Skip to content

andreped/GradientAccumulator

Repository files navigation

gradient-accumulator

GradientAccumulator

Seemless gradient accumulation for TensorFlow 2

Pip Downloads PyPI version License DOI

GradientAccumulator was developed by SINTEF Health due to the lack of an easy-to-use method for gradient accumulation in TensorFlow 2.

The package is available on PyPI and is compatible with and have been tested against TensorFlow 2.2-2.15 and Python 3.6-3.11, and works cross-platform (Ubuntu, Windows, macOS).

Build Type Status
Code coverage codecov
Documentations Documentation Status
Unit tests CI

Stable release from PyPI:

pip install gradient-accumulator

Or from source:

pip install git+https://github.com/andreped/GradientAccumulator

Disclaimer: Note that different TensorFlow versions supports different Python versions. Therefore, be sure to check which Python versions are supported when you install a specific TensorFlow version. GradientAccumulator does not support Python >3.11, as TensorFlow has yet to add support for them.

A simple example to add gradient accumulation to an existing model is by:

from gradient_accumulator import GradientAccumulateModel
from tensorflow.keras.models import Model

model = Model(...)
model = GradientAccumulateModel(accum_steps=4, inputs=model.input, outputs=model.output)

Then simply use the model as you normally would!

In practice, using gradient accumulation with a custom pipeline might require some extra overhead and tricks to get working.

For more information, see documentations which are hosted at gradientaccumulator.readthedocs.io

Gradient accumulation (GA) enables reduced GPU memory consumption through dividing a batch into smaller reduced batches, and performing gradient computation either in a distributing setting across multiple GPUs or sequentially on the same GPU. When the full batch is processed, the gradients are then accumulated to produce the full batch gradient.

Note that how we implemented gradient accumulation is slightly different from this illustration, as our design does not require having the entire batch in CPU memory. More information on what goes under the hood can be seen in the documentations.

In TensorFlow 2, there did not exist a plug-and-play method to use gradient accumulation with any custom pipeline. Hence, we have implemented two generic TF2-compatible approaches:

Method Usage
GradientAccumulateModel model = GradientAccumulateModel(accum_steps=4, inputs=model.input, outputs=model.output)
GradientAccumulateOptimizer opt = GradientAccumulateOptimizer(accum_steps=4, optimizer=tf.keras.optimizers.SGD(1e-2))

Both approaches control how frequently the weigths are updated but in their own way. Approach (1) overrides the train_step method of a given Model, whereas approach (2) wraps the optimizer. (1) is only compatible with single-GPU usage, whereas (2) also supports distributed training (multi-GPU).

Our implementations enable theoretically infinitely large batch size, with identical memory consumption as for a regular mini batch. If a single GPU is used, this comes at the cost of increased training runtime. Multiple GPUs could be used to improve runtime performance.

Technique Usage
Batch Normalization layer = AccumBatchNormalization(accum_steps=4)
Adaptive Gradient Clipping model = GradientAccumulateModel(accum_steps=4, agc=True, inputs=model.input, outputs=model.output)
Mixed precision model = GradientAccumulateModel(accum_steps=4, mixed_precision=True, inputs=model.input, outputs=model.output)
  • As batch normalization (BN) is not natively compatible with GA, we have implemented a custom BN layer which can be used as a drop-in replacement.
  • Support for adaptive gradient clipping has been added as an alternative to BN.
  • Mixed precision can also be utilized on both GPUs and TPUs.
  • Multi-GPU distributed training using generic optimizer wrapper.

For more information on usage, supported techniques, and examples, refer to the documentations.

The gradient accumulator model wrapper is based on the implementation presented in this thread on stack overflow. The adaptive gradient clipping method is based on the implementation by @sayakpaul. The optimizer wrapper is derived from the implementation by @fsx950223 and @stefan-falk.

The documentations hosted here was made possible by the incredible Read The Docs team which offer free documentation hosting!

If you used this package or found the project relevant in your research, please, include the following citation:

@software{andre_pedersen_2023_7905351,
  author       = {André Pedersen and Tor-Arne Schmidt Nordmo and Javier Pérez de Frutos and David Bouget},
  title        = {andreped/GradientAccumulator: v0.5.0},
  month        = may,
  year         = 2023,
  publisher    = {Zenodo},
  version      = {v0.5.0},
  doi          = {10.5281/zenodo.7905351},
  url          = {https://doi.org/10.5281/zenodo.7905351}
}