02、VGG16预训练好的参数加载与模型网络构建_笔记 - 东方耀AI技术分享

# -*- coding: utf-8 -*-
__author__ = 'dongfangyao'
__date__ = '2019/3/12 下午5:33'
__product__ = 'PyCharm'
__filename__ = '2_image_style_con'
import tensorflow as tf
from tensorflow import logging
import os
import time
import numpy as np
logging.set_verbosity(logging.INFO)
# logging.info('dfy_88888')
# vgg net 中写死的归一化的数据预处理
VGG_MEAN = [103.939, 116.779, 123.68]
class VGGNet:
"""
构建VGG16的网络结构并从预训练好的模型提取参数加载
"""
def __init__(self, data_dict):
self.data_dict = data_dict
def get_conv_kernel(self, name):
# 卷积核的参数：w 0 b 1
return tf.constant(self.data_dict[name][0], name='conv')
def get_fc_weight(self, name):
return tf.constant(self.data_dict[name][0], name='fc')
def get_bias(self, name):
return tf.constant(self.data_dict[name][1], name='bias')
def conv_layer(self, inputs, name):
"""
构建一个卷积计算层
:param inputs: 输入的feature_map
:param name: 卷积层的名字也是获得参数的key 不能出错
:return:
"""
with tf.name_scope(name):
"""
多使用name_scope的好处：1、防止参数命名冲突 2、tensorboard可视化时很规整
如果scope里面有变量需要训练时则用tf.variable_scope
"""
conv_w = self.get_conv_kernel(name)
conv_b = self.get_bias(name)
# tf.layers.conv2d() 这是一个封装更高级的api
# 里面并没有提供接口来输入卷积核参数这里不能用平时训练cnn网络时非常好用
result = tf.nn.conv2d(input=inputs, filter=conv_w, strides=[1, 1, 1, 1], padding='SAME', name=name)
result = tf.nn.bias_add(result, conv_b)
result = tf.nn.relu(result)
return result
def pooling_layer(self, inputs, name):
# tf.layers.max_pooling2d()
# tf.nn.max_pool 这里的池化层没有参数两套api都可以用
return tf.nn.max_pool(inputs, [1, 2, 2, 1], [1, 2, 2, 1], padding='SAME', name=name)
def fc_layer(self, inputs, name, activation=tf.nn.relu):
"""
构建全连接层
:param inputs: 输入
:param name:
:param activation: 是否有激活函数的封装
:return:
"""
with tf.name_scope(name):
fc_w = self.get_fc_weight(name)
fc_b = self.get_bias(name)
# fc: wx+b 线性变换
result = tf.nn.bias_add(tf.matmul(inputs, fc_w), fc_b)
if activation is None:
# vgg16的最后是不需relu激活的
return result
else:
return activation(result)
def flatten_op(self, inputs, name):
# 展平操作为了后续的fc层必须将维度展平
with tf.name_scope(name):
# [NHWC]---> [N, H*W*C]
x_shape = inputs.get_shape().as_list()
dim = 1
for d in x_shape[1:]:
dim *= d
inputs = tf.reshape(inputs, shape=[-1, dim])
# 直接用现成api也是可以的
# return tf.layers.flatten(inputs)
return inputs
def build(self, input_rgb):
"""
构建vgg16网络结构抽取特征 FP过程
:param input_rgb: [1, 224, 224, 3]
:return:
"""
start_time = time.time()
logging.info('building start...')
# 在通道维度上分离深度可分离卷积中也需要用到这个api
r, g, b = tf.split(input_rgb, num_or_size_splits=3, axis=3)
# 在通道维度上拼接
# 输入vgg网络的图像是bgr的（与OpenCV一样倒序的）而不是rgb
x_bgr = tf.concat(values=[
b - VGG_MEAN[0],
g - VGG_MEAN[1],
r - VGG_MEAN[2],
], axis=3)
assert x_bgr.get_shape().as_list()[1:] == [224, 224, 3]
# 构建网络
# stage 1
self.conv1_1 = self.conv_layer(x_bgr, 'conv1_1')
self.conv1_2 = self.conv_layer(self.conv1_1, 'conv1_2')
self.pool1 = self.pooling_layer(self.conv1_2, 'pool1')
# stage 2
self.conv2_1 = self.conv_layer(self.pool1, 'conv2_1')
self.conv2_2 = self.conv_layer(self.conv2_1, 'conv2_2')
self.pool2 = self.pooling_layer(self.conv2_2, 'pool2')
# stage 3
self.conv3_1 = self.conv_layer(self.pool2, 'conv3_1')
self.conv3_2 = self.conv_layer(self.conv3_1, 'conv3_2')
self.conv3_3 = self.conv_layer(self.conv3_2, 'conv3_3')
self.pool3 = self.pooling_layer(self.conv3_3, 'pool3')
# stage 4
self.conv4_1 = self.conv_layer(self.pool3, 'conv4_1')
self.conv4_2 = self.conv_layer(self.conv4_1, 'conv4_2')
self.conv4_3 = self.conv_layer(self.conv4_2, 'conv4_3')
self.pool4 = self.pooling_layer(self.conv4_3, 'pool4')
# stage 5
self.conv5_1 = self.conv_layer(self.pool4, 'conv5_1')
self.conv5_2 = self.conv_layer(self.conv5_1, 'conv5_2')
self.conv5_3 = self.conv_layer(self.conv5_2, 'conv5_3')
self.pool5 = self.pooling_layer(self.conv5_3, 'pool5')
# flatten_op
self.flatten = self.flatten_op(self.pool5, 'flatten_op')
# fc
self.fc6 = self.fc_layer(self.flatten, 'fc6')
self.fc7 = self.fc_layer(self.fc6, 'fc7')
self.fc8 = self.fc_layer(self.fc7, 'fc8', activation=None)
self.logits = tf.nn.softmax(self.fc8, name='logits')
logging.info('building end... 耗时%3d秒' % (time.time() - start_time))
vgg16_npy_path = './vgg16.npy'
vgg16_data = np.load(vgg16_npy_path, encoding='latin1')
print(type(vgg16_data))
data_dict = vgg16_data.item()
vgg16_for_result = VGGNet(data_dict)
image_rgb = tf.placeholder(dtype=tf.float32, shape=[1, 224, 224, 3], name='image_rgb')
vgg16_for_result.build(image_rgb)
print(vgg16_for_result.conv1_1)
print(vgg16_for_result.fc6)

复制代码