a
    kernel_4a_1 = weight_variable([1,1,480,192], name = 'kernel_4a_1')
    kernel_4a_2_1 = weight_variable([1,1,480,96], name = 'kernel_4a_2_1')
    kernel_4a_2_2 = weight_variable([3,3,96,208], name = 'kernel_4a_2_2')
    kernel_4a_3_1 = weight_variable([1,1,480,16], name = 'kernel_4a_3_1')
    kernel_4a_3_2 = weight_variable([5,5,16,48], name = 'kernel_4a_3_2')
    kernel_4a_4 = weight_variable([1,1,480,64], name = 'kernel_4a_4')
    bias_4a_1 = bias_variable([192], name = 'bias_4a_1')
    bias_4a_2_1 = bias_variable([96], name = 'bias_4a_2_1')
    bias_4a_2_2 = bias_variable([208], name = 'bias_4a_2_2')
    bias_4a_3_1 = bias_variable([16], name = 'bias_4a_3_1')
    bias_4a_3_2 = bias_variable([48], name = 'bias_4a_3_2')
    bias_4a_4 = bias_variable([64], name = 'bias_4a_4')
    layer_4a_1 = tf.nn.relu(conv2d(MaxPool_4, kernel_4a_1, [1,1,1,1], name = 'conv_layer_4a_1') + bias_4a_1, name = 'layer_4a_1') # size = (14,14,192)
    layer_4a_2_1 = tf.nn.relu(conv2d(MaxPool_4, kernel_4a_2_1, [1,1,1,1], name = 'conv_layer_4a_2_1') + bias_4a_2_1, name = 'layer_4a_2_1') # size = (14,14,96)
    layer_4a_2_2 = tf.nn.relu(conv2d(layer_4a_2_1, kernel_4a_2_2, [1,1,1,1], name = 'conv_layer_4a_2_2') + bias_4a_2_2, name = 'layer_4a_2_2') # size = (14,14,208)
    layer_4a_3_1 = tf.nn.relu(conv2d(MaxPool_4, kernel_4a_3_1, [1,1,1,1], name = 'conv_layer_4a_3_1') + bias_4a_3_1, name = 'layer_4a_3_1') # size = (14,14,16)
    layer_4a_3_2 = tf.nn.relu(conv2d(layer_4a_3_1, kernel_4a_3_2, [1,1,1,1], name = 'conv_layer_4a_3_2') + bias_4a_3_2, name = 'layer_4a_3_2') # size = (14,14,48)
    max_pool_4a_1 = tf.nn.max_pool(MaxPool_4, ksize=[1,3,3,1], strides=[1,1,1,1], padding="SAME", name='maxpool_4a_1') # size = (14,14,480)
    layer_4a_4 = tf.nn.relu(conv2d(max_pool_4a_1, kernel_4a_4, [1,1,1,1], name = 'conv_layer_4a_4') + bias_4a_4, name = 'layer_4a_4') # size = (14,14,64)
    layer_4a = tf.concat([layer_4a_1, layer_4a_2_2, layer_4a_3_2, layer_4a_4], 3) # size = (14,14,512)

    # Part4b
    kernel_4b_1 = weight_variable([1,1,512,160], name = 'kernel_4b_1')
    kernel_4b_2_1 = weight_variable([1,1,512,112], name = 'kernel_4b_2_1')
    kernel_4b_2_2 = weight_variable([3,3,112,224], name = 'kernel_4b_2_2')
    kernel_4b_3_1 = weight_variable([1,1,512,24], name = 'kernel_4b_3_1')
    kernel_4b_3_2 = weight_variable([5,5,24,64], name = 'kernel_4b_3_2')
    kernel_4b_4 = weight_variable([1,1,512,64], name = 'kernel_4b_4')
    bias_4b_1 = bias_variable([160], name = 'bias_4b_1')
    bias_4b_2_1 = bias_variable([112], name = 'bias_4b_2_1')
    bias_4b_2_2 = bias_variable([224], name = 'bias_4b_2_2')
    bias_4b_3_1 = bias_variable([24], name = 'bias_4b_3_1')
    bias_4b_3_2 = bias_variable([64], name = 'bias_4b_3_2')
    bias_4b_4 = bias_variable([64], name = 'bias_4b_4')
    layer_4b_1 = tf.nn.relu(conv2d(layer_4a, kernel_4b_1, [1,1,1,1], name = 'conv_layer_4b_1') + bias_4b_1, name = 'layer_4b_1') # size = (14,14,160)
    layer_4b_2_1 = tf.nn.relu(conv2d(layer_4a, kernel_4b_2_1, [1,1,1,1], name = 'conv_layer_4b_2_1') + bias_4b_2_1, name = 'layer_4b_2_1') # size = (14,14,112)
    layer_4b_2_2 = tf.nn.relu(conv2d(layer_4b_2_1, kernel_4b_2_2, [1,1,1,1], name = 'conv_layer_4b_2_2') + bias_4b_2_2, name = 'layer_4b_2_2') # size = (14,14,224)
    layer_4b_3_1 = tf.nn.relu(conv2d(layer_4a, kernel_4b_3_1, [1,1,1,1], name = 'conv_layer_4b_3_1') + bias_4b_3_1, name = 'layer_4b_3_1') # size = (14,14,24)
    layer_4b_3_2 = tf.nn.relu(conv2d(layer_4b_3_1, kernel_4b_3_2, [1,1,1,1], name = 'conv_layer_4b_3_2') + bias_4b_3_2, name = 'layer_4b_3_2') # size = (14,14,64)
    max_pool_4b_1 = tf.nn.max_pool(layer_4a, ksize=[1,3,3,1], strides=[1,1,1,1], padding="SAME", name='maxpool_4b_1') # size = (14,14,512)
    layer_4b_4 = tf.nn.relu(conv2d(max_pool_4b_1, kernel_4b_4, [1,1,1,1], name = 'conv_layer_4b_4') + bias_4b_4, name = 'layer_4b_4') # size = (14,14,64)
    layer_4b = tf.concat([layer_4b_1, layer_4b_2_2, layer_4b_3_2, layer_4b_4], 3) # size = (14,14,512)

    # Part4c
    kernel_4c_1 = weight_variable([1,1,512,128], name = 'kernel_4c_1')
    kernel_4c_2_1 = weight_variable([1,1,512,128], name = 'kernel_4c_2_1')
    kernel_4c_2_2 = weight_variable([3,3,128,256], name = 'kernel_4c_2_2')
    kernel_4c_3_1 = weight_variable([1,1,512,24], name = 'kernel_4c_3_1')
    kernel_4c_3_2 = weight_variable([5,5,24,64], name = 'kernel_4c_3_2')
    kernel_4c_4 = weight_variable([1,1,512,64], name = 'kernel_4c_4')
    bias_4c_1 = bias_variable([128], name = 'bias_4c_1')
    bias_4c_2_1 = bias_variable([128], name = 'bias_4c_2_1')
    bias_4c_2_2 = bias_variable([256], name = 'bias_4c_2_2')
    bias_4c_3_1 = bias_variable([24], name = 'bias_4c_3_1')
    bias_4c_3_2 = bias_variable([64], name = 'bias_4c_3_2')
    bias_4c_4 = bias_variable([64], name = 'bias_4c_4')
    layer_4c_1 = tf.nn.relu(conv2d(layer_4b, kernel_4c_1, [1,1,1,1], name = 'conv_layer_4c_1') + bias_4c_1, name = 'layer_4c_1') # size = (14,14,128)
    layer_4c_2_1 = tf.nn.relu(conv2d(layer_4b, kernel_4c_2_1, [1,1,1,1], name = 'conv_layer_4c_2_1') + bias_4c_2_1, name = 'layer_4c_2_1') # size = (14,14,128)
    layer_4c_2_2 = tf.nn.relu(conv2d(layer_4c_2_1, kernel_4c_2_2, [1,1,1,1], name = 'conv_layer_4c_2_2') + bias_4c_2_2, name = 'layer_4c_2_2') # size = (14,14,256)
    layer_4c_3_1 = tf.nn.relu(conv2d(layer_4b, kernel_4c_3_1, [1,1,1,1], name = 'conv_layer_4c_3_1') + bias_4c_3_1, name = 'layer_4c_3_1') # size = (14,14,24)
    layer_4c_3_2 = tf.nn.relu(conv2d(layer_4c_3_1, kernel_4c_3_2, [1,1,1,1], name = 'conv_layer_4c_3_2') + bias_4c_3_2, name = 'layer_4c_3_2') # size = (14,14,64)
    max_pool_4c_1 = tf.nn.max_pool(layer_4b, ksize=[1,3,3,1], strides=[1,1,1,1], padding="SAME", name='maxpool_4c_1') # size = (14,14,512)
    layer_4c_4 = tf.nn.relu(conv2d(max_pool_4c_1, kernel_4c_4, [1,1,1,1], name = 'conv_layer_4c_4') + bias_4c_4, name = 'layer_4c_4') # size = (14,14,64)
    layer_4c = tf.concat([layer_4c_1, layer_4c_2_2, layer_4c_3_2, layer_4c_4], 3) # size = (14,14,512)
    
