search

Deeplite / neutrino

Public documentation of Deeplite Neutrino™ engine.
Creative Commons Zero v1.0 Universal
64 stars 5 forks source link

YOLO V3 #6

Open adrianosantospb opened 3 years ago

adrianosantospb commented 3 years ago

Hello,

I have a YOLOV3 model already trained and used on production. I'd like to optimize it and make it faster, smaller and energy-efficient as the Neutrino says. I have some questions and I will be happy if someone helps me with:

a) Machine' requirements to deploy: I'd like to know what are the machine requirements to deploy a YOLO V3 model after of neutrino process. Could I use it in Raspi for example? b) Machine' requirements to optimize: Do I need some specific hardware or technologies to run it? For example CUDA, AVX-512 or VNNI? c) Do I need to retrain a new model or could I use the model that I already have?

jason-deeplite commented 3 years ago

Hi Adriano,

Thank you for your interest in Neutrino. Below are the answers to you questions:

a)Neutrino is hardware agnostic. If your trained YOLOV3 in production is currently running on a Raspi, then it should be able after being optimize. If fitting it on a Raspi is your end goal, then it would depend on how successful Neutrino is on the compression.

b)You can use the same hardware that was used to train the model. The minimum dependencies for optimization machine are Ubuntu >=16.04, Python 3.6+ and torch ==1.4 with at least 1 CUDA compatible GPU. You can also activate horovod to use multi-GPU and/or multi-server to speed up the process.

c)You can use the model that you already have, and you will need the training data that was used to train the model.

Hopefully these will help you on your way to using Neutrino.

Let us know if you have any more questions.