This document discusses building machine learning models for mobile apps using TensorFlow. It describes the process of gathering training data, training a model using Cloud ML Engine, optimizing the model for mobile, and integrating it into an Android app. Key steps involve converting video training data to images, retraining an InceptionV3 model, optimizing the model size with graph transformations, and loading the model into an Android app. TensorFlow allows developing machine learning models that can run efficiently on mobile devices.
14. @YufengG
From training to app
Gather training data
Load into Android app
?
Convert to images
Folders of images
Storage
Cloud Storage
Object storage with global
edge-caching
Training
Cloud ML Engine
Fast, scalable, and
easy-to-use ML services
24. @YufengG
From training to app
Gather training data
Load into Android app
Convert to images
Folders of images
Storage
Cloud Storage
Object storage with global
edge-caching
Training
Cloud ML Engine
Fast, scalable, and
easy-to-use ML services
Optimize for mobile
27. @YufengG
From training to app
Gather training data Convert to images
Training
Cloud ML Engine
Fast, scalable, and
easy-to-use ML services
Storage
Cloud Storage
Object storage with global
edge-caching
Folders of images
Optimize for mobile
Load into Android app
31. @YufengG
Community
● 475+ non-Google contributors to TensorFlow 1.0
● 14,000+ commits in 14 months
● Many community created tutorials, models,
translations, and projects
○ ~5,500 GitHub repositories with ‘TensorFlow’
in the title
33. @YufengG
From training to app
Gather training data Convert to images
Training
Cloud ML Engine
Fast, scalable, and
easy-to-use ML services
Storage
Cloud Storage
Object storage with global
edge-caching
Folders of images
Optimize for mobile
Load into Android app
34. @YufengG
Codelab to retrain your own InceptionV3 network
bit.ly/tf-retrain
Mobile Tensorflow
tensorflow.org/mobile
Cloud Machine Learning Engine
cloud.google.com/ml
Thank you!
Yufeng Guo
Developer Advocate
yufengg.com
35. @YufengG
Mobile machine learning, now with extra fast
Just-In-Time Compilation
via XLA, "Accelerated Linear Algebra" compiler
0x00000000 movq (%rdx), %rax
0x00000003 vmovaps (%rax), %xmm0
0x00000007 vmulps %xmm0, %xmm0, %xmm0
0x0000000b vmovaps %xmm0, (%rdi)
...
TF graphs go in
Optimized & specialized
assembly comes out.