optimum-deepsparse 0.1.0.dev1

Optimum DeepSparse is an extension of the Hugging Face Transformers library that integrates the DeepSparse inference runtime. DeepSparse offers GPU-class performance on CPUs, making it possible to run Transformers and other deep learning models on commodity hardware with sparsity. Optimum DeepSparse provides a framework for developers to easily integrate DeepSparse into their applications, regardless of the hardware platform.

optimum-deepsparse

Accelerated inference of 🤗 models on CPUs using the DeepSparse Inference Runtime.

Install

Optimum DeepSparse is a fast-moving project, and you may want to install from source.

pip install git+https://github.com/neuralmagic/optimum-deepsparse.git

Installing in developer mode

If you are working on the optimum-deepsparse code then you should use an editable install by cloning and installing optimum and optimum-deepsparse:

git clone https://github.com/huggingface/optimum
git clone https://github.com/neuralmagic/optimum-deepsparse
pip install -e optimum -e optimum-deepsparse

Now whenever you change the code, you'll be able to run with those changes instantly.

How to use it?

To load a model and run inference with DeepSparse, you can just replace your AutoModelForXxx class with the corresponding DeepSparseModelForXxx class.

import requests
from PIL import Image

- from transformers import AutoModelForImageClassification
+ from optimum.deepsparse import DeepSparseModelForImageClassification
from transformers import AutoFeatureExtractor, pipeline

url = "http://images.cocodataset.org/val2017/000000039769.jpg"
image = Image.open(requests.get(url, stream=True).raw)

model_id = "microsoft/resnet-50"
- model = AutoModelForImageClassification.from_pretrained(model_id)
+ model = DeepSparseModelForImageClassification.from_pretrained(model_id, export=True, input_shapes="[1,3,224,224]")
feature_extractor = AutoFeatureExtractor.from_pretrained(model_id)
cls_pipe = pipeline("image-classification", model=model, feature_extractor=feature_extractor)
outputs = cls_pipe(image)

Supported Task	Model Class
"image-classification"	DeepSparseModelForImageClassification
"text-classification"/"sentiment-analysis"	DeepSparseModelForSequenceClassification
"audio-classification"	DeepSparseModelForAudioClassification
"question-answering"	DeepSparseModelForQuestionAnswering
"image-segmentation"	DeepSparseModelForSemanticSegmentation

If you find any issue while using those, please open an issue or a pull request.