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

Question

I would like to create a number of tf.data.Dataset using the from_generator() function. I would like to send an argument to the generator function ( raw_data_gen ). The idea is that the generator function will yield different data depending on the argument sent. In this way I would like raw_data_gen to be able to provide either training, validation or test data.

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]))

The error message I get when I try to call from_generator() in this way is:

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

Here is the raw_data_gen function although I'm not sure if you will need this as my hunch is that the problem is with the call of from_generator() :

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

You need to define a new function based on raw_data_gen that doesn't take any arguments. You can use the lambda keyword to do this.

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

Now, we are passing a function to from_generator that doesn't take any arguments, but that will simply act as raw_data_gen with the argument set to 1. You can use the same scheme for the validation and test sets, passing 2 and 3 respectively.

Answer 2

For 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]))