kfold交叉验证显示出比i exp更高的准确性

Question

我在训练面部表情识别数据（CK+） 我在Keras上使用sklearn.model\u selection中的KFold

def load_data_kfold(X, Y, k):
    X_train = X
    y_train = Y
    print("*********", y_train.shape)
    folds = list(KFold(n_splits=k, shuffle=True, random_state=1).split(X_train, y_train))
    return folds, X_train, y_train

这是为kfold加载数据的代码

    for j, (train_idx, val_idx) in enumerate(folds):
        print('\nFOld ' , j)
        X_train_cv = X_train[train_idx]
        y_train_cv = y_train[train_idx]
        X_valid_cv = X_train[val_idx]
        y_valid_cv = y_train[val_idx]

顺便说一下，我设置了10k倍验证 准确度结果是这样的

FOld  9

  Training set =  (1581, 3, 128, 128, 3)
Validation set =  (175, 3, 128, 128, 3)
Train on 1756 samples, validate on 175 samples
Epoch 1/10
 - 27s - loss: 0.0208 - acc: 0.9932 - val_loss: 0.0016 - val_acc: 1.0000

Epoch 00001: saving model to app_cnn6/model/appearance_lbp_20190829T1710/3dconv_appearance_9fold_0001.h5
Epoch 2/10
 - 27s - loss: 0.0511 - acc: 0.9943 - val_loss: 0.0486 - val_acc: 0.9943

Epoch 00002: saving model to app_cnn6/model/appearance_lbp_20190829T1710/3dconv_appearance_9fold_0002.h5
Epoch 3/10
 - 27s - loss: 0.0492 - acc: 0.9909 - val_loss: 0.0029 - val_acc: 1.0000

Epoch 00003: saving model to app_cnn6/model/appearance_lbp_20190829T1710/3dconv_appearance_9fold_0003.h5
Epoch 4/10
 - 27s - loss: 0.0269 - acc: 0.9937 - val_loss: 6.4302e-04 - val_acc: 1.0000

Epoch 00004: saving model to app_cnn6/model/appearance_lbp_20190829T1710/3dconv_appearance_9fold_0004.h5
Epoch 5/10
 - 27s - loss: 0.0366 - acc: 0.9954 - val_loss: 0.0215 - val_acc: 0.9943

Epoch 00005: saving model to app_cnn6/model/appearance_lbp_20190829T1710/3dconv_appearance_9fold_0005.h5
Epoch 6/10
 - 27s - loss: 0.0284 - acc: 0.9932 - val_loss: 0.0293 - val_acc: 0.9886

Epoch 00006: saving model to app_cnn6/model/appearance_lbp_20190829T1710/3dconv_appearance_9fold_0006.h5
Epoch 7/10
 - 26s - loss: 0.0442 - acc: 0.9915 - val_loss: 0.0025 - val_acc: 1.0000

Epoch 00007: saving model to app_cnn6/model/appearance_lbp_20190829T1710/3dconv_appearance_9fold_0007.h5
Epoch 8/10
 - 27s - loss: 0.0644 - acc: 0.9886 - val_loss: 9.6634e-04 - val_acc: 1.0000

Epoch 00008: saving model to app_cnn6/model/appearance_lbp_20190829T1710/3dconv_appearance_9fold_0008.h5
Epoch 9/10
 - 27s - loss: 0.0900 - acc: 0.9863 - val_loss: 0.0129 - val_acc: 0.9943

Epoch 00009: saving model to app_cnn6/model/appearance_lbp_20190829T1710/3dconv_appearance_9fold_0009.h5
Epoch 10/10
 - 27s - loss: 0.0382 - acc: 0.9915 - val_loss: 0.0013 - val_acc: 1.0000

Epoch 00010: saving model to app_cnn6/model/appearance_lbp_20190829T1710/3dconv_appearance_9fold_0010.h5
175/175 [==============================] - 1s 5ms/step
Test Accuracy: 100.00%
Test Loss:0.00%

包括这个折叠9， 其他折叠的精确度也比我预期的要高。 你知道这件事吗？ 有可能吗