Libraries for applying sparsification recipes to neural networks with a few lines of code, enabling faster and smaller models
SparseML is a toolkit that includes APIs, CLIs, scripts and libraries that apply state-of-the-art sparsification algorithms such as pruning and quantization to any neural network. General, recipe-driven approaches built around these algorithms enable the simplification of creating faster and smaller models for the ML performance community at large.
This repository contains integrations within the PyTorch, Keras, and TensorFlow V1 ecosystems, allowing for seamless model sparsification.
Sparsification is the process of taking a trained deep learning model and removing redundant information from the overprecise and over-parameterized network resulting in a faster and smaller model. Techniques for sparsification are all encompassing including everything from inducing sparsity using pruning and quantization to enabling naturally occurring sparsity using activation sparsity or winograd/FFT. When implemented correctly, these techniques result in significantly more performant and smaller models with limited to no effect on the baseline metrics. For example, pruning plus quantization can give over noticeable improvements in performance while recovering to nearly the same baseline accuracy.
The Deep Sparse product suite builds on top of sparsification enabling you to easily apply the techniques to your datasets and models using recipe-driven approaches. Recipes encode the directions for how to sparsify a model into a simple, easily editable format.
- Download a sparsification recipe and sparsified model from the SparseZoo.
- Alternatively, create a recipe for your model using Sparsify.
- Apply your recipe with only a few lines of code using SparseML.
- Finally, for GPU-level performance on CPUs, deploy your sparse-quantized model with the DeepSparse Engine.
Full Deep Sparse product flow:
To enable flexibility, ease of use, and repeatability, sparsifying a model is generally done using a recipe. The recipes encode the instructions needed for modifying the model and/or training process as a list of modifiers. Example modifiers can be anything from setting the learning rate for the optimizer to gradual magnitude pruning. The files are written in YAML and stored in YAML or markdown files using YAML front matter. The rest of the SparseML system is coded to parse the recipes into a native format for the desired framework and apply the modifications to the model and training pipeline.
A sample recipe for pruning a model generally looks like the following:
version: 0.1.0
modifiers:
- !EpochRangeModifier
start_epoch: 0.0
end_epoch: 70.0
- !LearningRateModifier
start_epoch: 0
end_epoch: -1.0
update_frequency: -1.0
init_lr: 0.005
lr_class: MultiStepLR
lr_kwargs: {'milestones': [43, 60], 'gamma': 0.1}
- !GMPruningModifier
start_epoch: 0
end_epoch: 40
update_frequency: 1.0
init_sparsity: 0.05
final_sparsity: 0.85
mask_type: unstructured
params: ['sections.0.0.conv1.weight', 'sections.0.0.conv2.weight', 'sections.0.0.conv3.weight']More information on the available recipes, formats, and arguments can be found here. Additionally, all code implementations of the modifiers under the optim packages for the frameworks are documented with example YAML formats.
Pre-configured recipes and the resulting models can be explored and downloaded from the SparseZoo. Also, Sparsify enables autoML style creation of optimization recipes for use with SparseML.
For a more in-depth read, check out SparseML documentation.
The PyTorch sparsification libraries are located under the sparseml.pytorch.optim package.
Inside are APIs designed to make model sparsification as easy as possible by integrating seamlessly into PyTorch training pipelines.
The integration is done using the ScheduledOptimizer class.
It is intended to wrap your current optimizer and its step function.
The step function then calls into the ScheduledModifierManager class which can be created from a recipe file.
With this setup, the training process can then be modified as desired to sparsify the model.
To enable all of this, the integration code you'll need to write is only a handful of lines:
from sparseml.pytorch.optim import ScheduledModifierManager, ScheduledOptimizer
model = None # your model definition
optimizer = None # your optimizer definition
num_train_batches = len(train_data) / batch_size # your number of batches per training epoch
manager = ScheduledModifierManager.from_yaml("/PATH/TO/recipe.yaml")
optimizer = ScheduledOptimizer(optimizer, model, manager, steps_per_epoch=num_train_batches)
# PyTorch training code...The Keras sparsification libraries are located under the sparseml.keras.optim package.
Inside are APIs designed to make model sparsification as easy as possible by integrating seamlessly into Keras training pipelines.
The integration is done using the ScheduledModifierManager class which can be created from a recipe file.
This class handles modifying the Keras objects for the desired algorithms using the modify method.
The edited model, optimizer, and any callbacks necessary to modify the training process are returned.
The model and optimizer can be used normally and the callbacks must be passed into the fit or fit_generator function.
If using train_on_batch, the callbacks must be invoked after each call.
After training is completed, call into the manager's finalize method to clean up the graph for exporting.
