Dynamic Image size ONNX conversion

[YCAyca]

Hello, I am converting my model to onnx using
python export.py --weights runs/train/haribo_v32/weights/best.pt --img-size 416 416 --grid --dynamic
code and then I get the output layer names as output_names ['output', '540', '562'] where predictions=model_output[0] which refers to the "output" layer gives me predictions shape (3, 52, 40, 124) which I see in neutron.app a shape of float32[batch,3,y,x,124]. I understand the batch=1, and I reshape this output to predictions = predictions.reshape(-1, 124)

Then I process the output as well as I was doing for static onnx converted model.
obj_conf = predictions[:, 4]
scores = np.max(predictions[:, 5:], axis=1)
boxes=predictions[:, :4]

But when I print, I see that the boxes are wrong:

static onnx model boxes: (one box example)
[ 45.47592, 239.41583, 86.19278, 290.94891]
dynamic onnx model boxes: (one box example)
[ -0.55995 -0.72869 2.397 2.0979]

therefore when I rescale and draw the boxes, they are not correctly places since before rescaling coordinates are already wrong. What should I do? Anyone has any idea?