代码和图片来源：https://blog.csdn.net/index20001/article/details/76623749
前言
2014年，GoogleNet模型在ILSVRC取得分类任务第一名，GoogleNet模型已经达到了22层的深度，而且在网络架构中引入了inception单元；这两个重要的改变说明，通过使用inception单元构造的深层次卷积神经网络模型，能够进一步提高模型整体的性能
在介绍inception单元之前，我们先来了解下NIN（network in network）中1*1卷积层的意义和作用，我们都知道卷积层在用来做数据特征提取的，而且最常用的卷积核滑动窗口的高度和宽度一般都是3*3或者5*5，那么在卷积核中使用高度和宽度为1*1的滑动窗又是什么意思呢？答案是能够完成特征图通道的聚合和发散
举个例子来说，假设我们有一个50*50*100的特征图，接下来我们将这个特征图输入1*1的卷积层中，定义卷积层使用的卷积核的滑动窗维度是1*1*100，如果我们想输出一个深度为90的特征图，那么我们只需进行九十次的卷积操作就完成了维度为50*50*100的特征图，这就是特征图通道的聚合；同理，我们输出一个50*50*110的特征图也就实现了特征图通道的发散
比如讲一个32*32*10的特征图输入3*3的卷积层当中，要求最后输出的深度为20，那么在这个过程中需要用到的卷积参数是1800个，但是我们如果先将这个特征图输入到1*1*10的卷积核当中，使其变成32*32*5的特征图，然后再将其输入到3*3的卷积层上，这样的话最终参数就是1*1*10*5 + 3*3*5*20，也就是950个参数
总而言之，在GoogleNet模型中加入了inception单元，使卷积神经网络模型和的搭建实现了模块化，如果我们想要增加或者减少GoogleNet模型的深度，则只需要增添或者减少相应的inception单元就可以了，另外为了避免出现深层次模型中的梯度消失的问题，在GoogleNet模型结构中还增加了两个额外的辅助softmax函数，用于前向传播梯度
# -*- coding:utf-8 -*-
#
# inception_v3 net
# default_image_size = 299
import tensorflow as tf
slim = tf.contrib.slim
def inception_v3_base(inputs, num_classes, scope=None):
end_points = {}
with tf.variable_scope(scope, 'inception_v3', [inputs]):
with scopes.arg_scope([slim.conv2d, slim.fc, slim.batch_norm, slim.dropout],
is_training=is_training):
# First part:5 conv layer
with slim.arg_scope([slim.conv2d, slim.max_pool2d, slim.avg_pool2d],
stride=1, padding='VALID'):
# 299 x 299 x 3
end_points['Conv2d_1a'] = slim.conv2d(inputs, 32, [3, 3], stride=2, scope='Conv2d_1a')
# 149 x 149 x 32
end_points['Conv2d_2a'] = slim.conv2d(end_points['Conv2d_1a'], 32, [3, 3], scope='Conv2d_2a')
# 147 x 147 x 32
end_points['Conv2d_2b'] = slim.conv2d(end_points['Conv2d_2a'], 64, [3, 3], padding='SAME', scope='Conv2d_2b')
# 147 x 147 x 64
end_points['MaxPool_3a'] = slim.max_pool2d(end_points['Conv2d_2b'], [3, 3], stride=2, scope='MaxPool_3a')
# 73 x 73 x 64
end_points['Conv2d_3b'] = slim.conv2d(end_points['MaxPool_3a'], 80, [1, 1], scope='Conv2d_3b')
# 73 x 73 x 80
end_points['Conv2d_4a'] = slim.conv2d(end_points['Conv2d_3b'], 192, [3, 3], scope='Conv2d_4a')
# 71 x 71 x 192
end_points['MaxPool_5a'] = slim.max_pool2d(end_points['Conv2d_4a'], [3, 3], stride=2, scope='MaxPool_5a')
# 35 x 35 x 192
net = end_points['MaxPool_5a']
# Inception blocks
with slim.arg_scope([slim.conv2d, slim.max_pool2d, slim.avg_pool2d],
stride=1, padding='SAME'):
# mixed_0: 35 x 35 x 256
with tf.variable_scope('Mixed_5b'):
with tf.variable_scope('Branch_0'):
branch_0 = slim.conv2d(net, 64, [1, 1], scope='Conv_1x1')
with tf.variable_scope('Branch_1'):
branch_1 = slim.conv2d(net, 48, [1, 1], scope='Conv_1x1')
branch_1 = slim.conv2d(branch_1, 64, [5, 5], scope='Conv_5x5')
with tf.variable_scope('Branch_2'):
branch_2 = slim.conv2d(net, 64, [1, 1], scope='Conv_1x1')
branch_2 = slim.conv2d(branch_2, 96, [3, 3], scope='Conv_3x3_0')
branch_2 = slim.conv2d(branch_2, 96, [3, 3], scope='Conv_3x3_1')
