A better approach to data loading for Deep Learning. API-transparent caching to disk, GCS, or S3.
At Determined AI, we help many customers perform high-performance data loading for deep learning models. We believe every data loader should have two layers: the random-access layer and the sequential layer.
The random-access layer is critical for good training infrastructure. Direct random access to any record enables:
- Shuffling (potentially every epoch)
- Pausing/continuing training mid-epoch
- Sharding the dataset efficiently for distributed training
The sequential layer starts as soon as you decide the order in which you will access the records in the dataset. Often the transition is implicit, in which case it starts as soon as you are done modifying the order of access (i.e. via shuffling, sharding, or splitting). This layer is vital to performance optimizations because it enables:
- Prefetching data loading to hide latency costs
- Parallelizing data loading to hide compute costs
Here is a simple code snippet to illustrate what the transition from random-access layer to sequential layer looks like:
# Start of random-access layer.
indices = list(range(100))
indices = indices[skip:]
indices=np.random.shuffle(indices)
# Start of sequential layer.
def record_gen():
for i in indices:
yield read_file_at_index(i)
record_ds = tf.data.Dataset.from_generator(record_gen, ...)
final_ds = record_ds.prefetch(...)
Notice that in the above example, the tf.data
API is used, but only in the sequential layer.
This is because tf.data
has no concept of the random access layer. As a result:
tf.data.Dataset.shuffle()
can only approximate a shuffle. Calling.shuffle(N)
will readN
records into a buffer and choose samples randomly from thoseN
records, while a true shuffle chooses samples randomly from the entire dataset. This shortcoming forces you to choose between memory footprint and the quality of your shuffle. The only true shuffle with tf.data.Dataset.shuffle() is to read the entire dataset into memory.tf.data.Dataset.skip(N)
is as inefficient as possible. Each of theN
skipped records will still be read from disk and processed normally, according to all of the operations preceeding the.skip()
call, making.skip()
prohibitively expensive for most use cases.- Pausing and continuing training is only possible by saving the state of a
tf.data.Iterator
. However, saving atf.data.Iterator
does not work with all datasets. In particular, it does not work with datasets created usingfrom_generator()
, which is the easiest way to create atf.data.Dataset
.
We have seen countless instances where tf.data.Dataset
shortcomings have made life harder for
deep learning practitioners, so we set out to build something better. We set out to build a new
data layer which could augment an existing tf.data.Dataset
data loader with the properties should
come standard with every data loader.
At the same time, we wanted this new data layer to relieve another key pain point: high-performance dataset caching and dataset versioning.
We designed yogadl
to be two things: a standalone caching layer to imbue existing data loaders
with the properties that come from a random-access layer, and a better interface for defining data
loaders in general.
Since tf.data.Dataset
-based datasets have no random-access layer, yogadl
caches them to disk in
a random-access-friendly way. The storage mechanism is, in fact, nearly identical to how
TensorPack caches datasets to disk,
only with some additional abstractions to allow dataset versioning, cloud storage, and all of the
wonderful features that a data loader with a random-access layer ought to have.
What does all this do for you? A few things:
- Better training: A
yogadl
-cachedtf.data.Dataset
will have better shuffling than a nativetf.data.Dataset
. Additionally, pausing and continuing training mid-epoch will be simple and robust, and efficient sharding for distributed training comes standard. - Faster data loading: Slow data loader? Don't waste your time optimizing it.
yogadl
will save it in a high-performance cache the first time it is used, and all future uses will be fast and efficient. - API-transparent: Not all operations in the data loader are cacheable. Data augmentation
must be done at run time.
yogadl
allows you to keep your existing data augmentation code.
At the core of yogadl
is the DataRef
interface, which creates an explicit boundary between the
random-access layer and the sequential layer.
We are not the first people to think of this: PyTorch separates the DataSet
(the random-access
layer) from the Sampler
(which defines the sequential layer). Keras has a Sequence
object
which defines the random-access layer, leaving the order of access (the sequential layer) to be
decided by the arguments to model.fit()
. Both DataSet
and Sequence
are already 100%
compatible with yogadl
's DataRef
interface (although yogadl
does not yet include those
adapters).
And yet, the world is still full of data loaders which are lacking. At Determined AI, we are
dedicated to advancing the state of the art for training Deep Learning models, and we believe that
a better interface for data loading is a critical piece of that goal. Any data loader which
implements the DataRef
interface is capable of proper shuffling, pausing and continuing training
mid-epoch, and efficient multi-machine distributed training.
yogadl
is not a data manipulation API.
This
world
has
more
than
enough
of
those.
Instead, yogadl
seeks to be API-transparent so that you can continue to use your existing data
loading code, but with all the benefits of a high-performance, random-access cache. If you have
data augmentation steps which cannot be cached, that code should continue to work without any
modifications.
yogadl
does not (at this time) work with any data frameworks other than tf.data.Dataset.
First-class support for (tf.)Keras Sequence
objects, PyTorch DataSet
objects, and TensorPack
DataFlow
objects is on the near-term roadmap.
yogadl
offers basic dataset versioning, but it is not (at this time) a full-blown version control
for datasets. Offering something like version control for datasets is on the roadmap as well.
TODO: code examples here.