tensorflow demo 手写数字识别

Posted 2022-01-14 _刘文凯_

记录下使用tensorflow进行多分类任务，也就是识别0-9这10个数字

环境

python3.7
tensorflow1.15.0

代码

import tensorflow as tf
import numpy as np

from sklearn import datasets

from tensorflow.examples.tutorials.mnist import input_data

#  读取拟南芥数据
def read_infile():
    mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)
    return mnist.train.images, mnist.train.labels, mnist.test.images, mnist.test.labels

#  输入、生成权重w, 偏差b
def w_biases_placeholder(n_dim, n_clasess):
    X = tf.placeholder(tf.float32, [None, n_dim])
    Y = tf.placeholder(tf.float32, [None,n_clasess])

    w = tf.Variable(tf.random_normal([n_dim, n_clasess],stddev=0.01), name= 'w')

    b = tf.Variable(tf.random_normal([n_clasess]), name='w')
    return X, Y, w, b


def forward_pass(w,b,X):
    return tf.matmul(X,w)+b


def multiclass_cost(cout,Y):
    # 计算交叉熵损失
    return tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=cout, labels=Y))

# 初始化变量
def init():
    return tf.global_variables_initializer()

#  定义参数优化方法
def train_op(learning_rate, cost):
    return tf.python.train.GradientDescentOptimizer(learning_rate).minimize(cost)

##  开始训练模型 ##
def train_model(learning_rate=0.01, epochs=1000):
    trainx, trainy, testx, testy = read_infile()
    X, Y, w, b = w_biases_placeholder(trainx.shape[1], trainy.shape[1])

    out = forward_pass(w, b, X)

    cost = multiclass_cost(out,Y)

    op_train = train_op(learning_rate, cost)

    init_ = init()
    loss_trace = []
    acc_trace = []

    with tf.Session() as sess:
        sess.run(init_)

        for i in range(epochs):
            sess.run(op_train, feed_dict=X: trainx, Y: trainy)
            loss_ = sess.run(cost, feed_dict=X: trainx, Y: trainy)
            acc_ = np.mean(np.argmax(sess.run(out, feed_dict=X: trainx, Y: trainy),axis=1) == np.argmax(trainy, axis=1))

            loss_trace.append(loss_)
            acc_trace.append(acc_)
            if (i+1)%100 == 0 and (i+1)//100 >=1:
                print('acc:',acc_)

        loss_test = sess.run(cost, feed_dict=X: testx, Y: testy)
        pred = np.argmax(sess.run(out, feed_dict=X: testx, Y: testy), axis=1)

        acc_test = np.mean(pred == np.argmax(testy, axis=1))

        print(loss_test)  #
        print(acc_test)  # 得到acc准确率

        import matplotlib.pyplot as plt
        print('\\n')
        print('True:', np.argmax(testy[0:10], axis=1))
        print('Pred:', pred[0:10])

        f, a = plt.subplots(1,10,figsize=(10, 2))
        for i in range(10):
            a[i].imshow(np.reshape(testx[i],(28,28)))


if __name__ == '__main__':
    train_model()