预训练好的卷积神经网络CNN模型：vgg16.npy vgg19.npy 下载

东方耀 · 发表于 2019-2-3 22:43:15

预训练好的卷积神经网络CNN模型：vgg16.npy vgg19.npy 下载

文件比较大保存到了百度网盘方便大家转存 16层与19层的VGG网络

下载地址【回复本帖可见】：

游客，如果您要查看本帖隐藏内容请回复

# _*_ coding: utf-8 _*_
import numpy as np
vgg16_data = np.load('vgg16.npy', encoding='bytes')
print(type(vgg16_data))
# print(vgg16_data)
data_dict = vgg16_data.item()
print(type(data_dict))
print(data_dict.keys())
print(len(data_dict))
conv1_1 = data_dict[b'conv1_1']
print(len(conv1_1))
w, b = conv1_1
print(w.shape)
print(b.shape)
fc6 = data_dict[b'fc6']
print(len(fc6))
w, b = fc6
print(w.shape)
print(b.shape)

复制代码

# _*_ coding: utf-8 _*_
import os
import math
import numpy as np
import tensorflow as tf
import time
from PIL import Image
VGG_MEAN = [103.939, 116.779, 123.68]
class VGGNet:
"""
构建VGG16的网络结构并从预处理模型中加载训练好的参数
"""
def __init__(self, data_dict):
self.data_dict = data_dict
def get_conv_kernel(self, name):
# 卷积核的参数 w:0 bias：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"""
# 多使用name_scope的好处：1、防止参数命名冲突 2、可视化的显示规整
with tf.name_scope(name):
conv_w = self.get_conv_kernel(name)
conv_b = self.get_bias(name)
# tf.layers.conv2d() 里面没有参数的不能用了
result = tf.nn.conv2d(inputs, conv_w, [1, 1, 1, 1], padding='SAME')
result = tf.nn.bias_add(result, conv_b)
result = tf.nn.relu(result)
return result
def pooling_layer(self, inputs, name):
"""构建一个池化层 tf.layers.max_pooling2d()"""
result = tf.nn.max_pool(inputs, [1, 2, 2, 1], [1, 2, 2, 1], padding='SAME', name=name)
return result
def fc_layer(self, inputs, name, activation=tf.nn.relu):
"""构建全连接层的计算"""
with tf.name_scope(name):
fc_w = self.get_fc_weight(name)
fc_b = self.get_bias(name)
result = tf.nn.bias_add(tf.matmul(inputs, fc_w), fc_b)
if activation is None:
return result
else:
return activation(result)
def flatten_op(self, inputs, name):
"""展平操作 tf.layers.flatten()"""
with tf.name_scope(name):
# [N H W C]---> [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])
return inputs
def build(self, input_rgb):
"""构建vgg16网络提取特征 FP过程
参数：
--input_rgb: [1, 224, 224, 3]
"""
start_time = time.time()
print('building start...')
# r, g, b = tf.split(value=input_rgb, num_or_size_splits=[1, 1, 1], axis=3)
r, g, b = tf.split(value=input_rgb, num_or_size_splits=3, axis=3)
# 输入vgg的图像是bgr的顺序（跟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]
# 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')
# stage 6
self.flatten5 = self.flatten_op(self.pool5, 'flatten5')
self.fc6 = self.fc_layer(self.flatten5, 'fc6')
self.fc7 = self.fc_layer(self.fc6, 'fc7')
self.fc8 = self.fc_layer(self.fc7, 'fc8', activation=None)
self.prob = tf.nn.softmax(self.fc8, name='prob')
print(self.prob.shape)
print('buliding finished 耗时：%4ds' % (time.time() - start_time))
pass
vgg16_npy_path = './vgg16.npy'
vgg16_data = np.load(vgg16_npy_path, encoding='latin1')
data_dict = vgg16_data.item()
print(data_dict.keys())
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)

