即使在 Keras CNN 中使用相同的数据，验证精度也与训练精度有很大不同

Question

我正在尝试使用迁移学习来训练图像 model，我目前面临一个我不明白的问题。 为了验证我的网络，我将相同的数据推送到训练和验证，但是，我得到了非常不同的训练和验证准确性。 谁能帮我调试问题？

base_model = applications.MobileNetV2(weights='imagenet', include_top=False,
                                               input_shape=(IMAGE_LENGTH, IMAGE_WIDTH, 3))

CLASSES = 2
print('================='+str(len(base_model.layers))+"==================")
x = base_model.output
x = GlobalAveragePooling2D()(x)

x = Dense(1024, activation='relu')(x)

predictions = Dense(CLASSES, activation='softmax')(x)

model = Model(inputs=base_model.input, outputs=predictions)

for layer in model.layers[:100]:
    layer.trainable = False
for layer in model.layers[100:]:
    layer.trainable = True

model.compile(optimizer=optimizers.Adam(lr=CNN_parameters.learning_rate), loss='categorical_crossentropy', metrics=['accuracy'])
        model.compile(optimizer=optimizers.SGD(lr=0.005), loss='categorical_crossentropy', metrics=['accuracy'])
        train_datagen = ImageDataGenerator(
            #preprocessing_function=preprocess_input,
            rescale=CNN_parameters.rescale,
            #rotation_range=CNN_parameters.rotation_range,
            width_shift_range=CNN_parameters.width_shift_range,
            height_shift_range=CNN_parameters.height_shift_range,
            shear_range=CNN_parameters.shear_range,
            zoom_range=CNN_parameters.zoom_range,
            #horizontal_flip=CNN_parameters.horizontal_flip,
            fill_mode='nearest'
            )

train_generator = train_datagen.flow(
        x_train, y_train,
        batch_size=CNN_parameters.batch_size)

history = model.fit_generator(train_generator,
                                  steps_per_epoch=64,
                                  epochs=CNN_parameters.epochs,
                                  validation_data=train_generator,
                                  validation_steps=64)

但是，当我运行代码时，我得到了

64/64 [==============================] - 35s 554ms/step - loss: 0.4321 - accuracy: 0.8132 - val_loss: 0.8156 - val_accuracy: 0.5632
Epoch 2/50
64/64 [==============================] - 33s 523ms/step - loss: 0.3161 - accuracy: 0.8750 - val_loss: 0.9090 - val_accuracy: 0.5437
Epoch 3/50
64/64 [==============================] - 33s 515ms/step - loss: 0.2945 - accuracy: 0.8817 - val_loss: 0.7585 - val_accuracy: 0.5869
Epoch 4/50
64/64 [==============================] - 34s 531ms/step - loss: 0.2992 - accuracy: 0.8760 - val_loss: 0.9599 - val_accuracy: 0.5305
Epoch 5/50
64/64 [==============================] - 33s 515ms/step - loss: 0.2800 - accuracy: 0.8825 - val_loss: 0.8358 - val_accuracy: 0.5557
Epoch 6/50
64/64 [==============================] - 34s 528ms/step - loss: 0.2720 - accuracy: 0.8923 - val_loss: 0.9077 - val_accuracy: 0.5327
Epoch 7/50
64/64 [==============================] - 33s 515ms/step - loss: 0.2592 - accuracy: 0.8960 - val_loss: 0.7999 - val_accuracy: 0.5601
Epoch 8/50
64/64 [==============================] - 34s 536ms/step - loss: 0.2665 - accuracy: 0.8898 - val_loss: 0.8241 - val_accuracy: 0.5483
Epoch 9/50
64/64 [==============================] - 33s 520ms/step - loss: 0.2527 - accuracy: 0.8991 - val_loss: 0.7893 - val_accuracy: 0.5691
Epoch 10/50
64/64 [==============================] - 34s 533ms/step - loss: 0.2430 - accuracy: 0.9011 - val_loss: 0.8538 - val_accuracy: 0.5403
Epoch 11/50
64/64 [==============================] - 34s 526ms/step - loss: 0.2552 - accuracy: 0.8945 - val_loss: 0.8837 - val_accuracy: 0.5295
Epoch 12/50
64/64 [==============================] - 34s 539ms/step - loss: 0.2422 - accuracy: 0.9067 - val_loss: 0.8678 - val_accuracy: 0.5410
Epoch 13/50
64/64 [==============================] - 33s 520ms/step - loss: 0.2396 - accuracy: 0.9026 - val_loss: 0.7401 - val_accuracy: 0.5991
Epoch 14/50
64/64 [==============================] - 35s 541ms/step - loss: 0.2334 - accuracy: 0.9045 - val_loss: 0.7285 - val_accuracy: 0.5938
Epoch 15/50
64/64 [==============================] - 33s 517ms/step - loss: 0.2420 - accuracy: 0.9014 - val_loss: 0.8004 - val_accuracy: 0.5518

由于我使用相同的数据进行训练和验证，我希望在训练和验证上看到非常相似的准确性。 我不明白为什么他们如此不同。 我做错什么了吗？ 谢谢