    # Part4d
    kernel_4d_1 = weight_variable([1,1,512,112], name = 'kernel_4d_1')
    kernel_4d_2_1 = weight_variable([1,1,512,144], name = 'kernel_4d_2_1')
    kernel_4d_2_2 = weight_variable([3,3,144,288], name = 'kernel_4d_2_2')
    kernel_4d_3_1 = weight_variable([1,1,512,32], name = 'kernel_4d_3_1')
    kernel_4d_3_2 = weight_variable([5,5,32,64], name = 'kernel_4d_3_2')
    kernel_4d_4 = weight_variable([1,1,512,64], name = 'kernel_4d_4')
    bias_4d_1 = bias_variable([112], name = 'bias_4d_1')
    bias_4d_2_1 = bias_variable([144], name = 'bias_4d_2_1')
    bias_4d_2_2 = bias_variable([288], name = 'bias_4d_2_2')
    bias_4d_3_1 = bias_variable([32], name = 'bias_4d_3_1')
    bias_4d_3_2 = bias_variable([64], name = 'bias_4d_3_2')
    bias_4d_4 = bias_variable([64], name = 'bias_4d_4')
    layer_4d_1 = tf.nn.relu(conv2d(layer_4c, kernel_4d_1, [1,1,1,1], name = 'conv_layer_4d_1') + bias_4d_1, name = 'layer_4d_1') # size = (14,14,112)
    layer_4d_2_1 = tf.nn.relu(conv2d(layer_4c, kernel_4d_2_1, [1,1,1,1], name = 'conv_layer_4d_2_1') + bias_4d_2_1, name = 'layer_4d_2_1') # size = (14,14,144)
    layer_4d_2_2 = tf.nn.relu(conv2d(layer_4d_2_1, kernel_4d_2_2, [1,1,1,1], name = 'conv_layer_4d_2_2') + bias_4d_2_2, name = 'layer_4d_2_2') # size = (14,14,288)
    layer_4d_3_1 = tf.nn.relu(conv2d(layer_4c, kernel_4d_3_1, [1,1,1,1], name = 'conv_layer_4d_3_1') + bias_4d_3_1, name = 'layer_4d_3_1') # size = (14,14,32)
    layer_4d_3_2 = tf.nn.relu(conv2d(layer_4d_3_1, kernel_4d_3_2, [1,1,1,1], name = 'conv_layer_4d_3_2') + bias_4d_3_2, name = 'layer_4d_3_2') # size = (14,14,64)
    max_pool_4d_1 = tf.nn.max_pool(layer_4c, ksize=[1,3,3,1], strides=[1,1,1,1], padding="SAME", name='maxpool_4d_1') # size = (14,14,512)
    layer_4d_4 = tf.nn.relu(conv2d(max_pool_4d_1, kernel_4d_4, [1,1,1,1], name = 'conv_layer_4d_4') + bias_4d_4, name = 'layer_4d_4') # size = (14,14,64)
    layer_4d = tf.concat([layer_4d_1, layer_4d_2_2, layer_4d_3_2, layer_4d_4], 3) # size = (14,14,528)
    
    # Part4e
    kernel_4e_1 = weight_variable([1,1,528,256], name = 'kernel_4e_1')
    kernel_4e_2_1 = weight_variable([1,1,528,160], name = 'kernel_4e_2_1')
    kernel_4e_2_2 = weight_variable([3,3,160,320], name = 'kernel_4e_2_2')
    kernel_4e_3_1 = weight_variable([1,1,528,32], name = 'kernel_4e_3_1')
    kernel_4e_3_2 = weight_variable([5,5,32,128], name = 'kernel_4e_3_2')
    kernel_4e_4 = weight_variable([1,1,528,128], name = 'kernel_4e_4')
    bias_4e_1 = bias_variable([256], name = 'bias_4e_1')
    bias_4e_2_1 = bias_variable([160], name = 'bias_4e_2_1')
    bias_4e_2_2 = bias_variable([320], name = 'bias_4e_2_2')
    bias_4e_3_1 = bias_variable([32], name = 'bias_4e_3_1')
    bias_4e_3_2 = bias_variable([128], name = 'bias_4e_3_2')
    bias_4e_4 = bias_variable([128], name = 'bias_4e_4')
    layer_4e_1 = tf.nn.relu(conv2d(layer_4d, kernel_4e_1, [1,1,1,1], name = 'conv_layer_4e_1') + bias_4e_1, name = 'layer_4e_1') # size = (14,14,256)
    layer_4e_2_1 = tf.nn.relu(conv2d(layer_4d, kernel_4e_2_1, [1,1,1,1], name = 'conv_layer_4e_2_1') + bias_4e_2_1, name = 'layer_4e_2_1') # size = (14,14,160)
    layer_4e_2_2 = tf.nn.relu(conv2d(layer_4e_2_1, kernel_4e_2_2, [1,1,1,1], name = 'conv_layer_4e_2_2') + bias_4e_2_2, name = 'layer_4e_2_2') # size = (14,14,320)
    layer_4e_3_1 = tf.nn.relu(conv2d(layer_4d, kernel_4e_3_1, [1,1,1,1], name = 'conv_layer_4e_3_1') + bias_4e_3_1, name = 'layer_4e_3_1') # size = (14,14,32)
    layer_4e_3_2 = tf.nn.relu(conv2d(layer_4e_3_1, kernel_4e_3_2, [1,1,1,1], name = 'conv_layer_4e_3_2') + bias_4e_3_2, name = 'layer_4e_3_2') # size = (14,14,128)
    max_pool_4e_1 = tf.nn.max_pool(layer_4d, ksize=[1,3,3,1], strides=[1,1,1,1], padding="SAME", name='maxpool_4e_1') # size = (14,14,528)
    layer_4e_4 = tf.nn.relu(conv2d(max_pool_4e_1, kernel_4e_4, [1,1,1,1], name = 'conv_layer_4e_4') + bias_4e_4, name = 'layer_4e_4') # size = (14,14,128)
    layer_4e = tf.concat([layer_4e_1, layer_4e_2_2, layer_4e_3_2, layer_4e_4], 3) # size = (14,14,832)

    # Part5
    MaxPool_5 = tf.nn.max_pool(layer_4e, ksize=[1,3,3,1], strides=[1,2,2,1], padding="SAME", name='MaxPool_5') # size = (7,7,832)
    
    # Part5a
    kernel_5a_1 = weight_variable([1,1,832,256], name = 'kernel_5a_1')
    kernel_5a_2_1 = weight_variable([1,1,832,160], name = 'kernel_5a_2_1')
    kernel_5a_2_2 = weight_variable([3,3,160,320], name = 'kernel_5a_2_2')
    kernel_5a_3_1 = weight_variable([1,1,832,32], name = 'kernel_5a_3_1')
    kernel_5a_3_2 = weight_variable([5,5,32,128], name = 'kernel_5a_3_2')
    kernel_5a_4 = weight_variable([1,1,832,128], name = 'kernel_5a_4')
    bias_5a_1 = bias_variable([256], name = 'bias_5a_1')
    bias_5a_2_1 = bias_variable([160], name = 'bias_5a_2_1')
    bias_5a_2_2 = bias_variable([320], name = 'bias_5a_2_2')
    bias_5a_3_1 = bias_variable([32], name = 'bias_5a_3_1')
    bias_5a_3_2 = bias_variable([128], name = 'bias_5a_3_2')
    bias_5a_4 = bias_variable([128], name = 'bias_5a_4')
    layer_5a_1 = tf.nn.relu(conv2d(MaxPool_5, kernel_5a_1, [1,1,1,1], name = 'conv_layer_5a_1') + bias_5a_1, name = 'layer_5a_1') # size = (7,7,256)
    layer_5a_2_1 = tf.nn.relu(conv2d(MaxPool_5, kernel_5a_2_1, [1,1,1,1], name = 'conv_layer_5a_2_1') + bias_5a_2_1, name = 'layer_5a_2_1') # size = (7,7,160)
    layer_5a_2_2 = tf.nn.relu(conv2d(layer_5a_2_1, kernel_5a_2_2, [1,1,1,1], name = 'conv_layer_5a_2_2') + bias_5a_2_2, name = 'layer_5a_2_2') # size = (7,7,320)
    layer_5a_3_1 = tf.nn.relu(conv2d(MaxPool_5, kernel_5a_3_1, [1,1,1,1], name = 'conv_layer_5a_3_1') + bias_5a_3_1, name = 'layer_5a_3_1') # size = (7,7,32)
    layer_5a_3_2 = tf.nn.relu(conv2d(layer_5a_3_1, kernel_5a_3_2, [1,1,1,1], name = 'conv_layer_5a_3_2') + bias_5a_3_2, name = 'layer_5a_3_2') # size = (7,7,128)
    max_pool_5a_1 = tf.nn.max_pool(MaxPool_5, ksize=[1,3,3,1], strides=[1,1,1,1], padding="SAME", name='maxpool_5a_1') # size = (7,7,832)
    layer_5a_4 = tf.nn.relu(conv2d(max_pool_5a_1, kernel_5a_4, [1,1,1,1], name = 'conv_layer_5a_4') + bias_5a_4, name = 'layer_5a_4') # size = (7,7,128)
    layer_5a = tf.concat([layer_5a_1, layer_5a_2_2, layer_5a_3_2, layer_5a_4], 3) # size = (7,7,832)
    
