PyTorch Front-End Features: Named Tensors and Type Promotion

on YouTube ↗08m

" summary

PyTorch 1.3 introduces named tensors that attach dimension names (e.g., N C H W) so code can sum and align dimensions by name instead of position, enabling a single image-normalization function that works across formats via the align_as helper. It ships experimental eager-mode support, allows mixing named and unnamed tensors, and propagates names through ops though gradients are unnamed. The release also adds type promotion that auto-promotes mixed dtypes and scalar/tensor ops using numpy-like rules.

AI-written summary · about

" chapters (6)

1. 0:00Introduction
2. 1:41Code Example
3. 3:03Image Normalization
4. 5:07Name Tensors
5. 6:10Type Promotion
6. 7:50Outro

" related talks

• 28mOn TensorsNishant Sinha · The Fifth Elephant 2020
• 06mMaking PyTorch More "NumPy Compatible"Mike Ruberry · PyTorch Developer Day 2020
• 11mPyTorch 2.0: Dynamic Shapes SupportHorace He · PyTorch Conference 2022
• 24mWhat's New for Dynamic Shapes in PyTorch 2.1Edward Yang · PyTorch Conference 2023
• 19mGet to know PyTorch - a cutting edge open source AI framework for Machine LearningSahdev P. Zala · All Things Open 2023
• 54mKeynote: PyTorch 2.1 Technical Deep DiveJoe Spisak, Peng Wu, Mergen Nachin, Mark Saroufim, Yanan Cao, Will Constable, Mario Lezcano · PyTorch Conference 2023
• 19mIncreasing the Impact of AI Through Better Software (Pytorch Updates)Soumith Chintala · MLconf NYC 2019
• 45mTensorly: A Flexible Python Framework for Machine LearningAnima Anandkumar · PyData Los Angeles 2018