Let’s say the NodeMCU board with 8 megabytes of flash memory, it is possible to connect external flash memory via SPI.
I understand that this probably sounds crazy. But in theory?
I have long been looking for AI projects that would have the widest possible application and the ability to be installed on a chip or programmable controller.
If you found a way to compile/run ctranslate2 and synthesize a student model Training efficient neural network models for Firefox Translations - Mozilla Hacks - the Web developer blog, why not, theoretically possible, practically difficult.
I’m also unsure of the performance.
Thank you very much! This will help me. Indeed, performance problems may arise. It will be necessary to measure what power is required. I’m still thinking about using an ESP32 controller, maybe it will be enough…
Running on an Arduino will definitely be practically difficult but if you have any success I’d love to hear an update.
I think the CTranslate2 models require >100MB of memory to run.
I have a demonstration of using CTranslate2 directly from C++ if that helps:
Interesting. I see that during initialization a model with the extension “…argosmodel” is downloaded. Size 437 MB.
Can I just take the model (file extension “.bin”) from the “firefox-translations-models” repository and drop it into the “models” folder? Or is there some other format? 