DMVR: DeepMind Video Readers

DMVR is a library providing a framework for easily reading raw data and producing tf.data.Dataset objects ready to be consumed by models.

Design principles

Data processing graph

The main idea of the framework is to build a customizable and reusable data processing graph that when applied to raw data files, will produce final dataset objects. Building blocks called Builders are used to interact with the graph by adding, removing or replacing data processing blocks.

Dataset providers can write a Factory with a default data processing graph for each dataset. Dataset consumers can customize the graph to their needs either by creating a child Factory or just appending a given instance. Factory objects expose instances of Builders allowing control of the multiple phases of the data processing graph. The Factory is then able to generate tf.data.Dataset objects.

Phases

The data processing graph is split in multipple phases. This abstraction is purely semantic, which makes code easier to reuse. The phases are:

• Parse
• Sample
• Decode
• Preprocess
• Postprocess

Modalities

In order to easily add different modalities to the dataset from the raw data, sub graphs for some modalities with default processing (e.g. sample, decode and crop for images) is provided. These sub graphs can be added by simply calling the corresponding methods for the Builders.

Usage

Dataset providers

Dataset providers should implement a factory populating the default graph.

Example:

• Data is stored in TFRecords as tf.train.SequenceExample objects.

from typing import List

from dmvr import modalities
from dmvr import video_dataset

class Kinetics700Factory(video_dataset.BaseVideoDatasetFactory):

  _NUM_CLASSES = 700

  def __init__(self, subset: str):
    self._is_training = subset == 'train'
    shards: List[str] = path_to_the_data(subset)
    super().__init__(shards)

  def _build(self,
             # Video related parameters.
             num_frames: int = 32,
             stride: int = 1,
             num_test_clips: int = 1,
             min_resize: int = 224,
             crop_size: int = 200,
             zero_centering_image: bool = False,
             # Label related parameters.
             one_hot_label: bool = True,
             add_label_name: bool = False):
    """Build default data processing graph."""
    modalities.add_image(parser_builder=self.parser_builder,
                         sampler_builder=self.sampler_builder,
                         decoder_builder=self.decoder_builder,
                         preprocessor_builder=self.preprocessor_builder,
                         postprocessor_builder=self.postprocessor_builder,
                         is_training=self._is_training,
                         num_frames=num_frames,
                         stride=stride,
                         min_resize=min_resize,
                         crop_size=crop_size,
                         zero_centering_image=zero_centering_image)

    modalities.add_label(parser_builder=self.parser_builder,
                         decoder_builder=self.decoder_builder,
                         preprocessor_builder=self.preprocessor_builder,
                         one_hot_label=one_hot_label,
                         num_classes=self._NUM_CLASSES,
                         add_label_name=add_label_name)

Dataset consumers

Dataset consumers can create tf.data.Dataset objects from a factory instance.

Example:

factory = Kinetics700Factory('train')
factory.configure(num_frames=16)
ds = factory.make_dataset(batch_size=8)

The user can also customize the data processing graph by adding more functions:

from dmvr import builders
from dmvr import processors

factory = Kinetics700Factory('train')
factory.configure(num_frames=16)

factory.preprocess_builder.add_fn(processors.scale_jitter_augm,
                                  feature_name=builders.IMAGE_FEATURE_NAME)
factory.preprocess_builder.add_fn(processors.color_default_augm,
                                  feature_name=builders.IMAGE_FEATURE_NAME)

ds = factory.make_dataset(batch_size=8)

Installation

DMVR can be installed with pip directly from github, with the following command:

pip install git+git://github.com/deepmind/dmvr.git

Python 3.9+ is required in order for all features to be available.