NVIDIA深度學習教育機構 (DLI): Object detection with jetson
•
2 likes•1,420 views
運用 NVIDIA JETSON TX1 開發者套件進行圖像分類與物體檢測 課程目標與介紹 學習構建一個從訓練數據庫到智慧終端深度學習流程。在本課程終您將實際操作開發技能，不僅訓練深層神經網絡，而且還將如何在生產環境中進行部署。在本實驗中，您可以將預先訓練的圖像分類和物體檢測的神經網絡，並將其部署在 Jetson TX1 或 TX2 開發者套件上，接著使用內建攝影機測試這些網絡，以分類和檢測幾個真實世界的對象。網絡將部署在各種編程環境中，我們甚至將介紹如何使用 NVIDIA 的 TensorRT 推理引擎庫在運行時優化分類和檢測性能。
Recommended
Recommended
More Related Content
What's hot
What's hot (20)
Similar to NVIDIA深度學習教育機構 (DLI): Object detection with jetson
Similar to NVIDIA深度學習教育機構 (DLI): Object detection with jetson (20)
More from NVIDIA Taiwan
More from NVIDIA Taiwan (20)
Recently uploaded
Recently uploaded (20)
NVIDIA深度學習教育機構 (DLI): Object detection with jetson
- 1. 1 Instructor DLI Robotics Workshop “Pixels to Action” Wednesday, May 10, 2017
- 2. 2 DEEP LEARNING INSTITUTE DLI Mission Helping people solve challenging problems using AI and deep learning. • Developers, data scientists and engineers • Self-driving cars, healthcare and robotics • Training, optimizing, and deploying deep neural networks
- 3. 3 DLI TRAINING OFFERINGS 5-DAY COURSES INSTRUCTOR-LED WORKSHOPS SELF-PACED LABS Online labs offer on- demand 24/7 access to introductory concepts Prerequisites: Developer, Data Scientist or Researcher. Not designed for non-developers Coming later in 2017 Industry-specific courses teach students to fine-tune a neural network to deploy on a specific platform (e.g. NVIDIA Drive PX2) Prerequisites - varies by industry Both beginner and intermediate labs are offered - typically “Getting Started” and two more advanced labs constitute a day-long workshop Prerequisites - same as self-paced labs.
- 4. 4 ITINERARY 04:00 – 04:30 Deep Learning Introduction and Pre-Work demo 04:30 – 05:00 Image Classification deployed to Jetson TX1 05:00 – 05:20 JetBot Presentation & Demo 05:20 – 05:50 Object Detection on Jetson TX1 05:50 – 06:00 Wrap-up / Q&A / Additional Jetson time
- 6. 6 WHAT THIS LAB IS • An introduction to: • Deep Learning • Workflow of an end-to-end Deep Learning pipeline • Deploying trained DNNs • Hands-on exercises using Caffe/TensorRT on Jetson TX1
- 7. 7 WHAT THIS LAB IS NOT • Intro to machine learning from first principles • Rigorous mathematical formalism of neural networks • Survey of all the features and options of Caffe, DIGITS, or other tools • A deep dive into the hardware of Jetson TX1
- 8. 8 ASSUMPTIONS • No background in Deep Learning needed • No robotics experience needed • Understand how to: • Work in a Linux command-line environment • Basic programming skills in Python and C/C++
- 9. 9 TAKE AWAYS • Understanding of the workflow of Deep Learning • Ability to deploy a trained convolutional neural network • Comprehend “pixels to actions” on a Jetson TX1 platform
- 10. 10 DEEP LEARNING DEVELOPMENT CYCLE
- 11. 11 IMAGE CLASSIFICATION — Classify an entire image as one class — Works better for close up images — We’ll be using AlexNet initially “trained a large, deep convolutional neural network to classify the 1.3 million high- resolution images in the LSVRC-2010 ImageNet training set into the 1000 different classes. On the test data, we achieved top-1 and top-5 error rates of 39.7% and 18.9% which is considerably better than the previous state-of-the-art results”
- 12. 12 CREATING OUR MODEL - SUMMARY
- 13. 13 LOAD DATASET – IMAGENET DATA Different options will be presented based upon the task
- 14. 14 CHOOSE MODEL – ALEXNET IN CAFFE Differences may exist between model tasks Can anneal the learning rate Define custom layers with Python
- 15. 15 TRAINED MODEL Annealed learning rate Loss function and accuracy during training
- 16. 16 DEPLOYING OUR MODEL – YOUR TURN