with tf.variable_scope('Branch_3'):
branch_3 = slim.avg_pool2d(net, [3, 3], scope='AvgPool_3x3')
branch_3 = slim.conv2d(branch_3, 32, [1, 1], scope='Conv_1x1')
# 256 = 64 + 64 + 96 + 32
net = tf.concat(axis=3, values=[branch_0, branch_1, branch_2, branch_3])
end_points['Mixed_5b'] = net
# mixed_1: 35 x 35 x 288
with tf.variable_scope('Mixed_5c'):
with tf.variable_scope('Branch_0'):
branch_0 = slim.conv2d(net, 64, [1, 1], scope='Conv_1x1')
with tf.variable_scope('Branch_1'):
branch_1 = slim.conv2d(net, 48, [1, 1], scope='Conv_1x1')
branch_1 = slim.conv2d(branch_1, 64, [5, 5], scope='Conv_5x5')
with tf.variable_scope('Branch_2'):
branch_2 = slim.conv2d(net, 64, [1, 1], scope='Conv_1x1')
branch_2 = slim.conv2d(branch_2, 96, [3, 3], scope='Conv_3x3_0')
branch_2 = slim.conv2d(branch_2, 96, [3, 3], scope='Conv_3x3_1')
with tf.variable_scope('Branch_3'):
branch_3 = slim.avg_pool2d(net, [3, 3], scope='AvgPool_3x3')
branch_3 = slim.conv2d(branch_3, 64, [1, 1], scope='Conv_1x1')
# 288 = 64 + 64 + 96 + 64
net = tf.concat(axis=3, values=[branch_0, branch_1, branch_2, branch_3])
end_points['Mixed_5c'] = net
# mixed_2: 35 x 35 x 288
with tf.variable_scope('Mixed_5d'):
with tf.variable_scope('Branch_0'):
branch_0 = slim.conv2d(net, 64, [1, 1], scope='Conv_1x1')
with tf.variable_scope('Branch_1'):
branch_1 = slim.conv2d(net, 48, [1, 1], scope='Conv_1x1')
branch_1 = slim.conv2d(branch_1, 64, [5, 5], scope='Conv_5x5')
with tf.variable_scope('Branch_2'):
branch_2 = slim.conv2d(net, 64, [1, 1], scope='Conv_1x1')
branch_2 = slim.conv2d(branch_2, 96, [3, 3], scope='Conv_3x3_0')
branch_2 = slim.conv2d(branch_2, 96, [3, 3], scope='Conv_3x3_1')
with tf.variable_scope('Branch_3'):
branch_3 = slim.avg_pool2d(net, [3, 3], scope='AvgPool_3x3')
branch_3 = slim.conv2d(branch_3, 64, [1, 1], scope='Conv_1x1')
net = tf.concat(axis=3, values=[branch_0, branch_1, branch_2, branch_3])
end_points['Mixed_5d'] = net
# mixed_3: 17 x 17 x 768
with tf.variable_scope('Mixed_6a'):
with tf.variable_scope('Branch_0'):
branch_0 = slim.conv2d(net, 384, [3, 3], stride=2, padding='VALID', scope='Conv_1x1')
with tf.variable_scope('Branch_1'):
branch_1 = slim.conv2d(net, 64, [1, 1], scope='Conv_1x1')
branch_1 = slim.conv2d(branch_1, 96, [3, 3], scope='Conv_3x3')
branch_1 = slim.conv2d(branch_1, 96, [3, 3], stride=2, padding='VALID', scope='Conv_1x1')
with tf.variable_scope('Branch_2'):
branch_2 = slim.max_pool2d(net, [3, 3], stride=2, padding='VALID', scope='MaxPool_3x3')
# 768 = 384 + 96 + 288
net = tf.concat(axis=3, values=[branch_0, branch_1, branch_2])
end_points['Mixed_6a'] = net
# mixed_4: 17 x 17 x 768
with tf.variable_scope('Mixed_6b'):
with tf.variable_scope('Branch_0'):
branch_0 = slim.conv2d(net, 192, [1, 1], scope='Conv_1x1')
with tf.variable_scope('Branch_1'):
branch_1 = slim.conv2d(net, 128, [1, 1], scope='Conv_1x1')
