Python 使用和不使用数据生成器获得不同的训练损失

我一直在训练keras模型来检测面部标志。数据集大约为500MB，包含64x64个图像

我将数据生成器类实现为，而不是将整个数据集加载到内存中。我发现我的损失（SmoothL1）减少到5.9，然后停止进一步减少。通过测试36幅图像进行预测，该模型每次都预测出完全相同的地标，如下所示，这很奇怪。

我的数据生成器类别如下：

导入tensorflow作为tf 将numpy作为np导入 进口cv2

class Data_Generator(tf.keras.utils.Sequence):
    def __init__(self, data_csv, data_path, dim=(224, 224), grey_scale=False, batch_size=16, preprocess=None, shuffle=False, info={}):
        self.data_csv = data_csv
        self.PATH = data_path
        self.batch_size = batch_size
        self.preprocess = preprocess
        self.shuffle = shuffle
        self.info = info
        self.w = dim[0]
        self.h = dim[1]
        if grey_scale:
          self.channels = 1
        else:
          self.channels = 3
        self.on_epoch_end()

    def __len__(self):
        return int(np.floor(len(self.data_csv) / self.batch_size))

    def on_epoch_end(self):
        self.indexes = np.arange(len(self.data_csv))
        if self.shuffle:
            np.random.shuffle(self.indexes)

    def preprocess(self, img):
        # preprocessing goes here
        #  ...
        return img

    def __getitem__(self, idx):
        indexes = self.indexes[idx * self.batch_size: (idx + 1) * self.batch_size]
        data_temp = self.data_csv.iloc[indexes, :]

        imgs_batch = np.empty((self.batch_size, self.h, self.w, self.channels), dtype=np.float32)
        points_batch = np.empty((self.batch_size, 40))

        for index in range(len(data_temp)):
            img_path = df.iloc[index, :]["image_name"]
            self.info[index * self.batch_size + index] = img_path

            points = df.iloc[index, :]["landmarks"]
            points = np.fromstring(points, dtype=float, sep=' ').flatten()

            # only selecting mounth (upper and lower lip)
            points = points[96:]

            if self.channels == 1:
              img = cv2.imread(os.path.join(self.PATH, img_path), cv2.IMREAD_GRAYSCALE)
              img = cv2.resize(img, (self.w, self.h), interpolation=cv2.INTER_AREA)              
              img = np.expand_dims(img, axis=-1)
            else:
              img = cv2.imread(os.path.join(self.PATH, img_path))
              img = cv2.resize(img, (self.w, self.h), interpolation=cv2.INTER_AREA)
            
            imgs_batch[index, ] = img
            points_batch[index, ] = points

        if self.preprocess:
            imgs_batch = self.preprocess(imgs_batch)
        return imgs_batch, points_batch

train_data_generator = Data_Generator(train, data_path=PATH_TO_JPEGIMAGES, dim=input_shape, grey_scale=True, batch_size=64, shuffle=True)
validation_data_generator = Data_Generator(test, data_path=PATH_TO_JPEGIMAGES, grey_scale=True, dim=input_shape, batch_size=64)

model.fit(train_data_generator, 
          epochs=200,
          verbose=1,
          validation_data=validation_data_generator, 
          callbacks=[reduce_lr_loss, cb_checkpt, tensorboard_callback]
)

然后，我进行了第二次训练，但这次我将整个数据集加载到内存中，通过非常好的预测，我的损失减少到了0.0173，这非常棒

现在我的问题是，为什么使用data generator进行的培训显示出相同超参数和相同数据集的奇怪结果？我的数据生成器类有什么问题吗？我想知道，因为加载整个数据集当然对大型数据集不起作用。请帮助任何人