So, I'm trying to get into using AI with BlueIris but I'm discovering there is quite a bit of a learning curve. I'm setting up a new BlueIris instance on my system and I have an AMD Radeon RX 9070 XT (16MB) and what I've read said that I should use YOLOv5 .NET for AMD, but is that true? Should I use YOLOv5 .NET or is YOLOv8 the best and brightest now (that would work with a AMD GPU)?
Which YOLO to use?
- Thread starter Deadeye
- Start date
Having a little over 24 hours with AI enabled in BI6, I am pretty raw in all this too. I used the BI downloadable yolov8s.onnx You have 3 choices, n,m,s after you download & install the BI internal modules. N being the lightweight least accurate, M (medium) being the more intense heavy CPU but higher accuracy & "S" is the middle of the two. If you are concerned with performance, use the BI downloadable modules for BI6 (available via the tab on the AI tab in Settings) start with "S", if it is too hard on resources drop down to "N", if there is no performance impact & you want better accuracy, then go up to "M". I assume AMD CPU & AMD GPU, so it will just use the AMD CPU for all this as it only supports Nvidia or Intel GPU, so you would choose "No GPU" in that AI settings main tab.
From my understanding .net is mostly for CPU only. Yolo v5 for older graphics cards, and yolo v8 is better for newer cards.
OakleyFreak
Getting the hang of it
- Jun 2, 2016
- 59
- 43
def want custom module either ipcam-combined or ipcam-general. just run one in each camera not both
im running older BI 5.7 with CPAI 2.95
Running YOLO8 with a custom module for animals. on a GTX 4050 Super card
try the yolo8 your card should work fine as GPU setting
im running older BI 5.7 with CPAI 2.95
Running YOLO8 with a custom module for animals. on a GTX 4050 Super card
try the yolo8 your card should work fine as GPU setting
MikeLud1
IPCT Contributor
.NET can be use with all GPU makes including iGPU, out of all the testing I have done the .NET is the fastest.
Ohh I didn't know that. I've been running .net CPU medium for years (i7 6700 on a old dell). Average 400-500ms response times. Just for giggle I switched to GPU with the embeded GPU. Not much difference.
I do actually have a Nvidia Quadro P1000 on the way for this machine. I was initially planning on testing it out with v8 medium in hopes of a little better accuracy and fewer false detections.
I tried YOLO8 and it would not switch over to use my GPU. Turns out it's only for NVIDIA cards. If you have an AMD, you have to use the .net version as it's the only one with DIRECTML for non CUDA video cards (AMD). I guess I have my answer. Took a lot of tries to find this out.