PyTorch Mobile and Android Neural Networks API

on YouTube ↗06m

" summary

PyTorch mobile adds direct offload to the Android Neural Networks API, supporting over 100 ops on Android 11 (including LSTM, float and quantized types, and control flow) and providing a nightly build plus a simple torchscript conversion flow that sets NHWC and saves an NNAPI model. Measured wins include ML Kit 9x latency, Google Lens ~4x latency, and Facebook Messenger 2x latency and 2x power reductions; roadmap lists Mask R-CNN and more operators.

AI-written summary · about

" chapters (4)

1. 0:00Intro
2. 0:42Android Neural Networks API
3. 2:43PyTorch Mobile
4. 3:47Realtime Segmentation

" related talks

• 38mNEURAL NETWORKS ON YOUR MOBILE PHONE. DIVING INTO THE NN APIJürgen Fey, Androidian · droidcon Berlin 2018
• 09mPyTorch MobileDavid Reiss · PyTorch Developer Conference 2019
• 25mOn-Device LLM Inference on Android With ExecuTorch and Qualcomm QNNShivay Lamba, Kartikey Rawat · PyTorch Conference Europe 2026
• 39mAI/ML on mobile/android with TensorFlowKarim Beguir · Droidcon Dubai 2018
• 09mPyTorch Mobile: Past, present and 2.x futureRAZIEL ALVAREZ GUEVARA · PyTorch Conference 2022
• 46mFire Up Your Smartphone Neurons with On-Device Deep LearningQian Jin, Yoann Benoit, Sylvain Lequeux · XebiCon'17
• 43mWhat’s new in Android machine learningHoi Lam, Laurence Moroney, Matej Pfajfar‎, Dong Chen · Google I/O 2019
• 12mWhat’s new in PyTorch 1.3Lin Qiao · PyTorch Developer Conference 2019