pnnx直接转pt模型报错

[libofei2004]

https://github.com/ultralytics/assets/releases官网下载的yolov8.pt模型用pnnx yolov8n.pt inputshape=[1,3,320,320]命令转换报错，是否不支持直接转yolov8.pt？
报：
`pnnxparam = yolov8n.pnnx.param
pnnxbin = yolov8n.pnnx.bin
pnnxpy = yolov8n_pnnx.py
pnnxonnx = yolov8n.pnnx.onnx
ncnnparam = yolov8n.ncnn.param
ncnnbin = yolov8n.ncnn.bin
ncnnpy = yolov8n_ncnn.py
fp16 = 1
optlevel = 2
device = cpu
inputshape = [1,3,320,320]f32
inputshape2 =
customop =
moduleop =
Load torchscript failed: PytorchStreamReader failed locating file constants.pkl: file not found
Exception raised from valid at /project/pytorch-v2.8.0/caffe2/serialize/inline_container.cc:238 (most recent call first):
frame #0: /home/libofei/.local/lib/python3.10/site-packages/pnnx/pnnx() [0x35f1af7]
frame #1: /home/libofei/.local/lib/python3.10/site-packages/pnnx/pnnx() [0x35f1370]
frame #2: /home/libofei/.local/lib/python3.10/site-packages/pnnx/pnnx() [0x48730c]
frame #3: /home/libofei/.local/lib/python3.10/site-packages/pnnx/pnnx() [0x2ecb9d2]
frame #4: /home/libofei/.local/lib/python3.10/site-packages/pnnx/pnnx() [0x2ecbfe4]
frame #5: /home/libofei/.local/lib/python3.10/site-packages/pnnx/pnnx() [0x2ecc0e7]
frame #6: /home/libofei/.local/lib/python3.10/site-packages/pnnx/pnnx() [0x3357008]
frame #7: /home/libofei/.local/lib/python3.10/site-packages/pnnx/pnnx() [0x3350511]
frame #8: /home/libofei/.local/lib/python3.10/site-packages/pnnx/pnnx() [0x3351f61]
frame #9: /home/libofei/.local/lib/python3.10/site-packages/pnnx/pnnx() [0x3354dc5]
frame #10: /home/libofei/.local/lib/python3.10/site-packages/pnnx/pnnx() [0x3354f79]
frame #11: /home/libofei/.local/lib/python3.10/site-packages/pnnx/pnnx() [0x3355025]
frame #12: /home/libofei/.local/lib/python3.10/site-packages/pnnx/pnnx() [0x6cc2a1]
frame #13: /home/libofei/.local/lib/python3.10/site-packages/pnnx/pnnx() [0x48fd52]
frame #14: + 0x29d90 (0x7bb36dc29d90 in /lib/x86_64-linux-gnu/libc.so.6)
frame #15: __libc_start_main + 0x80 (0x7bb36dc29e40 in /lib/x86_64-linux-gnu/libc.so.6)
frame #16: /home/libofei/.local/lib/python3.10/site-packages/pnnx/pnnx() [0x4f308e]

Please export model to torchscript as follows

import torch
import torchvision.models as models

net = models.resnet18(pretrained=True)
net = net.eval()

x = torch.rand(1, 3, 224, 224)
mod = torch.jit.trace(net, x)
mod.save("resnet18.pt")

############# pass_level2`

[dosubot]

Hi @libofei2004! I'm Dosu and I’m helping the ncnn team.

pnnx 不支持直接转换从 Ultralytics 下载的 yolov8.pt 模型，因为这些 .pt 文件通常是 PyTorch 的 state_dict 或自定义格式，而 pnnx 只支持 torchscript 格式（通过 torch.jit.trace 或 torch.jit.script 导出的 .pt 文件）。你遇到的 constants.pkl: file not found 报错就是因为模型不是 torchscript 格式 官方文档说明。