To enable all of this, the integration code you'll need to write is only a handful of lines:
from sparseml.keras.optim import ScheduledModifierManager
model = None # your model definition
optimizer = None # your optimizer definition
num_train_batches = len(train_data) / batch_size # your number of batches per training epoch
manager = ScheduledModifierManager.from_yaml("/PATH/TO/recipe.yaml")
model, optimizer, callbacks = manager.modify(
model, optimizer, steps_per_epoch=num_train_batches
)
# Keras compilation and training code...
# Be sure to compile model after calling modify and pass the callbacks into the fit or fit_generator function.
# Note, if you are using train_on_batch, then you will need to invoke the callbacks after every step.
model.compile(...)
model.fit(..., callbacks=callbacks)
# finalize cleans up the graph for export
save_model = manager.finalize(model)The TensorFlow sparsification libraries for TensorFlow version 1.X are located under the sparseml.tensorflow_v1.optim package.
Inside are APIs designed to make model sparsification as easy as possible by integrating seamlessly into TensorFlow V1 training pipelines.
The integration is done using the ScheduledModifierManager class which can be created from a recipe file.
This class handles modifying the TensorFlow graph for the desired algorithms.
With this setup, the training process can then be modified as desired to sparsify the model.
Estimator-based pipelines are simpler to integrate with as compared to session-based pipelines.
The ScheduledModifierManager can override the necessary callbacks in the estimator to modify the graph using the modify_estimator function.
from sparseml.tensorflow_v1.optim import ScheduledModifierManager
estimator = None # your estimator definition
num_train_batches = len(train_data) / batch_size # your number of batches per training epoch
manager = ScheduledModifierManager.from_yaml("/PATH/TO/config.yaml")
manager.modify_estimator(estimator, steps_per_epoch=num_train_batches)
# Normal estimator training code...Session-based pipelines need a little bit more as compared to estimator-based pipelines; however,
it is still designed to require only a few lines of code for integration.
After graph creation, the manager's create_ops method must be called.
This will modify the graph as needed for the algorithms and return modifying ops and extras.
After creating the session and training normally, call into session.run with the modifying ops after each step.
Modifying extras contain objects such as tensorboard summaries of the modifiers to be used if desired.
Finally, once completed, complete_graph must be called to remove the modifying ops for saving and export.
from sparseml.tensorflow_v1.utils import tf_compat
from sparseml.tensorflow_v1.optim import ScheduledModifierManager
with tf_compat.Graph().as_default() as graph:
# Normal graph setup....
num_train_batches = len(train_data) / batch_size # your number of batches per training epoch
# Modifying graphs, be sure his is called after graph is created and before session is created
manager = ScheduledModifierManager.from_yaml("/PATH/TO/config.yaml")
mod_ops, mod_extras = manager.create_ops(steps_per_epoch=num_train_batches)
with tf_compat.Session() as sess:
# Normal training code...
# Call sess.run with the mod_ops after every batch update
sess.run(mod_ops)
# Call into complete_graph after training is done
manager.complete_graph()ONNX is a generic representation for neural network graphs that most ML frameworks can be converted to. Some inference engines such as DeepSparse natively take in ONNX for deployment pipelines, so convenience functions for conversion and export are provided for the supported frameworks.
ONNX is built into the PyTorch system natively.
The ModuleExporter class under the sparseml.pytorch.utils package features an export_onnx function built on top of this native support.
Example code:
import os
import torch
from sparseml.pytorch.models import mnist_net
from sparseml.pytorch.utils import ModuleExporter
model = mnist_net()
exporter = ModuleExporter(model, output_dir=os.path.join(".", "onnx-export"))
exporter.export_onnx(sample_batch=torch.randn(1, 1, 28, 28))ONNX is not built into the Keras system, but is supported through an ONNX official tool keras2onnx. The ModelExporter class under the sparseml.keras.utils package features an export_onnx function built on top of keras2onnx.
Example code:
import os
from sparseml.keras.utils import ModelExporter
model = None # fill in with your model
exporter = ModelExporter(model, output_dir=os.path.join(".", "onnx-export"))
exporter.export_onnx()ONNX is not built into the TensorFlow system, but it is supported through an ONNX official tool
tf2onnx.
The GraphExporter class under the sparseml.tensorflow_v1.utils package features an
export_onnx function built on top of tf2onnx.
Note that the ONNX file is created from the protobuf graph representation, so export_pb must be called first.
Example code:
import os
from sparseml.tensorflow_v1.utils import tf_compat, GraphExporter
from sparseml.tensorflow_v1.models import mnist_net
exporter = GraphExporter(output_dir=os.path.join(".", "mnist-tf-export"))
with tf_compat.Graph().as_default() as graph:
inputs = tf_compat.placeholder(
tf_compat.float32, [None, 28, 28, 1], name="inputs"
)
logits = mnist_net(inputs)
input_names = [inputs.name]
output_names = [logits.name]
with tf_compat.Session() as sess:
sess.run(tf_compat.global_variables_initializer())
exporter.export_pb(outputs=[logits])
exporter.export_onnx(inputs=input_names, outputs=output_names)This repository is tested on Python 3.6+, and Linux/Debian systems. It is recommended to install in a virtual environment to keep your system in order.
