如何使用 tf.data.Dataset.from_generator() 向生成器函数发送参数？

Question

我想使用from_generator() 函数创建多个tf.data.Dataset。我想向生成器函数 (raw_data_gen) 发送一个参数。这个想法是生成器函数将根据发送的参数产生不同的数据。通过这种方式，我希望raw_data_gen 能够提供训练、验证或测试数据。

training_dataset = tf.data.Dataset.from_generator(raw_data_gen, (tf.float32, tf.uint8), ([None, 1], [None]), args=([1]))

validation_dataset = tf.data.Dataset.from_generator(raw_data_gen, (tf.float32, tf.uint8), ([None, 1], [None]), args=([2]))

test_dataset = tf.data.Dataset.from_generator(raw_data_gen, (tf.float32, tf.uint8), ([None, 1], [None]), args=([3]))

当我尝试以这种方式调用from_generator() 时收到的错误消息是：

TypeError: from_generator() got an unexpected keyword argument 'args'

这里是raw_data_gen 函数，虽然我不确定你是否需要这个函数，因为我的直觉是问题出在from_generator() 的调用上：

def raw_data_gen(train_val_or_test):

    if train_val_or_test == 1:        
        #For every filename collected in the list
        for filename, lab in training_filepath_label_dict.items():
            raw_data, samplerate = soundfile.read(filename)
            try: #assume the audio is stereo, ready to be sliced
                raw_data = raw_data[:,0] #raw_data is a np.array, just take first channel with slice
            except IndexError:
                pass #this must be mono audio
            yield raw_data, lab

    elif train_val_or_test == 2:
        #For every filename collected in the list
        for filename, lab in validation_filepath_label_dict.items():
            raw_data, samplerate = soundfile.read(filename)
            try: #assume the audio is stereo, ready to be sliced
                raw_data = raw_data[:,0] #raw_data is a np.array, just take first channel with slice
            except IndexError:
                pass #this must be mono audio
            yield raw_data, lab

    elif train_val_or_test == 3:
        #For every filename collected in the list
        for filename, lab in test_filepath_label_dict.items():
            raw_data, samplerate = soundfile.read(filename)
            try: #assume the audio is stereo, ready to be sliced
                raw_data = raw_data[:,0] #raw_data is a np.array, just take first channel with slice
            except IndexError:
                pass #this must be mono audio
            yield raw_data, lab

    else:
        print("generator function called with an argument not in [1, 2, 3]")
        raise ValueError()

Answer 1

您需要基于raw_data_gen 定义一个不带任何参数的新函数。您可以使用lambda 关键字来执行此操作。

training_dataset = tf.data.Dataset.from_generator(lambda: raw_data_gen(train_val_or_test=1), (tf.float32, tf.uint8), ([None, 1], [None]))
...

现在，我们将一个不带任何参数的函数传递给from_generator，但它只是充当raw_data_gen，参数设置为1。您可以对验证集和测试集使用相同的方案，分别通过 2 和 3。

Answer 2

对于 TensorFlow 2.4：

training_dataset = tf.data.Dataset.from_generator(
     raw_data_gen, 
     args=(1), 
     output_types=(tf.float32, tf.uint8), 
     output_shapes=([None, 1], [None]))