网页站点,wordpress app插件,网站建设 模板,东莞有哪些好企业一、CNN训练模型 模型尺寸分析#xff1a;卷积层全都采用了补0#xff0c;所以经过卷积层长和宽不变#xff0c;只有深度加深。池化层全都没有补0#xff0c;所以经过池化层长和宽均减小#xff0c;深度不变。http://download.tensorflow.org/example_images/flower_photo…一、CNN训练模型 模型尺寸分析卷积层全都采用了补0所以经过卷积层长和宽不变只有深度加深。池化层全都没有补0所以经过池化层长和宽均减小深度不变。http://download.tensorflow.org/example_images/flower_photos.tgz 模型尺寸变化100×100×3-100×100×32-50×50×32-50×50×64-25×25×64-25×25×128-12×12×128-12×12×128-6×6×128 CNN训练代码如下 ?123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174from skimage import io,transformimport globimport osimport tensorflow as tfimport numpy as npimport time#数据集地址pathE:/data/datasets/flower_photos/#模型保存地址model_pathE:/data/model/flower/model.ckpt#将所有的图片resize成100*100w100h100c3#读取图片def read_img(path):    cate[pathx for x in os.listdir(path) if os.path.isdir(pathx)]    imgs[]    labels[]    for idx,folder in enumerate(cate):        for im in glob.glob(folder/*.jpg):            print(reading the images:%s%(im))            imgio.imread(im)            imgtransform.resize(img,(w,h))            imgs.append(img)            labels.append(idx)    return np.asarray(imgs,np.float32),np.asarray(labels,np.int32)data,labelread_img(path)#打乱顺序num_exampledata.shape[0]arrnp.arange(num_example)np.random.shuffle(arr)datadata[arr]labellabel[arr]#将所有数据分为训练集和验证集ratio0.8snp.int(num_example*ratio)x_traindata[:s]y_trainlabel[:s]x_valdata[s:]y_vallabel[s:]#-----------------构建网络----------------------#占位符xtf.placeholder(tf.float32,shape[None,w,h,c],namex)y_tf.placeholder(tf.int32,shape[None,],namey_)def inference(input_tensor, train, regularizer):    with tf.variable_scope(layer1-conv1):        conv1_weights tf.get_variable(weight,[5,5,3,32],initializertf.truncated_normal_initializer(stddev0.1))        conv1_biases tf.get_variable(bias, [32], initializertf.constant_initializer(0.0))        conv1 tf.nn.conv2d(input_tensor, conv1_weights, strides[1, 1, 1, 1], paddingSAME)        relu1 tf.nn.relu(tf.nn.bias_add(conv1, conv1_biases))    with tf.name_scope(layer2-pool1):        pool1 tf.nn.max_pool(relu1, ksize [1,2,2,1],strides[1,2,2,1],paddingVALID)    with tf.variable_scope(layer3-conv2):        conv2_weights tf.get_variable(weight,[5,5,32,64],initializertf.truncated_normal_initializer(stddev0.1))        conv2_biases tf.get_variable(bias, [64], initializertf.constant_initializer(0.0))        conv2 tf.nn.conv2d(pool1, conv2_weights, strides[1, 1, 1, 1], paddingSAME)        relu2 tf.nn.relu(tf.nn.bias_add(conv2, conv2_biases))    with tf.name_scope(layer4-pool2):        pool2 tf.nn.max_pool(relu2, ksize[1, 2, 2, 1], strides[1, 2, 2, 1], paddingVALID)    with tf.variable_scope(layer5-conv3):        conv3_weights tf.get_variable(weight,[3,3,64,128],initializertf.truncated_normal_initializer(stddev0.1))        conv3_biases tf.get_variable(bias, [128], initializertf.constant_initializer(0.0))        conv3 tf.nn.conv2d(pool2, conv3_weights, strides[1, 1, 1, 1], paddingSAME)        relu3 tf.nn.relu(tf.nn.bias_add(conv3, conv3_biases))    with tf.name_scope(layer6-pool3):        pool3 tf.nn.max_pool(relu3, ksize[1, 2, 2, 1], strides[1, 2, 2, 1], paddingVALID)    with