Deep Learning Use Cases using OpenPOWER systems
- 1. Developing Deep Learning Applications on OpenPOWER Dr. Abhinandan S Prasad Associate Prof, NIE Mysuru, India
- 2. AGENDA • Introduction • Diabetic Retinopathy • Covid-19 2
- 3. 3 Introduction to Deep Learning Application Development
- 4. INTRODUCTION 4 The picture can't be displayed.
- 7. 7 Diabetic Retinopathy using AI Jetson Nano
- 8. EDGE COMPUTING 8 Fig. 1: An illustration of edge computing resources to gain AI insight. Notably, edge intelligence has garnered much attention from both the industry and academia. on edg In t for ar learnin A. Ar Wh tremen coined intelli do. T refer f techno ligenc https://arxiv.org/abs/1905.10083
- 9. EDGE AI 9
- 11. EDGE DEVICES 11 Google Edge TPU Jetson Nano Boards https://developer.nvidia.com/embedded/jetson-nano-developer-kit https://cloud.google.com/edge-tpu/
- 12. INTRODUCTION • Diabetic Retinopathy is the field to diagnose the level of diabetes based on the retina image • There are five categories of diabetes • No DR • Mild • Moderate • Severe • Proliferative DR • Goal: Can we classify the diabetic level solely on retina image? 12
- 13. SOLUTION OVERVIEW 13 1.DR High Resolutio n Labeled Image set 3.Power9 System model training 2.Image Preprocessing and Data Augmentation 4.Trained Model 5.Inferencing output from the trained model on Jetson Nano 6.Retina Image Capturing Device Feed to train the mod el agai n with new imag es Inferencing could be done on cloud, sending image to cloud or model could itself be deployed on mobile device to do the inference
- 14. SOLUTION OVERVIEW • High resolution image date set has been taken with around 35000 images and labelled • Data pre-processing: Image rotation • Hardware: POWER9 systems with 4 Tesla V100 SXM2 GPU’s • Generated models are deployed on Jetson NANO board • Retina image capturing device captures the image and send the image to cloud or the local device which runs the model, It also sends images with marking label to Power9 system 14
- 15. IMPLEMENTATION • Convolutional Neural Network model has been created trained on 35000 labelled retina images • Model uses VGG16 network architecture and get trained from scratch on Power9 • Original 35000 images are augmented to 1L images out which 20% images are validation set images used to validate the training accuracy during training • Model is trained till 50 epochs to get around 97% accuracy • Notebook code could be accessed at https://github.com/arshad2101/retinopathy/blob/ master/Retinopathy_VGG16.ipynb 15
- 16. 16 Covid-19
- 17. MOTIVATION • Corona virus is one of the most deadly virus with highest number of causalities across the world • We don’t have the enough mechanism to detect the Corona from the X-Ray images • Idea: Can we build ML models to detect positive cases using X-ray images? 17
- 18. MOTIVATION 18
- 19. APPLICATION FLOW 19 C UI in Python Flask or DJango Server in Python Flask User enters credentials Authentication happens at UI end only Inference result and base64 encoding of the image C Model made inference and send back result C User uploads image Image to server Inference result on UI C