    # Part5b
    kernel_5b_1 = weight_variable([1,1,832,384], name = 'kernel_5b_1')
    kernel_5b_2_1 = weight_variable([1,1,832,192], name = 'kernel_5b_2_1')
    kernel_5b_2_2 = weight_variable([3,3,192,384], name = 'kernel_5b_2_2')
    kernel_5b_3_1 = weight_variable([1,1,832,48], name = 'kernel_5b_3_1')
    kernel_5b_3_2 = weight_variable([5,5,48,128], name = 'kernel_5b_3_2')
    kernel_5b_4 = weight_variable([1,1,832,128], name = 'kernel_5b_4')
    bias_5b_1 = bias_variable([384], name = 'bias_5b_1')
    bias_5b_2_1 = bias_variable([192], name = 'bias_5b_2_1')
    bias_5b_2_2 = bias_variable([384], name = 'bias_5b_2_2')
    bias_5b_3_1 = bias_variable([48], name = 'bias_5b_3_1')
    bias_5b_3_2 = bias_variable([128], name = 'bias_5b_3_2')
    bias_5b_4 = bias_variable([128], name = 'bias_5b_4')
    layer_5b_1 = tf.nn.relu(conv2d(layer_5a, kernel_5b_1, [1,1,1,1], name = 'conv_layer_5b_1') + bias_5b_1, name = 'layer_5b_1') # size = (7,7,384)
    layer_5b_2_1 = tf.nn.relu(conv2d(layer_5a, kernel_5b_2_1, [1,1,1,1], name = 'conv_layer_5b_2_1') + bias_5b_2_1, name = 'layer_5b_2_1') # size = (7,7,192)
    layer_5b_2_2 = tf.nn.relu(conv2d(layer_5b_2_1, kernel_5b_2_2, [1,1,1,1], name = 'conv_layer_5b_2_2') + bias_5b_2_2, name = 'layer_5b_2_2') # size = (7,7,384)
    layer_5b_3_1 = tf.nn.relu(conv2d(layer_5a, kernel_5b_3_1, [1,1,1,1], name = 'conv_layer_5b_3_1') + bias_5b_3_1, name = 'layer_5b_3_1') # size = (7,7,48)
    layer_5b_3_2 = tf.nn.relu(conv2d(layer_5b_3_1, kernel_5b_3_2, [1,1,1,1], name = 'conv_layer_5b_3_2') + bias_5b_3_2, name = 'layer_5b_3_2') # size = (7,7,128)
    max_pool_5b_1 = tf.nn.max_pool(layer_5a, ksize=[1,3,3,1], strides=[1,1,1,1], padding="SAME", name='maxpool_5b_1') # size = (7,7,832)
    layer_5b_4 = tf.nn.relu(conv2d(max_pool_5b_1, kernel_5b_4, [1,1,1,1], name = 'conv_layer_5b_4') + bias_5b_4, name = 'layer_5b_4') # size = (7,7,128)
    layer_5b = tf.concat([layer_5b_1, layer_5b_2_2, layer_5b_3_2, layer_5b_4], 3) # size = (7,7,1024)

    # Softmax_assist_0
    Avg_assist_0 = tf.nn.avg_pool(layer_4a, ksize=[1,5,5,1], strides=[1,3,3,1], padding="VALID", name='Avg') # size = (4,4,512)
    kernel_assist_0 = weight_variable([1,1,512,1024], name = 'kernel_assist_0')
    bias_assist_0 = bias_variable([1024], name = 'bias_assist_0')
    conv_layer_assist_0 = tf.nn.relu(conv2d(Avg_assist_0, kernel_assist_0, [1,1,1,1]) + bias_assist_0,name = 'conv_layer_assist_0') # size = (4,4,1024)
    line_assist_0 = tf.reshape(conv_layer_assist_0, [-1, 16384])
    fc_assist_0_weight_1 = weight_variable([16384,1000], name = 'fc_assist_0_weight_1')
    fc_assist_0_bias_1 = bias_variable([1000], name = 'fc_assist_0_bias_1')
    fc_assist_0_weight_2 = weight_variable([1000,1000], name = 'fc_assist_0_weight_2')
    fc_assist_0_bias_2 = bias_variable([1000], name = 'fc_assist_0_bias_2')
    layer_assist_0_dropout_1 = tf.nn.dropout(line_assist_0, keep_prob, name = 'layer_assist_0_dropout_1')
    layer_assist_0_1 = tf.nn.relu(tf.matmul(layer_assist_0_dropout_1, fc_assist_0_weight_1, name = 'matmul_layer_assist_0_1') + fc_assist_0_bias_1, name = 'layer_assist_0_1')
    layer_assist_0_dropout_2 = tf.nn.dropout(layer_assist_0_1, keep_prob, name = 'layer_assist_0_dropout_2')
    layer_assist_0_2 = tf.nn.relu(tf.matmul(layer_assist_0_dropout_2, fc_assist_0_weight_2, name = 'matmul_layer_assist_0_2') + fc_assist_0_bias_2, name = 'layer_assist_0_2')

    output_assist_0 = tf.nn.softmax(layer_assist_0_2, name = 'output_assist_0')

    # Softmax_assist_1
    Avg_assist_1 = tf.nn.avg_pool(layer_4d, ksize=[1,5,5,1], strides=[1,3,3,1], padding="VALID", name='Avg') # size = (4,4,528)
    kernel_assist_1 = weight_variable([1,1,528,1024], name = 'kernel_assist_1')
    bias_assist_1 = bias_variable([1024], name = 'bias_assist_1')
    conv_layer_assist_1 = tf.nn.relu(conv2d(Avg_assist_1, kernel_assist_1, [1,1,1,1]) + bias_assist_1,name = 'conv_layer_assist_1') # size = (4,4,1024)
    line_assist_1 = tf.reshape(conv_layer_assist_1, [-1, 16384])
    fc_assist_1_weight_1 = weight_variable([16384,1000], name = 'fc_assist_1_weight_1')
    fc_assist_1_bias_1 = bias_variable([1000], name = 'fc_assist_1_bias_1')
    fc_assist_1_weight_2 = weight_variable([1000,1000], name = 'fc_assist_1_weight_2')
    fc_assist_1_bias_2 = bias_variable([1000], name = 'fc_assist_1_bias_2')
    layer_assist_1_dropout_1 = tf.nn.dropout(line_assist_1, keep_prob, name = 'layer_assist_1_dropout_1')
    layer_assist_1_1 = tf.nn.relu(tf.matmul(layer_assist_1_dropout_1, fc_assist_1_weight_1, name = 'matmul_layer_assist_1_1') + fc_assist_1_bias_1, name = 'layer_assist_1_1')
    layer_assist_1_dropout_2 = tf.nn.dropout(layer_assist_1_1, keep_prob, name = 'layer_assist_1_dropout_2')
    layer_assist_1_2 = tf.nn.relu(tf.matmul(layer_assist_1_dropout_2, fc_assist_1_weight_2, name = 'matmul_layer_assist_1_2') + fc_assist_1_bias_2, name = 'layer_assist_1_2')

    output_assist_1 = tf.nn.softmax(layer_assist_1_2, name = 'output_assist_1')

    # Softmax
    Avg = tf.nn.avg_pool(layer_5b, ksize=[1,7,7,1], strides=[1,1,1,1], padding="VALID", name='Avg') # size = (1,1,1024)
    line = tf.reshape(Avg, [-1, 1024])
    fc_weight = weight_variable([1024,1000], name = 'fc_weight')
    fc_bias = bias_variable([1000], name = 'fc_bias')
    layer_dropout = tf.nn.dropout(line, keep_prob, name = 'layer_dropout')
    layer = tf.nn.relu(tf.matmul(layer_dropout, fc_weight, name = 'matmul_layer') + fc_bias, name = 'layer')
    
    output = tf.nn.softmax(layer, name = 'output')

    return output, output_assist_0, output_assist_1

def backward(datasets, label, test_data, test_label):
    X = tf.placeholder(tf.float32, [None, 224,224,3], name = "Input")
    Y_ = tf.placeholder(tf.float32, [None, 1], name = 'Estimation')
    LEARNING_RATE_BASE = 0.0001  # 最初学习率
    LEARNING_RATE_DECAY = 0.99  # 学习率的衰减率
    LEARNING_RATE_STEP = 1000  # 喂入多少轮BATCH-SIZE以后，更新一次学习率。一般为总样本数量/BATCH_SIZE
    gloabl_steps = tf.Variable(0, trainable=False)  # 计数器，用来记录运行了几轮的BATCH_SIZE，初始为0，设置为不可训练
    learning_rate = tf.train.exponential_decay(LEARNING_RATE_BASE, gloabl_steps, LEARNING_RATE_STEP, LEARNING_RATE_DECAY, staircase=True)
    keep_prob = tf.placeholder(tf.float32, name = "keep_prob")
    y = GoogLeNet_v1(X)
    global_step = tf.Variable(0, trainable=False)
    loss_mse = tf.reduce_mean(-tf.reduce_sum(Y_ * tf.log(y), reduction_indices=[1]))
    # loss_mse = tf.reduce_mean(tf.square(y-Y_)) 
    # train_step = tf.train.GradientDescentOptimizer(0.001).minimize(loss_mse)
    # train_step=tf.train.MomentumOptimizer(0.001,0.9).minimize(loss_mse) #其他方法
    train_step=tf.train.AdamOptimizer(0.001).minimize(loss_mse)  
    saver = tf.train.Saver()
    with tf.Session() as sess:
        init_op = tf.global_variables_initializer()
        sess.run(init_op)
        # 训练模型。
        STEPS = 500001
        min_loss = 1
        for i in range(STEPS):
            start = (i*BATCH_SIZE) % 862
            end = start + BATCH_SIZE
            sess.run(train_step, feed_dict={X: datasets[start:end], Y_: label[start:end]})
            if i % 100 == 0:
                train_loss = sess.run(loss_mse, feed_dict={X: datasets, Y_: label, keep_prob:1})
                total_loss = sess.run(loss_mse, feed_dict={X: test_data, Y_: test_label, keep_prob:1})
                if total_loss < min_loss:
                    min_loss = total_loss
                f = open('./text/loss.txt', 'a')
                f.write("After %d training step(s), loss_mse on train data is %g, loss_mse on val data is %g, min_loss is %g\n" % (i, train_loss, total_loss, min_loss))
                print("After %d training step(s), loss_mse on train data is %g, loss_mse on val data is %g, min_loss is %g" % (i, train_loss, total_loss, min_loss))
                f.close()
            if i % 10000 == 0:
                saver.save(sess, './checkpoint/variable', global_step = i)

def main():
    # datasets, label, test_data, test_label = reload_all_data()
    datasets, label, test_data, test_label = Sequential_disruption()
    backward(datasets, label, test_data, test_label)

if __name__ == '__main__':
    main()
    