Install with pip using:
pip install sparsemlThen if you would like to explore any of the scripts, notebooks, or integrations clone the repository and install any additional dependencies as required.
The currently supported framework versions are:
- PyTorch supported versions:
>= 1.1.0, < 1.8.0 - Keras supported versions:
2.3.0-tf(through the TensorFlow2.2package; as of Feb 1st, 2021,keras2onnxhas not been tested for TensorFlow >=2.3). - TensorFlow V1 supported versions: >=
1.8.0(TensorFlow >=2.Xis not currently supported)
Additionally, optional dependencies can be installed based on the framework you are using.
PyTorch:
pip install sparseml[torch]Keras:
pip install sparseml[tf_keras]TensorFlow V1:
pip install sparseml[tf_v1]TensorFlow V1 with GPU operations enabled:
pip install sparseml[tf_v1_gpu]Depending on your device and CUDA version, you may need to install additional dependencies for using TensorFlow V1 with GPU operations. You can find these steps here.
Note, TensorFlow V1 is no longer being built for newer operating systems such as Ubuntu 20.04. Therefore, SparseML with TensorFlow V1 is unsupported on these operating systems as well.
- SparseZoo Documentation
- SparseML Documentation
- Sparsify Documentation
- DeepSparse Documentation
- Neural Magic Blog, Resources, Website
We appreciate contributions to the code, examples, integrations, and documentation as well as bug reports and feature requests! Learn how here.
For user help or questions about Sparsify, use our GitHub Discussions. Everyone is welcome!
You can get the latest news, webinar and event invites, research papers, and other ML Performance tidbits by subscribing to the Neural Magic community.
For more general questions about Neural Magic, please email us at [email protected] or fill out this form.
The project is licensed under the Apache License Version 2.0.
Official builds are hosted on PyPi
- stable: sparseml
- nightly (dev): sparseml-nightly
Additionally, more information can be found via GitHub Releases.
Find this project useful in your research or other communications? Please consider citing:
@InProceedings{
pmlr-v119-kurtz20a,
title = {Inducing and Exploiting Activation Sparsity for Fast Inference on Deep Neural Networks},
author = {Kurtz, Mark and Kopinsky, Justin and Gelashvili, Rati and Matveev, Alexander and Carr, John and Goin, Michael and Leiserson, William and Moore, Sage and Nell, Bill and Shavit, Nir and Alistarh, Dan},
booktitle = {Proceedings of the 37th International Conference on Machine Learning},
pages = {5533--5543},
year = {2020},
editor = {Hal Daumé III and Aarti Singh},
volume = {119},
series = {Proceedings of Machine Learning Research},
address = {Virtual},
month = {13--18 Jul},
publisher = {PMLR},
pdf = {http://proceedings.mlr.press/v119/kurtz20a/kurtz20a.pdf},
url = {http://proceedings.mlr.press/v119/kurtz20a.html},
abstract = {Optimizing convolutional neural networks for fast inference has recently become an extremely active area of research. One of the go-to solutions in this context is weight pruning, which aims to reduce computational and memory footprint by removing large subsets of the connections in a neural network. Surprisingly, much less attention has been given to exploiting sparsity in the activation maps, which tend to be naturally sparse in many settings thanks to the structure of rectified linear (ReLU) activation functions. In this paper, we present an in-depth analysis of methods for maximizing the sparsity of the activations in a trained neural network, and show that, when coupled with an efficient sparse-input convolution algorithm, we can leverage this sparsity for significant performance gains. To induce highly sparse activation maps without accuracy loss, we introduce a new regularization technique, coupled with a new threshold-based sparsification method based on a parameterized activation function called Forced-Activation-Threshold Rectified Linear Unit (FATReLU). We examine the impact of our methods on popular image classification models, showing that most architectures can adapt to significantly sparser activation maps without any accuracy loss. Our second contribution is showing that these these compression gains can be translated into inference speedups: we provide a new algorithm to enable fast convolution operations over networks with sparse activations, and show that it can enable significant speedups for end-to-end inference on a range of popular models on the large-scale ImageNet image classification task on modern Intel CPUs, with little or no retraining cost.}
}@misc{
singh2020woodfisher,
title={WoodFisher: Efficient Second-Order Approximation for Neural Network Compression},
author={Sidak Pal Singh and Dan Alistarh},
year={2020},
eprint={2004.14340},
archivePrefix={arXiv},
primaryClass={cs.LG}
}