- 17. 17 TASK 1 1. Open a terminal on the Jetson 2. Go to ~/01-classification 3. Run ./static_classify.sh grannysmith.jpg — Try other images in the folder 4. Try nvidia.jpg — What is different about this image? Static Image Classification
- 18. 19 TASK 2 1. Run ./webCamClassify.py — Test some convenient items 2. Find IDs with their associated class names in synset_words.txt 3. Modify webCamClassify.py and change the object IDs — Function is alert — Try using “computer mouse” or “granny smith” or “bottle cap” 4. Re-run ./webCamClassify.py — Should see alerts when holding up recognized object Live Camera Classification
- 19. 20 HOW TO STOP TERMINATOR Bananas
- 20. 21 OBJECT DETECTION — Draws bounding boxes around objects within an image — We will use a DetectNet-based model trained to recognize “bottles” — Trained on Azure K80 GPUs — MS Coco dataset converted to KITTI format for DIGITS to use
- 21. 22 TASK 3 1. cd ~/02-detection; mkdir bottlenet 2. Extract ~/bottlenet.tgz to ~/detection/bottlenet 1. cd bottlenet; tar xzf ~/bottlenet.tgz — This simulates downloading model from Azure and deploying 3. From ~/02-detection, run ./imageDetect.py sodagroup.jpg 4. You should see an image with bounding boxes drawn on bottles. Static Object Detection
- 22. 23 TASK 4 1. Modify imageDetect.py to do something different with its inference. — For example: — Print the number of bottles — Calculate the % of the image that consists of bottles. — Display the results differently than big blue opaque rectangles. — Come up with your own algorithm! Detect Multiple Bottles
- 23. 24 TASK 5 1. Run ./webCamDetect.py — Hold up bottle in front of camera and look for rectangles on screen and printout indicating whether bottles are detected. 2. Modify code to change the output based on number of bottles — If really ambitious, can changed output based on size of bounding box Live Camera Object Detection
- 24. TENSORRT
- 25. 2626 NVIDIA DEEP LEARNING SOFTWARE PLATFORM NVIDIA DEEP LEARNING SDK TensorRT Embedded Automotive Data center TRAINING FRAMEWORK Training Data Training Data Management Model Assessment Trained Neural Network developer.nvidia.com/deep-learning-software
- 26. 2727 NVIDIA TensorRT High-performance deep learning inference for production deployment developer.nvidia.com/tensorrt High performance neural network inference engine for production deployment Generate optimized and deployment-ready models for datacenter, embedded and automotive platforms Deliver high-performance, low-latency inference demanded by real-time services Deploy faster, more responsive and memory efficient deep learning applications with INT8 and FP16 optimized precision support 0 1,000 2,000 3,000 4,000 5,000 6,000 7,000 2 8 128 CPU-Only Tesla P40 + TensorRT (FP32) Tesla P40 + TensorRT (INT8) Up to 36x More Image/sec Batch Size GoogLenet, CPU-only vs Tesla P40 + TensorRT CPU: 1 socket E4 2690 v4 @2.6 GHz, HT-on GPU: 2 socket E5-2698 v3 @2.3 GHz, HT off, 1 P40 card in the box Images/Second
- 27. 2828 TENSORRT • Image Classification (AlexNet, GoogleNet, VGG, ResNet) • Object Detection • Segmentation Networks Supported Not Yet Supported • RNN/LSTM • 3D convolutions • Custom user layers
- 28. 2929 TENSORRT • Convolution: Currently only 2D convolutions • Activation: ReLU, tanh and sigmoid • Pooling: max and average • Scale: similar to Caffe Power layer (shift+scale*x)^p • ElementWise: sum, product or max of two tensors • LRN: cross-channel only • Fully-connected: with or without bias • SoftMax: cross-channel only • Deconvolution Layers Types Supported
- 29. 30 TENSORRT Workflow Training Framework OPTIMIZATION USING TensorRT RUNTIME USING TensorRT PLANNEURAL NETWORK developer.nvidia.com/tensorrt
- 30. 3131 TENSORRT Optimizations • Fuse network layers • Eliminate concatenation layers • Kernel specialization • Auto-tuning for target platform • Tuned for given batch size TRAINED NEURAL NETWORK OPTIMIZED INFERENCE RUNTIME developer.nvidia.com/tensorrt
