为什么Keras的Evaluate_generator和Forecast_generator会产生不同的准确性得分？

Question

我正在尝试使用生成器在Keras中训练转换网络。但是，当我使用valuate_generator和predict_generator时，它们会产生不同的准确性得分（简单准确性，（TP + TN）/（TP + FP + TN + FN））。

我看到一些帖子说这是关于多处理的，将worker设置为1并将pickle_safe设置为True可以，但是没有。这是我的代码：

batch_size = 16
classes_list = [str(i) for i in range(num_classes)]

data_flow_train_param = {
    "dataframe": data_train,
    "x_col": "path",
    "y_col": "diagnosis",
    "classes": classes_list,
    "batch_size": batch_size,
    "target_size": IMAGE_SIZE,
    "shuffle": True
}
data_flow_train = image_generator.flow_from_dataframe(**data_flow_train_param)

data_flow_val_param = {
    "dataframe": data_val,
    "x_col": "path",
    "y_col": "diagnosis",
    "classes": classes_list,
    "batch_size": 1,
    "target_size": IMAGE_SIZE,
    "shuffle": False
}
data_flow_val = image_generator.flow_from_dataframe(**data_flow_val_param)

fit_params = {
    "generator": data_flow_train,
    "steps_per_epoch": data_train.shape[0] // batch_size,
    "epochs": 10,
    "verbose": 1,
    "validation_data": data_flow_val,
    "validation_steps": data_val.shape[0],
    "callbacks": [checkpoint, lr_decay, early_stopping]
}

print("Training the model...")
history = model.fit_generator(**fit_params)
print("Done!")

model.evaluate_generator(data_flow_val, steps=data_val.shape[0], verbose=1, pickle_safe=True)

y_val_true = np.array(data_val["diagnosis"]).astype("int")
y_val_pred = np.argmax(model.predict_generator(data_flow_val, steps=data_val.shape[0], verbose=1, pickle_safe=True), axis=-1)
print(accuracy_score(y_val_true, y_val_pred))

valuate_generator函数使我的准确度达到80％左右，而predict_generator只有30％。