验证准确性和预测准确性不同

Question

我正在执行图像分类任务，以对眼底图像的可分级性进行3类分类：好，可用和不良。问题是在训练过程中，验证准确度在最后一个时期显示为0.8115，但是当我尝试使用model.predict时，即使在如下图所示的相同验证数据集上，结果也很差

这是我编写的使用ImageGenerator加载图像的代码

# Generator
train_datagen = ImageDataGenerator(preprocessing_function=preprocess_input, validation_split=0.2)
# Load image
size_expected = (224,224)
train_generator = train_datagen.flow_from_directory('./train_gradability',
                                                batch_size = 100,
                                                shuffle=True,
                                                class_mode = 'categorical',
                                                target_size = size_expected,
                                                subset = 'training',
                                                seed = 42)
 val_generator = train_datagen.flow_from_directory('./train_gradability',
                                                batch_size = 100,
                                                class_mode = 'categorical',
                                                target_size = size_expected,
                                                subset = 'validation',
                                                seed = 42)

用于创建模型的代码，我使用了InceptionResNetV2架构：

incp_resnet = InceptionResNetV2(input_shape = (size_expected[0], 
size_expected[1], 3),
          include_top = False,
          weights='imagenet')
# Freeze layers
for layer in incp_resnet.layers:
    layer.trainable = False

# Add customize layers
model = tf.keras.models.Sequential([
        incp_resnet,
        Flatten(),
        # First hidden layer
        Dense(4096, activation='relu'),
        Dropout(0.5),
        # Second hidden layer
        Dense(2048, activation='relu'),
        Dropout(0.5),
        # Output layer
        Dense(3, activation = 'softmax')
])
# Training
model.compile(optimizer = 'Adam', loss = 'categorical_crossentropy', metrics= 
['categorical_accuracy'])
history = model.fit(train_generator, 
                epochs=1,
                steps_per_epoch = (train_generator.samples//100+1),
                validation_data = val_generator,
                validation_steps = (val_generator.samples//100+1),
                verbose = 1)

预测部分：

pred = model.predict(val_generator, steps=(val_generator.samples//100 + 1), verbose =1)
pred_classes = np.argmax(pred, axis = 1)
ground_truth = val_generator.classes
print(confusion_matrix(ground_truth, pred_classes))
print(classification_report(ground_truth, pred_classes))

我使用Tensorflow 2.1.0版和Keras 2.3.1版