branch_1 = slim.conv2d(branch_1, 128, [1, 7], scope='Conv_1x7')
branch_1 = slim.conv2d(branch_1, 192, [7, 1], scope='Conv_7x1')
with tf.variable_scope('Branch_2'):
branch_2 = slim.conv2d(net, 128, [1, 1], scope='Conv_1x1_a')
branch_2 = slim.conv2d(branch_2, 128, [7, 1], scope='Conv_7x1_b')
branch_2 = slim.conv2d(branch_2, 128, [1, 7], scope='Conv_1x7_c')
branch_2 = slim.conv2d(branch_2, 128, [7, 1], scope='Conv_7x1_d')
branch_2 = slim.conv2d(branch_2, 192, [1, 7], scope='Conv_1x7_e')
with tf.variable_scope('Branch_3'):
branch_3 = slim.avg_pool2d(net, [3, 3], scope='AvgPool_3x3')
branch_3 = slim.conv2d(branch_3, 192, [1, 1], scope='Conv_1x1')
# 768 = 192 + 192 + 192 + 192
net = tf.concat(axis=3, values=[branch_0, branch_1, branch_2, branch_3])
end_points['Mixed_6b'] = net
# mixed_5: 17 x 17 x 768
with tf.variable_scope('Mixed_6c'):
with tf.variable_scope('Branch_0'):
branch_0 = slim.conv2d(net, 192, [1, 1], scope='Conv_1x1')
with tf.variable_scope('Branch_1'):
branch_1 = slim.conv2d(net, 160, [1, 1], scope='Conv_1x1')
branch_1 = slim.conv2d(branch_1, 160, [1, 7], scope='Conv_1x7')
branch_1 = slim.conv2d(branch_1, 192, [7, 1], scope='Conv_7x1')
with tf.variable_scope('Branch_2'):
branch_2 = slim.conv2d(net, 160, [1, 1], scope='Conv_1x1_a')
branch_2 = slim.conv2d(branch_2, 160, [7, 1], scope='Conv_7x1_b')
branch_2 = slim.conv2d(branch_2, 160, [1, 7], scope='Conv_1x7_c')
branch_2 = slim.conv2d(branch_2, 160, [7, 1], scope='Conv_7x1_d')
branch_2 = slim.conv2d(branch_2, 192, [1, 7], scope='Conv_1x7_e')
with tf.variable_scope('Branch_3'):
branch_3 = slim.avg_pool2d(net, [3, 3], scope='AvgPool_3x3')
branch_3 = slim.conv2d(branch_3, 192, [1, 1], scope='Conv_1x1')
# 768 = 192 + 192 + 192 + 192
net = tf.concat(axis=3, values=[branch_0, branch_1, branch_2, branch_3])
end_points['Mixed_6c'] = net
# mixed_6: 17 x 17 x 768
with tf.variable_scope('Mixed_6d'):
with tf.variable_scope('Branch_0'):
branch_0 = slim.conv2d(net, 192, [1, 1], scope='Conv_1x1')
with tf.variable_scope('Branch_1'):
branch_1 = slim.conv2d(net, 160, [1, 1], scope='Conv_1x1')
branch_1 = slim.conv2d(branch_1, 160, [1, 7], scope='Conv_1x7')
branch_1 = slim.conv2d(branch_1, 192, [7, 1], scope='Conv_7x1')
with tf.variable_scope('Branch_2'):
branch_2 = slim.conv2d(net, 160, [1, 1], scope='Conv_1x1_a')
branch_2 = slim.conv2d(branch_2, 160, [7, 1], scope='Conv_7x1_b')
branch_2 = slim.conv2d(branch_2, 160, [1, 7], scope='Conv_1x7_c')
branch_2 = slim.conv2d(branch_2, 160, [7, 1], scope='Conv_7x1_d')
branch_2 = slim.conv2d(branch_2, 192, [1, 7], scope='Conv_1x7_e')
with tf.variable_scope('Branch_3'):
branch_3 = slim.avg_pool2d(net, [3, 3], scope='AvgPool_3x3')
branch_3 = slim.conv2d(branch_3, 192, [1, 1], scope='Conv_1x1')
# 768 = 192 + 192 + 192 + 192
net = tf.concat(axis=3, values=[branch_0, branch_1, branch_2, branch_3])
end_points['Mixed_6d'] = net
