tensorflow.keras：valuate_generator与predict_generator不同的结果/准确性

Question

我的代码在tensorflow.keras中。当我运行“ evaluate_generator”时，结果（准确性）与“ predict_generator”不同。 （evaluate_generator似乎提供了正确的准确性）

下面是几行代码：

model.load_weights(bestmodel_w)

val_score = model.evaluate_generator(valid_generator,valid_generator.n)
print("Validation Accuracy = ",val_score[1])

vs。

pred = model.predict_generator(test_generator,steps=STEP_SIZE_TEST,
            workers=0,
            use_multiprocessing=False,
            verbose=0)

p_labels = np.argmax(pred, axis=1)

print('Recognition Rate:', 100*float(sum(real_labels == p_labels))/float(len(p_labels)))

这是：

valid_generator = datagen_train.flow_from_directory(
    directory=train_dir,
    target_size=(img_rows, img_cols),
    color_mode="rgb",
    batch_size=batch_size,
    class_mode="categorical",
    #class_mode="binary",
    shuffle=False,
    subset='validation',
    seed=42
)

test_generator = datagen_test.flow_from_directory(
        directory=wsi_dir,
        target_size=(img_rows, img_cols),
        color_mode="rgb",
        batch_size=1,
        class_mode=None,
        #class_mode="categorical",
        shuffle=False,
        seed=42
    )