Skip to main content

Neural network model repository for highly sparse and sparse-quantized models with matching sparsification recipes

Project description

tool icon  SparseZoo

Neural network model repository for highly sparse and sparse-quantized models with matching sparsification recipes

Documentation Main GitHub release GitHub Contributor Covenant


SparseZoo is a constantly-growing repository of sparsified (pruned and pruned-quantized) models with matching sparsification recipes for neural networks. It simplifies and accelerates your time-to-value in building performant deep learning models with a collection of inference-optimized models and recipes to prototype from. Read more about sparsification.

Available via API and hosted in the cloud, the SparseZoo contains both baseline and models sparsified to different degrees of inference performance vs. baseline loss recovery. Recipe-driven approaches built around sparsification algorithms allow you to use the models as given, transfer-learn from the models onto private datasets, or transfer the recipes to your architectures.

The GitHub repository contains the Python API code to handle the connection and authentication to the cloud.

SparseZoo Flow

🚨 New SparseZoo Models 🚨

🎃 October 2023 🎃

Generative AI



This repository is tested on Python 3.8-3.11, 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 sparsezoo

Quick Tour

The SparseZoo Python API enables you to search and download sparsified models. Code examples are given below. We encourage users to load SparseZoo models by copying a stub directly from a model page.

Introduction to Model Class Object

The Model is a fundamental object that serves as a main interface with the SparseZoo library. It represents a SparseZoo model, together with all its directories and files.

Creating a Model Class Object From SparseZoo Stub

from sparsezoo import Model

stub = "zoo:cv/classification/resnet_v1-50/pytorch/sparseml/imagenet/pruned95_quant-none"

model = Model(stub)

>> Model(stub=zoo:cv/classification/resnet_v1-50/pytorch/sparseml/imagenet/pruned95_quant-none)

Creating a Model Class Object From Local Model Directory

from sparsezoo import Model

directory = ".../.cache/sparsezoo/eb977dae-2454-471b-9870-4cf38074acf0"

model = Model(directory)

>> Model(directory=.../.cache/sparsezoo/eb977dae-2454-471b-9870-4cf38074acf0)

Manually Specifying the Model Download Path

Unless specified otherwise, the model created from the SparseZoo stub is saved to the local sparsezoo cache directory. This can be overridden by passing the optional download_path argument to the constructor:

from sparsezoo import Model

stub = "zoo:cv/classification/resnet_v1-50/pytorch/sparseml/imagenet/pruned95_quant-none"
download_directory = "./model_download_directory"

model = Model(stub, download_path = download_directory)

Downloading the Model Files

Once the model is initialized from a stub, it may be downloaded either by calling the download() method or by invoking a path property. Both pathways are universal for all the files in SparseZoo. Invoking the path property will always trigger file download unless the file has already been downloaded.

# method 1 

# method 2 
model_path = model.path

Inspecting the Contents of the SparseZoo Model

We call the available_files method to inspect which files are present in the SparseZoo model. Then, we select a file by calling the appropriate attribute:


>> {'training': Directory(name=training), 
>> 'deployment': Directory(name=deployment), 
>> 'sample_inputs': Directory(name=sample_inputs.tar.gz), 
>> 'sample_outputs': {'framework': Directory(name=sample_outputs.tar.gz)}, 
>> 'sample_labels': Directory(name=sample_labels.tar.gz), 
>> 'model_card': File(, 
>> 'recipes': Directory(name=recipe), 
>> 'onnx_model': File(name=model.onnx)}

Then, we might take a closer look at the contents of the SparseZoo model:

model_card = model.model_card

>> File(
model_card_path = model.model_card.path

>> .../.cache/sparsezoo/eb977dae-2454-471b-9870-4cf38074acf0/

Model, Directory, and File

In general, every file in the SparseZoo model shares a set of attributes: name, path, URL, and parent:

  • name serves as an identifier of the file/directory
  • path points to the location of the file/directory
  • URL specifies the server address of the file/directory in question
  • parent points to the location of the parent directory of the file/directory in question

A directory is a unique type of file that contains other files. For that reason, it has an additional files attribute.


>> File(name=model.onnx)

print(f"File name: {}\n"
      f"File path: {model.onnx_model.path}\n"
      f"File URL: {model.onnx_model.url}\n"
      f"Parent directory: {model.onnx_model.parent_directory}")
>> File name: model.onnx
>> File path: .../.cache/sparsezoo/eb977dae-2454-471b-9870-4cf38074acf0/model.onnx
>> File URL:
>> Parent directory: .../.cache/sparsezoo/eb977dae-2454-471b-9870-4cf38074acf0

>> Directory(name=recipe)

print(f"File name: {}\n"
      f"Contains: {[ for file in]}\n"
      f"File path: {}\n"
      f"File URL: {}\n"
      f"Parent directory: {}")
>> File name: recipe
>> Contains: ['', '']
>> File path: /home/user/.cache/sparsezoo/eb977dae-2454-471b-9870-4cf38074acf0/recipe
>> File URL: None
>> Parent directory: /home/user/.cache/sparsezoo/eb977dae-2454-471b-9870-4cf38074acf0

Selecting Checkpoint-Specific Data

A SparseZoo model may contain several checkpoints. The model may contain a checkpoint that had been saved before the model was quantized - that checkpoint would be used for transfer learning. Another checkpoint might have been saved after the quantization step - that one is usually directly used for inference.

