深度學習解決手寫數字的圖片識別
- 2019 年 11 月 7 日
- 筆記
本篇使用TensorFlow框架，利用MNIST手寫數字數據集來演示深度學習的入門概念。其訓練集共有60000個樣本（圖片和標籤），測試集有10000個樣本。手寫數字的圖片都是尺寸為28*28的二值圖：
我們先導入必要的庫：
import tensorflow as tf import tensorflow.examples.tutorials.mnist.input_data as input_data import os
設置全連接神經網絡的參數：神經網絡的結構為784*500*10 （輸入層784節點，1層500個節點的隱藏層，除輸出層外每層的激活函數都使用ReLU, 輸出層10個節點, 最後使用tf.argmax()函數求出輸出層節點中最大的數的索引，範圍0~9，該索引值即為手寫數字的估計值）
註：上述圖片僅做示意，每層節點數，以及隱藏層的層數以代碼為準
#模型路徑 MODEL_SAVE_PATH ="/model_path/" MODEL_NAME = "MNIST_model1.ckpt" INPUT_NODE = 28*28 #圖片28*28像素，展平為784=28*28個輸入節點 OUTPUT_NODE = 10 #輸出特徵為10個，對應0~9的量 BATCH_SIZE =100 # 訓練批次的size LEARNING_RATE_BASE = 0.8 #基礎學習率 LEARNING_RATE_DECAY = 0.99 #學習率縮減係數 REGULARIZATION_RATE = 0.0001 # 正則率 TRAINING_STEPS = 30000 #總的訓練步數 MOVING_AVERAGE_DECAY = 0.99 #移動平均縮減係數 #神經網絡的結構，784*500*10 （輸入層784節點，1層500個節點的隱藏層，輸出層10個節點） layer_dimension = [INPUT_NODE,500,OUTPUT_NODE] #也可以是多個隱藏層，如 layer_dimension = [INPUT_NODE,50,100，20，UTPUT_NODE] n_layers = len(layer_dimension) #神經網絡總的層數
前向傳播：
def inference(input_x, avg_class, reuse = True): '''Forward propagation''' current_layer = input_x in_dimension = INPUT_NODE with tf.variable_scope("layers", reuse =reuse): for i in range(1, n_layers):#循環 out_dimension = layer_dimension[i] # weight, bias can't be local variables, or will not be updated!!!!!!!!! weight = tf.get_variable("weight_"+str(i), [in_dimension, out_dimension], initializer = tf.truncated_normal_initializer(stddev = 0.1)) tf.add_to_collection('losses', tf.contrib.layers.l2_regularizer(REGULARIZATION_RATE)(weight))# L2 regularization bias = tf.get_variable("bias_"+str(i), [out_dimension], initializer = tf.constant_initializer(0.0)) if avg_class == None: if i == n_layers - 1: current_layer = tf.matmul(current_layer, weight) + bias else: current_layer = tf.nn.relu(tf.matmul(current_layer, weight) +bias) else: if i == n_layers - 1: current_layer = tf.matmul(current_layer, avg_class.average(weight)) + avg_class.average(bias) else: current_layer = tf.nn.relu(tf.matmul(current_layer, avg_class.average(weight)) + avg_class.average(bias)) in_dimension = out_dimension return current_layer # output
神經網絡訓練，即反向傳播，以梯度下降算法優化各個權重W張量和各個偏置B張量。
def train(mnist): '''training''' x = tf.placeholder(tf.float32, [None, INPUT_NODE], name='x-input') y_ = tf.placeholder(tf.float32,[None, OUTPUT_NODE], name = 'y-input') y = inference(x, None, reuse = False) global_step = tf.Variable(0, trainable = False) #滑動平均模型 variable_averages = tf.train.ExponentialMovingAverage(MOVING_AVERAGE_DECAY, global_step) variables_averages_op = variable_averages.apply(tf.trainable_variables()) # moving average applied average_y = inference(x, variable_averages, reuse = True) # loss cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(logits = y, labels = tf.argmax(y_, 1)) cross_entropy_mean = tf.reduce_mean(cross_entropy) tf.add_to_collection('losses', cross_entropy_mean) loss = tf.add_n(tf.get_collection('losses')) #loss = cross_entropy_mean #learning rate with decay learning_rate = tf.train.exponential_decay(LEARNING_RATE_BASE, global_step,mnist.train.num_examples / BATCH_SIZE, LEARNING_RATE_DECAY, staircase = True) #learning_rate = 0.01 train_step = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss, global_step = global_step) train_op = tf.group(train_step, variables_averages_op) correct_prediction = tf.equal(tf.argmax(average_y, 1), tf.argmax(y_, 1)) accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))#tf.cast做類型轉換 with tf.Session() as sess: tf.global_variables_initializer().run() validate_feed = {x: mnist.validation.images, y_ : mnist.validation.labels} test_feed = {x: mnist.test.images, y_ : mnist.test.labels} steps = [] # only for plot accs = [] # only for plot losses = [] # only for plot for i in range(TRAINING_STEPS): xs, ys = mnist.train.next_batch(BATCH_SIZE) _, loss_value, step = sess.run([train_op, loss, global_step], feed_dict = {x : xs, y_: ys}) if i % 200 == 