You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
This repository was archived by the owner on Jun 3, 2025. It is now read-only.
@@ -44,42 +44,55 @@ limitations under the License.
44
44
45
45
## Overview
46
46
47
-
SparseZoo is a constantly-growing repository of optimized models and optimization recipes for neural networks.
47
+
SparseZoo is a constantly-growing repository of highly sparse and sparse-quantized models with matching sparsification recipes for neural networks.
48
48
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.
49
49
50
-
Available via API and hosted in the cloud, the SparseZoo contains both baseline and models optimized to different degrees of inference performance vs baseline loss recovery. Optimizations on neural networks include approaches such as [pruning](https://neuralmagic.com/blog/pruning-overview/) and [quantization](https://arxiv.org/abs/1609.07061)
51
-
allowing for significantly faster models with limited to no effect on their baseline metrics such as accuracy.
52
-
Recipe-driven approaches built around these optimizations allow you to take the models as given, transfer learn from the models onto private datasets, or transfer the recipes to your architectures.
50
+
Available via API and hosted in the cloud, the SparseZoo contains both baseline and models optimized to different degrees of inference performance vs. baseline loss recovery.
51
+
Recipe-driven approaches built around sparsification algorithms allow you to take the models as given, transfer-learn from the models onto private datasets, or transfer the recipes to your architectures.
53
52
54
53
This repository contains the Python API code to handle the connection and authentication to the cloud.
55
54
56
-
### Related Products
57
55
58
-
-[DeepSparse](https://github.com/neuralmagic/deepsparse): CPU inference engine that delivers unprecedented performance for sparse models
59
-
-[SparseML](https://github.com/neuralmagic/sparseml): Libraries for state-of-the-art deep neural network optimization algorithms, enabling simple pipelines integration with a few lines of code
60
-
-[Sparsify](https://github.com/neuralmagic/sparsify): Easy-to-use autoML interface to optimize deep neural networks for better inference performance and a smaller footprint
56
+
## Sparsification
57
+
58
+
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.
59
+
Techniques for sparsification are all encompassing including everything from inducing sparsity using [pruning](https://neuralmagic.com/blog/pruning-overview/) and [quantization](https://arxiv.org/abs/1609.07061) to enabling naturally occurring sparsity using [activation sparsity](http://proceedings.mlr.press/v119/kurtz20a.html) or [winograd/FFT](https://arxiv.org/abs/1509.09308).
60
+
When implemented correctly, these techniques result in significantly more performant and smaller models with limited to no effect on the baseline metrics.
61
+
For example, pruning plus quantization can give over [7x improvements in performance](https://neuralmagic.com/blog/benchmark-resnet50-with-deepsparse) while recovering to nearly the same baseline accuracy.
62
+
63
+
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.
64
+
Recipes encode the directions for how to sparsify a model into a simple, easily editable format.
65
+
- Download a sparsification recipe and sparsified model from the [SparseZoo](https://github.com/neuralmagic/sparsezoo).
66
+
- Alternatively, create a recipe for your model using [Sparsify](https://github.com/neuralmagic/sparsify).
67
+
- Apply your recipe with only a few lines of code using [SparseML](https://github.com/neuralmagic/sparseml).
68
+
- Finally, for GPU-level performance on CPUs, deploy your sparse-quantized model with the [DeepSparse Engine](https://github.com/neuralmagic/deepsparse).
| DOMAIN | The type of solution the model is architected and trained for | cv, nlp |
73
-
| SUB_DOMAIN | The sub type of solution the model is architected and trained for | classification, segmentation |
74
-
| ARCHITECTURE | The name of the guiding setup for the network's graph | resnet_v1, mobilenet_v1 |
75
-
| SUB_ARCHITECTURE | (optional) The scaled version of the architecture such as width or depth | 50, 101, 152 |
76
-
| FRAMEWORK | The machine learning framework the model was defined and trained in | pytorch, tensorflow_v1 |
77
-
| REPO | The model repository the model and baseline weights originated from | sparseml, torchvision |
78
-
| DATASET | The dataset the model was trained on | imagenet, cifar10 |
79
-
| TRAINING_SCHEME | (optional) A description on how the model was trained | augmented, lower_lr |
80
-
|OPTIM_NAME| An overview of what was done to optimize the model | base, pruned, quant (quantized), pruned_quant, arch (architecture modified) |
81
-
|OPTIM_CATEGORY | Descriptor on the degree to which the model is optimized as compared with the baseline metric | none, conservative (100% baseline), moderate (>= 99% baseline), aggressive (< 99%) |
82
-
|OPTIM_TARGET| (optional) Descriptor for the target environment the model was optimized for | disk, edge, deepsparse, gpu |
| DOMAIN | The type of solution the model is architected and trained for | cv, nlp |
86
+
| SUB_DOMAIN | The sub type of solution the model is architected and trained for | classification, segmentation |
87
+
| ARCHITECTURE | The name of the guiding setup for the network's graph | resnet_v1, mobilenet_v1 |
88
+
| SUB_ARCHITECTURE | (optional) The scaled version of the architecture such as width or depth | 50, 101, 152 |
89
+
| FRAMEWORK | The machine learning framework the model was defined and trained in | pytorch, tensorflow_v1 |
90
+
| REPO | The model repository the model and baseline weights originated from | sparseml, torchvision |
91
+
| DATASET | The dataset the model was trained on | imagenet, cifar10 |
92
+
| TRAINING_SCHEME | (optional) A description on how the model was trained | augmented, lower_lr |
93
+
|SPARSE_NAME| An overview of what was done to sparsify the model| base, pruned, quant (quantized), pruned_quant, arch (architecture modified) |
94
+
|SPARSE_CATEGORY | Descriptor on the degree to which the model is sparsified as compared with the baseline metric | none, conservative (100% baseline), moderate (>= 99% baseline), aggressive (< 99%) |
95
+
|SPARSE_TARGET| (optional) Descriptor for the target environment the model was sparsified for | disk, edge, deepsparse, gpu |
83
96
84
97
The contents of each model are made up of the following:
85
98
@@ -89,8 +102,8 @@ The contents of each model are made up of the following:
89
102
Currently ONNX does not support sparse tensors and quantized sparse tensors well for compression.
