Tensorflow for IoT
- 1. Tensorflow for IoT GEETA CHAUHAN JUNE, 2017
- 2. Agenda Era of AI First: What does it mean for IoT? What is Deep Learning? Use Cases for IoT Tensorflow for IoT Optimizations for IoT Common Problem Solutions Look into the Future References
- 3. Era of AI First Billions of connected devices Intelligence at the Edge Increasing Computation power Edison: 500 MHz, 1 GB RAM RPi3: 1.2 GHz Quad-core Deep neural networks running on the IoT device Local inferencing → compressed insights to cloud
- 4. What is Deep Learning? AI Neural Networks composed of many layers Learn like humans Automated Feature Learning Layers are like Image Filters
- 5. Use Cases for IoT Cucumber sorter Real time arrival prediction for Caltrains Wearable assistant for Blind (Horus) Smart cameras, smart door lock Human line counter Real time exercise score on smart watch Creative Arts, Music Generators Self driving cars, Intelligent Robots
- 7. Caltrain Rider Realtime Caltrain arrival prediction Audio Visual pipeline on Raspberry pi Image classification for Caltrains
- 8. Tensorflow for IoT Compile from source Android Things Tensorflow Inference Library (in preview now) Tensorflow Lite (Announced @Google IO for later this year)
- 9. Tensorflow IoT Pipeline Inference on DeviceTrain Model in Cloud
- 10. Tensorflow Android Things Add Tensorflow Inferencing library Create Inference object, load model feed() image, run() to predict fetch() to get classification result
- 11. Demo Intel Edison + Seed Studio Kit Raspberry Pi + Pi Camera
- 12. Transfer Learning Transfer Learning – New Classifier Inception V3
- 13. Optimizations Graph Transform Tool Freeze graph (variables to constants) Quantize weights (20 M weights for IV3) Quantization (32 bit float → 8 bit float) Memory Mapping Inception v3 93 MB → 1.5 MB
- 14. Common Problem Solutions Tensorboard is your friend – X-Ray vision Image size mismatch for the input tensor Output classifier specify correct number of classes from your model Model too large – Load in memory Missing ops -- Graph transform tool Device heating up under heavy processing load Split the model, do part detection on device, rest in cloud Reduce the frequency eg only do on movement detection
- 15. Look into the Future Hardware: Neural Network Chips Software: Tensorflow Lite (Android NN) Intel Fathom Neural Stick Nvidia Jetson
- 16. References Tensorflow for Android Things Sample Tensorboard hands-on Graph Transform Tool Tensorflow for Poets (Transfer learning) Tensorflow for Poets2 (Optimizations for Mobile/IoT) Cucumber Farmer Deep Learning Story Caltrain Rider story Intel Fathom Neural Stick Nvidia Jetson