深度学习-CNN卷积神经网络-目标分类、迁移学习

学习李伟老师《深度学习》课程

目标分类基本框架

数据准备：数据来源、数据扩充、数据规范

模型设计：模型类型（分类/回归/分类+回归/多目标分类）-> 现有模型->模型设计（局部更改/从头设计）

训练细节：Batch-Size、数据循环方式、平衡性考虑

迁移学习

概念

迁移学习(Transfer learning)就是把已学训练好的模型参数迁移到新的模型来帮助新模型训练。大部分数据或任务是存在相关性的，所以通过迁移学习我们可以将已经学到的模型参数（也可理解为模型学到的知识）通过某种方式来分享给新模型从而加快并优化模型的学习效率不用像大多数网络那样从零学习。

数据量	迁移学习
少量数据	全连接层更新
大量数据	全连接层+部分Conv层更新

学习率的处理

最低卷积层基本不变

中间卷积层看情况

最后全连接层参数情况、结构均改变

基础模型选择

ImageNet:VGG、GoogleNet、ResNet

基于VGG进行人脸表情识别

GitHub源码下载：https://github.com/YEN-GitHub/DeepLearning_BasicByTensorFlow/tree/master/CNN_targetClassification

VGG16参数下载：https://github.com/fchollet/deep-learning-models/releases



安装tflearn:

#coding=utf-8
# @Author: yangenneng
# @Time: 2018-02-10 10:56
# @Abstract：基于VGG进行人脸表情识别

import tflearn
from tflearn.data_preprocessing import ImagePreprocessing
from tflearn.data_utils import image_preloader
import os

#重新定义VGG
def vgg16(input,output):
activation_function='relu' #激活函数

# scope:define this layer scope (optional).
# activation: 激活函数
# trainable: bool量，是否可以被训练 trainable=False表示不会更新本层的参数
x = tflearn.conv_2d(input, 64, 3, activation=activation_function, scope='conv1_1', trainable=False)
x = tflearn.conv_2d(x, 64, 3, activation=activation_function, scope='conv1_2', trainable=False)
#最大池化操作
x = tflearn.max_pool_2d(x, 2, strides=2, name='maxpool1')

x = tflearn.conv_2d(x, 128, 3, activation=activation_function, scope='conv2_1', trainable=False)
x = tflearn.conv_2d(x, 128, 3, activation=activation_function, scope='conv2_2', trainable=False)
x = tflearn.max_pool_2d(x, 2, strides=2, name='maxpool2')

x = tflearn.conv_2d(x, 256, 3, activation=activation_function, scope='conv3_1', trainable=False)
x = tflearn.conv_2d(x, 256, 3, activation=activation_function, scope='conv3_2', trainable=False)
x = tflearn.conv_2d(x, 256, 3, activation=activation_function, scope='conv3_3', trainable=False)
x = tflearn.max_pool_2d(x, 2, strides=2, name='maxpool3')

x = tflearn.conv_2d(x, 512, 3, activation=activation_function, scope='conv4_1', trainable=False)
x = tflearn.conv_2d(x, 512, 3, activation=activation_function, scope='conv4_2', trainable=False)
x = tflearn.conv_2d(x, 512, 3, activation=activation_function, scope='conv4_3', trainable=False)
x = tflearn.max_pool_2d(x, 2, strides=2, name='maxpool4')

x = tflearn.conv_2d(x, 512, 3, activation=activation_function, scope='conv5_1')
x = tflearn.conv_2d(x, 512, 3, activation=activation_function, scope='conv5_2')
x = tflearn.conv_2d(x, 512, 3, activation=activation_function, scope='conv5_3')
x = tflearn.max_pool_2d(x, 2, strides=2, name='maxpool5')

# 全连接层
x = tflearn.fully_connected(x, 4096, activation=activation_function, scope='fc6')
# dropout操作
x = tflearn.dropout(x, 0.5, name='dropout1')

#此处原来是4096 现在改为2048，减小参数
x = tflearn.fully_connected(x, 2048, activation=activation_function, scope='fc7', restore=False)
x = tflearn.dropout(x, 0.5, name='dropout2')

x = tflearn.fully_connected(x, output, activation='softmax', scope='fc8', restore=False)

return x

#模型路径
model_path="."
#数据列表
files_list = "./train_fvgg_emo.txt"

#(X, Y): with X the images array and Y the labels array.
X,Y=image_preloader(files_list,image_shape=(224,224),mode='file',
categorical_labels=True,normalize=False,
files_extension=['.jpg', '.png'],filter_channel=True
)

output=7 #输出分类

# VGG 图像预处理
img_pre=ImagePreprocessing()
# 确定数据是规范的
img_pre.add_featurewise_zero_center(mean=[123.68, 116.779, 103.939],per_channel=True)

# VGG Network building
net=tflearn.input_data(shape=[None, 224, 224, 3],name='input',data_preprocessing=img_pre)

softmax=vgg16(net,output)

regression=tflearn.regression(softmax,optimizer='adam',
learning_rate=0.01,restore=False
)
#checkpoint_path：要评估的特定检查点的路径。如果为 None，则使用 model_dir 中的最新检查点。
model=tflearn.DNN(regression,checkpoint_path='vgg-finetuning',
max_checkpoints=3,tensorboard_verbose=2,
tensorboard_dir="./logs"
)

model_file=os.path.join(model_path,'vgg16_weights_tf_dim_ordering_tf_kernels.h5')

model.load(model_file,weights_only=True)

# Start finetuning
model.fit(X, Y, n_epoch=20, validation_set=0.1, shuffle=True,
show_metric=True, batch_size=64, snapshot_epoch=False,
snapshot_step=200, run_id='vgg-finetuning')

model.save('vgg_finetune_emo.tfmodel')