tf.variable_scope(layer7-conv4):        conv4_weights tf.get_variable(weight,[3,3,128,128],initializertf.truncated_normal_initializer(stddev0.1))        conv4_biases tf.get_variable(bias, [128], initializertf.constant_initializer(0.0))        conv4 tf.nn.conv2d(pool3, conv4_weights, strides[1, 1, 1, 1], paddingSAME)        relu4 tf.nn.relu(tf.nn.bias_add(conv4, conv4_biases))    with tf.name_scope(layer8-pool4):        pool4 tf.nn.max_pool(relu4, ksize[1, 2, 2, 1], strides[1, 2, 2, 1], paddingVALID)        nodes 6*6*128        reshaped tf.reshape(pool4,[-1,nodes])    with tf.variable_scope(layer9-fc1):        fc1_weights tf.get_variable(weight, [nodes, 1024],                                      initializertf.truncated_normal_initializer(stddev0.1))        if regularizer ! None: tf.add_to_collection(losses, regularizer(fc1_weights))        fc1_biases tf.get_variable(bias, [1024], initializertf.constant_initializer(0.1))        fc1 tf.nn.relu(tf.matmul(reshaped, fc1_weights) fc1_biases)        if train: fc1 tf.nn.dropout(fc1, 0.5)    with tf.variable_scope(layer10-fc2):        fc2_weights tf.get_variable(weight, [1024, 512],                                      initializertf.truncated_normal_initializer(stddev0.1))        if regularizer ! None: tf.add_to_collection(losses, regularizer(fc2_weights))        fc2_biases tf.get_variable(bias, [512], initializertf.constant_initializer(0.1))        fc2 tf.nn.relu(tf.matmul(fc1, fc2_weights) fc2_biases)        if train: fc2 tf.nn.dropout(fc2, 0.5)    with tf.variable_scope(layer11-fc3):        fc3_weights tf.get_variable(weight, [512, 5],                                      initializertf.truncated_normal_initializer(stddev0.1))        if regularizer ! None: tf.add_to_collection(losses, regularizer(fc3_weights))        fc3_biases tf.get_variable(bias, [5], initializertf.constant_initializer(0.1))        logit tf.matmul(fc2, fc3_weights) fc3_biases    return logit#---------------------------网络结束---------------------------regularizer tf.contrib.layers.l2_regularizer(0.0001)logits inference(x,False,regularizer)#(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tensorb tf.constant(value1,dtypetf.float32)logits_eval tf.multiply(logits,b,namelogits_eval) losstf.nn.sparse_softmax_cross_entropy_with_logits(logitslogits, labelsy_)train_optf.train.AdamOptimizer(learning_rate0.001).minimize(loss)correct_prediction tf.equal(tf.cast(tf.argmax(logits,1),tf.int32), y_)    acc tf.reduce_mean(tf.cast(correct_prediction, tf.float32))#定义一个函数按批次取数据def minibatches(inputsNone, targetsNone, batch_sizeNone, shuffleFalse):    assert len(inputs) len(targets)    if shuffle:        indices np.arange(len(inputs))        np.random.shuffle(indices)    for start_idx in range(0, len(inputs) - batch_size 1, batch_size):        if shuffle:            excerpt indices[start_idx:start_idx batch_size]        else:            excerpt slice(start_idx, start_idx batch_size)        yield inputs[excerpt], targets[excerpt]#训练和测试数据可将n_epoch设置更大一些n_epoch10                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                          batch_size64savertf.train.Saver()sesstf.Session()  