建议在inception这一块可以写一个函数，主要控制每个卷积核的个数，我这里为了直观就每写。

demo不能直接运行，如果要运行请修改最后一个main函数里面的代码段，直接把训练集和val集输入到background就行。这里输入一定是224*224*3的矩阵，输出是1000个的标签，如果想减小请在最后一个softmax区域把输出向量长度改小。

虽然有3个辅助分类器，最后输出三个，但是可以直接选择其中的一个，而不是3个共同训练。

这里再给出可视化demo：

import tensorflow as tf

# REGULARIZER = 0.01
BATCH_SIZE = 10


def weight_variable(shape, name=None):
    initial = tf.truncated_normal(shape, stddev=0.1)
    return tf.Variable(initial, name)


def bias_variable(shape, name=None):
    initial = tf.constant(0.1, shape=shape)
    return tf.Variable(initial, name)


def conv2d(input, filter, strides, padding="SAME", name=None):
    # filters with shape [filter_height * filter_width * in_channels, output_channels]
    # Must have strides[0] = strides[3] =1
    # For the most common case of the same horizontal and vertices strides, strides = [1, stride, stride, 1]
    '''
    Args:
        input: A Tensor. Must be one of the following types: float32, float64.
        filter: A Tensor. Must have the same type as input.
        strides: A list of ints. 1-D of length 4. The stride of the sliding window for each dimension of input.
        padding: A string from: "SAME", "VALID". The type of padding algorithm to use.
        use_cudnn_on_gpu: An optional bool. Defaults to True.
        name: A name for the operation (optional).
    '''
    return tf.nn.conv2d(input, filter, strides, padding="SAME", name=name)  # padding="SAME"用零填充边界


def GoogLeNet_v1(input, keep_prob=0.4):
    '''参考论文: Going deeper with convolutions'''
    # input_size:(224,224,3)
    # Part1
    with tf.name_scope('Conv_7x7'):
        kernel_1 = weight_variable([7, 7, 3, 64], name='kernel_1')
        bias_1 = bias_variable([64], name='bias_1')
        conv_layer_1 = conv2d(input, kernel_1, [1, 2, 2, 1], name='conv_layer_1') + bias_1  # size = (112,112,64)
        layer_1 = tf.nn.relu(conv_layer_1, name='layer_1')  # size = (112,112,64)
    with tf.name_scope('MaxPool_1'):
        maxpool_1 = tf.nn.max_pool(layer_1, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1], padding="SAME",
                                   name='maxpool_1')  # size = (56,56,64)
    with tf.name_scope('LocalRespNorm_1'):
        LocalRespNorm_1 = tf.nn.local_response_normalization(maxpool_1, name='LocalRespNorm_1')  # size = (56,56,64)

    # Part2
    with tf.name_scope('Conv_1x1'):
        kernel_2_reduce = weight_variable([1, 1, 64, 64], name='kernel_2_reduce')
        bias_2_reduce = bias_variable([64], name='bias_2_reduce')
        conv_layer_2_reduce = conv2d(LocalRespNorm_1, kernel_2_reduce, [1, 1, 1, 1], padding='VALID',
                                     name='conv_layer_2_reduce') + bias_2_reduce  # size = (56,56,64) 这里的padding不确定是不是SAME,论文上可能是VALID
        layer_2_reduce = tf.nn.relu(conv_layer_2_reduce, name='layer_2_reduce')  # size = (56,56,64)

    with tf.name_scope('Conv_3x3'):
        kernel_2 = weight_variable([3, 3, 64, 192], name='kernel_2')
        bias_2 = bias_variable([192], name='bias_2')
        conv_layer_2 = conv2d(layer_2_reduce, kernel_2, [1, 1, 1, 1], name='conv_layer_2') + bias_2  # size = (56,56,192)
        layer_2 = tf.nn.relu(conv_layer_2, name='layer_2')  # size = (56,56,192)

    with tf.name_scope('LocalRespNorm_2'):
        LocalRespNorm_2 = tf.nn.local_response_normalization(layer_2, name='LocalRespNorm_2')  # size = (56,56,192)

    with tf.name_scope('MaxPool_2'):
        max_pool_2 = tf.nn.max_pool(LocalRespNorm_2, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1], padding="SAME",
                                name='maxpool_2')  # size = (28,28,192)

    # Part3a
    with tf.name_scope('inception_3a'):
        with tf.name_scope('Part1'):
            kernel_3a_1 = weight_variable([1, 1, 192, 64], name='kernel_3a_1')
            bias_3a_1 = bias_variable([64], name='bias_3a_1')
            with tf.name_scope('conv1x1'):
                layer_3a_1 = tf.nn.relu(conv2d(max_pool_2, kernel_3a_1, [1, 1, 1, 1], name='conv_layer_3a_1') + bias_3a_1,
                                        name='layer_3a_1')  # size = (28,28,64)
        with tf.name_scope('Part2'):
            kernel_3a_2_1 = weight_variable([1, 1, 192, 96], name='kernel_3a_2_1')
            bias_3a_2_1 = bias_variable([96], name='bias_3a_2_1')
            kernel_3a_2_2 = weight_variable([3, 3, 96, 128], name='kernel_3a_2_2')
            bias_3a_2_2 = bias_variable([128], name='bias_3a_2_2')
            with tf.name_scope('conv1x1'):
                layer_3a_2_1 = tf.nn.relu(
                    conv2d(max_pool_2, kernel_3a_2_1, [1, 1, 1, 1], name='conv_layer_3a_2_1') + bias_3a_2_1,
                    name='layer_3a_2_1')  # size = (28,28,96)
            with tf.name_scope('conv3x3'):
                layer_3a_2_2 = tf.nn.relu(
                    conv2d(layer_3a_2_1, kernel_3a_2_2, [1, 1, 1, 1], name='conv_layer_3a_2_2') + bias_3a_2_2,
                    name='layer_3a_2_2')  # size = (28,28,128)
        with tf.name_scope('Part3'):
            kernel_3a_3_1 = weight_variable([1, 1, 192, 16], name='kernel_3a_3_1')
            bias_3a_3_1 = bias_variable([16], name='bias_3a_3_1')
            kernel_3a_3_2 = weight_variable([5, 5, 16, 32], name='kernel_3a_3_2')
            bias_3a_3_2 = bias_variable([32], name='bias_3a_3_2')
            with tf.name_scope('conv1x1'):
                layer_3a_3_1 = tf.nn.relu(
                    conv2d(max_pool_2, kernel_3a_3_1, [1, 1, 1, 1], name='conv_layer_3a_3_1') + bias_3a_3_1,
                    name='layer_3a_3_1')  # size = (28,28,16)
            with tf.name_scope('conv5x5'):
                layer_3a_3_2 = tf.nn.relu(
                    conv2d(layer_3a_3_1, kernel_3a_3_2, [1, 1, 1, 1], name='conv_layer_3a_3_2') + bias_3a_3_2,
                    name='layer_3a_3_2')  # size = (28,28,32)
        with tf.name_scope('Part4'):
            kernel_3a_4 = weight_variable([1, 1, 192, 32], name='kernel_3a_4')
            bias_3a_4 = bias_variable([32], name='bias_3a_4')
            with tf.name_scope('MaxPool_3x3'):
                max_pool_3a_1 = tf.nn.max_pool(max_pool_2, ksize=[1, 3, 3, 1], strides=[1, 1, 1, 1], padding="SAME",
                                               name='maxpool_3a_1')  # size = (28,28,192)
            with tf.name_scope('conv1x1'):
                layer_3a_4 = tf.nn.relu(conv2d(max_pool_3a_1, kernel_3a_4, [1, 1, 1, 1], name='conv_layer_3a_4') + bias_3a_4,
                                        name='layer_3a_4')  # size = (28,28,32)
        with tf.name_scope('DepthConcat'):
            layer_3a = tf.concat([layer_3a_1, layer_3a_2_2, layer_3a_3_2, layer_3a_4], 3)  # size = (28,28,256)