90
103
-[FRAMEWORK]/model.[EXTENSION]: The native ML framework file(s) for the model in which it was originally trained.
91
104
Such as PyTorch, Keras, TensorFlow V1
92
-
- recipes/optimization.[md|yaml]: The original optimization recipe used to create the model.
93
-
- recipes/[NAME].[md|yaml]: Additional optimization recipes that can be used with the model such as transfer learning.
105
+
- recipes/original.[md|yaml]: The original sparsification recipe used to create the model.
106
+
- recipes/[NAME].[md|yaml]: Additional sparsification recipes that can be used with the model such as transfer learning.
94
107
- sample-originals: The original sample data without any preprocessing for use with the model.
95
108
- sample-inputs: The sample data after pre processing for use with the model.
96
109
- sample-outputs: The outputs after running the sample inputs through the model.
@@ -210,7 +223,7 @@ clone the repository and install any additional dependencies as required.
210
223
211
224
## Available Models and Recipes
212
225
213
-
A number of pre-trained baseline and optimized models across domains and sub domains are available and constantly being added.
226
+
A number of pre-trained baseline and sparsified models across domains and sub domains are available and constantly being added.
214
227
For an up to date list, please consult the [available models listing](https://github.com/neuralmagic/sparsezoo/blob/main/docs/source/models.md).
Copy file name to clipboardExpand all lines: docs/source/index.rst
+38-36Lines changed: 38 additions & 36 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -17,7 +17,7 @@
17
17
SparseZoo |version|
18
18
===================
19
19
20
-
Neural network model repository for highly sparse models and optimization recipes
20
+
Neural network model repository for highly sparse and sparse-quantized models with matching sparsification recipes
21
21
22
22
.. raw:: html
23
23
@@ -48,51 +48,53 @@ Neural network model repository for highly sparse models and optimization recipe
48
48
Overview
49
49
========
50
50
51
-
SparseZoo is a constantly-growing repository of optimized models and optimization recipes for neural networks.
52
-
It simplifies and accelerates your time-to-value in building performant deep learning models with a
53
-
collection of inference-optimized models and recipes to prototype from.
51
+
SparseZoo is a constantly-growing repository of highly sparse and sparse-quantized models with matching sparsification recipes for neural networks.
52
+
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.
54
53
55
-
Available via API and hosted in the cloud, the SparseZoo contains both baseline and models optimized
56
-
to different degrees of inference performance vs baseline loss recovery.
57
-
Optimizations on neural networks include approaches such as
58
-
`pruning <https://neuralmagic.com/blog/pruning-overview />`_ and `quantization <https://arxiv.org/abs/1609.07061 />`_
59
-
allowing for significantly faster models with limited to no effect on their baseline metrics such as accuracy.
60
-
Recipe-driven approaches built around these optimizations allow you to take the models as given,
61
-
transfer learn from the models onto private datasets, or transfer the recipes to your architectures.
54
+
Available via API and hosted in the cloud, the SparseZoo contains both baseline and models optimized to different degrees of inference performance vs. baseline loss recovery.
55
+
Recipe-driven approaches built around sparsification algorithms allow you to take the models as given, transfer-learn from the models onto private datasets, or transfer the recipes to your architectures.
62
56
63
-
This repository contains the Python API code to handle the connection and authentication to the cloud.
57
+
`This repository<https://github.com/neuralmagic/sparsezoo>`_ contains the Python API code to handle the connection and authentication to the cloud.
Easy-to-use autoML interface to optimize deep neural networks for
75
-
better inference performance and a smaller footprint
62
+
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.
63
+
Techniques for sparsification are all encompassing including everything from inducing sparsity using `pruning <https://neuralmagic.com/blog/pruning-overview/>`_ and `quantization <https://arxiv.org/abs/1609.07061>`_ to enabling naturally occurring sparsity using `activation sparsity <http://proceedings.mlr.press/v119/kurtz20a.html>`_ or `winograd/FFT <https://arxiv.org/abs/1509.09308>`_.
64
+
When implemented correctly, these techniques result in significantly more performant and smaller models with limited to no effect on the baseline metrics.
65
+
For example, pruning plus quantization can give over `7x improvements in performance <https://neuralmagic.com/blog/benchmark-resnet50-with-deepsparse>`_ while recovering to nearly the same baseline accuracy.
66
+
67
+
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.
68
+
Recipes encode the directions for how to sparsify a model into a simple, easily editable format.
69
+
- Download a sparsification recipe and sparsified model from the `SparseZoo <https://github.com/neuralmagic/sparsezoo>`_.
70
+
- Alternatively, create a recipe for your model using `Sparsify <https://github.com/neuralmagic/sparsify>`_.
71
+
- Apply your recipe with only a few lines of code using `SparseML <https://github.com/neuralmagic/sparseml>`_.
72
+
- Finally, for GPU-level performance on CPUs, deploy your sparse-quantized model with the `DeepSparse Engine <https://github.com/neuralmagic/deepsparse>`_.
0 commit comments