sess.run(tf.global_variables_initializer())for epoch in range(n_epoch):    start_time time.time()    #training    train_loss, train_acc, n_batch 0, 0, 0    for x_train_a, y_train_a in minibatches(x_train, y_train, batch_size, shuffleTrue):        _,err,acsess.run([train_op,loss,acc], feed_dict{x: x_train_a, y_: y_train_a})        train_loss err; train_acc ac; n_batch 1    print(   train loss: %f % (np.sum(train_loss)/ n_batch))    print(   train acc: %f % (np.sum(train_acc)/ n_batch))    #validation    val_loss, val_acc, n_batch 0, 0, 0    for x_val_a, y_val_a in minibatches(x_val, y_val, batch_size, shuffleFalse):        err, ac sess.run([loss,acc], feed_dict{x: x_val_a, y_: y_val_a})        val_loss err; val_acc ac; n_batch 1    print(   validation loss: %f % (np.sum(val_loss)/ n_batch))    print(   validation acc: %f % (np.sum(val_acc)/ n_batch))saver.save(sess,model_path)sess.close()二、调用模型进行预测 调用模型进行花卉的预测代码如下 ?12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455from skimage import io,transformimport tensorflow as tfimport numpy as nppath1 E:/data/datasets/flower_photos/daisy/5547758_eea9edfd54_n.jpgpath2 E:/data/datasets/flower_photos/dandelion/7355522_b66e5d3078_m.jpgpath3 E:/data/datasets/flower_photos/roses/394990940_7af082cf8d_n.jpgpath4 E:/data/datasets/flower_photos/sunflowers/6953297_8576bf4ea3.jpgpath5 E:/data/datasets/flower_photos/tulips/10791227_7168491604.jpgflower_dict {0:dasiy,1:dandelion,2:roses,3:sunflowers,4:tulips}w100h100c3def read_one_image(path):    img io.imread(path)    img transform.resize(img,(w,h))    return np.asarray(img)with tf.Session() as sess:    data []    data1 read_one_image(path1)    data2 read_one_image(path2)    data3 read_one_image(path3)    data4 read_one_image(path4)    data5 read_one_image(path5)    data.append(data1)    data.append(data2)    data.append(data3)    data.append(data4)    data.append(data5)    saver tf.train.import_meta_graph(E:/data/model/flower/model.ckpt.meta)    saver.restore(sess,tf.train.latest_checkpoint(E:/data/model/flower/))    graph tf.get_default_graph()    x graph.get_tensor_by_name(x:0)    feed_dict {x:data}    logits graph.get_tensor_by_name(logits_eval:0)    classification_result sess.run(logits,feed_dict)    #打印出预测矩阵    print(classification_result)    #打印出预测矩阵每一行最大值的索引    print(tf.argmax(classification_result,1).eval())    #根据索引通过字典对应花的分类    output []    output tf.argmax(classification_result,1).eval()    for i in range(len(output)):        print(第,i1,朵花预测:flower_dict[output[i]])运行结果 ?1234567891011[[  5.76620245   3.18228579  -3.89464641  -2.81310582   1.40294015] [ -1.01490593   3.55570269  -2.76053429   2.93104005  -3.47138596] [ -8.05292606  -7.26499033  11.70479774   0.59627819   2.15948296] [ -5.12940931   2.18423128  -3.33257103   9.0591135    5.03963232] [ -4.25288343  -0.95963973  -2.33347392   1.54485476   5.76069307]][0 1 2 3 4]第 1 朵花预测:dasiy第 2 朵花预测:dandelion第 3 朵花预测:roses第 4 朵花预测:sunflowers第 5 朵花预测:tulips预测结果和调用模型代码中的五个路径相比较是完全准确的。 本文的模型对于花卉的分类准确率大概在70%左右采用迁移学习调用Inception-v3模型对本文中的花卉数据集分类准确率在95%左右。主要的原因在于本文的CNN模型较于简单而且花卉数据集本身就比mnist手写数字数据集分类难度就要大一点同样的模型在mnist手写数字的识别上准确率要比花卉数据集准确率高不少。