What's Inside the Box?

A presentation at PAPIs.io in October 2018 in Cambridge, MA, USA by Ray Deck

What’s Inside the Box? Comparing Data Storage Across Frameworks @ray_deck

Code as Data

PyTorch Keras TensorFlow TensorFlow Lite CoreML

Serialization pytorch.org/ docs/ stable/ notes/ serialization.html

Pickle docs.python.org/3/library/pickle.html

JSON + HDF5

HDFView portal.hdfgroup.org/display/support/Download+HDFView Panoply www.giss.nasa.gov/tools/panoply/

JSON in HDF5

https://sparktoro.com/fake-followers/ray_deck

developers.google.com/protocol-buffers

https://www.slideshare.net/SergeyPodolsky/google-protocol-buffers-56085699

github.com/ tensorflow/ tensorflow/ tree/ master/ tensorflow/ core/ protobuf

Protobuf Viewer MacOS App Store Protobuf Editor sourceforge.net/projects/protobufeditor/

FlatBuf google.github.io/ flatbuffers/

FlatBuf github.com/ tensorflow/ tensorflow/ blob/ master/ tensorflow/ contrib/ lite/ schema/ schema_v3.fbs

github.com/ apple/ coremltools/ tree/ master/ mlmodel/ format

let c = MLModel.compile(url: u) let m = MLModel(url: c) let d:[String:Any] = ["input": image] let dfpin = MLDictionaryFeatureProvider(dictionary: d) let fp = m.predict(dfp) let dfpout = fp.featureValue(for: "labels") let dic = dfpout.dictionaryValue let topResult = dic.first()

Quo Vadis?

Thank You @ray_deck github.com/rhdeck/papis-2018

Ray Deck
@raydeck

1 / 41

Protocol Buffer. HDF5. TFLite. Pickle. They all hold model weights and/or code, and in many cases they can and do hold the same ones. Understanding how they work and how to unpack them outside of our tools gives us insights into how the tools themselves work, and how we can pipe our models around, and make them more useful for our customers - often app developers!

Video