简单CNN 测试例
1.训练数据:
import tensorflow as tf import cv2 import os import numpy as np import time import matplotlib.pyplot as plt from sklearn.metrics import confusion_matrix, classification_report drop_prob = 0.4 input_imgs = tf.placeholder(dtype=tf.float32,shape=[None,128,64,3],name='input_imgs') input_label = tf.placeholder(dtype=tf.float32,shape=[None,2],name='input_label') # 初始化权重(卷积核) def weight_init(shape): weight = tf.truncated_normal(shape, stddev=0.1, dtype=tf.float32) return tf.Variable(weight) # 初始化偏置项 def bias_init(shape): bias = tf.random_normal(shape, dtype=tf.float32) return tf.Variable(bias) # 全连接层 def fch_init(layer1, layer2, const=1): min = -const * (6.0 / (layer1 + layer2)) max = -min weight = tf.random_uniform([layer1, layer2], minval=min, maxval=max, dtype=tf.float32) return tf.Variable(weight) # 卷积层 def conv2d(images, weight): return tf.nn.conv2d(images, weight, strides=[1, 1, 1, 1], padding='SAME') # 最大池化层 def max_pool2x2(images, tname): return tf.nn.max_pool(images, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME', name=tname) # 卷积核3*3*3 16个 第一层卷积 w1 = weight_init([3, 3, 3, 16]) b1 = bias_init([16]) # 结果 NHWC N H W C conv_1 = conv2d(input_imgs, w1) + b1 relu_1 = tf.nn.relu(conv_1, name='relu_1') max_pool_1 = max_pool2x2(relu_1, 'max_pool_1') # 卷积核3*3*16 32个 第二层卷积 w2 = weight_init([3, 3, 16, 32]) b2 = bias_init([32]) conv_2 = conv2d(max_pool_1, w2) + b2 # 激活层 relu_2 = tf.nn.relu(conv_2, name='relu_2') # 亚采样层 max_pool_2 = max_pool2x2(relu_2, 'max_pool_2') w3 = weight_init([3, 3, 32, 64]) b3 = bias_init([64]) conv_3 = conv2d(max_pool_2, w3) + b3 relu_3 = tf.nn.relu(conv_3, name='relu_3') max_pool_3 = max_pool2x2(relu_3, 'max_pool_3') print(max_pool_3.shape, '-------扁平-------',max_pool_3.shape[1]*max_pool_3.shape[2]*max_pool_3.shape[3]) f_input = tf.reshape(max_pool_3, [-1, max_pool_3.shape[1]*max_pool_3.shape[2]*max_pool_3.shape[3]]) print('-=--=', f_input.shape) # 全连接第一层 31*31*32,512 f_w1 = fch_init(8192, 512) f_b1 = bias_init([512]) f_r1 = tf.matmul(f_input, f_w1) + f_b1 f_relu_r1 = tf.nn.relu(f_r1) # 抛弃一部分神经元,防止过拟合 f_dropout_r1 = tf.nn.dropout(f_relu_r1, drop_prob) print('f_dropout_r1.shape-----------', f_dropout_r1.shape)
#第二层 f_w2 = fch_init(512, 128) f_b2 = bias_init([128]) f_r2 = tf.matmul(f_dropout_r1, f_w2) + f_b2 f_relu_r2 = tf.nn.relu(f_r2) f_dropout_r2 = tf.nn.dropout(f_relu_r2, drop_prob) # 全连接第三层 512,2 f_w3 = fch_init(128, 2) f_b3 = bias_init([2]) f_r3 = tf.matmul(f_dropout_r2, f_w3) + f_b3 # print(f_r3.shape, '-=-===============') f_softmax = tf.nn.softmax(f_r3, name='f_softmax') # 定义交叉熵 cross_entry = tf.reduce_mean(tf.reduce_sum(-input_label * tf.log(f_softmax))) optimizer = tf.train.AdamOptimizer(0.0001).minimize(cross_entry) # 计算准确率 arg1 = tf.argmax(input_label, 1) arg2 = tf.argmax(f_softmax, 1) cos = tf.equal(arg1, arg2) acc = tf.reduce_mean(tf.cast(cos, dtype=tf.float32)) sess = tf.Session() sess.run(tf.global_variables_initializer()) train_img = [] train_labels = [] test_img = [] test_labels = [] images = [] labels = [] # for root, dirs, files in os.walk('../img/pos/'): # images.append(os.path.join('../img.pos',)) for fileName in os.listdir('../img/pos'): images.append([os.path.join('../img/pos', fileName)]) labels.append([1,0]) for fileName in os.listdir('../img/neg'): images.append([os.path.join('../img/neg', fileName)]) labels.append([0,1]) images = np.array(images) labels = np.array(labels) permutation = np.random.permutation(labels.shape[0]) images = images[permutation,:] labels = labels[permutation,:]
#获取训练数据或者测试数据 def get_train_data(batch,isTrain=True): global test_labels,test_img if isTrain: train_num = int(labels.shape[0]*0.8) train_img = images[:train_num,:] train_labels = labels[:train_num,:] test_img = images[train_num:,:] test_labels = labels[train_num:,:] # print(train_img[batch:batch+20], train_labels[batch:batch+20]) return train_img[batch*20:(batch+1)*20], train_labels[batch*20:(batch+1)*20] else: return test_img[batch*20:(batch+1)*20], test_labels[batch*20:(batch+1)*20] # get_train_data(20) def read_img(train_img): # print(train_img) imgs = [] for i in range(20): img = cv2.imread(train_img[i][0]) imgs.append(img) # cv2.imshow('12',img) # cv2.waitKey(0) imgs = np.array(imgs) return imgs # print(imgs) Cost = [] Accuracy=[] start_time = time.time() for i in range(100): train_img, train_labels = get_train_data(i) imgs = read_img(train_img) result,acc1,cross_entry_r,cos1,f_softmax1,relu_1_r= sess.run([optimizer,acc,cross_entry,cos,f_softmax,relu_1],feed_dict={input_imgs:imgs,input_label:train_labels}) print("rpoch: {}, accurate: {} , cross_loss:{}".format(i,acc1,cross_entry_r)) Cost.append(cross_entry_r) Accuracy.append(acc1) print('total time:%d'%(time.time()-start_time)) # 代价函数曲线 fig1,ax1 = plt.subplots(figsize=(10,7)) plt.plot(Cost) print('---------cost-----------',Cost) ax1.set_xlabel('Epochs') ax1.set_ylabel('Cost') plt.title('Cross Loss') plt.grid() plt.show() # 准确率曲线 fig7,ax7 = plt.subplots(figsize=(10,7)) plt.plot(Accuracy) ax7.set_xlabel('Epochs') ax7.set_ylabel('Accuracy Rate') plt.title('Train Accuracy Rate') plt.grid() plt.show() #测试 test_img,test_labels = get_train_data(1,False) test_img = read_img(test_img) arg2_r = sess.run(arg2,feed_dict={input_imgs:test_img,input_label:test_labels}) arg1_r = sess.run(arg1,feed_dict={input_imgs:test_img,input_label:test_labels}) print (classification_report(arg1_r, arg2_r)) #保存模型 global_step:训练模型的命名 saver = tf.train.Saver() saver.save(sess, './model/my-gender-v1.0',global_step=123)
2. 从保存的模型中读取数据
