tensorflow学習--fashion-mnistデータセットをボリュームニューラルネットワークで分類する

#       
import tensorflow as tf
import numpy as np
from tensorflow import keras

#    one_hot  
def to_onehot(y,num):
    lables = np.zeros([num,len(y)])
    for i in range(len(y)):
        lables[int(y[i]),i] = 1
    return lables.T

#      
mnist = keras.datasets.fashion_mnist
(train_images,train_lables),(test_images,test_lables) = mnist.load_data()
#      ，        
train_images = train_images / 255.0
test_images = test_images / 255.0
#    one_hot  
train_lables = to_onehot(train_lables,10)
test_lables = to_onehot(test_lables,10)
#       ，               channel，fashion-mnist    channel=1 [batch,28,28,28] -> [batch,28,28,1]
train_images_exp = np.expand_dims(train_images,axis=3)
test_images_exp = np.expand_dims(test_images,axis=3)

input_nodes = 784    #    width*height
output_nodes = 10    #       
image_size = 28     #      （=  ）
channels = 1
labels = 10
conv1_deep = 32    #          ，               [*,*,32]
conv1_size = 5     #       
conv2_deep = 64
conv2_size = 5
fc_size = 512      #         

#     
def inference(input_tensor,train,regularizer):
    with tf.variable_scope('layer1_conv1',reuse=tf.AUTO_REUSE):
        conv1_weights = tf.get_variable('weight',[conv1_size,conv1_size,channels,conv1_deep],initializer=tf.truncated_normal_initializer(stddev=0.1))
        conv1_bias = tf.get_variable('bias',[conv1_deep],initializer=tf.constant_initializer(0.0))
        conv1 = tf.nn.conv2d(input_tensor,conv1_weights,strides=[1,1,1,1],padding="SAME")   #   1 ：strides    1，4      （      1，       ），                 
        relu1 = tf.nn.relu((tf.nn.bias_add(conv1,conv1_bias)))
        
    with tf.variable_scope('layer2_pool1',reuse=tf.AUTO_REUSE):
        pool1 = tf.nn.max_pool(relu1,ksize=[1,2,2,1],strides=[1,2,2,1],padding="SAME")      #    ，  2：padding = same       padding                
                                                                                            #ksize=[1,2,2,1]        2*2 
    with tf.variable_scope('layer3_conv2',reuse=tf.AUTO_REUSE):
        conv2_weights = tf.get_variable("weights",[conv2_size,conv2_size,conv1_deep,conv2_deep],initializer=tf.truncated_normal_initializer(stddev = 0.1))
        conv2_bias = tf.get_variable("bias",[conv2_deep],initializer=tf.constant_initializer(0.0))
        
        conv2 = tf.nn.conv2d(pool1,conv2_weights,strides=[1,1,1,1],padding="SAME")
        relu2 = tf.nn.relu(tf.nn.bias_add(conv2,conv2_bias))
    
    with tf.variable_scope('layer4_pool2',reuse=tf.AUTO_REUSE):
        pool2 = tf.nn.max_pool(relu2,ksize=[1,2,2,1],strides=[1,2,2,1],padding="SAME")
    pool2_shape = pool2.get_shape().as_list()
    nodes = pool2_shape[1] * pool2_shape[2] * pool2_shape[3]
    #         shape[0]       
    reshaped = tf.reshape(pool2,[-1,nodes])   
    with tf.variable_scope('layer5_fc1',reuse=tf.AUTO_REUSE):
        fc1_weights = tf.get_variable("weights",[nodes,fc_size],initializer=tf.truncated_normal_initializer(stddev=0.1))
        #   3 ：              
        if regularizer!=None:
            tf.add_to_collection("losses",regularizer(fc1_weights))
        fc1_bias = tf.get_variable("bias",[fc_size],initializer=tf.constant_initializer(0.0))
        fc1 = tf.nn.relu(tf.matmul(reshaped,fc1_weights) + fc1_bias)
        if train:
            fc1 = tf.nn.dropout(fc1,0.2)
           
    with tf.variable_scope('layer6_fc2',reuse=tf.AUTO_REUSE):
        fc2_weights = tf.get_variable("weight",[fc_size,labels],initializer=tf.truncated_normal_initializer(stddev=0.1))
        #               
        if regularizer!=None:
            tf.add_to_collection('losses',regularizer(fc2_weights))
        fc2_bias = tf.get_variable("bias",[labels],initializer=tf.constant_initializer(0.0))
        
        logit = tf.matmul(fc1,fc2_weights) + fc2_bias
   #        softmax  ，          logit      ，         ，    softmax   
    return logit

epchos = 1000      #       
learning_rate = 0.001
batch_size = 128
model_save_path = "model_conv_1/model.ckpt"

def train_model(data,train=True):
    
    trX,trY,teX,teY = data
    X = tf.placeholder(tf.float32,[None,image_size,image_size,channels],name="x_input")
    Y_ = tf.placeholder(tf.float32,[None,labels],name = 'y_input')
    regularizer = tf.contrib.layers.l2_regularizer(0.0000001)
   
    y_hat = inference(X,train,regularizer)
    #     
    if train:
        #     ，logits       softmax   ，           softmax  ，labels         1（  ）
        cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits=y_hat,labels=Y_))
        loss = cross_entropy + tf.add_n(tf.get_collection("losses"))
        train_step = tf.train.GradientDescentOptimizer(learning_rate=learning_rate).minimize(loss)
        acc = tf.reduce_mean(tf.cast(tf.equal(tf.argmax(Y_,1),tf.argmax(y_hat,1)),tf.float32))
        saver = tf.train.Saver()
        
        with tf.Session() as sess:
            tf.global_variables_initializer().run()
            
            for i in range(epchos):
                total_loss = 0
                for j in range(int(trX.shape[0] / batch_size + 1)):
                    #         
                    x_batch = trX[j*batch_size:min(trX.shape[0],(j+1)*batch_size),:]
                    y_batch = trY[j*batch_size:min(trY.shape[0],(j+1)*batch_size),:]
                   
                    batch_loss,_=sess.run([cross_entropy,train_step],feed_dict={X:x_batch,Y_:y_batch})
                    total_loss+=batch_loss
                total_loss /= int(trX.shape[0] / batch_size + 1)
                
                
                test_acc = sess.run(acc,feed_dict={X:teX,Y_:teY})
                print("test acc:",test_acc)
                if i % 100 == 0:
                    saver.save(sess,model_save_path)
                    print("      ！")

train_model((train_images_exp,train_lables,test_images_exp,test_lables))

実際に本に基づいてコードを書くとき、ブロガーは4つの問題が解決されていないことに遭遇しました.まずコードの下に置いて、さらに研究した後、解答を更新します.皆さんがコメントエリアにあなたたちの見解を残してください.

stridesのパラメータのうち1,4の位置のパラメータは決定され、ここではなぜ1なのか、プール化層のksizeのような他のパラメータも同様である.

paddingの「same」は誰sameと、ボリューム層とプール化層のpaddingが具体的にどのように操作されているのか.

正規化項が全接続層にのみ加算されるのはなぜですか.他のパラメータは正規化できますか.

は、batch内とbatch間のデータの順序をどのように乱してランダム性を実現しますか?