    # Part3b
    with tf.name_scope('inception_3b'):
        with tf.name_scope('Part1'):
            kernel_3b_1 = weight_variable([1, 1, 256, 128], name='kernel_3b_1')
            bias_3b_1 = bias_variable([128], name='bias_3b_1')
            with tf.name_scope('conv1x1'):
                layer_3b_1 = tf.nn.relu(conv2d(layer_3a, kernel_3b_1, [1, 1, 1, 1], name='conv_layer_3b_1') + bias_3b_1,
                                        name='layer_3b_1')  # size = (28,28,128)
        with tf.name_scope('Part2'):
            kernel_3b_2_1 = weight_variable([1, 1, 256, 128], name='kernel_3b_2_1')
            bias_3b_2_1 = bias_variable([128], name='bias_3b_2_1')
            kernel_3b_2_2 = weight_variable([3, 3, 128, 192], name='kernel_3b_2_2')
            bias_3b_2_2 = bias_variable([192], name='bias_3b_2_2')
            with tf.name_scope('conv1x1'):
                layer_3b_2_1 = tf.nn.relu(conv2d(layer_3a, kernel_3b_2_1, [1, 1, 1, 1], name='conv_layer_3b_2_1') + bias_3b_2_1,
                                          name='layer_3b_2_1')  # size = (28,28,128)
            with tf.name_scope('conv3x3'):
                layer_3b_2_2 = tf.nn.relu(
                    conv2d(layer_3b_2_1, kernel_3b_2_2, [1, 1, 1, 1], name='conv_layer_3b_2_2') + bias_3b_2_2,
                    name='layer_3b_2_2')  # size = (28,28,192)
        with tf.name_scope('Part3'):
            kernel_3b_3_1 = weight_variable([1, 1, 256, 32], name='kernel_3b_3_1')
            bias_3b_3_1 = bias_variable([32], name='bias_3b_3_1')
            kernel_3b_3_2 = weight_variable([5, 5, 32, 96], name='kernel_3b_3_2')
            bias_3b_3_2 = bias_variable([96], name='bias_3b_3_2')
            with tf.name_scope('conv1x1'):
                layer_3b_3_1 = tf.nn.relu(conv2d(layer_3a, kernel_3b_3_1, [1, 1, 1, 1], name='conv_layer_3b_3_1') + bias_3b_3_1,
                                          name='layer_3b_3_1')  # size = (28,28,32)
            with tf.name_scope('conv5x5'):
                layer_3b_3_2 = tf.nn.relu(
                    conv2d(layer_3b_3_1, kernel_3b_3_2, [1, 1, 1, 1], name='conv_layer_3b_3_2') + bias_3b_3_2,
                    name='layer_3b_3_2')  # size = (28,28,96)
        with tf.name_scope('Part4'):
            kernel_3b_4 = weight_variable([1, 1, 256, 64], name='kernel_3b_4')
            bias_3b_4 = bias_variable([64], name='bias_3b_4')
            with tf.name_scope('MaxPool_3x3'):
                max_pool_3b_1 = tf.nn.max_pool(layer_3a, ksize=[1, 3, 3, 1], strides=[1, 1, 1, 1], padding="SAME",
                                               name='maxpool_3b_1')  # size = (28,28,256)
            with tf.name_scope('conv1x1'):
                layer_3b_4 = tf.nn.relu(conv2d(max_pool_3b_1, kernel_3b_4, [1, 1, 1, 1], name='conv_layer_3b_4') + bias_3b_4,
                                        name='layer_3b_4')  # size = (28,28,64)
        with tf.name_scope('DepthConcat'):
            layer_3b = tf.concat([layer_3b_1, layer_3b_2_2, layer_3b_3_2, layer_3b_4], 3)  # size = (28,28,480)

        # Part4
    with tf.name_scope('MaxPool_3'):
        MaxPool_3 = tf.nn.max_pool(layer_3b, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1], padding="SAME",
                               name='MaxPool_5')  # size = (14,14,480)

    # Part4a
    with tf.name_scope('inception_4a'):
        with tf.name_scope('Part1'):
            kernel_4a_1 = weight_variable([1, 1, 480, 192], name='kernel_4a_1')
            bias_4a_1 = bias_variable([192], name='bias_4a_1')
            with tf.name_scope('conv1x1'):
                layer_4a_1 = tf.nn.relu(conv2d(MaxPool_3, kernel_4a_1, [1, 1, 1, 1], name='conv_layer_4a_1') + bias_4a_1,
                                        name='layer_4a_1')  # size = (14,14,192)
        with tf.name_scope('Part2'):
            kernel_4a_2_1 = weight_variable([1, 1, 480, 96], name='kernel_4a_2_1')
            bias_4a_2_1 = bias_variable([96], name='bias_4a_2_1')
            kernel_4a_2_2 = weight_variable([3, 3, 96, 208], name='kernel_4a_2_2')
            bias_4a_2_2 = bias_variable([208], name='bias_4a_2_2')
            with tf.name_scope('conv1x1'):
                layer_4a_2_1 = tf.nn.relu(
                    conv2d(MaxPool_3, kernel_4a_2_1, [1, 1, 1, 1], name='conv_layer_4a_2_1') + bias_4a_2_1,
                    name='layer_4a_2_1')  # size = (14,14,96)
            with tf.name_scope('conv3x3'):
                layer_4a_2_2 = tf.nn.relu(
                    conv2d(layer_4a_2_1, kernel_4a_2_2, [1, 1, 1, 1], name='conv_layer_4a_2_2') + bias_4a_2_2,
                    name='layer_4a_2_2')  # size = (14,14,208)
        with tf.name_scope('Part3'):
            kernel_4a_3_1 = weight_variable([1, 1, 480, 16], name='kernel_4a_3_1')
            bias_4a_3_1 = bias_variable([16], name='bias_4a_3_1')
            kernel_4a_3_2 = weight_variable([5, 5, 16, 48], name='kernel_4a_3_2')
            bias_4a_3_2 = bias_variable([48], name='bias_4a_3_2')
            with tf.name_scope('conv1x1'):
                layer_4a_3_1 = tf.nn.relu(
                    conv2d(MaxPool_3, kernel_4a_3_1, [1, 1, 1, 1], name='conv_layer_4a_3_1') + bias_4a_3_1,
                    name='layer_4a_3_1')  # size = (14,14,16)
            with tf.name_scope('conv5x5'):
                layer_4a_3_2 = tf.nn.relu(
                    conv2d(layer_4a_3_1, kernel_4a_3_2, [1, 1, 1, 1], name='conv_layer_4a_3_2') + bias_4a_3_2,
                    name='layer_4a_3_2')  # size = (14,14,48)
        with tf.name_scope('Part4'):
            kernel_4a_4 = weight_variable([1, 1, 480, 64], name='kernel_4a_4')
            bias_4a_4 = bias_variable([64], name='bias_4a_4')
            with tf.name_scope('MaxPool_3x3'):
                max_pool_4a_1 = tf.nn.max_pool(MaxPool_3, ksize=[1, 3, 3, 1], strides=[1, 1, 1, 1], padding="SAME",
                                               name='maxpool_4a_1')  # size = (14,14,480)
            with tf.name_scope('conv1x1'):
                layer_4a_4 = tf.nn.relu(conv2d(max_pool_4a_1, kernel_4a_4, [1, 1, 1, 1], name='conv_layer_4a_4') + bias_4a_4,
                                        name='layer_4a_4')  # size = (14,14,64)
        with tf.name_scope('DepthConcat'):
            layer_4a = tf.concat([layer_4a_1, layer_4a_2_2, layer_4a_3_2, layer_4a_4], 3)  # size = (14,14,512)

    # Part4b
    with tf.name_scope('inception_4b'):
        with tf.name_scope('Part1'):
            kernel_4b_1 = weight_variable([1, 1, 512, 160], name='kernel_4b_1')
            bias_4b_1 = bias_variable([160], name='bias_4b_1')
            with tf.name_scope('conv1x1'):
                layer_4b_1 = tf.nn.relu(conv2d(layer_4a, kernel_4b_1, [1, 1, 1, 1], name='conv_layer_4b_1') + bias_4b_1,
                                        name='layer_4b_1')  # size = (14,14,160)
        with tf.name_scope('Part2'):
            kernel_4b_2_1 = weight_variable([1, 1, 512, 112], name='kernel_4b_2_1')
            bias_4b_2_1 = bias_variable([112], name='bias_4b_2_1')
            kernel_4b_2_2 = weight_variable([3, 3, 112, 224], name='kernel_4b_2_2')
            bias_4b_2_2 = bias_variable([224], name='bias_4b_2_2')
            with tf.name_scope('conv1x1'):
                layer_4b_2_1 = tf.nn.relu(conv2d(layer_4a, kernel_4b_2_1, [1, 1, 1, 1], name='conv_layer_4b_2_1') + bias_4b_2_1,
                                          name='layer_4b_2_1')  # size = (14,14,112)
            with tf.name_scope('conv3x3'):
                layer_4b_2_2 = tf.nn.relu(
                    conv2d(layer_4b_2_1, kernel_4b_2_2, [1, 1, 1, 1], name='conv_layer_4b_2_2') + bias_4b_2_2,
                    name='layer_4b_2_2')  # size = (14,14,224)

        with tf.name_scope('Part3'):
            kernel_4b_3_1 = weight_variable([1, 1, 512, 24], name='kernel_4b_3_1')
            bias_4b_3_1 = bias_variable([24], name='bias_4b_3_1')
            kernel_4b_3_2 = weight_variable([5, 5, 24, 64], name='kernel_4b_3_2')
            bias_4b_3_2 = bias_variable([64], name='bias_4b_3_2')
            with tf.name_scope('conv1x1'):
                layer_4b_3_1 = tf.nn.relu(conv2d(layer_4a, kernel_4b_3_1, [1, 1, 1, 1], name='conv_layer_4b_3_1') + bias_4b_3_1,
                                          name='layer_4b_3_1')  # size = (14,14,24)
            with tf.name_scope('conv5x5'):
                layer_4b_3_2 = tf.nn.relu(
                    conv2d(layer_4b_3_1, kernel_4b_3_2, [1, 1, 1, 1], name='conv_layer_4b_3_2') + bias_4b_3_2,
                    name='layer_4b_3_2')  # size = (14,14,64)

        with tf.name_scope('Part4'):
            kernel_4b_4 = weight_variable([1, 1, 512, 64], name='kernel_4b_4')
            bias_4b_4 = bias_variable([64], name='bias_4b_4')
            with tf.name_scope('MaxPool_3x3'):
                max_pool_4b_1 = tf.nn.max_pool(layer_4a, ksize=[1, 3, 3, 1], strides=[1, 1, 1, 1], padding="SAME",
                                               name='maxpool_4b_1')  # size = (14,14,512)
            with tf.name_scope('conv1x1'):
                layer_4b_4 = tf.nn.relu(conv2d(max_pool_4b_1, kernel_4b_4, [1, 1, 1, 1], name='conv_layer_4b_4') + bias_4b_4,
                                        name='layer_4b_4')  # size = (14,14,64)
        with tf.name_scope('DepthConcat'):
            layer_4b = tf.concat([layer_4b_1, layer_4b_2_2, layer_4b_3_2, layer_4b_4], 3)  # size = (14,14,512)