复制代码

# -*- coding: utf-8 -*-
__author__ = 'dongfangyao'
__date__ = '2019/2/4 下午3:33'
__product__ = 'PyCharm'
__filename__ = '10_image_style_conversion'
"""
TensorFlow实现图像风格转换 V1算法
"""
import os
import numpy as np
import tensorflow as tf
import time
from PIL import Image
import matplotlib.pyplot as plt
VGG_MEAN = [103.939, 116.779, 123.68]
class VGGNet:
"""
构建VGG16的网络结构并从预处理模型中加载训练好的参数
"""
def __init__(self, data_dict):
self.data_dict = data_dict
def get_conv_kernel(self, name):
# 卷积核的参数 w:0 bias：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"""
# 多使用name_scope的好处：1、防止参数命名冲突 2、可视化的显示规整
with tf.name_scope(name):
conv_w = self.get_conv_kernel(name)
conv_b = self.get_bias(name)
# tf.layers.conv2d() 里面没有参数的不能用了
result = tf.nn.conv2d(inputs, conv_w, [1, 1, 1, 1], padding='SAME')
result = tf.nn.bias_add(result, conv_b)
result = tf.nn.relu(result)
return result
def pooling_layer(self, inputs, name):
"""构建一个池化层 tf.layers.max_pooling2d()"""
result = tf.nn.max_pool(inputs, [1, 2, 2, 1], [1, 2, 2, 1], padding='SAME', name=name)
return result
def fc_layer(self, inputs, name, activation=tf.nn.relu):
"""构建全连接层的计算"""
with tf.name_scope(name):
fc_w = self.get_fc_weight(name)
fc_b = self.get_bias(name)
result = tf.nn.bias_add(tf.matmul(inputs, fc_w), fc_b)
if activation is None:
return result
else:
return activation(result)
def flatten_op(self, inputs, name):
"""展平操作 tf.layers.flatten()"""
with tf.name_scope(name):
# [N H W C]---> [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])
return inputs
def build(self, input_rgb):
"""构建vgg16网络提取特征 FP过程
参数：
--input_rgb: [1, 224, 224, 3]
"""
start_time = time.time()
print('building start...')
# r, g, b = tf.split(value=input_rgb, num_or_size_splits=[1, 1, 1], axis=3)
r, g, b = tf.split(value=input_rgb, num_or_size_splits=3, axis=3)
# 输入vgg的图像是bgr的顺序（跟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]
# 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')
# stage 6
# self.flatten5 = self.flatten_op(self.pool5, 'flatten5')
# self.fc6 = self.fc_layer(self.flatten5, 'fc6')
# self.fc7 = self.fc_layer(self.fc6, 'fc7')
# self.fc8 = self.fc_layer(self.fc7, 'fc8', activation=None)
# self.prob = tf.nn.softmax(self.fc8, name='prob')
# print(self.prob.shape)
print('buliding finished 耗时：%4ds' % (time.time() - start_time))
pass
vgg16_npy_path = './vgg16.npy'
# vgg16_data = np.load(vgg16_npy_path, encoding='latin1')
# data_dict = vgg16_data.item()
#
# print(data_dict.keys())
#
# 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)
# 224 * 224
content_img_path = './img/content.jpeg'
style_img_path = './img/style.jpeg'
num_steps = 100
learning_rate = 10
lambda_c = 0.1
lambda_s = 500
output_dir = './img/output_img'
if not os.path.exists(output_dir):
os.mkdir(output_dir)
def initial_image(shape, mean, stddev):
initial_image = tf.truncated_normal(shape=shape, mean=mean, stddev=stddev, dtype=tf.float32)
return tf.Variable(initial_image)
initial_image_result = initial_image([1, 224, 224, 3], mean=127.5, stddev=20)
def read_img(image_name):
img = Image.open(image_name)
# [224 224 3]
np_img = np.array(img)
# np_img = tf.reshape(np_img, shape=[1, 224, 224, 3])
# [1 224 224 3]
# np_img = np.reshape(np_img, newshape=[1, 224, 224, 3])
np_img = np.asarray([np_img], dtype=np.float32)
print(np_img.shape)
return np_img
# plt.imshow(read_img(style_img_path)[0])
# plt.show()
content_img_arr_val = read_img(content_img_path)
style_img_arr_val = read_img(style_img_path)
content_img = tf.placeholder(dtype=tf.float32, shape=[1, 224, 224, 3], name='content_img')
style_img = tf.placeholder(dtype=tf.float32, shape=[1, 224, 224, 3], name='style_img')
# initial_image_result content_img style_img 三张图片进入vgg网络提取特征
vgg16_data = np.load(vgg16_npy_path, encoding='latin1')
data_dict = vgg16_data.item()
print(data_dict.keys())
vgg16_for_initial_result = VGGNet(data_dict)
vgg16_for_content_img = VGGNet(data_dict)
