long picture detection wrong

[guotao]

Hello, I use yolov5 for long picture detection like this :

when I use python (detect.py), the result is right:

while in mobile device,I use _objectModel.getImagePrediction and then _objectModel.renderBoxesOnImage, the result is wrong,

This is the result in example code, just change the model and label files,and also change "nc" config inPytorchLite.loadObjectDetectionModel function call

For nomal size image, for example 600* 400, the result is fine， Is there any config can fix this issue? thanks 。

[abdelaziz-mahdy]

Can you send the test images and the model so I can test to find the issue and try to fix it ?

[guotao]

Can you send the test images and the model so I can test to find the issue and try to fix it ?

Of course https://drive.google.com/file/d/1gp6wucf_sNBLBEQu2lI4oxXr9MKWt5gw/view?usp=drive_link this is the shared link

Also invated you as collaborator to team, incase this link not work 😁

[abdelaziz-mahdy]

Sorry, the file you have requested does not exist.

Make sure that you have the correct URL and the file exists.

For the drive,

And GitHub shows

[guotao]

I create a repo which upload the model files, and invate you to join

[abdelaziz-mahdy]

is this a yolov8 or yolov5?

[abdelaziz-mahdy]

i found out its yolov5, in detect.py did you use the same model torch script like the one in the app?

[guotao]

i found out its yolov5, in detect.py did you use the same model torch script like the one in the app?

yes, exactly, you can run detect.py --weight 'path to b2.torchscript' --source to see the result, and compare with the one in example，thanks。

[abdelaziz-mahdy]

running !python detect.py --weight '../b2.torchscript' --source ../test.PNG

the model doest even run and outputs

detect: weights=['../b2.torchscript'], source=../test.PNG, data=data/coco128.yaml, imgsz=[640, 640], conf_thres=0.25, iou_thres=0.45, max_det=1000, device=, view_img=False, save_txt=False, save_csv=False, save_conf=False, save_crop=False, nosave=False, classes=None, agnostic_nms=False, augment=False, visualize=False, update=False, project=runs/detect, name=exp, exist_ok=False, line_thickness=3, hide_labels=False, hide_conf=False, half=False, dnn=False, vid_stride=1
YOLOv5 🚀 v7.0-230-g53efd07 Python-3.10.12 torch-2.1.0+cu118 CUDA:0 (Tesla T4, 15102MiB)

Loading ../b2.torchscript for TorchScript inference...
Traceback (most recent call last):
  File "/content/yolov5/detect.py", line 285, in <module>
    main(opt)
  File "/content/yolov5/detect.py", line 280, in main
    run(**vars(opt))
  File "/usr/local/lib/python3.10/dist-packages/torch/utils/_contextlib.py", line 115, in decorate_context
    return func(*args, **kwargs)
  File "/content/yolov5/detect.py", line 101, in run
    model = DetectMultiBackend(weights, device=device, dnn=dnn, data=data, fp16=half)
  File "/content/yolov5/models/common.py", line 364, in __init__
    model = torch.jit.load(w, _extra_files=extra_files, map_location=device)
  File "/usr/local/lib/python3.10/dist-packages/torch/jit/_serialization.py", line 162, in load
    cpp_module = torch._C.import_ir_module(cu, str(f), map_location, _extra_files, _restore_shapes)  # type: ignore[call-arg]
RuntimeError: xnnpack::convolution not available! Reason: The provided (weight, bias, padding, stride, dilation, groups, transposed, output_min, output_max) parameters are either invalid individually or their combination is not supported by XNNPACK.

[kainy]

RuntimeError: xnnpack::convolution not available! Reason: The provided (weight, bias, padding, stride, dilation, groups, transposed, output_min, output_max) parameters are either invalid individually or their combination is not supported by XNNPACK

Saw this issue ultralytics/ultralytics#2465, which might provide infomation helped @guotao

[guotao]

running !python detect.py --weight '../b2.torchscript' --source ../test.PNG

the model doest even run and outputs

detect: weights=['../b2.torchscript'], source=../test.PNG, data=data/coco128.yaml, imgsz=[640, 640], conf_thres=0.25, iou_thres=0.45, max_det=1000, device=, view_img=False, save_txt=False, save_csv=False, save_conf=False, save_crop=False, nosave=False, classes=None, agnostic_nms=False, augment=False, visualize=False, update=False, project=runs/detect, name=exp, exist_ok=False, line_thickness=3, hide_labels=False, hide_conf=False, half=False, dnn=False, vid_stride=1
YOLOv5 🚀 v7.0-230-g53efd07 Python-3.10.12 torch-2.1.0+cu118 CUDA:0 (Tesla T4, 15102MiB)

Loading ../b2.torchscript for TorchScript inference...
Traceback (most recent call last):
  File "/content/yolov5/detect.py", line 285, in <module>
    main(opt)
  File "/content/yolov5/detect.py", line 280, in main
    run(**vars(opt))
  File "/usr/local/lib/python3.10/dist-packages/torch/utils/_contextlib.py", line 115, in decorate_context
    return func(*args, **kwargs)
  File "/content/yolov5/detect.py", line 101, in run
    model = DetectMultiBackend(weights, device=device, dnn=dnn, data=data, fp16=half)
  File "/content/yolov5/models/common.py", line 364, in __init__
    model = torch.jit.load(w, _extra_files=extra_files, map_location=device)
  File "/usr/local/lib/python3.10/dist-packages/torch/jit/_serialization.py", line 162, in load
    cpp_module = torch._C.import_ir_module(cu, str(f), map_location, _extra_files, _restore_shapes)  # type: ignore[call-arg]
RuntimeError: xnnpack::convolution not available! Reason: The provided (weight, bias, padding, stride, dilation, groups, transposed, output_min, output_max) parameters are either invalid individually or their combination is not supported by XNNPACK.

Thanks for repling, I found this may help ,would you have a try?
TORCH_XNNPACK_DISABLE=1 python detect.py --weights runs/train/exp/best.torchscript -- source /images/

And also , I have upload a model exported without argument --optimize, you can have a try if above method not work, thanks。

[abdelaziz-mahdy]

Does the models exported work and detect the image like how you want?? On detect.py

Since I think the exporting to torch script reduces accuracy by small factor, which is why I think it's the problem

[guotao]

Does the models exported work and detect the image like how you want?? On detect.py

Since I think the exporting to torch script reduces accuracy by small factor, which is why I think it's the problem

No, as you can see

these were result on detect.py and on pytorch_lite, there were big difference。
When I use normal size picture, the difference is small。
I just report the problem, If you think the problem is acceptable, you may close this issue。Thanks for your time。

[abdelaziz-mahdy]

The point is I followed the official java example , and both of my implementations give me the same results, this is why I am thinking it's a model problem that's all,

Sorry If I was not able to help, will check it again when I got time

[abdelaziz-mahdy]

can you check again using the last version, some stuff was changed and i dont know if the problem is still there or not