# mixed_7: 17 x 17 x 768
with tf.variable_scope('Mixed_6e'):
with tf.variable_scope('Branch_0'):
branch_0 = slim.conv2d(net, 192, [1, 1], scope='Conv_1x1')
with tf.variable_scope('Branch_1'):
branch_1 = slim.conv2d(net, 192, [1, 1], scope='Conv_1x1')
branch_1 = slim.conv2d(branch_1, 192, [1, 7], scope='Conv_1x7')
branch_1 = slim.conv2d(branch_1, 192, [7, 1], scope='Conv_7x1')
with tf.variable_scope('Branch_2'):
branch_2 = slim.conv2d(net, 192, [1, 1], scope='Conv_1x1_a')
branch_2 = slim.conv2d(branch_2, 192, [7, 1], scope='Conv_7x1_b')
branch_2 = slim.conv2d(branch_2, 192, [1, 7], scope='Conv_1x7_c')
branch_2 = slim.conv2d(branch_2, 192, [7, 1], scope='Conv_7x1_d')
branch_2 = slim.conv2d(branch_2, 192, [1, 7], scope='Conv_1x7_e')
with tf.variable_scope('Branch_3'):
branch_3 = slim.avg_pool2d(net, [3, 3], scope='AvgPool_3x3')
branch_3 = slim.conv2d(branch_3, 192, [1, 1], scope='Conv_1x1')
# 768 = 192 + 192 + 192 + 192
net = tf.concat(axis=3, values=[branch_0, branch_1, branch_2, branch_3])
end_points['Mixed_6e'] = net
# Auxiliary Head logits
aux_logits = tf.identity(end_points['Mixed_6e'])
with tf.variable_scope('AuxLogits'):
aux_logits = slim.avg_pool2d(aux_logits, [5, 5], stride=3, padding='VALID', scope='AvgPool_5x5')
aux_logits = slim.conv2d(aux_logits, 128, [1, 1], scope='Conv_1x1')
# Shape of feature map before the final layer.
shape = aux_logits.get_shape().as_list()
aux_logits = slim.conv2d(aux_logits, 768, shape[1:3], weights_initializer=trunc_normal(0.01),
padding='VALID', scope='Conv2d_2a_{}x{}'.format(shape[1],shape[2]))
aux_logits = slim.conv2d(aux_logits, num_classes, [1, 1], activation_fn=None,
normalizer_fn=None, weights_initializer=trunc_normal(0.001), scope='Conv_1x1')
aux_logits = tf.squeeze(aux_logits, [1, 2], name='SpatialSqueeze')
end_points['AuxLogits'] = aux_logits
# mixed_8: 8 x 8 x 1280
with tf.variable_scope('Mixed_7a'):
with tf.variable_scope('Branch_0'):
branch_0 = slim.conv2d(net, 192, [1, 1], scope='Conv_1x1')
branch_0 = slim.conv2d(branch_0, 320, [3, 3], stride=2, padding='VALID', scope='Conv_3x3')
with tf.variable_scope('Branch_1'):
branch_1 = slim.conv2d(net, 192, [1, 1], scope='Conv_1x1')
branch_1 = slim.conv2d(branch_1, 192, [1, 7], scope='Conv_1x7')
branch_1 = slim.conv2d(branch_1, 192, [7, 1], scope='Conv_7x1')
branch_1 = slim.conv2d(branch_1, 192, [3, 3], stride=2, padding='VALID', scope='Conv_3x3')
with tf.variable_scope('Branch_2'):
branch_2 = slim.max_pool2d(net, [3, 3], stride=2, padding='VALID', scope='MaxPool_3x3')
# 1280 = 320 + 192 + 768
net = tf.concat(axis=3, values=[branch_0, branch_1, branch_2])
end_points['Mixed_7a'] = net
# mixed_9: 8 x 8 x 2048
with tf.variable_scope('Mixed_7b'):
with tf.variable_scope('Branch_0'):
branch_0 = slim.conv2d(net, 320, [1, 1], scope='Conv_1x1')
with tf.variable_scope('Branch_1'):
branch_1 = slim.conv2d(net, 384, [1, 1], scope='Conv_1x1')