- 31. 32 GRAPH OPTIMIZATION Unoptimized network concat max pool input next input 3x3 conv. relu bias 1x1 conv. relu bias 1x1 conv. relu bias 1x1 conv. relu bias concat 1x1 conv. relu bias 5x5 conv. relu bias
- 32. 33 GRAPH OPTIMIZATION Vertical fusion concat max pool input next input concat 1x1 CBR 3x3 CBR 5x5 CBR 1x1 CBR 1x1 CBR 1x1 CBR
- 33. 34 GRAPH OPTIMIZATION Horizontal fusion concat max pool input next input concat 3x3 CBR 5x5 CBR 1x1 CBR 1x1 CBR
- 34. 35 GRAPH OPTIMIZATION Concat elision max pool input next input 3x3 CBR 5x5 CBR 1x1 CBR 1x1 CBR
- 35. 3636 INT8 PRECISION New in TensorRT ACCURACYEFFICIENCYPERFORMANCE 0 1000 2000 3000 4000 5000 6000 7000 2 4 128 FP32 INT8 Up To 3x More Images/sec with INT8 Precision Batch Size GoogLenet, FP32 vs INT8 precision + TensorRT on Tesla P40 GPU, 2 Socket Haswell E5-2698 v3@2.3GHz with HT off Images/Second 0 200 400 600 800 1000 1200 1400 2 4 128 FP32 INT8 Deploy 2x Larger Models with INT8 Precision Batch Size Memory(MB) 0% 20% 40% 60% 80% 100% Top 1 Accuracy Top 5 Accuracy FP32 INT8 Deliver full accuracy with INT8 precision %Accuracy
- 36. 37 TENSORRT AVAILABILITY http://developer.nvidia.com/tensorrrt Register to download Sign up for early access testing Learn More TensorRT 1 stable version and testing for TensorRT 2
- 37. 38 TASK 6-1 1. Change to ~/03-detection-RT/jetson-inference/data/networks 2. mkdir bottlenet; cd bottlenet 3. Extract bottlenet: tar xzf ~/bottlenet.tgz 4. Remove the Python layer: patch -p0 < ../../../../deploy_bottlenet.patch 5. If you want to see what the patch did: cat ../../../../deploy_bottlenet.patch Use TensorRT for Better Performance
- 38. 39 TASK 6-2 1. Change to ~/03-detection/jetson-inference/build/aarch64/bin 2. Run ./detectnet-camera bottlenet 1. Notice the 5x speedup in performance! 3. Modify ~/03-detection/jetson-inference/detectnet-camera/detectnet-camera.cpp 1. Change display algorithm as previous tasks 2. Go to ~/detectnet/jetson-inference/build and type make to re-compile Use TensorRT for Better Performance
- 39. 40 UPDATING YOUR DATASET — Find images/video that do not work correctly — Label them — Use tools such as Sloth, MathWorks, or others for labeling images — Upload new data to Azure — Retrain network to improve accuracy — Deploy
- 40. 41 WHAT’S NEXT • Use / practice what you learned • Move your data to Azure and start training • Buy a Jetson! • Discuss with peers practical applications of DNN • Reach out to Microsoft and the Deep Learning Institute
- 41. 4242 WHAT’S NEXT WITH DLI …for the chance to win an NVIDIA SHIELD TV. Check your email for a link. TAKE SURVEY Check your email for details to access more DLI training online. ACCESS ONLINE LABS Visit www.nvidia.com/dli for workshops in your area. ATTEND WORKSHOP Visit https://developer.nvidia.com/join for more. JOIN DEVELOPER PROGRAM
- 42. 4343*Limit five per person **Limit one per student/instructor JETSON TX2 DEVELOPER KIT GTC Show Special: Just $399* EDU Discount: Just $299** Available at the GTC Gear Store all week
- 43. 4444 May 8 - 11, 2017 | Silicon Valley | #GTC17 www.gputechconf.com Enjoy the world’s most important event for GPU developers May 8 – 11, 2017 in Silicon Valley INNOVATE Hear about disruptive innovations from startups DISCOVER See how GPUs are creating amazing breakthroughs in important fields such as deep learning and AI CONNECT Connect with technology experts from NVIDIA and other leading organizations LEARN Gain insight and valuable hands-on training through hundreds of sessions and research posters
- 44. 4545 FINAL TASK 1. cd /home 2. sudo /usr/local/clean_home/restore.sh (password is ubuntu) Restore Jetson