ONNX转化RKNN报错

使用YOLOV7训练的人脸检测的模型，输出一共包含五个分支，维度分别为[1,102000,21]，[1,3,160,160,21]，[1,3,80,80,21]，[1,3,40,40,21]，[1,3,20,20,21]，其中21为6个box信息，加上15个关键点坐标信息：5*{x,y,v}。在转化时报错：
D RKNN: [08:48:43.304] >>>>>> start: N4rknn20RKNNMultiSurfacePassED RKNN: [08:48:43.304] <<<<<<<< end: N4rknn20RKNNMultiSurfacePassED RKNN: [08:48:43.304] >>>>>> start: N4rknn14RKNNTilingPassEW RKNN: [08:48:43.308] Failed to config this layer using 3Core fallback to single core mode,W RKNN: [08:48:43.308] core_num 3 ori_Ih 20 ori_Iw 20 ori_Ic 416 ori_Ib 1 W RKNN: [08:48:43.308] ori_Kh 1 ori_Kw 1 ori_Kk 416 ori_Kc 416 ori_Ksx 1 ori_Ksy 1 W RKNN: [08:48:43.308] ori_Oh 20 oriOw 20 oriOc 416 pad_t 0 pad_b 0 pad_l 0 pad_r 0,W RKNN: [08:48:43.308] Please help report this bug!W RKNN: [08:48:43.311] Failed to config this layer using 3Core fallback to single core mode,W RKNN: [08:48:43.311] core_num 3 ori_Ih 20 ori_Iw 20 ori_Ic 432 ori_Ib 1 W RKNN: [08:48:43.311] ori_Kh 1 ori_Kw 1 ori_Kk 416 ori_Kc 432 ori_Ksx 1 ori_Ksy 1 W RKNN: [08:48:43.311] ori_Oh 20 oriOw 20 oriOc 416 pad_t 0 pad_b 0 pad_l 0 pad_r 0,W RKNN: [08:48:43.311] Please help report this bug!D RKNN: [08:48:43.312] <<<<<<<< end: N4rknn14RKNNTilingPassED RKNN: [08:48:43.312] >>>>>> start: N4rknn20RKNNAddSecondaryNodeED RKNN: [08:48:43.314] <<<<<<<< end: N4rknn20RKNNAddSecondaryNodeED RKNN: [08:48:43.314] >>>>>> start: N4rknn23RKNNProfileAnalysisPassED RKNN: [08:48:43.314] <<<<<<<< end: N4rknn23RKNNProfileAnalysisPassED RKNN: [08:48:43.314] >>>>>> start: OpEmitE build: Catch exception when building RKNN model!E build: Traceback (most recent call last):E build:   File "rknn/api/rknn_base.py", line 1580, in rknn.api.rknn_base.RKNNBase.buildE build:   File "rknn/api/rknn_base.py", line 341, in rknn.api.rknn_base.RKNNBase._generate_rknnE build:   File "rknn/api/rknn_base.py", line 307, in rknn.api.rknn_base.RKNNBase._build_rknnE build: IndexError: vector::_M_range_check: __n (which is 4) >= this->size() (which is 4)

所使用pytorch版本为1.11.0，opset为12。另外在此之前转化成功过相同维度的模型,其关键点排列方式为{x,y,v, x,y,v, x,y,v...},而现在仅是将排列方式改为了{x,x,x,x,x, y,y,y,y,y, v,v,v,v,v}就无法进行转换了，所以看起来也不像是版本的问题，但又实现找不到解决的办法，请教一下大家。

最好不要出现5维的数据，

aurora 发表于 2023-10-31 10:06
最好不要出现5维的数据，

后续尝试仅保留了[1,102000,21]一个维度的数据，转化后还是出现此问题，找不到原因