Skip to content
/ qsparse Public

Train neural networks with joint quantization and pruning on both weights and activations using any pytorch modules

License

Notifications You must be signed in to change notification settings

mlzxy/qsparse

Repository files navigation

QSPARSE

License: MIT

QSPARSE provides the open source implementation of the quantization and pruning methods proposed in Learning Low-Precision Structured Subnetworks Using Joint Layerwise Channel Pruning and Uniform Quantization. This library was developed to support and demonstrate strong performance and flexibility among various experiments.

Full Precision Joint Quantization 4bit and Channel Pruning 75%
import torch.nn as nn
net = nn.Sequential(
    nn.Conv2d(3, 32, 5),
    nn.ReLU(),
    nn.ConvTranspose2d(32, 3, 5, stride=2)
)
import torch.nn as nn
from qsparse import prune, quantize, convert
net = nn.Sequential(
    quantize(nn.Conv2d(3, 32, 5), bits=4), 
    nn.ReLU(),
    prune(sparsity=0.75, dimensions={1}), 
    quantize(bits=8),  
    quantize(nn.ConvTranspose2d(32, 3, 5, stride=2), bits=4)
)
# Automatic conversion is available via `convert`.
# Please refer to documentation for more details.

Installation

QSPARSE can be installed from PyPI:

pip install qsparse

Usage

Documentation can be accessed from Read the Docs.

Examples of applying QSPARSE to different tasks are provided at examples and mdpi2022.

Citing

If you find this open source release useful, please reference in your paper:

Zhang, X.; Colbert, I.; Das, S. Learning Low-Precision Structured Subnetworks Using Joint Layerwise Channel Pruning and Uniform Quantization. Appl. Sci. 2022, 12, 7829. https://doi.org/10.3390/app12157829

@Article{app12157829,
	AUTHOR = {Zhang, Xinyu and Colbert, Ian and Das, Srinjoy},
	TITLE = {Learning Low-Precision Structured Subnetworks Using Joint Layerwise Channel Pruning and Uniform Quantization},
	JOURNAL = {Applied Sciences},
	VOLUME = {12},
	YEAR = {2022},
	NUMBER = {15},
	ARTICLE-NUMBER = {7829},
	URL = {https://www.mdpi.com/2076-3417/12/15/7829},
	ISSN = {2076-3417}
}

About

Train neural networks with joint quantization and pruning on both weights and activations using any pytorch modules

Topics

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Languages