我建议在使用 Keras 之前运行一个脚本，以便将您的数据清理为更标准的格式，例如.h5. 然后，您可以将此数据分批成具有键 ID 中给定索引的批次。然后DataGenerator在 Keras 中编写变得非常容易。

我这样做的方法是使用以下类型的键：

dataset_identifier_ix

其中 dataset 是train、validation或test，具体取决于您希望使用的集合。在您的情况下，标识符将是image和label。最后，ix是批次索引。在每个批次中，您都有$n$用于训练模型的实例。

如果这是你的结构.h5，那么DataGenerator代码可以像

class DataGenerator(keras.utils.Sequence):

    'Generates data for Keras'
    def __init__(self, dataset, filename, X_identifier, Y_identifier, 
                 batch_size, percent_data_use = 1, shuffle=True):
        super(DataGenerator,self).__init__()

        with h5py.File(filename+'.h5', 'r') as hf:
            keys = list(hf.keys())
            # Get dimensions of the input space
            temp = [i for i in keys if dataset+'_'+X_identifier in i]
            x_dims = hf[temp[0]].shape[1::]
            num_files = len(temp)
            # Get dimensions of the output space
            temp = [i for i in keys if dataset+'_'+Y_identifier in i]
            y_dims = hf[temp[0]].shape[1::]

        # The batch size
        self.batch_size = batch_size
        self.num_files = num_files
        # Assumes the file ids are always from 0 to num_files
        self.file_ids = list(range(int(percent_data_use*num_files)))
        # Calculate the number of batches
        self.num_batches = int(percent_data_use*num_files)

        self.filename = filename
        self.dataset = dataset 
        self.X_identifier = X_identifier
        self.Y_identifier = Y_identifier

        # Dimensions of the input and the output
        self.x_dims = x_dims
        self.y_dims = y_dims
        self.input_shape = (int(x_dims[0]*sampling_percent),) + x_dims[1:]

        self.sampling_percent = sampling_percent
        self.shuffle = shuffle
        if shuffle is True: self.on_epoch_end()

    def __len__(self):
        """Number of batch in the Sequence.

        Returns
            The number of batches in the Sequence.
        """
        return self.num_batches

    def __getitem__(self, index):
        """Gets batch at position `index`.
        Arguments
            index: position of the batch in the Sequence.
        Returns
            A batch
        """        
        X = np.zeros((self.batch_size,) + self.x_dims)
        Y = np.zeros((self.batch_size,) + self.y_dims)

        with h5py.File(self.filename+'.h5', 'r') as hf:

            # Input and output identifiers
            x_id = self.dataset + '_' + \
                    self.X_identifier + '_' + str(self.file_ids[index])
            y_id = self.dataset + '_' + \
                    self.Y_identifier + '_' + str(self.file_ids[index])

            X = np.asarray(hf[x_id])
            Y = np.asarray(hf[y_id]) 

        if len(X) == 0:
            return None

        ## Restructure X, and Y for use in the Keras network
        return X, Y

    def on_epoch_end(self):
        'Updates indexes after each epoch'
        if self.shuffle == True:
            np.random.shuffle(self.file_ids)