【Kaggle竞赛】AlexNet模型定义及实现

# coding:utf-8
# filename:alexnet.py
# Environment:windows10,python3,numpy,TensorFlow1.9,numpy,time
# Function:负责定义设计AlexNet网络

import os
import numpy as np
import tensorflow as tf
# 准备数据程序模块
import input_data

IMG_W=227
IMG_H=227
IMG_C=3
BATCH_SIZE = 20
n_epoch = 10000
n_classes = 2
# 本地电脑训练对应路径地址
train_dir = "F:/Software/Python_Project/Classification-cat-dog/train/"
logs_train_dir = "F:/Software/Python_Project/Classification-cat-dog/logs/"

# 云服务器训练对应路径地址
# train_dir = '/data/Dogs-Cats-Redux-Kernels-Edition/train/'
# logs_train_dir = '/data/Dogs-Cats-Redux-Kernels-Edition/logs/

#----------------------------------------------定义构建网络-------------------------------------
# 占位符,用于得到传递进来的真实训练样本,输入数据为猫狗识别数据集batch
# x  = tf.placeholder(tf.float32,shape=[BATCH_SIZE,IMG_W,IMG_H,IMG_C],name='x')
# y_ = tf.placeholder(tf.int32,shape=[BATCH_SIZE,],name='y_')

def inference(input_tensor,train,batch_size,regularizer,n_classes):
    # conv1,输出矩阵大小(55,55,96)
    with tf.variable_scope('layer1-conv1'):
        conv1_weights = tf.get_variable("weight",[11,11,3,96],initializer=tf.truncated_normal_initializer(stddev=0.1))
        conv1_biases = tf.get_variable("bias", [96], initializer=tf.constant_initializer(0.0))     # 定义偏置变量并初始化
        conv1 = tf.nn.conv2d(input_tensor, conv1_weights, strides=[1, 4, 4, 1], padding='SAME')   # 'VALID'不添加
        relu1 = tf.nn.relu(tf.nn.bias_add(conv1, conv1_biases))                                    # 使用ReLu激活函数，完成去线性化

    # pool1 && norm1,输出矩阵大小(27,27,96)
    with tf.name_scope("layer2-pool1"):
        pool1 = tf.nn.max_pool(relu1, ksize = [1,3,3,1],strides=[1,2,2,1],padding="VALID")
        norm1 = tf.nn.lrn(pool1, depth_radius=4, bias=1.0, alpha=0.001/9.0,beta=0.75, name='norm1') # lrn层，局部响应归一化
        
    # conv2,输出矩阵大小(27,27,256)
    with tf.variable_scope("layer3-conv2"):
        conv2_weights = tf.get_variable("weight",[5,5,96,256],initializer=tf.truncated_normal_initializer(stddev=0.1))
        conv2_biases = tf.get_variable("bias", [256], initializer=tf.constant_initializer(0.0))
        conv2 = tf.nn.conv2d(norm1, conv2_weights, strides=[1, 1, 1, 1], padding='SAME')
        relu2 = tf.nn.relu(tf.nn.bias_add(conv2, conv2_biases))                                      # 使用ReLu激活函数，完成去线性化

    # pool2 && norm2,输出矩阵大小(13,13,256)
    with tf.name_scope("layer4-pool2"):
        pool2 = tf.nn.max_pool(relu2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='VALID')
        norm2 = tf.nn.lrn(pool2, depth_radius=4, bias=1.0, alpha=0.001/9.0,beta=0.75, name='norm1') # lrn层，局部响应归一化
        
    # conv3,输出矩阵大小(13,13,384)
    with tf.variable_scope("layer5-conv3"):
        conv3_weights = tf.get_variable("weight",[3,3,256,384],initializer=tf.truncated_normal_initializer(stddev=0.1))
        conv3_biases = tf.get_variable("bias", [384], initializer=tf.constant_initializer(0.0))
        conv3 = tf.nn.conv2d(norm2, conv3_weights, strides=[1, 1, 1, 1], padding='SAME')
        relu3 = tf.nn.relu(tf.nn.bias_add(conv3, conv3_biases))                                      # 使用ReLu激活函数，完成去线性化

    # conv4,输出矩阵大小(13,13,384)
    with tf.variable_scope("layer6-conv4"):
        conv4_weights = tf.get_variable("weight",[3,3,384,384],initializer=tf.truncated_normal_initializer(stddev=0.1))
        conv4_biases = tf.get_variable("bias", [384], initializer=tf.constant_initializer(0.0))
        conv4 = tf.nn.conv2d(relu3, conv4_weights, strides=[1, 1, 1, 1], padding='SAME')
        relu4 = tf.nn.relu(tf.nn.bias_add(conv4, conv4_biases))                                      # 使用ReLu激活函数，完成去线性化

    # conv5,输出矩阵大小(13,13,256)
    with tf.variable_scope("layer7-conv5"):
        conv5_weights = tf.get_variable("weight",[3,3,384,256],initializer=tf.truncated_normal_initializer(stddev=0.1))
        conv5_biases = tf.get_variable("bias", [256], initializer=tf.constant_initializer(0.0))
        conv5 = tf.nn.conv2d(relu4, conv5_weights, strides=[1, 1, 1, 1], padding='SAME')
        relu5 = tf.nn.relu(tf.nn.bias_add(conv5, conv5_biases))                                      # 使用ReLu激活函数，完成去线性化