正确流程是：先用 PyTorch 加载 yolov8.pt，然后用 torch.jit.trace 或 torch.jit.script 导出为 torchscript 格式，再用 pnnx 转换。例如：

import torch
from ultralytics import YOLO

model = YOLO('yolov8n.pt')  # 加载模型
model.model.eval()
dummy_input = torch.randn(1, 3, 320, 320)
traced = torch.jit.trace(model.model, dummy_input)
traced.save('yolov8n_ts.pt')

然后用 pnnx 转换：

pnnx yolov8n_ts.pt inputshape=[1,3,320,320]

如果模型包含自定义模块，可能还需要用 moduleop 参数指定模块路径，参考 pnnx README。

详细 torchscript 导出和 pnnx 用法见 官方文档。

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}   

[libofei2004]

@dosubot 我运行了代码,

import torch
from ultralytics import YOLO

model = YOLO('yolov8n.pt')  # 加载模型
model.model.eval()
dummy_input = torch.randn(1, 3, 320, 320)
traced = torch.jit.trace(model.model, dummy_input)
traced.save('yolov8n_ts.pt')

但报错 :
/home/libofei/.local/lib/python3.10/site-packages/ultralytics/nn/modules/head.py:163: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
if self.format != "imx" and (self.dynamic or self.shape != shape):
/home/libofei/.local/lib/python3.10/site-packages/ultralytics/utils/tal.py:372: TracerWarning: Iterating over a tensor might cause the trace to be incorrect. Passing a tensor of different shape won't change the number of iterations executed (and might lead to errors or silently give incorrect results).
for i, stride in enumerate(strides):
/home/libofei/.local/lib/python3.10/site-packages/torch/jit/_trace.py:1278: TracerWarning: Encountering a list at the output of the tracer might cause the trace to be incorrect, this is only valid if the container structure does not change based on the module's inputs. Consider using a constant container instead (e.g. for list, use a tuple instead. for dict, use a NamedTuple instead). If you absolutely need this and know the side effects, pass strict=False to trace() to allow this behavior.
module._c._create_method_from_trace(
Traceback (most recent call last):
File "/mnt/d/项目/识别/yolo/daochu.py", line 23, in
traced = torch.jit.trace(model.model, dummy_input)
File "/home/libofei/.local/lib/python3.10/site-packages/torch/jit/_trace.py", line 1002, in trace
traced_func = _trace_impl(
File "/home/libofei/.local/lib/python3.10/site-packages/torch/jit/_trace.py", line 698, in _trace_impl
return trace_module(
File "/home/libofei/.local/lib/python3.10/site-packages/torch/jit/_trace.py", line 1306, in trace_module
_check_trace(
File "/home/libofei/.local/lib/python3.10/site-packages/torch/utils/_contextlib.py", line 116, in decorate_context
return func(*args, **kwargs)
File "/home/libofei/.local/lib/python3.10/site-packages/torch/jit/_trace.py", line 592, in _check_trace
raise TracingCheckError(*diag_info)
torch.jit._trace.TracingCheckError: Tracing failed sanity checks!
ERROR: Graphs differed across invocations!
Graph diff:
graph(%self.1 : torch.ultralytics.nn.tasks.DetectionModel,
%x.1 : Tensor):
%model : torch.torch.nn.modules.container.Sequential = prim::GetAttrname="model"
%_22 : torch.ultralytics.nn.modules.head.Detect = prim::GetAttrname="22"
%model.45 : torch.torch.nn.modules.container.Sequential = prim::GetAttrname="model"
...

