Lukas Geiger

deep learning scientist | astroparticle physicist | software engineer

[Lukas Geiger @_lgeiger]

@KyleRayKelley @rahuldave @zeddotdev Feels like it's becoming a habit...
The first lines I write in any new language absolutely need to be a PR for @KyleRayKelley to review 🤣
Last time it was JS/CoffeScript this time Rust 🦀😎

[Lukas Geiger @_lgeiger]

@KyleRayKelley @zeddotdev @AtomEditor @nteractio Just looked at some old screenshots. Looks like we were copying what ipython did.
However, ipython only showed `In[*]` starting version in 8. Before that it would look something like `<ipython-input-1-6fcf9dfbd479> in <module>()` so even ipython didn't have line numbers.

[Lukas Geiger @_lgeiger]

@KyleRayKelley @nathansobo @zeddotdev @cursor_ai @KyleRayKelley our last chat actually got me thinking about the same thing. Zed+AI+Hydrogen would be a pretty sweet combo 🤔

[Plumerai @plumerai]

We made our ultra-fast and ultra-small inference engine for Arm Cortex-M processors even smaller. We now require 49% less RAM for 8-bit MobileNetV2 than TensorFlow Lite for Microcontrollers with Arm’s CMSIS-NN kernels.
Read the article here: https://t.co/Y7llUkUveT

[Plumerai @plumerai]

🏆 We built the world’s fastest and most memory-efficient deep learning inference software for Arm Cortex-M microcontrollers 🏆 for both Binarized Neural Networks and for 8-bit deep learning models.

https://t.co/Y7llUkUveT

[Plumerai @plumerai]

Our latest paper has been accepted to the #MLSys2021 conference! Congratulations Tom, Arash, Adam, Lukas, Tim, Leon, Jelmer and Koen!

We wrote a short blog post about the paper and the importance of Larq Compute Engine: https://t.co/tKZvgQYnn3

[Plumerai @plumerai]

It was great to see so many people attending our #tinyml webcast earlier this week! In case you missed the event, a recording of the talk is now online along with a short blog post explaining the main takeaways from our presentation: https://t.co/rwqQPPD8oC

[Pete Warden @petewarden]

A fascinating talk on binarized neural networks for #tinyml by Lukas of @plumerai is now up at https://t.co/0mVeqnGtvk, with slides at https://t.co/Cj3J1gHrVb

[Plumerai @plumerai]

This is starting in a few hours, come join our #tinyML webcast!

[Plumerai @plumerai]

Come join us on Tuesday for our @tinyMLTalks webcast! @_lgeiger will explain how we use 🚀Binarized Neural Networks🚀 to enable real-time deep learning on cheap, low-power edge devices. You can register here: https://t.co/JcMdenewX2

[Plumerai @plumerai]

We've released v0.4 of Larq Compute Engine!

Some of our internal work optimizing BNN performance on microcontrollers has made its way to mobile and Raspberry Pi, as have a few major architectural improvements.

🚀 Full change log: https://t.co/7P1Cn85lbI

[kamil @kamil_k7k]

We just published blog post about open source tools for #DeepLearning presented at our social event at the @iclr_conf!
Shout-out to all the co-authors for their great contributions and efforts. Without your help, this would not have been possible!

https://t.co/fma6o6u0Nu

(1/n)

[Plumerai @plumerai]

We’ve released Larq Compute Engine v0.3, featuring even faster BNN kernels and support for Python 3.8, TensorFlow 2.2, and Windows in the converter! 🎉

💻 Full release notes: https://t.co/ykUgHfzaXo
📚 Docs: https://t.co/awSf15mt66

[Plumerai @plumerai]

The Larq Compute Engine team pushed out LCE v0.2.1 earlier today—same models, up to 15% lower latency. https://t.co/ffRKxpMAVS

[Lukas Geiger @_lgeiger]

@fchollet Improved contributor experience. Moving Keras out of the tf monorepo will hopefully help, but currently having to wait hours for tf to build on a laptop just for a small Python fix and then waiting weeks to get a PR review doesn't encourage community contributions.

[Lukas Geiger @_lgeiger]

@fchollet One recommended way to save and resume training
- tf.train.Checkpoint is great but requires too much custom code
- https://t.co/Cs86LtGGqY_weights and ModelCheckpoint don't reload optimizer state or epoch
- https://t.co/Cs86LtGGqY doesn't play well with tf.distribute

[Plumerai @plumerai]

We wrote a blog post about how we’re applying a software 2.0 approach in the new module for state-of-the-art binarized neural networks in Larq Zoo (`zoo.sota`). It also features some exciting improvements across the whole Larq stack: https://t.co/k532X5Enga

[Plumerai @plumerai]

We’ve just released Larq v0.9, featuring:

🔲 Non-zero padding
💾 Easier export for binary weights
📋 Improved model summaries

And lots more! 🎉 https://t.co/yuKkNgcZ03

[Plumerai @plumerai]

More about Larq Compute Engine:

📑 Docs: https://t.co/E8YnZ4UOVG
👩‍💻 GitHub: https://t.co/BzFuQLPudv
💱 MLIR converter: https://t.co/JUeaLwYSVl
📱 Android deployment: https://t.co/5J6s434xfX
🥧 Raspberry Pi deployment: https://t.co/a2qIcRX02p