这一部分几乎网上写的代码的框架都是差不多的，，，所以只要理解每一层是什么意思，，怎么用就行了，，当然可以继续深究下去，，理解每一层为什么这么做，，这么做的方法是什么等等，，
我们之前在这一部分投入的精力也最多，，出现的问题也是很多的，，其中我感觉最印象深刻的就是层数的叠加和那个神奇的 loss=0.69.....
层数的叠加就是说大部分的代码都是做的3层的神经网络，，但是很少有做多层的，，当时找了很多的资料才一点一点的大致会算怎么计算下一层
关于那一个神奇的 loss=0.69 ，，貌似不同的网络、需求下可能不会出现这种情况，，，但是如果出现不管比怎么调参出来的loss都保持在0.69附近的话，，可以试着查一下这一个问题，，我们最后的解决方法是再添加一层 batch nomalization层 ，，具体干什么的网上有很多讲解的，，，加了它就可以使loss快速收敛，，，
训练模型这一部分的框架都是差不多的，，变化可能有参数的变化，，优化器的不同等等
代码：
#!/usr/bin/python #coding=utf-8 ''' face detect convolution''' # pylint: disable=invalid-name import os import sys import logging as log import matplotlib.pyplot as plt import common import numpy as np from tensorflow.examples.tutorials.mnist import input_data import tensorflow as tf import cv2 from sklearn.model_selection import train_test_split import random SIZE = 64 x = tf.placeholder(tf.float32, [None, SIZE, SIZE, 3]) y_ = tf.placeholder(tf.float32, [None, None]) keep_prob_5 = tf.placeholder(tf.float32) keep_prob_75 = tf.placeholder(tf.float32) def weightVariable(shape): ''' build weight variable''' init = tf.random_normal(shape, stddev=0.01) #init = tf.truncated_normal(shape, stddev=0.01) return tf.Variable(init) def biasVariable(shape): ''' build bias variable''' init = tf.random_normal(shape) #init = tf.truncated_normal(shape, stddev=0.01) return tf.Variable(init) def conv2d(x, W): ''' conv2d by 1, 1, 1, 1''' return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME') def maxPool(x): ''' max pooling''' return tf.nn.max_pool(x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME') def dropout(x, keep): ''' drop out''' return tf.nn.dropout(x, keep) def batch_norm_layer(value,is_training=False,name='batch_norm'): ''' 批量归一化 返回批量归一化的结果 args: value:代表输入，第一个维度为batch_size is_training:当它为True，代表是训练过程，这时会不断更新样本集的均值与方差。当测试时，要设置成False，这样就会使用训练样本集的均值和方差。 默认测试模式 name：名称。 ''' if is_training is True: #训练模式 使用指数加权函数不断更新均值和方差 return tf.contrib.layers.batch_norm(inputs=value,decay=0.9,updates_collections=None,is_training = True) else: #测试模式 不更新均值和方差，直接使用 return tf.contrib.layers.batch_norm(inputs=value,decay=0.9,updates_collections=None,is_training = False) def cnnLayer(classnum, isTrue): ''' create cnn layer''' # 第一层 W1 = weightVariable([3, 3, 3, 32]) # 卷积核大小(3,3)， 输入通道(3)， 输出通道(32) b1 = biasVariable([32]) conv1 = tf.nn.relu(batch_norm_layer(conv2d(x, W1) + b1, isTrue)) pool1 = maxPool(conv1) # 减少过拟合，随机让某些权重不更新 drop1 = dropout(pool1, keep_prob_5) # 32 * 32 * 32 多个输入channel 被filter内积掉了 # 第二层 W2 = weightVariable([3, 3, 32, 64]) b2 = biasVariable([64]) conv2 = tf.nn.relu(batch_norm_layer(conv2d(drop1, W2) + b2, isTrue)) pool2 = maxPool(conv2) drop2 = dropout(pool2, keep_prob_5) # 64 * 16 * 16 # 第三层 W3 = weightVariable([3, 3, 64, 64]) b3 = biasVariable([64]) conv3 = tf.nn.relu(conv2d(drop2, W3) + b3) pool3 = maxPool(conv3) drop3 = dropout(pool3, keep_prob_5) # 64 * 8 * 8 # 全连接层 Wf = weightVariable([8*8*64,512]) bf = biasVariable([512]) drop3_flat = tf.reshape(drop3, [-1, 8*8*64]) dense = tf.nn.relu(tf.matmul(drop3_flat, Wf) + bf) dropf = dropout(dense, keep_prob_75) # 输出层 Wout = weightVariable([512, classnum]) bout = weightVariable([classnum]) #out = tf.matmul(dropf, Wout) + bout out = tf.add(tf.matmul(dropf, Wout), bout) return out # # 第三层 # W3 = weightVariable([3, 3, 64, 128]) # b3 = biasVariable([128]) # conv3 = tf.nn.relu(batch_norm_layer(conv2d(drop2, W3) + b3, True)) # pool3 = maxPool(conv3) # drop3 = dropout(pool3, keep_prob_5) # 128 * 8 * 8 # # 第四层 # W4 = weightVariable([3, 3, 128, 512]) # b4 = biasVariable([512]) # conv4 = tf.nn.relu(batch_norm_layer(conv2d(drop3, W4) + b4, True)) # pool4 = maxPool(conv4) # drop4 = dropout(pool4, keep_prob_5) # 512 * 4 * 4 # # 第五层 # W5 = weightVariable([3, 3, 512, 1024]) # b5 = biasVariable([1024]) # conv5 = tf.nn.relu(batch_norm_layer(conv2d(drop4, W5) + b5, True)) # pool5 = maxPool(conv5) # drop5 = dropout(pool5, keep_prob_5) # 1024 * 2 * 2 # # 第六层 # W6 = weightVariable([3, 3, 1024, 1024]) # b6 = biasVariable([1024]) # conv6 = tf.nn.relu(conv2d(drop5, W6) + b6) # pool6 = maxPool(conv6) # drop6 = dropout(pool6, keep_prob_5) # 2048 * 1 * 1 # # 全连接层 # Wf = weightVariable([1*1*1024, 2048]) # bf = biasVariable([2048]) # drop3_flat = tf.reshape(drop6, [-1, 1*1*1024]) # dense = tf.nn.relu(tf.matmul(drop3_flat, Wf) + bf) # # dense = tf.nn.relu(tf.matmul(max_pool22_flat, Wf) + bf) # dropf = dropout(dense, keep_prob_75) # # 输出层 # Wout = weightVariable([2048, classnum]) # bout = weightVariable([classnum]) # #out = tf.matmul(dropf, Wout) + bout # out = tf.add(tf.matmul(dropf, Wout), bout) # #return out def train(train_x, train_y, tfsavepath): ''' train''' ##### log.debug('train') # 随机划分测试集与训练集 train_x,test_x,train_y,test_y = train_test_split(train_x, train_y, test_size=0.05, random_state=random.randint(0,100)) # 得到卷积结果 out = cnnLayer(train_y.shape[1],True) # 参数：图片数据的总数，图片的高、宽、通道 train_x = train_x.reshape(train_x.shape[0], SIZE, SIZE, 3) test_x = test_x.reshape(test_x.shape[0], SIZE, SIZE, 3) print('train size:%s, test size:%s' % (len(train_x), len(test_x))) sys.stdout.flush() # 图片块，每次取32张图片 batch_size = 32 num_batch = len(train_x) // batch_size # 交叉熵 cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=out, labels=y_)) # Adam优化器，学习速率：0.001 train_step = tf.train.AdamOptimizer(0.001).minimize(cross_entropy) # train_step = tf.train.AdadeltaOptimizer(0.001).minimize(cross_entropy) # 比较标签是否相等，再求的所有数的平均值，tf.cast(强制转换类型) # 准确率计算公式 accuracy = tf.reduce_mean(tf.cast(tf.equal(tf.argmax(out, 1), tf.argmax(y_, 1)), tf.float32)) # 将loss与accuracy保存以供tensorboard使用 tf.summary.scalar('loss', cross_entropy) tf.summary.scalar('accuracy', accuracy) merged_summary_op = tf.summary.merge_all() # 数据保存器的初始化 saver = tf.train.Saver() with tf.Session() as sess: # tensorflow初始化 sess.run(tf.global_variables_initializer()) # tensorboard数据保存 summary_writer = tf.summary.FileWriter('./tmp', graph=tf.get_default_graph()) # 迭代80次 for n in range(80): # 每次取32(batch_size)张图片 for i in range(num_batch): # 训练集、测试集分块 batch_x = train_x[i*batch_size : (i+1)*batch_size] batch_y = train_y[i*batch_size : (i+1)*batch_size] # 开始训练数据，同时训练三个变量，返回三个数据 _,loss,summary = sess.run([train_step, cross_entropy, merged_summary_op], feed_dict={x:batch_x,y_:batch_y, keep_prob_5:0.5,keep_prob_75:0.75}) # tensorboard记录数据 summary_writer.add_summary(summary, n*num_batch+i) # 打印损失 print(n*num_batch+i, loss) sys.stdout.flush() # if (n*num_batch+i) % batch_size == 0: # # 获取测试数据的准确率 # acc = accuracy.eval({x:test_x, y_:test_y, keep_prob_5:1.0, keep_prob_75:1.0}) # print(n*num_batch+i, acc, '--', n) # accc = acc # # 准确率大于0.98时保存并退出 # if acc > 0.95 and n > 2: # # saver.save(sess, './train_faces.model', global_step=n*num_batch+i) # saver.save(sess, tfsavepath) # # saver.save(sess, tfsavepath) # sys.exit(0) # # saver.save(sess, './train_faces.model', global_step=n*num_batch+i) # # saver.save(sess, tfsavepath) # print('accuracy less 0.98, exited!') # 准确率计算表达式 acc = accuracy.eval({x:test_x, y_:test_y, keep_prob_5:1.0, keep_prob_75:1.0}) print('after 80 times run: accuracy is ', acc) sys.stdout.flush() # 模型保存 saver.save(sess, tfsavepath) if __name__ == '__main__': pass 识别这部分就是将训练好的模型加载，，然后实时的获取每一张相片处理等等，，，