    # Part4c
    with tf.name_scope('inception_4c'):
        with tf.name_scope('Part1'):
            kernel_4c_1 = weight_variable([1, 1, 512, 128], name='kernel_4c_1')
            bias_4c_1 = bias_variable([128], name='bias_4c_1')
            with tf.name_scope('conv1x1'):
                layer_4c_1 = tf.nn.relu(conv2d(layer_4b, kernel_4c_1, [1, 1, 1, 1], name='conv_layer_4c_1') + bias_4c_1,
                                        name='layer_4c_1')  # size = (14,14,128)
        with tf.name_scope('Part2'):
            kernel_4c_2_1 = weight_variable([1, 1, 512, 128], name='kernel_4c_2_1')
            bias_4c_2_1 = bias_variable([128], name='bias_4c_2_1')
            kernel_4c_2_2 = weight_variable([3, 3, 128, 256], name='kernel_4c_2_2')
            bias_4c_2_2 = bias_variable([256], name='bias_4c_2_2')
            with tf.name_scope('conv1x1'):
                layer_4c_2_1 = tf.nn.relu(conv2d(layer_4b, kernel_4c_2_1, [1, 1, 1, 1], name='conv_layer_4c_2_1') + bias_4c_2_1,
                                          name='layer_4c_2_1')  # size = (14,14,128)
            with tf.name_scope('conv3x3'):
                layer_4c_2_2 = tf.nn.relu(
                    conv2d(layer_4c_2_1, kernel_4c_2_2, [1, 1, 1, 1], name='conv_layer_4c_2_2') + bias_4c_2_2,
                    name='layer_4c_2_2')  # size = (14,14,256)
        with tf.name_scope('Part3'):
            kernel_4c_3_1 = weight_variable([1, 1, 512, 24], name='kernel_4c_3_1')
            bias_4c_3_1 = bias_variable([24], name='bias_4c_3_1')
            kernel_4c_3_2 = weight_variable([5, 5, 24, 64], name='kernel_4c_3_2')
            bias_4c_3_2 = bias_variable([64], name='bias_4c_3_2')
            with tf.name_scope('conv1x1'):
                layer_4c_3_1 = tf.nn.relu(conv2d(layer_4b, kernel_4c_3_1, [1, 1, 1, 1], name='conv_layer_4c_3_1') + bias_4c_3_1,
                                          name='layer_4c_3_1')  # size = (14,14,24)
            with tf.name_scope('conv5x5'):
                layer_4c_3_2 = tf.nn.relu(
                    conv2d(layer_4c_3_1, kernel_4c_3_2, [1, 1, 1, 1], name='conv_layer_4c_3_2') + bias_4c_3_2,
                    name='layer_4c_3_2')  # size = (14,14,64)
        with tf.name_scope('Part4'):
            kernel_4c_4 = weight_variable([1, 1, 512, 64], name='kernel_4c_4')
            bias_4c_4 = bias_variable([64], name='bias_4c_4')
            with tf.name_scope('MaxPool_3x3'):
                max_pool_4c_1 = tf.nn.max_pool(layer_4b, ksize=[1, 3, 3, 1], strides=[1, 1, 1, 1], padding="SAME",
                                               name='maxpool_4c_1')  # size = (14,14,512)
            with tf.name_scope('conv1x1'):
                layer_4c_4 = tf.nn.relu(conv2d(max_pool_4c_1, kernel_4c_4, [1, 1, 1, 1], name='conv_layer_4c_4') + bias_4c_4,
                                        name='layer_4c_4')  # size = (14,14,64)
        with tf.name_scope('DepthConcat'):
            layer_4c = tf.concat([layer_4c_1, layer_4c_2_2, layer_4c_3_2, layer_4c_4], 3)  # size = (14,14,512)

    # Part4d
    with tf.name_scope('inception_4d'):
        with tf.name_scope('Part1'):
            kernel_4d_1 = weight_variable([1, 1, 512, 112], name='kernel_4d_1')
            bias_4d_1 = bias_variable([112], name='bias_4d_1')
            with tf.name_scope('conv1x1'):
                layer_4d_1 = tf.nn.relu(conv2d(layer_4c, kernel_4d_1, [1, 1, 1, 1], name='conv_layer_4d_1') + bias_4d_1,
                                        name='layer_4d_1')  # size = (14,14,112)
        with tf.name_scope('Part2'):
            kernel_4d_2_1 = weight_variable([1, 1, 512, 144], name='kernel_4d_2_1')
            bias_4d_2_1 = bias_variable([144], name='bias_4d_2_1')
            kernel_4d_2_2 = weight_variable([3, 3, 144, 288], name='kernel_4d_2_2')
            bias_4d_2_2 = bias_variable([288], name='bias_4d_2_2')
            with tf.name_scope('conv1x1'):
                layer_4d_2_1 = tf.nn.relu(conv2d(layer_4c, kernel_4d_2_1, [1, 1, 1, 1], name='conv_layer_4d_2_1') + bias_4d_2_1,
                                          name='layer_4d_2_1')  # size = (14,14,144)
            with tf.name_scope('conv3x3'):
                layer_4d_2_2 = tf.nn.relu(
                    conv2d(layer_4d_2_1, kernel_4d_2_2, [1, 1, 1, 1], name='conv_layer_4d_2_2') + bias_4d_2_2,
                    name='layer_4d_2_2')  # size = (14,14,288)
        with tf.name_scope('Part3'):
            kernel_4d_3_1 = weight_variable([1, 1, 512, 32], name='kernel_4d_3_1')
            bias_4d_3_1 = bias_variable([32], name='bias_4d_3_1')
            kernel_4d_3_2 = weight_variable([5, 5, 32, 64], name='kernel_4d_3_2')
            bias_4d_3_2 = bias_variable([64], name='bias_4d_3_2')
            with tf.name_scope('conv1x1'):
                layer_4d_3_1 = tf.nn.relu(conv2d(layer_4c, kernel_4d_3_1, [1, 1, 1, 1], name='conv_layer_4d_3_1') + bias_4d_3_1,
                                          name='layer_4d_3_1')  # size = (14,14,32)
            with tf.name_scope('conv5x5'):
                layer_4d_3_2 = tf.nn.relu(
                    conv2d(layer_4d_3_1, kernel_4d_3_2, [1, 1, 1, 1], name='conv_layer_4d_3_2') + bias_4d_3_2,
                    name='layer_4d_3_2')  # size = (14,14,64)
        with tf.name_scope('Part4'):
            kernel_4d_4 = weight_variable([1, 1, 512, 64], name='kernel_4d_4')
            bias_4d_4 = bias_variable([64], name='bias_4d_4')
            with tf.name_scope('MaxPool_3x3'):
                max_pool_4d_1 = tf.nn.max_pool(layer_4c, ksize=[1, 3, 3, 1], strides=[1, 1, 1, 1], padding="SAME",
                                               name='maxpool_4d_1')  # size = (14,14,512)
            with tf.name_scope('conv1x1'):
                layer_4d_4 = tf.nn.relu(conv2d(max_pool_4d_1, kernel_4d_4, [1, 1, 1, 1], name='conv_layer_4d_4') + bias_4d_4,
                                        name='layer_4d_4')  # size = (14,14,64)
        with tf.name_scope('DepthConcat'):
            layer_4d = tf.concat([layer_4d_1, layer_4d_2_2, layer_4d_3_2, layer_4d_4], 3)  # size = (14,14,528)

