π¬ππ½πππππ is here!
Itβs an library focused on real-time vision models & ease of use. The key differentiating factors are:
Run inference with just a couple of lines of code
Ready with some real-time vision models
Runs in a livebook
Ready with CPU-capable models
If you are interested i origin story and lessons learnt along the way, please check the blog post.
If you would like to give it a try, please check out the hex package.
If you are interested in the code itself, please checkout the Github repo.
Great addition to the ecosystem! its a good time for me to remove the classification and segmentation functions from Image in favour of pointing people at ExVision.
The version for the exs import is wrong. You are now on 0.3.0, while it says to import 0.2.0 in there.
Only Elixir 1.16 and above should be used. I am not sure this is mentioned somewhere but I had to navigate the errors to fix this.
The module name has changed from MobileNetV3 to MobileNetV3Small, which I found digging through the library, because my load() function was crashing. This change is not reflected in the hex documentation additionally to being wrong from the example.
The inference took 5min 44sec to do a cat image with the example code. The example code being loading the model and using it on the picture.
The inference took 1.2sec with EXLA turned ON for the first attempt, and around 170ms after that. So EXLA monstrously improves performance.
Awesome library! I am off to detect some kittens and play around with it some more.
Cheers!