ERROR: Tensor-valued Constant nodes differed in value across invocations. This often indicates that the tracer has encountered untraceable code.
Node:
%984 : Tensor = prim::Constantvalue={2}, scope: __module.model.22 # /home/libofei/.local/lib/python3.10/site-packages/ultralytics/utils/tal.py:388:0
Source Location:
/home/libofei/.local/lib/python3.10/site-packages/ultralytics/utils/tal.py(388): dist2bbox
/home/libofei/.local/lib/python3.10/site-packages/ultralytics/nn/modules/head.py(202): decode_bboxes
/home/libofei/.local/lib/python3.10/site-packages/ultralytics/nn/modules/head.py(182): _inference
/home/libofei/.local/lib/python3.10/site-packages/ultralytics/nn/modules/head.py(123): forward
/home/libofei/.local/lib/python3.10/site-packages/torch/nn/modules/module.py(1726): _slow_forward
/home/libofei/.local/lib/python3.10/site-packages/torch/nn/modules/module.py(1747): _call_impl
/home/libofei/.local/lib/python3.10/site-packages/torch/nn/modules/module.py(1736): _wrapped_call_impl
/home/libofei/.local/lib/python3.10/site-packages/ultralytics/nn/tasks.py(180): _predict_once
/home/libofei/.local/lib/python3.10/site-packages/ultralytics/nn/tasks.py(157): predict
/home/libofei/.local/lib/python3.10/site-packages/ultralytics/nn/tasks.py(139): forward
/home/libofei/.local/lib/python3.10/site-packages/torch/nn/modules/module.py(1726): _slow_forward
/home/libofei/.local/lib/python3.10/site-packages/torch/nn/modules/module.py(1747): _call_impl
/home/libofei/.local/lib/python3.10/site-packages/torch/nn/modules/module.py(1736): _wrapped_call_impl
/home/libofei/.local/lib/python3.10/site-packages/torch/jit/_trace.py(1278): trace_module
/home/libofei/.local/lib/python3.10/site-packages/torch/jit/_trace.py(698): _trace_impl
/home/libofei/.local/lib/python3.10/site-packages/torch/jit/_trace.py(1002): trace
/mnt/d/项目/识别/yolo/daochu.py(23):
Comparison exception: The values for attribute 'shape' do not match: torch.Size([]) != torch.Size([1, 2100]).

[dosubot]

你遇到的 torch.jit.trace 报错（TracingCheckError、TracerWarning）是因为 YOLOv8 模型内部有动态控制流、tensor 迭代、Python 逻辑等，导致 trace 过程无法稳定捕获模型的计算图。这是 torch.jit.trace 的已知限制，尤其在 YOLOv8 这类复杂模型上很常见 官方说明。

你可以尝试以下几种方法：

1. 关闭 trace 检查
   trace 时加参数 check_trace=False，有时能跳过部分检查错误：

   traced = torch.jit.trace(model.model, dummy_input, check_trace=False)
   traced.save('yolov8n_ts.pt')

   但如果模型本身 forward 逻辑不稳定，这种方法可能还是不行。

2. 尝试 torch.jit.script
   对于包含动态控制流的模型，推荐用 script：

   scripted = torch.jit.script(model.model)
   scripted.save('yolov8n_ts.pt')

   但如果模型用到不支持的 Python 语法（比如 for 循环遍历 tensor、if 判断 tensor），script 也可能失败，需要你手动修改模型代码，把 list 换成 tuple、避免 tensor 参与 Python 逻辑等。

3. 导出 ONNX
   如果 torchscript 导出始终失败，可以尝试导出 ONNX：

   torch.onnx.export(model.model, dummy_input, 'yolov8n.onnx')

   然后用 pnnx 转换 ONNX 文件：

   pnnx yolov8n.onnx inputshape=[1,3,320,320]

   ONNX 导出也可能遇到类似问题，但有时比 torchscript 更兼容。

4. 代码兼容性修改
   如果 script/ONNX 都失败，通常需要修改 YOLOv8 源码，让 forward 逻辑更“静态”，比如把 for 循环里的 tensor 换成 tuple，避免 tensor 参与 if 判断等。

目前 pnnx/ncnn 只支持 torchscript 或 ONNX 格式，不能直接转换 Ultralytics 下载的 .pt 文件 参考。

相关官方文档和讨论见：ncnn wiki

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}   

[libofei2004]