    # pool5 && norm5,输出矩阵大小(6,6,256),13=(13-3)/2+1
    with tf.name_scope("layer8-pool5"):
        pool5 = tf.nn.max_pool(relu5, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1], padding='VALID')     # 'VALID'不添加

    # 将第5层池化层的输出转换为第6层全连接层的输入格式（向量）
    # pool_shape = pool5.get_shape().as_list()
    # nodes = pool_shape[1]*pool_shape[2]*pool_shape[3]    # 计算矩阵拉直成向量之后的长度,（pool_shape[1]*pool_shape[2]*pool_shape[3]）=(6*6*256)

    # fc1,输出矩阵大小(-1,4096)
    with tf.variable_scope('layer9-fc1'):
        reshaped = tf.reshape(pool5,[batch_size,-1])         # 通过tf.reshaped函数将第5层池化层的输出变成一个batch的向量
        nodes = reshaped.get_shape()[1].value
        fc1_weights = tf.get_variable("weight", [nodes, 4096],
                                      initializer=tf.truncated_normal_initializer(stddev=0.1))
        if regularizer != None: tf.add_to_collection('losses', regularizer(fc1_weights))
        fc1_biases = tf.get_variable("bias", [4096], initializer=tf.constant_initializer(0.1))

        fc1 = tf.nn.relu(tf.matmul(reshaped, fc1_weights) + fc1_biases)
        if train: fc1 = tf.nn.dropout(fc1, 0.5)                                                    # 使用丢失输出技巧dropout
    
    # fc2,输出矩阵大小(-1,4096)
    with tf.variable_scope('layer10-fc2'):
        fc2_weights = tf.get_variable("weight", [4096, 4096],
                                      initializer=tf.truncated_normal_initializer(stddev=0.1))
        if regularizer != None: tf.add_to_collection('losses', regularizer(fc2_weights))
        fc2_biases = tf.get_variable("bias", [4096], initializer=tf.constant_initializer(0.1))

        fc2 = tf.nn.relu(tf.matmul(fc1, fc2_weights) + fc2_biases)
        if train: fc2 = tf.nn.dropout(fc2, 0.5)                                                     # 使用丢失输出技巧dropout

    # fc3,输出层输出矩阵大小(-1,n_classes)=(-1,2)
    with tf.variable_scope('layer11-fc3'):
        fc3_weights = tf.get_variable("weight", [4096, n_classes],
                                      initializer=tf.truncated_normal_initializer(stddev=0.1))
        if regularizer != None: tf.add_to_collection('losses', regularizer(fc3_weights))
        fc3_biases = tf.get_variable("bias", [n_classes], initializer=tf.constant_initializer(0.1))
        logit = tf.matmul(fc2, fc3_weights) + fc3_biases

    # shape:(batch_size,num_classes)=(16,10)
    return logit

# # ------------------------------------网络结束，定义损失、评估模型-------------------------------------------------
# regularizer = tf.contrib.layers.l2_regularizer(0.0001)
# # logits是一个batch_size*10的二维数组
# logits = inference(x,True,regularizer,n_classes)
# print(logits)                                       # 仅供测试程序时用，迭代训练模型时建议注释掉
# # (小处理)将logits乘以1赋值给logits_eval，定义name，方便在后续调用模型时通过tensor名字调用输出tensor
# b = tf.constant(value=1,dtype=tf.float32)
# logits_eval = tf.multiply(logits,b,name='logits_eval') 

# #计算交叉熵作为刻画预测值和真实值之间差距的损失函数
# cross_entroy = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits, labels=y_)
# #计算在当前batch中所有样例的交叉熵平均值
# loss = tf.reduce_mean(cross_entroy,name='loss')+tf.add_n(tf.get_collection('losses'))
# #使用tf.train.AdamOptimizer优化算法来优化损失函数
# train_op = tf.train.AdamOptimizer(learning_rate=0.001).minimize(loss)

# #计算模型在一个batch数据上的正确率
# correct_prediction = tf.equal(tf.cast(tf.argmax(logits,1),tf.int32), y_)    
# acc = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
# print(loss,acc)                                     # 仅供测试程序时用，迭代训练模型时建议注释掉

# 定义损失函数（labels没有进行one-hot编码）
def losses(logits, labels):
    with tf.variable_scope("loss") as scope:
        cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits,
                                                                       labels=labels, name="xentropy_per_example")
        loss = tf.reduce_mean(cross_entropy, name="loss")# 计算张量的平均值
        tf.summary.scalar(scope.name + "loss", loss)     # 对loss汇总和记录
    return loss

# 定义训练函数
def trainning(loss, learning_rate):
    with tf.name_scope("optimizer"):
        # 使用Adam优化器
        optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)
        # 定义训练轮数变量
        global_step = tf.Variable(0, name="global_step", trainable=False)
        # 更新参数最小化损失
        train_op = optimizer.minimize(loss, global_step=global_step)
    return train_op
# 定义评估函数，计算准确率
def evaluation(logits,labels):
    with tf.variable_scope('accuracy') as scope:
        correct = tf.nn.in_top_k(logits,labels,1)
        correct = tf.cast(correct,tf.float16)           # 将布尔型数据转换为float16
        accuracy = tf.reduce_mean(correct)              # 计算correct均值得到准确率
        tf.summary.scalar(scope.name+'/accuracy',accuracy)    # 对准确率数据进行汇总
    return accuracy