Hi @kindred wharf @sick quail @pine jasper @vale sigil it gave 97% detection accuracy before it is exported to ONNX model which gave 45% accuracy
#Multi-Object Detection Template, detection accuracy is below 50% after it is exported in ONNX
tired mirage
Multi-Object Detection Template, detection accuracy is below 50% after it is exported in ONNX
!pip install protobuf<4.21.3
!pip install onnx>=1.9.0
!pip install onnx-simplifier>=0.3.6 --user
!python export.py --weights ./runs/train/$JOB_NAME/weights/best.pt
--grid --simplify --export-snapml
--img-size $IMG_SIZE $IMG_SIZE --max-wh $IMG_SIZE
steep hamlet
Hi! Is there any compression set on the model asset?
If you could share the model we can have an ML engineer take a look
tired mirage
https://we.tl/t-ZPcylpWitr ONNX file
tired mirage
@kindred wharf @steep hamlet i tried 2 models trained from above template and both gave exact results after export so please look into the template
steep hamlet
@tired mirage thanks for sharing the model, but i can not import it for some reason
Guide here https://docs.snap.com/lens-studio/references/templates/ml/multi-object-detection#importing-model asks you to set scale and bias when you import the model. Did you follow this step?
tired mirage
@steep hamlet yes i did
tired mirage
@steep hamlet sorry i sent the wrong model ,please find the model here https://we.tl/t-hTDUAwxyOP
steep hamlet
taking a look!
is this from default training notebook of from roboflow? output size is differentin this model
And just to clarify - do you use default template from LS homepage or the one provided on roboflow website?
steep hamlet
Anyways i've forwarded your questions!
tired mirage
@steep hamlet this is the default multi-object detection notebook from snapml templates from github, just i trained it with different dataset from roboflow also used default template from LS homepage ,this models detects the tetra packs
tired mirage
Hey @steep hamlet any updates on this?
steep hamlet
Hi @tired mirage ! ML engineer has advised to look into these questions : what’s the shape of input images used at training ? (LS project uses stretched image) Also what was the range of the input images used at training ([0,1], [-1,1] or [0,255])? Maybe you need different scale and bias. Also what is the number of classes detected by the model? So I can set up my project similarly
tired mirage
hey @steep hamlet
Input Shape: stretched
Input Images Range: [0,255] H:244 W:244
Number Of classes:1
Label: "carton"
Scale:0.0039215,Bias:0
steep hamlet
what about the training data, was it from -1 to 1 or 0-1? Do you mind sharing your training notebook and project? I tried different settings but could not make it work . You may share files with [email protected]
tired mirage
@steep hamlet sure
tired mirage
Hi @steep hamlet any updates on this?