vgg16_for_style_img = VGGNet(data_dict)
vgg16_for_content_img.build(content_img)
vgg16_for_style_img.build(style_img)
vgg16_for_initial_result.build(initial_image_result)
# 定义提取哪些层的特征 cnn的
# 内容特征越低层越精细
content_features = [
vgg16_for_content_img.conv1_2,
vgg16_for_content_img.conv2_2,
# vgg16_for_content_img.conv3_3,
# vgg16_for_content_img.conv4_3,
# vgg16_for_content_img.conv5_3
]
# 结果的内容特征必须与内容特征一致
result_content_features = [
vgg16_for_initial_result.conv1_2,
vgg16_for_initial_result.conv2_2,
# vgg16_for_initial_result.conv3_3,
# vgg16_for_initial_result.conv4_3,
# vgg16_for_initial_result.conv5_3
]
# 风格特征越高层越抽象
style_features = [
# vgg16_for_style_img.conv1_2,
# vgg16_for_style_img.conv2_2,
# vgg16_for_style_img.conv3_3,
vgg16_for_style_img.conv4_3,
vgg16_for_style_img.conv5_3
]
# 结果的风格特征必须与风格特征一致
result_style_features = [
# vgg16_for_initial_result.conv1_2,
# vgg16_for_initial_result.conv2_2,
# vgg16_for_initial_result.conv3_3,
vgg16_for_initial_result.conv4_3,
vgg16_for_initial_result.conv5_3
]
# 开始计算损失
content_loss = tf.zeros(shape=1, dtype=tf.float32)
# zip([1, 2], [3, 4]) ---> [(1, 3), (2, 4)] 两个数组变成一个数组
# c与c_的shape（卷积激励之后）：[1 height width channel] 在通道axis=[1, 2, 3]求平均
# loss = mse 平方差损失函数
for c, c_ in zip(content_features, result_content_features):
content_loss += tf.reduce_mean(tf.square(c - c_), axis=[1, 2, 3])
pass
# 计算Gram矩阵
def gram_matrix(x):
"""Gram矩阵计算 k个feature_map 两两之间的关联性相似度 k*k的矩阵
Args：
---x feature_map from Conv层 shape：[1 height width channels] """
b, h ,w, ch = x.get_shape().as_list()
features = tf.reshape(x, shape=[b, h*w, ch])
# [ch ch] = [ch h*w] 矩阵相乘 [h*w ch]
# 除以维度相乘是为了防止值过大除以统一的数
# features是三维的导致gram矩阵也是三维的
gram = tf.matmul(features, features, adjoint_a=True) / tf.constant(h*w*ch, tf.float32)
# features是2维的导致gram矩阵也是2维的
# gram = tf.matmul(tf.matrix_transpose(features[0]), features[0]) / tf.constant(h*w*ch, tf.float32)
return gram
pass
style_gram_matrix = [gram_matrix(feature) for feature in style_features]
result_style_gram_matrix = [gram_matrix(feature) for feature in result_style_features]
style_loss = tf.zeros(shape=1, dtype=tf.float32)
for s, s_ in zip(style_gram_matrix, result_style_gram_matrix):
# loss = mse 平方差损失函数
# axis=[0, 1, 2, 3] 这里会报错 gram矩阵输出是三维的
style_loss += tf.reduce_mean(tf.square(s - s_), axis=[1, 2])
# gram矩阵输出是二维的
# style_loss += tf.reduce_mean(tf.square(s - s_), axis=[0, 1])
pass
loss = content_loss * lambda_c + style_loss * lambda_s
with tf.name_scope('train'):
train_op = tf.train.AdamOptimizer(learning_rate).minimize(loss)
init_op = tf.global_variables_initializer()
with tf.Session() as sess:
sess.run(init_op)
for step in range(num_steps):
loss_value, content_loss_value, style_loss_value, _ = \
sess.run(fetches=[loss, content_loss, style_loss, train_op], feed_dict={
content_img: content_img_arr_val,
style_img: style_img_arr_val
})
print('step:%d loss_value:%8.4f content_loss_value:%8.4f style_loss_value:%8.4f'
% (step+1, loss_value[0], content_loss_value[0], style_loss_value[0]))
# result_image：shape [224, 224, 3]
result_image = initial_image_result.eval(sess)[0]
# np.clip值裁剪小于0的变为0 大于255的变为255
result_image = np.clip(result_image, 0, 255)
result_image = np.asarray(result_image, dtype=np.uint8)
# np_img = np.asarray([np_img], dtype=np.float32)
img = Image.fromarray(result_image)
result_image_path = os.path.join(output_dir, 'result-%05d.jpg' % (step + 1))
img.save(result_image_path)
pass

复制代码

123456dfg · 发表于 2019-9-14 18:41:38

多谢提供

qianrushi9 · 发表于 2019-11-5 11:35:03

nice多谢

yang2019 · 发表于 2019-11-16 18:54:48

非常好的资料，谢谢

mhy286664061 · 发表于 2020-3-6 19:47:25

haodongxi xiexie

一知羊 · 发表于 2020-3-7 12:47:43

菜菜 · 发表于 2020-4-9 15:34:23

想要下载资源～

sunhaichao · 发表于 2020-4-21 19:17:13

好的好好看看

Drg · 发表于 2020-6-23 19:22:48

谢谢大佬新手求教

ailearningfei · 发表于 2020-8-24 15:56:50

aggjlkj;klgyjudytuytui

		自动登录	找回密码
密码			立即注册

[课堂笔记] 预训练好的卷积神经网络CNN模型：vgg16.npy vgg19.npy 下载