how to export to onnx? about yolov9 HOT 15 CLOSED

need to export onnx model，please integrated them into this repo quickly!

from yolov9.

#20 should fix this 👍 (+ converted models can be found here)

from yolov9.

Hello!!
Fix the error shape = tuple((y[0] if isinstance(y, tuple) else y).shape) # model output shape
AttributeError: 'list' object has no attribute 'shape' "
Passing y directly to shape=y

y is a tuple but y[0] is a list. And this list has no .shape.

My export looked like this:

!python export.py --weights /content/drive/MyDrive/IAGeneration/Yolov9/yolov9/runs/train/exp3/weights/best.pt
--batch-size 1 --imgsz 640 --include torchscript onnx

My best.pt has 133.47 M
While my best.onnx has 262.22 M.
Is this normal?

The inference worked with:

!python detect.py --data /Yolov9/projdocyolov9red.yaml --weight best.onnx --conf 0.50 --source /data/imgs --device 0 --save-txt

But it took about 3 minutes to detect 100 images. This is normal?

from yolov9.

The re-parameterization functions are provided in yolov7 repo, but currently we have not integrated them into this repo.

from yolov9.

I can confirm that @xenova exports work

from yolov9.

Export functions are not yet supported. The models should do re-parameterization first, then do export.

from yolov9.

how to re-parameterization?

from yolov9.

how to tensorrt forward? I need to export to onnx，then tensorrt.

from yolov9.

need to export onnx model，please integrated them into this repo quickly!

Modify this code "if isinstance(m, (Detect, V6Detect))" to "if isinstance(m, (Detect, DualDDetect))" ，and you can get the onnx

from yolov9.

need to export onnx model，please integrated them into this repo quickly!

Modify this code "if isinstance(m, (Detect, V6Detect))" to "if isinstance(m, (Detect, DualDDetect))" ，and you can get the onnx

I don't think that really works

from yolov9.

need to export onnx model，please integrated them into this repo quickly!

Modify this code "if isinstance(m, (Detect, V6Detect))" to "if isinstance(m, (Detect, DualDDetect))" ，and you can get the onnx

I don't think that really works

This solution works for me.I have completed the ONNX export, as well as inference with TRT and ONNX Runtime (ORT).Furthermore, the ONNX used by TRT includes an NMS node.

from yolov9.

@xenova's solution works

from yolov9.

（修改代码 “if isinstance（m， （Detect， V6Detect））” 改为 “if isinstance（m， （Detect， DualDDetect））” ，就可以得到 onnx）在进行如下修改后又得到一个新问题

Fusing layers...
Model summary: 724 layers, 51141120 parameters, 0 gradients, 238.7 GFLOPs
Traceback (most recent call last):
File "export.py", line 607, in
main(opt)
File "export.py", line 602, in main
run(**vars(opt))
File "/home/mouxing/anaconda3/envs/V8/lib/python3.8/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context
return func(*args, **kwargs)
File "export.py", line 507, in run
shape = tuple((y[0] if isinstance(y, tuple) else y).shape) # model output shape
AttributeError: 'list' object has no attribute 'shape'

from yolov9.

You can found how to export onnx model with the following
https://github.com/AICVer/yolov9.infer

from yolov9.

@xinsuinizhuan I used it:
https://github.com/thaitc-hust/yolov9-tensorrt

I also first transfer weights.pt to ONNX with the NMS module, and then transfer to TensorRT Engine for inference.

from yolov9.