使用Pimia印第安人糖尿病数据集，我构建了以下顺序模型：
import matplotlib.pyplot as pltimport numpyfrom keras import callbacksfrom keras import optimizersfrom keras.layers import Densefrom keras.models import Sequentialfrom keras.callbacks import ModelCheckpointfrom sklearn.preprocessing import StandardScaler#TensorBoard callback for visualization of training historytb = callbacks.TensorBoard(log_dir='./logs/latest', histogram_freq=10, batch_size=32, write_graph=True, write_grads=True, write_images=False, embeddings_freq=0, embeddings_layer_names=None, embeddings_metadata=None)# Early stopping - Stop training before overfittingearly_stop = callbacks.EarlyStopping(monitor='val_loss', min_delta=0, patience=3, verbose=1, mode='auto')# fix random seed for reproducibilityseed = 42numpy.random.seed(seed)# load pima indians datasetdataset = numpy.loadtxt("pima-indians-diabetes.csv", delimiter=",")# split into input (X) and output (Y) variablesX = dataset[:, 0:8]Y = dataset[:, 8]# Standardize features by removing the mean and scaling to unit variancescaler = StandardScaler()X = scaler.fit_transform(X)#ADAM Optimizer with learning rate decayopt = optimizers.Adam(lr=0.0001, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.0001)## Create our modelmodel = Sequential()model.add(Dense(12, input_dim=8, kernel_initializer='uniform', activation='relu'))model.add(Dense(8, kernel_initializer='uniform', activation='relu'))model.add(Dense(1, kernel_initializer='uniform', activation='sigmoid'))# Compile the model using binary crossentropy since we are predicting 0/1model.compile(loss='binary_crossentropy', optimizer=opt, metrics=['accuracy'])# checkpointfilepath="./checkpoints/weights.best.hdf5"checkpoint = ModelCheckpoint(filepath, monitor='val_acc', verbose=1, save_best_only=True, mode='max')# Fit the modelhistory = model.fit(X, Y, validation_split=0.33, epochs=10000, batch_size=10, verbose=0, callbacks=[tb,early_stop,checkpoint])# list all data in historyprint(history.history.keys())# summarize history for accuracyplt.plot(history.history['acc'])plt.plot(history.history['val_acc'])plt.title('model accuracy')plt.ylabel('accuracy')plt.xlabel('epoch')plt.legend(['train', 'test'], loc='upper left')plt.show()# summarize history for lossplt.plot(history.history['loss'])plt.plot(history.history['val_loss'])plt.title('model loss')plt.ylabel('loss')plt.xlabel('epoch')plt.legend(['train', 'test'], loc='upper left')plt.show()我添加了提前停止、检查点和Tensorboard回调，得到以下结果：
Epoch 00000: val_acc improved from -inf to 0.67323, saving model to ./checkpoints/weights.best.hdf5Epoch 00001: val_acc did not improve...Epoch 00024: val_acc improved from 0.67323 to 0.67323, saving model to ./checkpoints/weights.best.hdf5...Epoch 00036: val_acc improved from 0.76378 to 0.76378, saving model to ./checkpoints/weights.best.hdf5...Epoch 00044: val_acc improved from 0.79921 to 0.80709, saving model to ./checkpoints/weights.best.hdf5...Epoch 00050: val_acc improved from 0.80709 to 0.80709, saving model to ./checkpoints/weights.best.hdf5...Epoch 00053: val_acc improved from 0.80709 to 0.81102, saving model to ./checkpoints/weights.best.hdf5...Epoch 00257: val_acc improved from 0.81102 to 0.81102, saving model to ./checkpoints/weights.best.hdf5...Epoch 00297: val_acc improved from 0.81102 to 0.81496, saving model to ./checkpoints/weights.best.hdf5Epoch 00298: val_acc did not improveEpoch 00299: val_acc did not improveEpoch 00300: val_acc did not improveEpoch 00301: val_acc did not improveEpoch 00302: val_acc did not improveEpoch 00302: early stopping根据日志，我的模型准确率是0.81496。奇怪的是，验证准确率高于训练准确率（0.81 vs 0.76），验证损失低于训练损失（0.41 vs 0.47）。
Q: 我遗漏了什么？我需要在代码中做哪些更改来解决这个问题？
回答：
如果你对数据进行洗牌，这个问题就解决了。
import matplotlib.pyplot as pltimport numpyfrom keras import callbacksfrom keras import optimizersfrom keras.layers import Densefrom keras.models import Sequentialfrom keras.callbacks import ModelCheckpointfrom sklearn.preprocessing import StandardScalerfrom sklearn.utils import shuffle# TensorBoard callback for visualization of training historytb = callbacks.TensorBoard(log_dir='./logs/4', histogram_freq=10, batch_size=32, write_graph=True, write_grads=True, write_images=False, embeddings_freq=0, embeddings_layer_names=None, embeddings_metadata=None)# Early stopping - Stop training before overfittingearly_stop = callbacks.EarlyStopping(monitor='val_loss', min_delta=0, patience=3, verbose=1, mode='auto')# fix random seed for reproducibilityseed = 42numpy.random.seed(seed)# load pima indians datasetdataset = numpy.loadtxt("../Downloads/pima-indians-diabetes.csv", delimiter=",")# split into input (X) and output (Y) variablesX = dataset[:, 0:8]Y = dataset[:, 8]# Standardize features by removing the mean and scaling to unit variancescaler = StandardScaler()X = scaler.fit_transform(X)# This is the important partX, Y = shuffle(X, Y)#ADAM Optimizer with learning rate decayopt = optimizers.Adam(lr=0.0001, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.0001)## Create our modelmodel = Sequential()model.add(Dense(12, input_dim=8, kernel_initializer='uniform', activation='relu'))model.add(Dense(8, kernel_initializer='uniform', activation='relu'))model.add(Dense(1, kernel_initializer='uniform', activation='sigmoid'))# Compile the model using binary crossentropy since we are predicting 0/1model.compile(loss='binary_crossentropy', optimizer=opt, metrics=['accuracy'])# checkpoint# filepath="./checkpoints/weights.best.hdf5"# checkpoint = ModelCheckpoint(filepath, monitor='val_acc', verbose=1, save_best_only=True, mode='max')# Fit the modelhistory = model.fit(X, Y, validation_split=0.33, epochs=1000, batch_size=10, verbose=0, callbacks=[tb,early_stop])# list all data in historyprint(history.history.keys())# summarize history for accuracyplt.plot(history.history['acc'])plt.plot(history.history['val_acc'])plt.title('model accuracy')plt.ylabel('accuracy')plt.xlabel('epoch')plt.legend(['train', 'test'], loc='upper left')plt.show()# summarize history for lossplt.plot(history.history['loss'])plt.plot(history.history['val_loss'])plt.title('model loss')plt.ylabel('loss')plt.xlabel('epoch')plt.legend(['train', 'test'], loc='upper left')plt.show()