import tensorflow as tf
import numpy as np
import matplotlib.pylab as plt

plt.rcParams["font.family"] = 'SimHei'                          # 將字型改為中文
plt.rcParams['axes.unicode_minus'] = False                      # 設定了中文字型預設後，座標的"-"號無法顯示，設定這個引數就可以避免

# 載入手寫數字資料
mnist = tf.keras.datasets.mnist
(train_x, train_y), (test_x, test_y) 
 
= mnist.load_data()

#   將0到9轉化為one-hot編碼
y_hot = np.zeros((10, 10))
for i in range(y_hot.shape[0]):
    y_hot[i, i] = 1
# print('y_hot:', y_hot)
#   將標記值轉化為one-hot編碼
train_Y = np.zeros((train_y.shape[0], 10))
for i in range(train_y.shape[0]):
    train_Y[i] = y_hot[train_y[i]]
print('train_Y:', train_Y, train_Y.shape)

 
#   將28*28展開為784*1
#   訓練集
train_X1 = np.ones((train_x.shape[0], 784))
ones = np.ones((train_x.shape[0], 1))
print('ones.shape:', ones.shape)
for i in range(train_x.shape[0]):
    train_X1[i] = train_x[i].reshape([1, -1])
print('train_X1.shape:', train_X1.shape)
train_X = tf.concat([train_X1, ones], axis=1)
 
#   測試集
test_X1 = np.ones((test_x.shape[0], 784))
ones = np.ones((test_x.shape[0], 1))
for i in range(test_x.shape[0]):
    test_X1[i] = test_x[i].reshape([1, -1])
test_X = tf.concat([test_X1, ones], axis=1)
#   將標記資料轉化為列向量
train_y = train_y.reshape(-1,1)
test_y = test_y.reshape(-1,1)
#   儲存準確值資料
acc_train = []
acc_test = []
#   設定超引數
iter = 1500                 #   迭代次數
learn_rate = 5e-12          #   學習率
#   初始化訓練引數
w = tf.Variable(np.random.randn(785, 10)*0.0001)
print('初試w:',w,w.shape)
for i in range(iter):
    with tf.GradientTape() as tape:
        y_p = 1/(1+tf.math.exp(-tf.matmul(train_X,w)))
        y_p_test = 1 / (1 + tf.math.exp(-tf.matmul(test_X, w)))
        loss = tf.reduce_sum(-(train_Y * tf.math.log(y_p)+(1 - train_Y)*tf.math.log(1-y_p)))
        # print('loss:',loss)
    dl_dw = tape.gradient(loss,w)
    w.assign_sub(learn_rate * dl_dw)
    if i % 20 == 0:
        print('i:{}, loss:{}, w:{}'.format(i,loss,w))
        # print('y_p:',y_p)
        #   訓練集準確率
        y_p_round = tf.round(y_p)                                           #   將預測資料進行四捨五入變成one-hot編碼格式
        p_y = tf.reshape(tf.argmax(y_p_round, 1), (-1, 1))                  #   將one-hot轉化為預測數字
        is_right = tf.equal(p_y, train_y)                                   #   比對是否預測正確
        right_int = tf.cast(is_right, tf.int8)                              #   將bool型轉化為0，1
        acc = tf.reduce_mean(tf.cast(right_int, dtype=tf.float32))          #   求準確陣列的平均值，也就是準確率
        acc_train.append(acc)
        print('acc:', acc)
        #   測試集準確率
        y_p_test_round = tf.round(y_p_test)
        p_y_test = tf.reshape(tf.argmax(y_p_test_round, 1), (-1, 1))
        is_right_test = tf.equal(p_y_test, test_y)
        right_int_test = tf.cast(is_right_test, tf.int8)
        acc2 = tf.reduce_mean(tf.cast(right_int_test, dtype=tf.float32))
        acc_test.append(acc2)
        print('acc2:', acc2)
        print()

#   畫出準確率的訓練折線圖
plt.plot(acc_train,label = '訓練集正確率')
plt.plot(acc_test,label = '測試集正確率')
plt.legend()
plt.show()