0: validate_acc = sess.run(accuracy, feed_dict = validate_feed) steps.append(step); accs.append(validate_acc*100); losses.append(loss_value) # only for plot #print("After %d training steps, validation accuracy using average model on training batch is %g%%, loss on training batch is %g"%(step, validate_acc*100,loss_value)) #saver.save(sess, os.path.join(MODEL_SAVE_PATH, MODEL_NAME), global_step =global_step) test_acc = sess.run(accuracy, feed_dict = test_feed) print("After %d training steps, test accuracy using average model is %g%%"% (TRAINING_STEPS, test_acc*100)) #僅用於tensorboard writer = tf.summary.FileWriter("E://TensorBoard//test",sess.graph) #保存神經網絡模型 saver = tf.train.Saver() saver.save(sess, r"E:Python36my tensorflowckpt filesmode_mnist.ckpt") #only for plot#不是必要步驟 from matplotlib import pyplot as plt import matplotlib.ticker as mtick plt.subplot(211) plt.plot(steps, losses,color="red") plt.scatter(steps, losses,s=20,color="red") plt.xlabel("step"); plt.ylabel("loss(including L2 regularization loss) on training set") plt.subplot(212) plt.plot(steps, accs,color="green") plt.scatter(steps, accs,s=20,color="green") yticks = mtick.FormatStrFormatter("%.3f%%") plt.gca().yaxis.set_major_formatter(yticks) plt.xlabel("step"); plt.ylabel("accuracy on training set") plt.show()
真正加載MNIST數據集，並訓練模型：
def main(argv = None): mnist = input_data.read_data_sets(r"E:Python36my tensorflowMNIST_data",one_hot =True) train(mnist) if __name__ == "__main__": tf.app.run()
可以看出5000步以後，模型已大致收斂：
30000步迭代之後，在測試集上的準確率已高達98.5%。
After 30000 training steps, test accuracy using average model is 98.5%
下面我們利用已訓練好的模型做預測：
import tensorflow as tf import tensorflow.examples.tutorials.mnist.input_data as input_data import matplotlib.pyplot as plt INPUT_NODE = 28*28 OUTPUT_NODE = 10 layer_dimension = [INPUT_NODE,500,OUTPUT_NODE] n_layers = len(layer_dimension) def inference(input_x, avg_class, reuse = True): current_layer = input_x in_dimension = INPUT_NODE with tf.variable_scope("layers", reuse =reuse): for i in range(1, n_layers): out_dimension = layer_dimension[i] # weight, bias can't be local variables, or will not be updated!!!!!!!!! w = tf.get_variable("weight_"+str(i), [in_dimension, out_dimension]) b = tf.get_variable("bias_"+str(i), [out_dimension]) if avg_class == None: if i == n_layers - 1: current_layer = tf.matmul(current_layer, w) + b else: current_layer = tf.nn.relu(tf.matmul(current_layer, w) +b) in_dimension = out_dimension return current_layer # output with tf.variable_scope("layers", reuse =tf.AUTO_REUSE): ####!!!!!!!! weight1 = tf.get_variable("weight_1", [INPUT_NODE, 500]) bias1= tf.get_variable("bias_1", [500]) weight2 = tf.get_variable("weight_2", [500, OUTPUT_NODE]) bias2= tf.get_variable("bias_2", [OUTPUT_NODE]) #saver = tf.train.import_meta_graph(r"E:Python36my tensorflowckpt filesmode_mnist.ckpt.meta") saver = tf.train.Saver() with tf.Session() as sess: saver.restore(sess, r"E:Python36my tensorflowckpt filesmode_mnist.ckpt") #graph = tf.get_default_graph() #for v in tf.global_variables(): #print(v.name) #print(v.eval()) image_rawdata = tf.gfile.FastGFile(r"E:Python36MNIST picturetest54.jpg","rb").read() img_data = tf.image.decode_jpeg(image_rawdata) if img_data.dtype != tf.float32: img_data = tf.image.convert_image_dtype(img_data, dtype = tf.float32) image_data_shaped2 = tf.reshape(img_data,(1,INPUT_NODE))# tftensor input_x = image_data_shaped2 y = inference(input_x, avg_class=None, reuse = True) print("y: ", sess.run(y)) print( "predict number: ", sess.run(tf.argmax(y, 1)))#output label image_data = img_data.eval() # return a numpy array image_data_shaped1 = image_data.reshape(image_data.shape[0],image_data.shape[1])#numpy array #print(image_data_shaped1) plt.imshow(image_data_shaped1,cmap='gray') plt.show()
注意，要保持神經網絡的結構和訓練時的一致。我們加載已訓練好的模型，用它來預測測試集中的第54張圖：
輸出結果是：
predict number: [6]
正確！
如果想要預測我們自己拍的照片，記得須先將照片轉化為28*28的二值圖, 用openCV實現起來很簡單，不再贅述。