    # Part4e
    with tf.name_scope('inception_4e'):
        with tf.name_scope('Part1'):
            kernel_4e_1 = weight_variable([1, 1, 528, 256], name='kernel_4e_1')
            bias_4e_1 = bias_variable([256], name='bias_4e_1')
            with tf.name_scope('conv1x1'):
                layer_4e_1 = tf.nn.relu(conv2d(layer_4d, kernel_4e_1, [1, 1, 1, 1], name='conv_layer_4e_1') + bias_4e_1,
                                        name='layer_4e_1')  # size = (14,14,256)
        with tf.name_scope('Part2'):
            kernel_4e_2_1 = weight_variable([1, 1, 528, 160], name='kernel_4e_2_1')
            bias_4e_2_1 = bias_variable([160], name='bias_4e_2_1')
            kernel_4e_2_2 = weight_variable([3, 3, 160, 320], name='kernel_4e_2_2')
            bias_4e_2_2 = bias_variable([320], name='bias_4e_2_2')
            with tf.name_scope('conv1x1'):
                layer_4e_2_1 = tf.nn.relu(conv2d(layer_4d, kernel_4e_2_1, [1, 1, 1, 1], name='conv_layer_4e_2_1') + bias_4e_2_1,
                                          name='layer_4e_2_1')  # size = (14,14,160)
            with tf.name_scope('conv3x3'):
                layer_4e_2_2 = tf.nn.relu(
                    conv2d(layer_4e_2_1, kernel_4e_2_2, [1, 1, 1, 1], name='conv_layer_4e_2_2') + bias_4e_2_2,
                    name='layer_4e_2_2')  # size = (14,14,320)
        with tf.name_scope('Part3'):
            kernel_4e_3_1 = weight_variable([1, 1, 528, 32], name='kernel_4e_3_1')
            bias_4e_3_1 = bias_variable([32], name='bias_4e_3_1')
            kernel_4e_3_2 = weight_variable([5, 5, 32, 128], name='kernel_4e_3_2')
            bias_4e_3_2 = bias_variable([128], name='bias_4e_3_2')
            with tf.name_scope('conv1x1'):
                layer_4e_3_1 = tf.nn.relu(conv2d(layer_4d, kernel_4e_3_1, [1, 1, 1, 1], name='conv_layer_4e_3_1') + bias_4e_3_1,
                                          name='layer_4e_3_1')  # size = (14,14,32)
            with tf.name_scope('conv5x5'):
                layer_4e_3_2 = tf.nn.relu(
                    conv2d(layer_4e_3_1, kernel_4e_3_2, [1, 1, 1, 1], name='conv_layer_4e_3_2') + bias_4e_3_2,
                    name='layer_4e_3_2')  # size = (14,14,128)
        with tf.name_scope('Part4'):
            kernel_4e_4 = weight_variable([1, 1, 528, 128], name='kernel_4e_4')
            bias_4e_4 = bias_variable([128], name='bias_4e_4')
            with tf.name_scope('MaxPool_3x3'):
                max_pool_4e_1 = tf.nn.max_pool(layer_4d, ksize=[1, 3, 3, 1], strides=[1, 1, 1, 1], padding="SAME",
                                               name='maxpool_4e_1')  # size = (14,14,528)
            with tf.name_scope('conv1x1'):
                layer_4e_4 = tf.nn.relu(conv2d(max_pool_4e_1, kernel_4e_4, [1, 1, 1, 1], name='conv_layer_4e_4') + bias_4e_4,
                                        name='layer_4e_4')  # size = (14,14,128)
        with tf.name_scope('DepthConcat'):
            layer_4e = tf.concat([layer_4e_1, layer_4e_2_2, layer_4e_3_2, layer_4e_4], 3)  # size = (14,14,832)

    # Part5
    with tf.name_scope('MaxPool_4'):
        MaxPool_4 = tf.nn.max_pool(layer_4e, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1], padding="SAME",
                                   name='MaxPool_5')  # size = (7,7,832)

    # Part5a
    with tf.name_scope('inception_5a'):
        with tf.name_scope('Part1'):
            kernel_5a_1 = weight_variable([1, 1, 832, 256], name='kernel_5a_1')
            bias_5a_1 = bias_variable([256], name='bias_5a_1')
            with tf.name_scope('conv1x1'):
                layer_5a_1 = tf.nn.relu(conv2d(MaxPool_4, kernel_5a_1, [1, 1, 1, 1], name='conv_layer_5a_1') + bias_5a_1,
                                        name='layer_5a_1')  # size = (7,7,256)
        with tf.name_scope('Part2'):
            kernel_5a_2_1 = weight_variable([1, 1, 832, 160], name='kernel_5a_2_1')
            bias_5a_2_1 = bias_variable([160], name='bias_5a_2_1')
            kernel_5a_2_2 = weight_variable([3, 3, 160, 320], name='kernel_5a_2_2')
            bias_5a_2_2 = bias_variable([320], name='bias_5a_2_2')
            with tf.name_scope('conv1x1'):
                layer_5a_2_1 = tf.nn.relu(
                    conv2d(MaxPool_4, kernel_5a_2_1, [1, 1, 1, 1], name='conv_layer_5a_2_1') + bias_5a_2_1,
                    name='layer_5a_2_1')  # size = (7,7,160)
            with tf.name_scope('conv3x3'):
                layer_5a_2_2 = tf.nn.relu(
                    conv2d(layer_5a_2_1, kernel_5a_2_2, [1, 1, 1, 1], name='conv_layer_5a_2_2') + bias_5a_2_2,
                    name='layer_5a_2_2')  # size = (7,7,320)
        with tf.name_scope('Part3'):
            kernel_5a_3_1 = weight_variable([1, 1, 832, 32], name='kernel_5a_3_1')
            bias_5a_3_1 = bias_variable([32], name='bias_5a_3_1')
            kernel_5a_3_2 = weight_variable([5, 5, 32, 128], name='kernel_5a_3_2')
            bias_5a_3_2 = bias_variable([128], name='bias_5a_3_2')
            with tf.name_scope('conv1x1'):
                layer_5a_3_1 = tf.nn.relu(
                    conv2d(MaxPool_4, kernel_5a_3_1, [1, 1, 1, 1], name='conv_layer_5a_3_1') + bias_5a_3_1,
                    name='layer_5a_3_1')  # size = (7,7,32)
            with tf.name_scope('conv5x5'):
                layer_5a_3_2 = tf.nn.relu(
                    conv2d(layer_5a_3_1, kernel_5a_3_2, [1, 1, 1, 1], name='conv_layer_5a_3_2') + bias_5a_3_2,
                    name='layer_5a_3_2')  # size = (7,7,128)
        with tf.name_scope('Part4'):
            kernel_5a_4 = weight_variable([1, 1, 832, 128], name='kernel_5a_4')
            bias_5a_4 = bias_variable([128], name='bias_5a_4')
            with tf.name_scope('MaxPool_3x3'):
                max_pool_5a_1 = tf.nn.max_pool(MaxPool_4, ksize=[1, 3, 3, 1], strides=[1, 1, 1, 1], padding="SAME",
                                               name='maxpool_5a_1')  # size = (7,7,832)
            with tf.name_scope('conv1x1'):
                layer_5a_4 = tf.nn.relu(conv2d(max_pool_5a_1, kernel_5a_4, [1, 1, 1, 1], name='conv_layer_5a_4') + bias_5a_4,
                                        name='layer_5a_4')  # size = (7,7,128)
        with tf.name_scope('DepthConcat'):
            layer_5a = tf.concat([layer_5a_1, layer_5a_2_2, layer_5a_3_2, layer_5a_4], 3)  # size = (7,7,832)

    # Part5b
    with tf.name_scope('inception_5b'):
        with tf.name_scope('Part1'):
            kernel_5b_1 = weight_variable([1, 1, 832, 384], name='kernel_5b_1')
            bias_5b_1 = bias_variable([384], name='bias_5b_1')
            with tf.name_scope('conv1x1'):
                layer_5b_1 = tf.nn.relu(conv2d(layer_5a, kernel_5b_1, [1, 1, 1, 1], name='conv_layer_5b_1') + bias_5b_1,
                                        name='layer_5b_1')  # size = (7,7,384)
        with tf.name_scope('Part2'):
            kernel_5b_2_1 = weight_variable([1, 1, 832, 192], name='kernel_5b_2_1')
            bias_5b_2_1 = bias_variable([192], name='bias_5b_2_1')
            kernel_5b_2_2 = weight_variable([3, 3, 192, 384], name='kernel_5b_2_2')
            bias_5b_2_2 = bias_variable([384], name='bias_5b_2_2')
            with tf.name_scope('conv1x1'):
                layer_5b_2_1 = tf.nn.relu(conv2d(layer_5a, kernel_5b_2_1, [1, 1, 1, 1], name='conv_layer_5b_2_1') + bias_5b_2_1,
                                          name='layer_5b_2_1')  # size = (7,7,192)
            with tf.name_scope('conv3x3'):
                layer_5b_2_2 = tf.nn.relu(
                    conv2d(layer_5b_2_1, kernel_5b_2_2, [1, 1, 1, 1], name='conv_layer_5b_2_2') + bias_5b_2_2,
                    name='layer_5b_2_2')  # size = (7,7,384)
        with tf.name_scope('Part3'):
            kernel_5b_3_1 = weight_variable([1, 1, 832, 48], name='kernel_5b_3_1')
            bias_5b_3_1 = bias_variable([48], name='bias_5b_3_1')
            kernel_5b_3_2 = weight_variable([5, 5, 48, 128], name='kernel_5b_3_2')
            bias_5b_3_2 = bias_variable([128], name='bias_5b_3_2')
            with tf.name_scope('conv1x1'):
                layer_5b_3_1 = tf.nn.relu(conv2d(layer_5a, kernel_5b_3_1, [1, 1, 1, 1], name='conv_layer_5b_3_1') + bias_5b_3_1,
                                          name='layer_5b_3_1')  # size = (7,7,48)
            with tf.name_scope('conv5x5'):
                layer_5b_3_2 = tf.nn.relu(
                    conv2d(layer_5b_3_1, kernel_5b_3_2, [1, 1, 1, 1], name='conv_layer_5b_3_2') + bias_5b_3_2,
                    name='layer_5b_3_2')  # size = (7,7,128)
        with tf.name_scope('Part4'):
            kernel_5b_4 = weight_variable([1, 1, 832, 128], name='kernel_5b_4')
            bias_5b_4 = bias_variable([128], name='bias_5b_4')
            with tf.name_scope('MaxPool_3x3'):
                max_pool_5b_1 = tf.nn.max_pool(layer_5a, ksize=[1, 3, 3, 1], strides=[1, 1, 1, 1], padding="SAME",
                                               name='maxpool_5b_1')  # size = (7,7,832)
            with tf.name_scope('conv1x1'):
                layer_5b_4 = tf.nn.relu(conv2d(max_pool_5b_1, kernel_5b_4, [1, 1, 1, 1], name='conv_layer_5b_4') + bias_5b_4,
                                        name='layer_5b_4')  # size = (7,7,128)
        with tf.name_scope('DepthConcat'):
            layer_5b = tf.concat([layer_5b_1, layer_5b_2_2, layer_5b_3_2, layer_5b_4], 3)  # size = (7,7,1024)