@dosubot 用

traced = torch.jit.trace(model.model, dummy_input, check_trace=False)
traced.save('yolov8n_ts.pt')
语句或
torch.onnx.export(model.model, dummy_input, 'yolov8n.onnx')
pnnx yolov8n.onnx inputshape=[1,3,320,320]

确实可以导出成功，但放在安卓程序框架(https://github.com/nihui/ncnn-android-yolov8)里运行后会出问题卡住，栈如下：
art_sigsegv_fault 0x0000007ebb1124d4
generate_proposals(const ncnn::Mat &, int, const ncnn::Mat &, float, std::__ndk1::vector<…> &) yolov8_det.cpp:180
generate_proposals(const ncnn::Mat &, const std::__ndk1::vector<…> &, const ncnn::Mat &, float, std::__ndk1::vector<…> &) yolov8_det.cpp:265
YOLOv8_det::detect(const cv::Mat &, std::__ndk1::vector<…> &) yolov8_det.cpp:324
......

页面弹出如下内存信息：
art_sigsegv_fault:
paciasp
sub sp, sp, #0x40
stp x29, x30, [sp, #0x20]
stp x20, x19, [sp, #0x30]
add x29, sp, #0x20
mrs x19, TPIDR_EL0
ldr x8, [x19, #0x28]
stur x8, [x29, #-0x8]
adrp x8, 1278
ldr x8, [x8, #0xe90]
ldrb w8, [x8, #0xd]
cbnz w8, 0x312528 ; <+84>
ldr x8, [x19, #0x28]

[dosubot]

你的模型在安卓端卡死/崩溃，主要原因很可能是 ncnn 的 generate_proposals 函数对输出 shape 和格式有严格要求。如果模型导出后 param/bin 文件的输出 shape（比如通道数、grid 排布）和 ncnn-android-yolov8 框架预期不一致，就会导致内存越界或访问非法地址，出现 SIGSEGV 错误。

ncnn 的 generate_proposals 期望输出为三维 tensor，通道数为 255（即 85*3，3 个 anchor，每个 anchor 85 个属性），shape 应为 [grid_w, grid_h, 85, 3]，其中 grid_w、grid_h 取决于输入尺寸。输出内容必须包含 center-x、center-y、box-w、box-h、置信度和每类分数等 "generate_proposals 代码注释"。

建议你：

1. 检查 pnnx 导出的 ncnn param 文件，确认输出层的 shape 是否为 [grid_w, grid_h, 255]，且 grid_w、grid_h 与输入尺寸对应。
2. 在安卓端推理前，打印输出层 shape，确保和 generate_proposals 预期一致。
3. 检查 anchors 设置和 stride 是否和 YOLOv8 官方模型一致。
4. 如果 shape 不对，需重新导出模型，或手动修改 param 文件，确保输出 shape 和 ncnn 框架一致。

如果 param 文件最后有 Reduction 层导致输出一维，可以尝试将其改为 Noop，避免 shape 错误 "param 文件修改建议"。

总之，重点排查模型输出 shape 和 anchors 配置，确保和 ncnn-android-yolov8 框架一致，否则 generate_proposals 会崩溃。

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}   

[libofei2004]

@dosubot 如何用param 确认输出层的 shape 是否为 [grid_w, grid_h, 255]，原框架自带的yolov8n.ncnn.param最后是Concat cat_16 3 1 213 215 217 out0 0=0，而我转出来的是Concat cat_18 2 1 263 264 out0 0=0，确实要少一列，但原框架的yolov8n.ncnn.param是怎么生成的呢？

[dosubot]

ncnn 的 .param 文件本身不直接显示输出 shape，比如 [grid_w, grid_h, 255]，只能通过网络结构推断或实际推理时打印输出 shape。最后一层 Concat 的输入数量通常对应 YOLOv8 检测头（一般是3个，分别对应 stride 8/16/32），比如 Concat cat_16 3 1 ...，而你导出的只有2个输入，说明少了一个检测头，导致输出 shape 不对，ncnn 框架就会崩溃。

