NCNN not work: layer torch.topk not exists or registered

[yl1994yl]

when i exported model to ncnn, it cannot work. The error message is "layer torch.topk not exists or registered", but i try use ncnn in yolov8 which also has the layer torch.topk and it succeed, so i think the layer error is not the key. Relative infomation is followed, thank you for suggestions!

---

Ultralytics YOLOv8.1.34 🚀 Python-3.9.19 torch-2.0.1+cu117 CPU (Intel Xeon E5-2699 v4 2.20GHz) YOLOv10n summary (fused): 285 layers, 2762608 parameters, 0 gradients, 8.6 GFLOPs

PyTorch: starting from 'yolov10n.pt' with input shape (1, 3, 640, 640) BCHW and output shape(s) (1, 300, 6) (10.9 MB)

TorchScript: starting export with torch 2.0.1+cu117... TorchScript: export success ✅ 30.6s, saved as 'yolov10n.torchscript' (11.1 MB)

NCNN: starting export with NCNN 1.0.20240410... NCNN: running './pnnx yolov10n.torchscript ncnnparam=yolov10n_ncnn_model/model.ncnn.param ncnnbin=yolov10n_ncnn_model/model.ncnn.bin ncnnpy=yolov10n_ncnn_model/model_ncnn.py pnnxparam=yolov10n_ncnn_model/model.pnnx.param pnnxbin=yolov10n_ncnn_model/model.pnnx.bin pnnxpy=yolov10n_ncnn_model/model_pnnx.py pnnxonnx=yolov10n_ncnn_model/model.pnnx.onnx fp16=0 device=cpu inputshape="[1, 3, 640, 640]"' pnnxparam = yolov10n_ncnn_model/model.pnnx.param pnnxbin = yolov10n_ncnn_model/model.pnnx.bin pnnxpy = yolov10n_ncnn_model/model_pnnx.py pnnxonnx = yolov10n_ncnn_model/model.pnnx.onnx ncnnparam = yolov10n_ncnn_model/model.ncnn.param ncnnbin = yolov10n_ncnn_model/model.ncnn.bin ncnnpy = yolov10n_ncnn_model/model_ncnn.py fp16 = 0 optlevel = 2 device = cpu inputshape = [1,3,640,640]f32 inputshape2 = customop = moduleop = ############# pass_level0 inline module = torch.nn.modules.linear.Identity inline module = ultralytics.nn.modules.block.Attention inline module = ultralytics.nn.modules.block.Bottleneck inline module = ultralytics.nn.modules.block.C2f inline module = ultralytics.nn.modules.block.C2fCIB inline module = ultralytics.nn.modules.block.CIB inline module = ultralytics.nn.modules.block.DFL inline module = ultralytics.nn.modules.block.PSA inline module = ultralytics.nn.modules.block.RepVGGDW inline module = ultralytics.nn.modules.block.SCDown inline module = ultralytics.nn.modules.block.SPPF inline module = ultralytics.nn.modules.conv.Concat inline module = ultralytics.nn.modules.conv.Conv inline module = ultralytics.nn.modules.head.v10Detect inline module = torch.nn.modules.linear.Identity inline module = ultralytics.nn.modules.block.Attention inline module = ultralytics.nn.modules.block.Bottleneck inline module = ultralytics.nn.modules.block.C2f inline module = ultralytics.nn.modules.block.C2fCIB inline module = ultralytics.nn.modules.block.CIB inline module = ultralytics.nn.modules.block.DFL inline module = ultralytics.nn.modules.block.PSA inline module = ultralytics.nn.modules.block.RepVGGDW inline module = ultralytics.nn.modules.block.SCDown inline module = ultralytics.nn.modules.block.SPPF inline module = ultralytics.nn.modules.conv.Concat inline module = ultralytics.nn.modules.conv.Conv inline module = ultralytics.nn.modules.head.v10Detect

---

############# pass_level1 ############# pass_level2 ############# pass_level3 ############# pass_level4 ############# pass_level5 ############# pass_ncnn BinaryOp remainder not supported yet BinaryOp floor_divide not supported yet ignore torch.topk torch.topk_56 param dim=-1 ignore torch.topk torch.topk_56 param k=300 ignore torch.topk torch.topk_56 param largest=True ignore torch.topk torch.topk_56 param sorted=True ignore torch.gather torch.gather_37 param dim=1 ignore torch.gather torch.gather_38 param dim=1 ignore torch.topk torch.topk_57 param dim=-1 ignore torch.topk torch.topk_57 param k=300 ignore torch.topk torch.topk_57 param largest=True ignore torch.topk torch.topk_57 param sorted=True ignore torch.gather torch.gather_39 param dim=1 NCNN: export success ✅ 4.4s, saved as 'yolov10n_ncnn_model' (9.0 MB)

Export complete (66.0s) Results saved to /mnt/disk5/yinliang/notebooks/yolov10-main Predict: yolo predict task=detect model=yolov10n_ncnn_model imgsz=640
Validate: yolo val task=detect model=yolov10n_ncnn_model imgsz=640 data=coco.yaml
Visualize: https://netron.app WARNING ⚠️ Unable to automatically guess model task, assuming 'task=detect'. Explicitly define task for your model, i.e. 'task=detect', 'segment', 'classify','pose' or 'obb'. Loading yolov10n_ncnn_model for NCNN inference... layer torch.topk not exists or registered

[yl1994yl]

Here is the code.

from ultralytics import YOLO

model = YOLO("yolov10n.pt") model.export(format="ncnn") ncnn_model = YOLO("./yolov10n_ncnn_model") results = ncnn_model("bus.jpg") print()

[petchies]

Hello jameslahm Wang Ao] 你好。 最近什么样 ？你好吗 ？ I would like to write python about with Ai and Data Center. Please you advice me about with a kind of program run on HTML

多谢谢 Thank you so much

[jameslahm]

Thanks for your interest! It seems that there is no torch.topk in yolov8. Could you please confirm that? Thank you!

[yl1994yl]

Thanks for your interest! It seems that there is no torch.topk in yolov8. Could you please confirm that? Thank you!

I confirmed that there is no topk layer in ncnn model of yolov8, but it do have topk in yolov10 v10postprocess. Following the code:

[jameslahm]

Thanks. This issue may rely on ncnn to add the support for torch.topk.

[shitianyi]

要么按照ncnn的教程添加topk算子，要么就像yolov8把后处理部分全砍了，一步一步自己手写实现。。

[FLamefiREz]

Maybe,it can help you.https://zhuanlan.zhihu.com/p/699632697

[yl1994yl]

Maybe,it can help you.https://zhuanlan.zhihu.com/p/699632697

Great thanks for suggestions! I tried and it works!

[yl1994yl]

Maybe,it can help you.https://zhuanlan.zhihu.com/p/699632697

But the inference speed is not faster than yolov5, did you get the same result? Thanks for replying!