    # Softmax_assist_0
    with tf.name_scope('Auxiliary_classifier_1'):
        with tf.name_scope('AveragePool'):
            Avg_assist_0 = tf.nn.avg_pool(layer_4a, ksize=[1, 5, 5, 1], strides=[1, 3, 3, 1], padding="VALID",
                                          name='Avg')  # size = (4,4,512)
        with tf.name_scope('Conv'):
            kernel_assist_0 = weight_variable([1, 1, 512, 1024], name='kernel_assist_0')
            bias_assist_0 = bias_variable([1024], name='bias_assist_0')
            conv_layer_assist_0 = tf.nn.relu(conv2d(Avg_assist_0, kernel_assist_0, [1, 1, 1, 1]) + bias_assist_0,
                                             name='conv_layer_assist_0')  # size = (4,4,1024)
        with tf.name_scope('FC1'):
            line_assist_0 = tf.reshape(conv_layer_assist_0, [-1, 16384])
            fc_assist_0_weight_1 = weight_variable([16384, 1000], name='fc_assist_0_weight_1')
            fc_assist_0_bias_1 = bias_variable([1000], name='fc_assist_0_bias_1')
            layer_assist_0_dropout_1 = tf.nn.dropout(line_assist_0, keep_prob, name='layer_assist_0_dropout_1')
            layer_assist_0_1 = tf.nn.relu(
                tf.matmul(layer_assist_0_dropout_1, fc_assist_0_weight_1,
                          name='matmul_layer_assist_0_1') + fc_assist_0_bias_1,
                name='layer_assist_0_1')

        with tf.name_scope('FC2'):
            fc_assist_0_weight_2 = weight_variable([1000, 1000], name='fc_assist_0_weight_2')
            fc_assist_0_bias_2 = bias_variable([1000], name='fc_assist_0_bias_2')
            layer_assist_0_dropout_2 = tf.nn.dropout(layer_assist_0_1, keep_prob, name='layer_assist_0_dropout_2')
            layer_assist_0_2 = tf.nn.relu(
                tf.matmul(layer_assist_0_dropout_2, fc_assist_0_weight_2, name='matmul_layer_assist_0_2') + fc_assist_0_bias_2,
                name='layer_assist_0_2')
        with tf.name_scope('softmax0'):
            output_assist_0 = tf.nn.softmax(layer_assist_0_2, name='output_assist_0')

    # Softmax_assist_1
    with tf.name_scope('Auxiliary_classifier_2'):
        with tf.name_scope('AveragePool'):
            Avg_assist_1 = tf.nn.avg_pool(layer_4d, ksize=[1, 5, 5, 1], strides=[1, 3, 3, 1], padding="VALID",
                                          name='Avg')  # size = (4,4,528)
        with tf.name_scope('Conv'):
            kernel_assist_1 = weight_variable([1, 1, 528, 1024], name='kernel_assist_1')
            bias_assist_1 = bias_variable([1024], name='bias_assist_1')
            conv_layer_assist_1 = tf.nn.relu(conv2d(Avg_assist_1, kernel_assist_1, [1, 1, 1, 1]) + bias_assist_1,
                                             name='conv_layer_assist_1')  # size = (4,4,1024)
        with tf.name_scope('FC1'):
            line_assist_1 = tf.reshape(conv_layer_assist_1, [-1, 16384])
            fc_assist_1_weight_1 = weight_variable([16384, 1000], name='fc_assist_1_weight_1')
            fc_assist_1_bias_1 = bias_variable([1000], name='fc_assist_1_bias_1')
            layer_assist_1_dropout_1 = tf.nn.dropout(line_assist_1, keep_prob, name='layer_assist_1_dropout_1')
            layer_assist_1_1 = tf.nn.relu(
                tf.matmul(layer_assist_1_dropout_1, fc_assist_1_weight_1,
                          name='matmul_layer_assist_1_1') + fc_assist_1_bias_1,
                name='layer_assist_1_1')

        with tf.name_scope('FC2'):
            fc_assist_1_weight_2 = weight_variable([1000, 1000], name='fc_assist_1_weight_2')
            fc_assist_1_bias_2 = bias_variable([1000], name='fc_assist_1_bias_2')
            layer_assist_1_dropout_2 = tf.nn.dropout(layer_assist_1_1, keep_prob, name='layer_assist_1_dropout_2')
            layer_assist_1_2 = tf.nn.relu(
                tf.matmul(layer_assist_1_dropout_2, fc_assist_1_weight_2, name='matmul_layer_assist_1_2') + fc_assist_1_bias_2,
                name='layer_assist_1_2')
        with tf.name_scope('softmax1'):
            output_assist_1 = tf.nn.softmax(layer_assist_1_2, name='output_assist_1')

    # Softmax
    with tf.name_scope('Classifier'):
        with tf.name_scope('AveragePool'):
            Avg = tf.nn.avg_pool(layer_5b, ksize=[1, 7, 7, 1], strides=[1, 1, 1, 1], padding="VALID",
                                 name='Avg')  # size = (1,1,1024)
        with tf.name_scope('FC'):
            line = tf.reshape(Avg, [-1, 1024])
            fc_weight = weight_variable([1024, 1000], name='fc_weight')
            fc_bias = bias_variable([1000], name='fc_bias')
            layer_dropout = tf.nn.dropout(line, keep_prob, name='layer_dropout')
            layer = tf.nn.relu(tf.matmul(layer_dropout, fc_weight, name='matmul_layer') + fc_bias, name='layer')

    with tf.name_scope('Output'):
        output = tf.nn.softmax(layer, name='output')

    return output, output_assist_0, output_assist_1


def backward(datasets, label, test_data, test_label):
    with tf.name_scope('Input_data'):
        X = tf.placeholder(tf.float32, [None, 224, 224, 3], name="Input")
        Y_ = tf.placeholder(tf.float32, [None, 1], name='Estimation')
    LEARNING_RATE_BASE = 0.00001  # 最初学习率
    LEARNING_RATE_DECAY = 0.99  # 学习率的衰减率
    LEARNING_RATE_STEP = 1000  # 喂入多少轮BATCH-SIZE以后，更新一次学习率。一般为总样本数量/BATCH_SIZE
    gloabl_steps = tf.Variable(0, trainable=False)  # 计数器，用来记录运行了几轮的BATCH_SIZE，初始为0，设置为不可训练
    learning_rate = tf.train.exponential_decay(LEARNING_RATE_BASE, gloabl_steps, LEARNING_RATE_STEP,
                                               LEARNING_RATE_DECAY, staircase=True)
    keep_prob = tf.placeholder(tf.float32, name="keep_prob")
    y = VGG16(X)
    global_step = tf.Variable(0, trainable=False)
    with tf.name_scope('Accuracy'):
        correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(Y_, 1))
        accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float"))
    with tf.name_scope('loss'):
        loss_mse = tf.reduce_mean(-tf.reduce_sum(Y_ * tf.log(y), reduction_indices=[1]))
        tf.summary.scalar('loss', loss)
    with tf.name_scope('train'):
        train_step = tf.train.AdamOptimizer(0.001).minimize(loss_mse)
    saver = tf.train.Saver()
    with tf.Session() as sess:
        init_op = tf.global_variables_initializer()
        writer = tf.summary.FileWriter("./logs", sess.graph)
        sess.run(init_op)
        # 训练模型。
        STEPS = 500001
        min_loss = 1
        for i in range(STEPS):
            start = (i * BATCH_SIZE) % len(datasets)
            end = start + BATCH_SIZE
            if i % 100 == 0:
                run_options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)
                run_metadata = tf.RunMetadata()
                summary_str, step, _ = sess.run([merged, gloabl_steps, train_step],
                                                feed_dict={X: datasets[start:end], Y_: label[start:end], keep_prob:1.0},
                                                options=run_options, run_metadata=run_metadata)
                writer.add_summary(summary_str, i)
                writer.add_run_metadata(run_metadata, 'step%d' % (i))
                test_accuracy = accuracy.eval(feed_dict={X: test_data, Y_: test_label, keep_prob: 1.0})
                print("After %d training step(s), accuracy is %g" % (i, test_accuracy))
                saver.save(sess, './logs/variable', global_step=i)
            else:
                summary_str, step, _ = sess.run([merged, gloabl_steps, train_step],
                                                feed_dict={X: datasets[start:end], Y_: label[start:end]})
                writer.add_summary(summary_str, i)


def main():
    with tf.name_scope('Input_data'):
        X = tf.placeholder(tf.float32, [None, 224, 224, 3], name="Input")
    y = GoogLeNet_v1(X)
    sess = tf.Session()
    writer = tf.summary.FileWriter("./logs", sess.graph)
    writer.close()


if __name__ == '__main__':
    main()

然后我们用tensorboard打开图观察：

这里与真实的模型的结构相当 ，验证我们的demo编程的正确性。