您好，大佬，如何把人脸检测的模型转成torchscript模型啊

[wuxiaolianggit]

您好，大佬，如何把人脸检测的模型转成torchscript模型啊 @djmmoss @muhammad-tayyab @hylrh2008 @bobo0810 @wjtan99

[bobo0810]

您好，大佬，如何把人脸检测的模型转成torchscript模型啊 @djmmoss @muhammad-tayyab @hylrh2008 @bobo0810 @wjtan99

参考https://github.com/bobo0810/Classification/blob/main/Models/Backend/torchscript.py

[wuxiaolianggit]

@好的，大佬，多谢啊，我试一下啊 ，大佬我好像有你的微信。@ @bobo0810

[wuxiaolianggit]

Traceback (most recent call last): File "D:/python_code/yolov5-face-master/detect_face_jit.py", line 152, in detect_one(model, opt.image, device) File "D:/python_code/yolov5-face-master/detect_face_jit.py", line 108, in detect_one pred = model(img)[0] File "E:\python_code\pytorch-image-classification\venv\lib\site-packages\torch\nn\modules\module.py", line 727, in _call_impl result = self.forward(*input, **kwargs) RuntimeError: The following operation failed in the TorchScript interpreter. Traceback of TorchScript, serialized code (most recent call last): File "code/torch/models/yolo.py", line 44, in forward _34 = (_4).forward(_33, ) _35 = (_2).forward((_3).forward(_34, ), _28, ) _36 = (_0).forward(_32, _34, (_1).forward(_35, ), ) ~~~~~~~~~~~ <--- HERE _37, _38, _39, _40, = _36 return (_40, [_37, _38, _39]) File "code/torch/models/yolo.py", line 76, in forward _53 = torch.sigmoid(torch.index(_49, _52)) _54 = torch.to(CONSTANTS.c0, torch.device("cuda"), 4, False, False, None) _55 = torch.view(_53, [3, 100, 100, 6]) ~~~~~~~~~~ <--- HERE _56 = torch.to(_54, dtype=4, layout=0, device=torch.device("cuda:0"), pin_memory=None, non_blocking=False, copy=False, memory_format=None) _57 = annotate(List[Optional[Tensor]], [None, None, None, None, _56])

Traceback of TorchScript, original code (most recent call last): D:\python_code\yolov5-face-master\models\yolo.py(88): forward E:\python_code\pytorch-image-classification\venv\lib\site-packages\torch\nn\modules\module.py(709): _slow_forward E:\python_code\pytorch-image-classification\venv\lib\site-packages\torch\nn\modules\module.py(725): _call_impl D:\python_code\yolov5-face-master\models\yolo.py(196): forward_once D:\python_code\yolov5-face-master\models\yolo.py(180): forward E:\python_code\pytorch-image-classification\venv\lib\site-packages\torch\nn\modules\module.py(709): _slow_forward E:\python_code\pytorch-image-classification\venv\lib\site-packages\torch\nn\modules\module.py(725): _call_impl E:\python_code\pytorch-image-classification\venv\lib\site-packages\torch\jit_trace.py(934): trace_module E:\python_code\pytorch-image-classification\venv\lib\site-packages\torch\jit_trace.py(733): trace D:/python_code/yolov5-face-master/export2.py(89): RuntimeError: shape '[3, 100, 100, 6]' is invalid for input of size 108000

转换成功，但是推理的时候会报错，这个怎么解决啊，大佬@ @bobo0810

[bobo0810]

尝试debug排查原因吧

[wuxiaolianggit]

好的，多谢大佬 @bobo0810

[muhammad-tayyab]

The model will generate face feature points (16 size vector) in addition to bounding box and non-max-suppression is a bit different than standard Yolov5. So if you plan to use torchscript model in another app, you need to add that functionality too, otherwise you will not be able to get good quality detections

[wuxiaolianggit]

Thank you for your reply. I understand. @muhammad-tayyab