branch_1 = tf.concat(axis=3, values=[
slim.conv2d(branch_1, 384, [1, 3], scope='Conv_1x3'),
slim.conv2d(branch_1, 384, [3, 1], scope='Conv_3x1')])
with tf.variable_scope('Branch_2'):
branch_2 = slim.conv2d(net, 448, [1, 1], scope='Conv_1x1')
branch_2 = slim.conv2d(branch_2, 384, [3, 3], scope='Conv_3x3')
branch_2 = tf.concat(axis=3, values=[
slim.conv2d(branch_2, 384, [1, 3], scope='Conv_1x3'),
slim.conv2d(branch_2, 384, [3, 1], scope='Conv_3x1')])
with tf.variable_scope('Branch_3'):
branch_3 = slim.avg_pool2d(net, [3, 3], scope='AvgPool_3x3')
branch_3 = slim.conv2d(branch_3, 192, [1, 1], scope='Conv_1x1')
# 2048 = 320 + 384*2 + 384*2 + 192
net = tf.concat(axis=3, values=[branch_0, branch_1, branch_2, branch_3])
end_points['Mixed_7b'] = net
# mixed_10: 8 x 8 x 2048
with tf.variable_scope('Mixed_7c'):
with tf.variable_scope('Branch_0'):
branch_0 = slim.conv2d(net, 320, [1, 1], scope='Conv_1x1')
with tf.variable_scope('Branch_1'):
branch_1 = slim.conv2d(net, 384, [1, 1], scope='Conv_1x1')
branch_1 = tf.concat(axis=3, values=[
slim.conv2d(branch_1, 384, [1, 3], scope='Conv_1x3'),
slim.conv2d(branch_1, 384, [3, 1], scope='Conv_3x1')])
with tf.variable_scope('Branch_2'):
branch_2 = slim.conv2d(net, 448, [1, 1], scope='Conv_1x1')
branch_2 = slim.conv2d(branch_2, 384, [3, 3], scope='Conv_3x3')
branch_2 = tf.concat(axis=3, values=[
slim.conv2d(branch_2, 384, [1, 3], scope='Conv_1x3'),
slim.conv2d(branch_2, 384, [3, 1], scope='Conv_3x1')])
with tf.variable_scope('Branch_3'):
branch_3 = slim.avg_pool2d(net, [3, 3], scope='AvgPool_3x3')
branch_3 = slim.conv2d(branch_3, 192, [1, 1], scope='Conv_1x1')
# 2048 = 320 + 384*2 + 384*2 + 192
net = tf.concat(axis=3, values=[branch_0, branch_1, branch_2, branch_3])
end_points['Mixed_7c'] = net
# Final pooling and prediction
with tf.variable_scope('Logits'):
shape = net.get_shape().as_list()
net = slim.avg_pool2d(net, shape[1:3], padding='VALID', scope='AvgPool_{}x{}'.format(shape[1],shape[2]))
# 1 x 1 x 2048
net = slim.dropout(net, keep_prob=dropout_keep_prob, scope='Dropout')
end_points['PreLogits'] = net
# 2048
logits = slim.conv2d(net, num_classes, [1, 1], activation_fn=None,
normalizer_fn=None, scope='Conv_1x1')
logits = tf.squeeze(logits, [1, 2], name='SpatialSqueeze')
# 1000
end_points['Logits'] = logits
end_points['Predictions'] = prediction_fn(logits, scope='Predictions')
return logits, end_points
def inception_v3(inputs, num_classes=1000, is_training=True, dropout_keep_prob=0.8,
prediction_fn=slim.softmax, reuse=None, scope='inception_v3'):
with tf.variable_scope(scope, 'inception_v3', [inputs, num_classes], reuse=reuse) as scope:
with slim.arg_scope([slim.batch_norm, slim.dropout], is_training=is_training):
net, end_points = inception_v3_base(inputs, scope=scope)
return logits, end_points
809
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
