[Training] Custom trained YOLOv7 network using ONNXRuntime in Java

[meeeee12]

Describe the issue

I am trying to use YOLOv7 to train a custom network and run it with ONNXRuntime in Java but I receive the following error:
ai.onnxruntime.OrtException: This tensor is not representable in Java, it's too big - shape = [0, 7].

When I convert https://github.com/WongKinYiu/yolov7/releases/download/v0.1/yolov7.pt to onnx using the suggested method from the YOLOv7 github page python export.py --weights yolov7-tiny.pt --grid --end2end --simplify --topk-all 100 --iou-thres 0.65 --conf-thres 0.35 --img-size 640 640 --max-wh 640 it works fine. I can also use re-parameterization on https://github.com/WongKinYiu/yolov7/releases/download/v0.1/yolov7_training.pt convert that to onnx and that also works. But when I try using transfer learning on yolov7_training.pt then re-parameterize and convert I receive the error. I have noticed that the working non custom trained versions have a shape of [3, 7]

I have found another issue with the same problem https://github.com/microsoft/onnxruntime/issues/7270 but that appears to be related to not using a square training/inference shape, however I am using a square. Another comment suggested simply changing the '0' value in the shape but I don't know how to do that.

To reproduce

1. Use the roboflow tutorial to train a custom YOLOv7 network: https://www.youtube.com/watch?v=5nsmXLyDaU4&t=865s&ab_channel=Roboflow (using a public training data set, such as the hard hat one)
2. Convert custom trained network to onnx using YOLOv7 github page suggested method: python export.py --weights yolov7.pt --grid --end2end --simplify --topk-all 100 --iou-thres 0.65 --conf-thres 0.35 --img-size 640 640 --max-wh 640
3. Attempt to perform inference using custom trained network with Java version of ONNXRuntime

Urgency

No response

ONNX Runtime Installation

Released Package

ONNX Runtime Version or Commit ID

1.13.1

PyTorch Version

Execution Provider

Default CPU

Execution Provider Library Version

No response

[Craigacp]

I must have forgot to fix that error message. Python returns an empty array, but Java computes the size value, sees that it is zero, gets confused and throws an exception. Either way there's no actual output, but the exception is rather unfriendly so I'll deal with it.

[Craigacp]

For the time being you can catch that exception as it's only emitted when no objects are detected.

[meeeee12]

ok, thanks.