Published On: Thu, Sep 7th, 2017

Facebook and Microsoft combine to facilitate conversions from PyTorch to Caffe2


Facebook and Microsoft announced ONNX, the Open Neural Network Exchange this morning in particular blog posts. The Exchange creates it easier for appurtenance training developers to modify models between PyTorch and Caffe2 to revoke a loiter time between investigate and productization.

Facebook has prolonged confirmed a eminence between a FAIR and AML appurtenance training groups. Facebook AI Research (FAIR) handles draining corner investigate while Applied Machine Learning (AML) brings comprehension to products.

Choice of low training horizon underlies this pivotal ideological distinction. FAIR is accustomed to operative with PyTorch — a low training horizon optimized for achieving state of a art formula in research, regardless of apparatus constraints.

Unfortunately in a genuine world, many of us are singular by a computational capabilities of a smartphones and computers. When AML wants to build something for deployment and scale, it opts for Caffe2. Caffe2 is also a low training horizon though it’s optimized for apparatus efficiency, utterly with honour to Caffe2Go that’s optimized for using appurtenance training models on underpowered mobile devices.

The collaborative work Facebook and Microsoft are announcing helps folks simply modify models built in PyTorch into Caffe2 models. By shortening a barriers to relocating between these dual frameworks, a dual companies can indeed urge a freeing of investigate and assistance speed adult a whole commercialization process.

Unfortunately not each association uses a same PyTorch and Caffe2 pairing. Plenty of investigate is still finished in TensorFlow and other pivotal frameworks. Outside of a investigate context, others have been operative to make it easier to modify appurtenance training models into formats optimized for specific devices.

Apple’s CoreML for instance helps developers modify a very limited series of models. At this indicate CoreML doesn’t even support TensorFlow and a routine of formulating tradition converters seems utterly difficult and expected to finish in disappointment. As companies like Google and Apple benefit some-more control over appurtenance training horizon optimization on tradition hardware, it’s going to be critical to continue to guard interoperability.

The Open Neural Network Exchange has been expelled on Github, we can find it here.

Featured Image: Bryce Durbin

About the Author

Leave a comment

XHTML: You can use these html tags: <a href="" title=""> <abbr title=""> <acronym title=""> <b> <blockquote cite=""> <cite> <code> <del datetime=""> <em> <i> <q cite=""> <s> <strike> <strong>