python - 在 keras.fit_generator 训练期间冻结 model.predict() 结果

Question

我试图了解 fit_generator 的工作原理，但我遇到了问题。

使用“fit”功能，我在测试模型时有这个输出：

...翻转：原始：-0.027306326664984226 - 预测：[-0.01349835]

非翻转：原始：0.02869351301342249 - 预测：[0.01401607]

翻转：原始：-0.02869351301342249 - 预测：[-0.00223063]

非翻转：原始：0.027306326664984226 - 预测：[0.01300101] ...

但是使用 fit_generator 我有这个：

...

翻转：原始：-0.01975388266146183 - 预测：[0.01830903]

非翻转：原始：0.018212568247690797 - 预测：[0.01830903]

翻转：原始：-0.018212568247690797 - 预测：[0.01830903]

...

即所有图像的结果相同（我认为这是过度拟合）

为什么？

我的代码恰到好处：

def flip_sequence(self, images, steer):

        complete_images = deque()

        for image in images:

            image = cv2.flip(image, 1)

            complete_images.append(image)

        if steer != 0.0:
           steer = -steer
        else:
           steer = 0.0

        return np.array(complete_images), steer

def EvalModel(self, model):

        for i in range(0, 5):

            data = np.load("{}/data_{}.npy".format(self._patch, i))

            random.shuffle(data)

            for image, _, speed_sequence, steer, throttle, brake in data:

                image2, steer2 = self.flip_sequence(image, steer)

                image = np.array(image).astype('float32')
                image = image / 255

                image2 = np.array(image2).astype('float32')
                image2 = image2 / 255

                image = np.expand_dims(image, axis=0)

                image2 = np.expand_dims(image2, axis=0)


                non_flipped_prediction = model.predict(image)[0][0]

                flipped_prediction = model.predict(image2)[0][0]

                print("NON-FLIPPED: ")

                print(f"ORIGINAL: {steer} - PREDICTED: {non_flipped_prediction}")

                print("FLIPPED: ")

                print(f"ORIGINAL: {steer2} - PREDICTED: {flipped_prediction}")


    def load_data(self, start, end):

      Vehicle_steers = deque()

      Vehicle_throttles = deque()

      Vehicle_brakes = deque()

      Radar_images = deque()

      Road_images = deque()


      for i in range(start, end + 1):

          index = random.randint(start, end+1)

          data = np.load("{}/data_{}.npy".format(self._patch, index))

          random.shuffle(data)

          if i % 10 == 0:
              print(f"Loading {index} - {i}")



          for image, road_image_sequence, speed_sequence, steer, throttle, brake in data:

              image = np.array(image).astype('float32')
              image = image / 255

              Radar_images.append(image)
              Vehicle_throttles.append(throttle)
              Vehicle_brakes.append(brake)
              Vehicle_steers.append(steer)

      return np.array(Radar_images), np.array(Vehicle_steers), np.array(Vehicle_brakes), np.array(Vehicle_throttles)


data_loader = DataLoader("some_patch")

RADAR_IMAGES, VEHICLE_STEERS, VEHICLE_BRAKES, VEHICLE_THROTTLES = data_loader.load_data(0, 50)

RADAR_INPUT = keras.layers.Input(shape=(5, 100, 100, 3), name='image_input')
...

...
model.fit(RADAR_IMAGES,
          [VEHICLE_STEERS, VEHICLE_BRAKES, VEHICLE_THROTTLES], epochs=5,
          callbacks=[checkpoint], batch_size=16, verbose=1, validation_split=0.025)

和 fit_generator 的代码：

...

    def load_data(self, start, end):

            Vehicle_steers = deque()

            Vehicle_throttles = deque()

            Vehicle_brakes = deque()

            Radar_images = deque()

            Road_images = deque()

            while True:

                index = random.randint(start, end+1)

                data = np.load("{}/data_{}.npy".format(self._patch, index))


                for image, road_image_sequence, speed_sequence, steer, throttle, brake in data:

                    image = np.array(image).astype('float32')
                    image = image / 255


                    Radar_images.append(image)
                    Vehicle_throttles.append(throttle)
                    Vehicle_brakes.append(brake)
                    Vehicle_steers.append(steer)

                    if np.array(Radar_images).shape[0] >= 16:

                        yield [np.array(Radar_images)], [np.array(Vehicle_steers), np.array(Vehicle_brakes), np.array(Vehicle_throttles)]

                        Vehicle_steers.clear()

                        Vehicle_throttles.clear()

                        Vehicle_brakes.clear()

                        Radar_images.clear()

                        Road_images.clear()

...

model.fit_generator(data_loader.load_data(0, 50),
                    steps_per_epoch=5000//16, epochs=5, verbose=1)

标签： pythontensorflowkerasneural-network