Керас Ручная перекрестная проверка - точность не соответствует

Question

Я бегу LSTM для классификации текста. Использование перекрестной проверки для получения ярлыков для каждого отзыва.

def k_fold_cross_validation(X, K):
    for k in range(K):
        training = [x for i, x in enumerate(X) if i % K != k]
        validation = [x for i, x in enumerate(X) if i % K == k]
        yield training, validation

data_x = k_fold_cross_validation(cleaned_reviews, K=10)
data_y = k_fold_cross_validation(ratings, K=10)

y_final = []

for i in range(10):
    print('Model', i+1 ,'of 10')
    X_train, X_test = next(data_x)
    y_train, y_test = next(data_y)

    max_features = 20000
    tokenizer = Tokenizer(num_words=max_features)
    tokenizer.fit_on_texts(list(X_train))
    list_tokenized_train = tokenizer.texts_to_sequences(X_train)
    list_tokenized_test = tokenizer.texts_to_sequences(X_test)

    maxlen = 100
    X_t = pad_sequences(list_tokenized_train, maxlen=maxlen)
    X_te = pad_sequences(list_tokenized_test, maxlen=maxlen)

    y_train = np.asarray(y_train)
    y_test = np.asarray(y_test)

    inp = Input(shape=(maxlen, ))
    embed_size = 128
    x = Embedding(max_features, embed_size)(inp)
    x = LSTM(200, return_sequences=True,name='lstm_layer')(x)
    x = GlobalMaxPool1D()(x)
    x = Dropout(0.1)(x)
    x = Dense(120, activation="relu")(x)
    x = Dropout(0.1)(x)
    x = Dense(60, activation="relu")(x)
    x = Dropout(0.1)(x)
    x = Dense(2, activation="sigmoid")(x)
    model = Model(inputs=inp, outputs=x)

    print("start compiling...")
    model.compile(loss='sparse_categorical_crossentropy', optimizer='adam', metrics=['accuracy'])

    print("start fitting...")
    model.fit(X_t,y_train, epochs=2, batch_size=32, validation_split=0.1)

    print("start evaluating")
    y_pred = model.predict(X_te, batch_size=1024)
    y_classes = y_pred.argmax(axis=-1)
    y_final += y_classes.tolist()

    accScore = metrics.accuracy_score(y_test,y_classes)

    lbl = [0,1]
    precision = metrics.precision_score(y_test,y_classes,average=None,labels=lbl)
    recall = metrics.recall_score(y_test,y_classes,average=None,labels=lbl)
    f1Score = metrics.f1_score(y_test,y_classes,average=None,labels=lbl)

    print("\nAcurracy: ",accScore,"\n")

    for i in range(len(lbl)):
        print("Precision of %s class: %f" %(lbl[i],precision[i]))
        print("Recall of %s class: %f" %(lbl[i],recall[i]))
        print("F1-Score of %s class: %f" %(lbl[i],f1Score[i]),"\n")

Все 10 моделей дают точность около 80%, но когда я пытаюсь объединить результаты всех этих моделей для проверки точности по всему набору данных, она падает до ~53%. Я думаю, что я делаю что-то не так, собирая все прогнозируемые метки в y_final, который используется для сравнения с начальными рейтингами для получения метрик производительности.

import copy

y_preds = copy.deepcopy(y_final)

accScore = metrics.accuracy_score(ratings, y_preds)

lbl = [0,1]
precision = metrics.precision_score(ratings, y_preds, average=None, labels=lbl)
recall = metrics.recall_score(ratings, y_preds, average=None, labels=lbl)
f1Score = metrics.f1_score(ratings, y_preds, average=None, labels=lbl)

print("\nOverall Acurracy: ",accScore,"\n")

for i in range(len(lbl)):
    print("Precision of %s class: %f" %(lbl[i],precision[i]))
    print("Recall of %s class: %f" %(lbl[i],recall[i]))
    print("F1-Score of %s class: %f" %(lbl[i],f1Score[i]),"\n")

Какие-либо советы, предложения или исправления?