正确的 yolov8n.ncnn.param 生成流程是：先用 Ultralytics 导出 torchscript，再用 pnnx 转换，然后需要手动修改 pnnx 导出的 python 脚本（如 yolov8n_pnnx.py），让三个 head 的输出都参与 concat，且 axis 设置正确，最后重新导出 torchscript，再用 pnnx 转换为 ncnn 格式。官方 yolov8.cpp 示例和注释有详细流程 "yolov8.cpp注释"。

要确认输出 shape，推荐用 ncnn C++ 代码加载 param/bin，推理一张图片后打印输出 blob 的 w/h/c，比如：

ncnn::Extractor ex = net.create_extractor();
ex.input("in0", in_pad);
ncnn::Mat out;
ex.extract("out0", out);
printf("out shape: w=%d h=%d c=%d\n", out.w, out.h, out.c);

如果 out.c（通道数）不是 255 或 144（YOLOv8新版），或者 out.h 不等于所有 grid cell 总数（比如 8400），说明 param 文件结构有问题，需要回到 torchscript/pnnx 导出环节修正 "官方流程"。

总之，Concat 层输入数量必须和 YOLOv8 head 数一致，param 文件结构要和官方 demo一致，必要时需手动调整 pnnx 导出脚本和 param 文件。

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}   

[libofei2004]

@dosubot 我不太明白 yolov8n.torchscript为什么要导出两次，第一次导出和第二次有什么关系？看步骤也比较繁琐，那么pnnx的转换方式相比ncnn命令的方式优势在哪里？

[dosubot]

yolov8n.torchscript 之所以要导出两次，是因为第一次导出只是把 YOLOv8 模型静态地转成 torchscript，pnnx 再把它转换成 Python 脚本（如 yolov8n_pnnx.py）。但 YOLOv8 的结构和后处理很复杂，直接导出的模型 shape、concat 方式、动态尺寸支持等都不一定符合 ncnn 框架的要求。

所以需要手动修改 pnnx 生成的 yolov8n_pnnx.py，比如把 view(1, 144, 6400) 改成 view(1, 144, -1).transpose(1, 2)，调整 concat 的 axis，去掉后处理部分等，确保输出 shape 和 ncnn demo一致。然后用修改后的 yolov8n_pnnx.py 再次导出 torchscript（第二次），再用 pnnx 转成最终的 ncnn param/bin 文件，这样才能保证在 ncnn-android-yolov8 框架里正常推理 详细流程和原因见 yolov8.cpp 注释。

pnnx 的优势在于它能生成可读可编辑的 Python 脚本，你可以灵活修正模型结构、支持动态 shape、优化算子，最终导出高兼容性的 ncnn 模型。相比直接用 ncnn 命令行转 ONNX，pnnx 更适合处理 PyTorch 复杂模型，能更好地还原原始算子和结构，避免 ONNX 导出时的 glue operator、shape 不一致等问题 pnnx 设计和优势详见 README。

简单对比：

pnnx 转换	ncnn 命令行转 ONNX
支持 PyTorch 原生 torchscript，算子还原度高	只支持 ONNX，部分 PyTorch 算子不兼容
可生成 Python 脚本，便于手动修正	无法编辑 ONNX，调试困难
支持 shape 推理、动态 shape、复杂模块	动态 shape、复杂控制流支持有限
适合 YOLOv8/v5 等复杂模型	适合结构简单的模型

如果模型结构复杂、需要动态 shape或自定义后处理，推荐用 pnnx 多步流程，虽然繁琐但更可靠。

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}   

[libofei2004]

@dosubot 多谢！

[dosubot]

不用谢！如果还有其他问题，随时告诉我。

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}