We are a group of PyTorch users and developers in and around Munich to gather every now and then for talks and general exchange.
Currently we use the Munich Applied DL and PyTorch meetup group to organize events. If you want to host an event or give a talk, do contact Thomas.
PyTorch grew to be one of the most popular libraries used for machine learning thanks to all of you. However, we’re not really done with it just yet. Even though many of the improvements might feel incremental, there are some changes that I think can have a profound impact on how we write and express the same computations in some time. In this talk I’ll cover the most exciting (in my opinion) improvement coming up — dispatcher changes — which doesn’t sound super interesting alone, but is actually fascinating and is at the very core of everything we do in PyTorch. After that, I’ll move on to a project that is a little bit independent of the core, but might make it a lot easier to deal with shapes in your programs.
Adam is one of the original authors of PyTorch. His background is in Computer Science and Mathematics.
Carla will walk us through the design and implementation of a Parking Space Detection System which will be linked to a Telegram Bot to provide the information about the detected available parking spots on the street. She will show us how by leveranging the features and tools made available by PyTorch, a smooth transition from the development to the deployment of Deep Learning models on edge devices can be achieved.
Bio: Carla uses Pytorch to develop and deploy ML algorithms in her current position at BMW Group. She has gathered experience in the field of ML by working in different companies. She was Team Lead during the development of Samsung's Voice Assistant for the German market and has developed multiple deep learning algorithms to solve problems in the fields of Computer Vision and Biometrics.
PyTorch is well-liked for developing and training models.
We look at the more recent developments to make PyTorch models even faster and deploy PyTorch models outside Python. One of the key features to help this is the PyTorch Just-in-Time-Compiler (JIT).
First we look at how you can use the JIT to make your model run faster in places where you would traditionally have to create custom (CUDA) kernels.
We then turn to exporting a PyTorch model to C++ using the JIT. While many models, such as ResNets or Neural Style can be traced out of the box, we also look at how more complex, cutting-edge models can be exported with ease and good style. Then you can embed your models in a C++ application, shedding the Python dependency.
Finally, we take a look at running PyTorch on Android. The traditional way to export models to devices is through ONNX, but with PyTorch's C++ library running on the device, things become much easier. We show how this can be done and also discuss some of the (current) limitations.
Thomas worked with and on PyTorch since early 2017. He is a prolific contributor to PyTorch with more than 80 features and bugfixes, from implementing CTC Loss, to speeding up Batch Norm, to enhancing the JIT capabilities. In 2018 he founded the consultancy MathInf and helping clients with mathematical modelling and AI implementation.
Piotr uses PyTorch for his work on all sorts of deep learning models. As the most active contributor to the PyTorch forums, he is one of the friendly faces of the PyTorch community and has helped hundreds of PyTorch users with their modelling questions. He holds a B.Sc. in Biomedical Engineering and a M.Sc. in Information Technology.