Python 为什么我的keras计算精度为80%，而手动计算的精度为50%？

我的模型有问题。 在训练并保存之后，我加载了我的模型并尝试预测图像。 为了测试它，我试着用我用来训练模型的相同图像

from tensorflow.keras.models import load_model
from keras.preprocessing.image import ImageDataGenerator
import csv

model = load_model('newmodel.h5')

#training and validation dataset
train_dir = "ROIClassifier/data/train/"
train_image_generator = ImageDataGenerator(rescale=1./255) # Generator for our training data
train_dataset = train_image_generator.flow_from_directory(directory=train_dir,
                                                          target_size= (128,128),
                                                          classes=['tumor', 'stroma'],
                                                          class_mode='binary',
                                                          )

validation_image_generator = ImageDataGenerator(rescale=1./255,
                                                )  # Generator for our training data
test_dir = "ROIClassifier/data/validation/"
validation_dataset = validation_image_generator.flow_from_directory(directory=test_dir,
                                                                    target_size= (128,128),
                                                                    classes=['tumor', 'stroma'],
                                                                    class_mode='binary')

results = model.predict_classes(train_dataset,  batch_size=None)
evaluation = model.evaluate(train_dataset)
print(train_dataset.class_indices)
# name of csv file
filename = "results.csv"
# field names
fields = ['ID', 'Class', 'Prediction']

with open(filename, 'w', newline='') as csvfile:
    # creating a csv writer object
    csvwriter = csv.writer(csvfile)
    # writing the fields
    csvwriter.writerow(fields)
    a = 0
    for i in range(0, 293):
        if train_dataset.classes[i]== int(*results[i]):
            a += 1
        csvwriter.writerow([i, train_dataset.classes[i], *results[i]])

我加载模型，生成数据集，预测类并评估模型。最后，我保存了一个csv（用SPSS进行一些统计）

模型评估：

1/10 [==>...........................] - ETA: 2s - loss: 0.6935 - accuracy: 0.6562
 2/10 [=====>........................] - ETA: 1s - loss: 0.6839 - accuracy: 0.6875
 3/10 [========>.....................] - ETA: 1s - loss: 0.6657 - accuracy: 0.7188
 4/10 [===========>..................] - ETA: 1s - loss: 0.6388 - accuracy: 0.7812
 5/10 [==============>...............] - ETA: 0s - loss: 0.6361 - accuracy: 0.7812
 6/10 [=================>............] - ETA: 0s - loss: 0.6315 - accuracy: 0.7865
 7/10 [====================>.........] - ETA: 0s - loss: 0.6202 - accuracy: 0.7991
 8/10 [=======================>......] - ETA: 0s - loss: 0.6241 - accuracy: 0.7891
 9/10 [==========================>...] - ETA: 0s - loss: 0.6256 - accuracy: 0.7951
10/10 [==============================] - 1s 147ms/step - loss: 0.6205 - accuracy: 0.7959

但只有50%的预测符合实际的类别。

---

为什么？

我在代码中找不到存储模型的步骤，您可以看到如何在这个线程中完成。这需要小心完成，否则您可能不会保存最终权重，并且当您加载模型时，它将作为未经培训的对象加载。我在模型拟合之后使用了“model.save（'newmodel.h5'）。实际上我编写了history=model.fit（…），这可能是问题吗？我发现了问题：我必须在来自\目录的flow\中使用“shuffle=off”