我有20个主题，当我训练使用Tensorflow实现的模型时，我想使用留一法交叉验证。我遵循一些指示，最后是我的伪代码：
for train_index,test_index in loo.split(data):
print("TRAIN:",train_index,"TEST:",test_index)
train_X=np.concatenate(np.array([data[ii][0] for ii in train_index]))
train_y=np.concatenate(np.array([data[ii][1] for ii in train_index]))
test_X=np.concatenate(np.array([data[ii][0] for ii in test_index]))
test_y=np.concatenate(np.array([data[ii][1] for ii in test_index]))
train_X,train_y = shuffle(train_X,train_y)
test_X,test_y = shuffle(test_X,test_y)
#Compile the model
model.compile(optimizer='adam',loss='sparse_categorical_crossentropy',metrics=['accuracy'])
#keras.initializers.RandomNormal(mean=0.0,stddev=0.05,seed=None)
#Train the model
model.fit(train_X,train_y,batch_size=16,epochs=30,validation_split=.10)#,callbacks=[])
#test accuracy
test_loss,test_acc = model.evaluate(test_X,test_y)
print('\nTest accuracy:',test_acc)
但是第一个主题之后的结果是这样的：
Epoch 30/30
3590/3590 [==============================] - 4s 1ms/sample - loss: 0.5976 -
**acc: 0.8872** - val_loss: 1.3873 - val_acc: 0.6591
255/255 [==============================] - 0s 774us/sample - loss: 1.8592 -
acc: 0.4471
Test accuracy: 0.44705883
第二次迭代（主题）：
TRAIN: [ 0 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17] TEST: [1]
Train on 3582 samples,validate on 398 samples
Epoch 1/30
3582/3582 [==============================] - 5s 1ms/sample - loss: 0.7252 -
**acc: 0.8238** - val_loss: 1.0627 - val_acc: 0.6859
听起来该模型使用了先前的重量！如果我们看一下第二次迭代的第一精度，则从acc：0.8238开始！
我的实现正确吗？还是我需要更多步骤来确定每个主题的初始体重？