#How to config ML Asset for Model Zoo onnx files

16 messages · Page 1 of 1 (latest)

honest thorn
#

When I add the house_detection.onnx Model Zoo file, I get a ML Asset with parameters that look more or less like defaults values. Do those all come from the ONNX file or do i need to know something about the expected resolution and output shapes from that model?

LensStudio Does someone have a project with house_detection.onnx working?

My next challenge is not knowing all the labels and to what channels they should be configured. Ultimately, I want to show a bunch of garment augmentations based on what's in the scene. I don't even need the specific bounding boxes, just multi-object detection.

Thanks in advance for an insight to figure this out..

Model Zoo in snap docs

Screenshot:

honest thorn
#

hi anyone tried this yet?

honest thorn
#

so with a little research I found the ONNX files do have a bunch of useful info about the shapes of the input and output. tried two free viewers. Netron was teh most light-weight and useful : https://github.com/lutzroeder/netron

halcyon shadow
honest thorn
#

Thanks for the reply!
Any examples using the house detection out there? I'm not seeing how to know the mappings from channels to labels ..

halcyon shadow
#

that model is not too fun, it's meant to be used with single object detection template. I would rather surrest to train your own multi object detection using Roboflow - then you''ll be able to differentiate between classes

#

if you want to classify only ( don't care about bounding boxes) you could easilly plug in some existing models into Multi-Class Classification template

#

for example Tiny Image Net

#

Or even use a built in Scan feature

honest thorn
#

thx. ill try !

honest thorn
#

In trying a few Classification models i get an error about 'QuantizeLinear' layer type not supported.. anything that can be done?

  • vgg16-12-int8.onnx
  • efficientnet-lite4-11-int8.onnx
honest thorn
#

Ok, so I got some basic tutorials working and trained the multi--object with various data sets. super fun stuff thanks!

I still have a couple question, @halcyon shadow :

Will a roboflow trained model work in snapdirectly or? is it more abouit using a working notebook and model with roboflow datas?

How would the Multi-Class Classification without bounding boxes come into snap lens? is that like transfer learning yolo7 with a different output?

thanks again!

halcyon shadow
#

Are you talking about Multi-Object Detection , not Multi-Class Classification ?

#

You should be able to import model from Roboflow to the template they provide on their website ( it is a bit altered)

#

and then for example disable visual representation of the boxes

#

when you have that project working we can look together!