Highlights of PyTorch ecosystem days
How one of the most popular DL frameworks builds the whole ecosystem of software and packages around itself.
What is all about
Recently Pytorch held an ecosystem day - event, that brings developers and practitioners together to share how they integrate their recent software/packages with Pytorch.
When I first saw it, I registered straight away. But boy, I must say, the schedule and organizational part disappointed me. The info and schedule were in Google docs (and yes, Google develops a direct competitor framework), it was big, overloaded, and hard to understand. In comparison, Google created a simple, easy to use and fancy webpage with the schedule of their similar Google I/O event. As a Pytorch fan, I feel a bit ashamed.
Still, the event is full of content! It’s great that Pytorch and the Community support each other in better tools development.
To the content!
Summary:
First of all, I see that products based on Pytorch are growing rapidly, and packages are becoming more and more domain-specific.
There are wrappers over PyTorch like Pytorch-lightning, Ignite, fastai, Catalyst - they meant to make high-level API with lots of SOTA features implemented.
The level of specification of pytorch ecosystem goes deeper each year - we now can find not only CV/NLP packages but also biomedical imaging, audio, time-series, reinforcement learning. 2D/3D Augmentation libraries, MLOps solutions. Some packages help to diagnose your models, like finding model drift, mitigate unfairness, compress them. I even saw a separate MLOps package for automated driving.
The whole ecosystem is here, and it quite interesting to dive into it. There are in total 60 (!) tools, libraries, and packages there. Soon, they will need an advanced filtering/categorization system.
It is amazing, how fast this ecosystem grows, curious to see the next step of this evolution.
Posters:
They are here - https://pytorch.org/ecosystem/pted/2021
I decided to start from the posters. and not from the videos, since imo posters are more interesting.
There are posters for such wide-known products like huggingface, pytorch lightning, etc. - I won’t mention them. Below is the list of something new/interesting that I personally want to mention.
PyTorch development in VS Code - they have profiler and tensorboard integration. I am using PyChram right now, but thinking more and more about switching to VS code.
AI Model Efficiency Toolkit (AIMET)- about model compression
Enabling PyTorch on AMD Instinct™ GPUs with the AMD ROCm™ Open Software Platform - nice to see that AMD is catching up finally
TorchStudio, a machine learning studio software based on PyTorch - product is not ready but looks interesting
High-fidelity performance metrics for generative models in PyTorch
UPIT: A fastai Package for Unpaired Image-to-Image Translation
CompressAI: a research library and evaluation platform for end-to-end compression
Videos:
There are opening talk videos for EMEA and APAC.
They are more or less the same. To be honest, I would recommend skipping those and watch cuts from them instead.
My Journey to PyTorch by Piotr Bialecki @Nvidia
You probably know Piotr Bialecki if you are using pytorch - he is the guy who answers most of the questions in pytorch forum. Piotr is the Technical Lead of The PyTorch Team @ NVIDIA.
He speaks about his path in ML/DL, how he started to use pytorch, he reflects on the past and looks forward.
You should watch that, if you need a little inspiration.
PyTorch Release by Joe Spisak
You should watch this video, If you want to learn more about latest pytorch release features from PyTorch Product Lead @Facebook AI. Joe speaks aboutpython code transformations with FX (it is a toolkit for pass writers to facilitate Python-to-Python transformation of nn.Module instances - not sure everyone will need this)
torch.linalg - provides NumPy-ish linear algebra operations support
torch.fft that cover discrete Fourier transforms and related functions
pytoch native profiler (yay!) with tensorboard plugin
distributed training (including support of AMD (sic!) GPUs)
PyTorch Partner Collaborations by Geeta Chauhan
Geeta leads AI Partnership Engineering at Facebook AI. Nice lecture for the ones who want to know recent collaboration features. She talks about:
more details on profiler, with use cases, integrations into partnering frameworks, etc.
scaling in production with torchserve (meant to be model serving framework for PyTorch that makes it easy to deploy trained PyTorch models performantly at scale without having to write custom code)
MLOps with Kubeflow (building pipelines)
MLOps with MLFlow (from model artifact serving to auto-tracking of pytorch training metrics, hyperparameters search)
Disney's Creative Genome by Miquel Farré
Miquel is Senior Technology Manager @Disney. He speaks about Creative Genome - a project aimed to provide curated time-based metadata. On top of this metadata, they are building their models to recognize their characters in movies/animations/comics/series, detect certain activities and events.
Community Updates by Suraj Subramanian - PyTorch Developer Advocate @Faceook AI
Well, the name speaks for itself. Suraj gives some statistics on contributions into PyTorch, and also speak a bit about the ecosystem.
Applications of AI and PyTorch in Asia Pacific by Ritchie Ng
Ritchie Ng is the CEO of Hessian Matrix, an AI systematic global hedge fund based in Singapore. He provides an overview of APAC, speaks about CV in e-commerce and retail, NLP in finance.
More resources:
Contributor Newsletter - Includes curated news including RFCs, feature roadmaps, notable PRs, editorials from developers, and more to support keeping track of everything that’s happening in our community.
Contributors Discussion Forum - Designed for contributors to learn and collaborate on the latest development across PyTorch.
PyTorch Developer Podcast (Beta) - Edward Yang, PyTorch Research Scientist, at Facebook AI shares bite-sized (10 to 20 mins) podcast episodes discussing topics about all sorts of internal development topics in PyTorch.
Adel