How to use keras.utils.Sequence data generator with tf.distribute.MirroredStrategy for multi-gpu model training in tensorflow?

Question

I want to train a model on several GPUs using tensorflow 2.0. In the tensorflow tutorial for distributed training (https://www.tensorflow.org/guide/distributed_training), the tf.data datagenerator is converted into a distributed dataset as follows:

dist_dataset = mirrored_strategy.experimental_distribute_dataset(dataset)

However, I want to use my own custom data generator instead (for example, the keras.utils.Sequence datagenerator, along with keras.utils.data_utils.OrderedEnqueuer for asynchronous batch generation). But the mirrored_strategy.experimental_distribute_dataset method supports only tf.data datagenerator. How do I use the keras datagenerator instead?

Thank you!

Brummett

Answer 1

I used tf.data.Dataset.from_generator with my keras.utils.sequence in the same situation, and it solved my issues!

train_generator = SegmentationMultiGenerator(datasets, folder) # My keras.utils.sequence object

def generator():
    multi_enqueuer = OrderedEnqueuer(train_generator, use_multiprocessing=True)
    multi_enqueuer.start(workers=10, max_queue_size=10)
    while True:
        batch_xs, batch_ys, dset_index = next(multi_enqueuer.get()) # I have three outputs
        yield batch_xs, batch_ys, dset_index

dataset = tf.data.Dataset.from_generator(generator,
                                         output_types=(tf.float64, tf.float64, tf.int64),
                                         output_shapes=(tf.TensorShape([None, None, None, None]),
                                                        tf.TensorShape([None, None, None, None]),
                                                        tf.TensorShape([None, None])))

strategy = tf.distribute.MirroredStrategy()

train_dist_dataset = strategy.experimental_distribute_dataset(dataset)

Note that this is my first working solution - at the moment I have found it most convenient to just put 'None' in the place of the real output shapes, which I have found to work.

Careerist

Answer 2

Without using an Enqueuer, here is another way, assuming you have a generator dg which yields samples in the form (feature, label) when called:

import tensorflow as tf
import numpy as np


def get_tf_data_Dataset(data_generator_settings_dict):
    length_req = data_generator_settings_dict["length"]
    x_d1 = data_generator_settings_dict['x_d1']
    x_d2 = data_generator_settings_dict['x_d2']
    x_d3 = data_generator_settings_dict['x_d3']
    y_d1 = data_generator_settings_dict['y_d1']
    x_d2 = data_generator_settings_dict['x_d2']
    y_d3 = data_generator_settings_dict['y_d3']
    list_of_x_arrays = [np.zeros((x_d1, x_d2, x_d3)) for _ in range(length_req)]
    list_of_y_arrays = [np.zeros((y_d1, y_d2, y_d3)) for _ in range(length_req)]
    list_of_tuple_samples = [(x, y) for (x, y) in dg()]
    list_of_x_samples = [x for (x, y) in list_of_tuple_samples]
    list_of_y_samples = [y for (x, y) in list_of_tuple_samples]
    for sample_index in range(length_req):
        list_of_x[sample_index][:] = list_of_x_samples[sample_index]
        list_of_y[sample_index][:] = list_of_y_samples[sample_index]
    return tf.data.Dataset.from_tensor_slices((list_of_x, list_of_y))

It is convoluted but guaranteed to work. This also implies that the __call__ method of dg is a for loop like (after the __init__ of course):

def __call__(self):
    for _ in self.length:
        # generate x (single sample of feature)
        # generate y (single matching sample of label)
        yield x, y

Fantasize