tensorflow2.0的手写数字识别

import tensorflow as tf import tensorflow.keras as keras import tensorflow.keras.layers as layers from tensorflow.keras import Sequential import matplotlib.pyplot as plt import pandas as pd from tensorflow.keras.utils import normalize from tensorflow.keras.utils import to_categorical import tensorflow.keras.optimizers as optimizers import tensorflow.keras.losses as losses import tensorflow.keras.metrics as metrics 导入数据 (x_train, y_train),(x_test, y_test) = tf.keras.datasets.mnist.load_data() x1_train = x_train.reshape(-1, 784) #改变形状，变一维 x1_test = x_test.reshape(-1, 784) x_train_nor = normalize(x1_train, axis=1)#归一化 x_test_nor = normalize(x1_test, axis=1) y_train_onehot = to_categorical(y_train)#标签转为onehot类型 y_test_onthot = to_categorical(y_test) 创建网络 network = Sequential() network.add(layers.Dense(units=256, kernel_initializer='normal', activation='relu')) network.add(layers.Dropout(0.5)) network.add(layers.Dense(units=10,kernel_initializer='normal',activation='softmax')) network.build(input_shape=(4,784))#此处的build是为了创建参数，所以需要设置 查看网络： network.summary()

训练参数设置以及训练 #设置优化方法， 损失函数， 以及评价指标 network.compile(optimizer=optimizers.Adam(), loss=losses.CategoricalCrossentropy(), metrics=['accuracy']) #设置训练数据以及标签， 验证集占比重， 循环次数， 批大小， 日志级别 train_history = network.fit(x=x_train_nor, y=y_train_onehot, validation_split=0.2, epochs=5, batch_size=200, verbose=2)

其中verbose是日志等级： verbose：日志显示 verbose = 0 为不在标准输出流输出日志信息 verbose = 1 为输出进度条记录 verbose = 2 为每个epoch输出一行记录 注意： 默认为 1

Train on 48000 samples, validate on 12000 samples Epoch 1/5 48000/48000 - 4s - loss: 0.9353 - accuracy: 0.7783 - val_loss: 0.3760 - val_accuracy: 0.9023 Epoch 2/5 48000/48000 - 1s - loss: 0.3890 - accuracy: 0.8895 - val_loss: 0.2828 - val_accuracy: 0.9208 Epoch 3/5 48000/48000 - 1s - loss: 0.3156 - accuracy: 0.9080 - val_loss: 0.2428 - val_accuracy: 0.9312 Epoch 4/5 48000/48000 - 1s - loss: 0.2757 - accuracy: 0.9200 - val_loss: 0.2166 - val_accuracy: 0.9382 Epoch 5/5 48000/48000 - 1s - loss: 0.2461 - accuracy: 0.9281 - val_loss: 0.1941 - val_accuracy: 0.9457

查看训练过程中相关指标的值：可用于后期画图使用 train_history.history

{‘loss’: [0.9352779900034268, 0.3889504156385859, 0.3156389035284519, 0.27566256585220494, 0.24609261999527612], ‘accuracy’: [0.7783125, 0.8895208, 0.9079792, 0.91997916, 0.928125], ‘val_loss’: [0.37601099784175557, 0.28281597085297105, 0.24280174300074578, 0.21655768944571416, 0.19407257350782553], ‘val_accuracy’: [0.9023333, 0.92083335, 0.93116665, 0.9381667, 0.9456667]}

画图： def show_train_history(history, train, validation): plt.plot(history.history[train]) plt.plot(history.history[validation]) plt.title('Train History') plt.xlabel('epoch') plt.ylabel('accuracy') plt.legend(['validation','train'],loc='upper left')

网络评估： scores = network.evaluate(x_test_nor, y_test_onthot) 类别预测 prediction = network.predict_classes(x_test_nor) 查看分类结果 分类表： pd.crosstab(y_test, prediction, rownames=['label'], colnames=['prediction'])

预测错误的样本位置：

df = pd.DataFrame({'label':y_test,'predict':prediction}) print(df[(df.label==5)&(df.predict==3)]) #查看是5预测错成3的数据位置