I'm a graduate student at IUPUI. I am working with pruning of Squeezenext architecture for my thesis. So I want to deploy my model on a Cortex-M. I am using Pytorch framework. So could anyone help me in using CMSIS with Pytorch?
Thanks in Advance!!!
Hi RavitejaChappa I have moved this to the Machine Learning forum for you.
Thank you so much!
Right now we're still working on support for PyTorch, especially for M class. I would recommend looking at Tensorflow Lite Micro for the time being.
Hi Karl Fezer
We are also highly relying on PyTorch in our development work. Can you share the latest state of the support for PyTorch.
Specifically, we are planning to work on the new NPU architecture on the IMX8-Plus board.
Is there a way to efficiently used PyTorch models on this platform already, ideally via the C++ frontend?
Hi, I am working on a converter from PyTorch to CMSIS-NN.
You can find the code at https://github.com/BCJuan/torch2cmsis
Contributions are more than welcome.
The support is really restricted right now since I have just begun the library. Hope to be able to improve it in the next months :)
Hey Juan, this is great! Can't believe this isn't provided by arm
In these years, deep learning models have become popular in the real-time embedded application, but there are many complexities for hardware deployment as of few resources such as memory, energy and computational power. Recent research in the field of deep learning focuses on reducing the model size of CNN by various compression techniques like Pruning, Architectural compression,and Quantization.
Wow, sorry I missed this for so long. I would say we are looking into it, but consistently recommend testing in PyTorch and then rebuilding in Tensorflow if TFLu is the end goal.
Watch "Bringing PyTorch Models to Cortex-M" at AI Virtual Tech Talk, www.youtube.com/watch
Watch "Bringing PyTorch Models to Cortex-M" at ARM AI Virtual Tech Talk, on www.youtube.com/watch
View all questions in Machine Learning forum