The recipes may also vary depending on the use case. We may want to access a recipe that was used to sparsify the dense model (recipe_original) or the one that enables us to sparse transfer learn from the already sparsified model (recipe_transfer).

There are two ways to access those specific files.

Accessing Recipes (Through Python API)

available_recipes =

>> ['original', 'transfer-classification']

transfer_recipe =["transfer-classification"]

>> File(

original_recipe = # recipe defaults to `original`
original_recipe_path = original_recipe.path # downloads the recipe and returns its path

>> .../.cache/sparsezoo/eb977dae-2454-471b-9870-4cf38074acf0/recipe/

Accessing Checkpoints (Through Python API)

In general, we are expecting the following checkpoints to be included in the model:

  • checkpoint_prepruning
  • checkpoint_postpruning
  • checkpoint_preqat
  • checkpoint_postqat

The checkpoint that the model defaults to is the preqat state (just before the quantization step).

from sparsezoo import Model

stub = "zoo:nlp/question_answering/bert-base/pytorch/huggingface/squad/pruned_quant_3layers-aggressive_84"

model = Model(stub)
available_checkpoints =

>> ['preqat']

preqat_checkpoint = # recipe defaults to `preqat`
preqat_checkpoint_path = preqat_checkpoint.path # downloads the checkpoint and returns its path

>> .../.cache/sparsezoo/0857c6f2-13c1-43c9-8db8-8f89a548dccd/training

[print( for file in preqat_checkpoint.files]

>> vocab.txt
>> special_tokens_map.json
>> pytorch_model.bin
>> config.json
>> training_args.bin
>> tokenizer_config.json
>> trainer_state.json
>> tokenizer.json

Accessing Recipes (Through Stub String Arguments)

You can also directly request a specific recipe/checkpoint type by appending the appropriate URL query arguments to the stub:

from sparsezoo import Model

stub = "zoo:cv/classification/resnet_v1-50/pytorch/sparseml/imagenet/pruned95_quant-none?recipe=transfer"

model = Model(stub)

# Inspect which files are present.
# Note that the available recipes are restricted
# according to the specified URL query arguments

>> ['transfer-classification']

transfer_recipe = # Now the recipes default to the one selected by the stub string arguments

>> File(

Accessing Sample Data

The user may easily request a sample batch of data that represents the inputs and outputs of the model.

sample_data = model.sample_batch(batch_size = 10)

>> (10, 3, 224, 224) # (batch_size, num_channels, image_dim, image_dim)

>> (10, 1000) # (batch_size, num_classes)

Model Search

The function search_models enables the user to quickly filter the contents of SparseZoo repository to find the stubs of interest:

from sparsezoo import search_models

args = {
    "domain": "cv",
    "sub_domain": "segmentation",
    "architecture": "yolact",

models = search_models(**args)
[print(model) for model in models]

>> Model(stub=zoo:cv/segmentation/yolact-darknet53/pytorch/dbolya/coco/pruned82_quant-none)
>> Model(stub=zoo:cv/segmentation/yolact-darknet53/pytorch/dbolya/coco/pruned90-none)
>> Model(stub=zoo:cv/segmentation/yolact-darknet53/pytorch/dbolya/coco/base-none)

Environmental Variables

Users can specify the directory where models (temporarily during download) and its required credentials will be saved in your working machine. SPARSEZOO_MODELS_PATH is the path where the downloaded models will be saved temporarily. Default ~/.cache/sparsezoo/ SPARSEZOO_CREDENTIALS_PATH is the path where credentials.yaml will be saved. Default ~/.cache/sparsezoo/

Console Scripts

In addition to the Python APIs, a console script entry point is installed with the package sparsezoo. This enables easy interaction straight from your console/terminal.


Download command help -h

Download ResNet-50 Model zoo:cv/classification/resnet_v1-50/pytorch/sparseml/imagenet/base-none

Download pruned and quantized ResNet-50 Model zoo:cv/classification/resnet_v1-50/pytorch/sparseml/imagenet/pruned_quant-moderate


Search command help

sparsezoo search -h

Searching for all classification MobileNetV1 models in the computer vision domain

sparsezoo search --domain cv --sub-domain classification --architecture mobilenet_v1

Searching for all ResNet-50 models

sparsezoo search --domain cv --sub-domain classification \
    --architecture resnet_v1 --sub-architecture 50

For a more in-depth read, check out SparseZoo documentation.


Learning More

Release History

Official builds are hosted on PyPI

Additionally, more information can be found via GitHub Releases.


The project is licensed under the Apache License Version 2.0.



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 SparseZoo, sign up or log in to our Neural Magic Community Slack. We are growing the community member by member and happy to see you there. Bugs, feature requests, or additional questions can also be posted to our GitHub Issue Queue.

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 fill out this form.

Project details

Release history Release notifications | RSS feed

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distributions

No source distribution files available for this release.See tutorial on generating distribution archives.

Built Distribution

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page