import tensorflow as tf import numpy as np import cv2 import matplotlib.pyplot as plt import os #取一张图片 # img/pos/758.jpg img = cv2.imread('../img/pos/760.jpg') # labels = train_data.labels[0:1] fig2,ax2 = plt.subplots(figsize=(2,2)) ax2.imshow(img) plt.show() img = np.reshape(img,[1,128,64,3]) sess = tf.Session() graph_path=os.path.abspath('./model/my-gender-v1.0-123.meta') model=os.path.abspath('./model/') server = tf.train.import_meta_graph(graph_path) server.restore(sess,tf.train.latest_checkpoint(model)) graph = tf.get_default_graph() #填充feed_dict x = graph.get_tensor_by_name('input_imgs:0') y = graph.get_tensor_by_name('input_label:0') feed_dict={x:img,y:[[1,0]]} #第一层卷积+池化 relu_1 = graph.get_tensor_by_name('relu_1:0') max_pool_1 = graph.get_tensor_by_name('max_pool_1:0') #第二层卷积+池化 relu_2 = graph.get_tensor_by_name('relu_2:0') max_pool_2 = graph.get_tensor_by_name('max_pool_2:0') #第三层卷积+池化 relu_3 = graph.get_tensor_by_name('relu_3:0') max_pool_3 = graph.get_tensor_by_name('max_pool_3:0') #全连接最后一层输出 f_softmax = graph.get_tensor_by_name('f_softmax:0') #relu_1_r,max_pool_1_,relu_2,max_pool_2,relu_3,max_pool_3,f_softmax=sess.run([relu_1,max_pool_1,relu_2,max_pool_2,relu_3,max_pool_3,f_softmax],feed_dict) #----------------------------------各个层特征可视化------------------------------- #conv1 特征 r1_relu = sess.run(relu_1,feed_dict) print('r1_relu',r1_relu.shape) # 将矩阵转置 r1_tranpose = sess.run(tf.transpose(r1_relu,[3,0,1,2])) print('r1_tranpose',r1_tranpose.shape) fig,ax = plt.subplots(nrows=1,ncols=16,figsize=(16,1)) for i in range(16): ax[i].imshow(r1_tranpose[i][0]) plt.title('Conv1 16*112*92') plt.show() #pool1特征 max_pool_1 = sess.run(max_pool_1,feed_dict) r1_tranpose = sess.run(tf.transpose(max_pool_1,[3,0,1,2])) fig,ax = plt.subplots(nrows=1,ncols=16,figsize=(16,1)) for i in range(16): ax[i].imshow(r1_tranpose[i][0]) plt.title('Pool1 16*56*46') plt.show() #conv2 特征 r2_relu = sess.run(relu_2,feed_dict) r2_tranpose = sess.run(tf.transpose(r2_relu,[3,0,1,2])) fig,ax = plt.subplots(nrows=1,ncols=32,figsize=(32,1)) for i in range(32): ax[i].imshow(r2_tranpose[i][0]) plt.title('Conv2 32*56*46') plt.show() #pool2 特征 max_pool_2 = sess.run(max_pool_2,feed_dict) tranpose = sess.run(tf.transpose(max_pool_2,[3,0,1,2])) fig,ax = plt.subplots(nrows=1,ncols=32,figsize=(32,1)) for i in range(32): ax[i].imshow(tranpose[i][0]) plt.title('Pool2 32*28*23') plt.show() #conv3 特征 r3_relu = sess.run(relu_3,feed_dict) tranpose = sess.run(tf.transpose(r3_relu,[3,0,1,2])) fig,ax = plt.subplots(nrows=1,ncols=64,figsize=(32,1)) for i in range(64): ax[i].imshow(tranpose[i][0]) plt.title('Conv3 64*28*23') plt.show() #pool3 特征 max_pool_3 = sess.run(max_pool_3,feed_dict) tranpose = sess.run(tf.transpose(max_pool_3,[3,0,1,2])) fig,ax = plt.subplots(nrows=1,ncols=64,figsize=(32,1)) for i in range(64): ax[i].imshow(tranpose[i][0]) plt.title('Pool3 64*14*12') plt.show() print(sess.run(f_softmax,feed_dict))
注意:
卷积神经网络:conv2d ->pool->relu(softmax二分类) 多层卷积神经网络的使用,注意使用卷积核的个数,步长及大小。