Problem with custom yolov5 model

liu

Hi, friends
I met a big problem in my project.
I use a yolov5 custom model trained with pytorch. Following the steps, 1convert to onnx, 2change the top output layer, 3convert openvino, 4convert to blob, 5add model *.json. The result is wrong. I can not find what happened after these converting steps. (From guidance of https://www.github.com/luxonis/depthai-ml-training/blob/master/colab-notebooks/YoloV5_training.ipynb)
Who have the REAL experience of produce a custom yolov5 model for oak-d in-device inference. help!

And the Depthai api 'YoloDetectionNetwork' is designed to yolov3&yolov4, it's not friendly to yolov5, and the guidance is too complicated and easy leading to some error.

Thanks

Is there some plan to upgrade 'YoloDetectionNetwork' module to fit yolov5?
After all yolo is a very popular framework especially yolov5, and v3/v4 is a little old.

barney2074

liu and the guidance is too complicated and easy leading to some error.

Hi Liu

yes, I agree with this.
I imagine it is difficult to keep documentation up to date with multiple versions of different applications, but all of the documentation I've looked at has some kind of errors- seems like it is out of date, or something has changed

Andrew

liu

barney2074
Hi barney
yes,
I suggest you read this one https://docs.luxonis.com/_/downloads/en/latest/pdf/
something like a little more logical and continues than html docs. (i just find 🙂)

liu

My source model is .pt, then export it to .onnx, and test it ok with same inference result, and then modify the .onnx with the output layer following guidance (https://www.github.com/luxonis/depthai-ml-training/blob/master/colab-notebooks/YoloV5_training.ipynb), a pruned.onnx produced.

From pruned.onnx to final .blob, I have tried 3 different ways following doc advice.
1 onnx-->OpenVINO IR(using optimizer mo.py)-->BLOB(using online converter) 41,704KB
2 onnx-->BLOB using online converter 41,737KB
3 onnx-->BLOB using 'blobconverter' module 41,714KB
but they produce different size and different inference result model...

liu

In third way using 'blobconverter' module, code:

import blobconverter
blobconverter.from_onnx(
model="C:\Project\rdd\oak\models\pruned.onnx",
output_dir="C:\Project\rdd\oak\models",
data_type="FP16",
shaves=6,
use_cache=False,
optimizer_params=[]
)

what 'optimizer_params' means, is there any example? I can't find in docs.

And my .json:

{
"nn_config":
{
"output_format" : "detection",
"NN_family" : "YOLO",
"input_size": "640x640",
"NN_specific_metadata" :
{
"classes" : 4,
"coordinates" : 4,
"anchors" : [10,13, 16,30, 33,23, 30,61, 62,45, 59,119, 116,90, 156,198, 373,326],
"anchor_masks" :
{
"side80" : [0,1,2],
"side40" : [3,4,5],
"side20" : [6,7,8]
},
"iou_threshold" : 0.5,
"confidence_threshold" : 0.2
}
},
"mappings":
{
"labels":
[
"D00",
"D10",
"D20",
"D40"
]
}
}