按照官方操作手册转换yolo11模型出现问题

[deyiluobo]

环境：
python环境：
ultralytics 8.4.52

docker环境 两个版本同样错误
pulsar2 6.0
pulsar2 7.0

报错日志：
`2026-07-31 11:22:21.482 | INFO | yamain.command.precision_analysis:quant_precision_analysis:403 - save precision analysis table to [tmp/quant/debug/precision_analysis_table.txt]
2026-07-31 11:22:21.489 | INFO | yamain.command.precision_analysis:quant_precision_analysis:411 - save precision analysis graph to [tmp/quant/debug/precision_analysis.mmd]
Building native ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 100% 0:00:00
2026-07-31 11:22:22.667 | INFO | yamain.command.build:compile_ptq_model:1169 - group 0 compiler transformation
2026-07-31 11:22:22.676 | WARNING | yamain.command.load_model:pre_process:622 - preprocess tensor [images]
2026-07-31 11:22:22.676 | INFO | yamain.command.load_model:pre_process:624 - tensor: images, (1, 640, 640, 3), U8
2026-07-31 11:22:22.676 | INFO | yamain.command.load_model:pre_process:625 - op: op:pre_dequant_1, AxDequantizeLinear, {'const_inputs': {'x_zeropoint': array(0, dtype=int32), 'x_scale': array(1., dtype=float32)}, 'output_dtype': <class 'numpy.float32'>, 'quant_method': 0}
2026-07-31 11:22:22.676 | INFO | yamain.command.load_model:pre_process:624 - tensor: tensor:pre_norm_1, (1, 640, 640, 3), FP32
2026-07-31 11:22:22.676 | INFO | yamain.command.load_model:pre_process:625 - op: op:pre_norm_1, AxNormalize, {'dim': 3, 'mean': [0.0, 0.0, 0.0], 'std': [255.0, 255.0, 255.0], 'output_dtype': FP32}
2026-07-31 11:22:22.676 | INFO | yamain.command.load_model:pre_process:624 - tensor: tensor:pre_transpose_1, (1, 640, 640, 3), FP32
2026-07-31 11:22:22.676 | INFO | yamain.command.load_model:pre_process:625 - op: op:pre_transpose_1, AxTranspose, {'perm': [0, 3, 1, 2]}
2026-07-31 11:22:22.677 | WARNING | yamain.command.load_model:post_process:633 - postprocess tensor [/model.23/Concat_output_0]
Traceback (most recent call last):
File "", line 59, in guard_context
File "", line 1174, in compile_ptq_model
File "", line 782, in ir_compiler_transformation
File "", line 647, in post_process
File "", line 347, in transpose_postprocess
IndexError: list assignment index out of range

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
File "", line 4, in
File "", line 281, in
File "", line 277, in pulsar2
File "", line 158, in wrapper
File "", line 21, in wrapper
File "", line 115, in build_error
File "", line 13, in wrapper
File "", line 625, in build
File "", line 1138, in compile_ptq_model
File "/usr/local/lib/python3.12/contextlib.py", line 158, in exit
self.gen.throw(value)
File "", line 61, in guard_context
File "", line 73, in error_func
yamain.common.error.CodeException: (<ErrorCode.FrontendError: 6>, IndexError('list assignment index out of range'))`

目前解决方案：
可以通过修改yolo11n.npu.json的方式来规避报错：
"dst_perm": [0, 1, 3, 2]“ 改为 "dst_perm": [0, 1, 2]"

生成的模型可以在板端推理，使用下面命令：
ax_run_model -m xxx.axmodel -r 100

但无法使用Maixpy 推理

报错:
`start running...

-- [I] [Maix Comm Protocol] auto set pinmap A21 to UART4_TX for uart4
-- [I] [Maix Comm Protocol] auto set pinmap A22 to UART4_RX for uart4
-- [I] [Maix Comm Protocol] listening on uart port: uart4(/dev/ttyS4)
-- [I] npu_ai_isp_on from config: 0 (off)
-- [I] init middleware SYSTEM
-- [I] maix multi-media driver used count: 1
-- [I] init middleware NPU ENGINE
-- [I] maix npu driver used count: 1
-- [I] load model from: /root/my_models/yolo11n_npu.axmodel
-- [I] input_cache : 1, flush input cache to ddr before run model: 0
-- [I] output_cache: 1, invalidate output cache after run model: 1
-- [I] model info:
type: yolo11
input type: rgb
mean:0.000000 0.000000 0.000000
scale:0.003922 0.003922 0.003922
labels num: 1
input size: 96x96

-- [E] output node shape error, please check 2
-- [I] Outputs:
LayerInfo(name='/model.23/Concat_output_0', dtype=float32, shape=[1, 64, 189])
LayerInfo(name='/model.23/Concat_1_output_0', dtype=float32, shape=[1, 1, 189])
LayerInfo(name='/model.23/Concat_2_output_0', dtype=float32, shape=[1, 4, 189])
Traceback (most recent call last):
File "/tmp/maixpy_run/main.py", line 4, in
detector = nn.YOLO11(model="/root/my_models/yolo11n.mud", dual_buff=True)
RuntimeError: Invalid arguments: load model failed`

[kingo0807]

您好，这个问题是ONNX 输出节点和output_processors配置不匹配，后面手动改dst_perm只是绕过了转换错误，但没有生成MaixPy nn.YOLO11能识别的模型输出格式
根据 MaixCAM2 转换文档，output_processors里的节点必须和裁剪后的export.onnx输出节点一致，dst_perm也必须和输出tensor的维度对应
当前模型实际输出是：
/model.23/Concat_output_0 [1, 64, 189]
/model.23/Concat_1_output_0 [1, 1, 189]
/model.23/Concat_2_output_0 [1, 4, 189]
这是3D输出，但原来的dst_perm是4维转置，例如[0, 1, 3, 2]，所以Pulsar2对3D tensor做4D transpose时会越界，从而报这个错误
IndexError: list assignment index out of range
把dst_perm改成[0, 1, 2]后，转换能继续，只能说明没再做错误的4D转置，但生成的模型输出不是MaixPy nn.YOLO11 detector的格式

MaixPy里nn.YOLO11如果看到detector有3个输出，会按三个尺度的4D特征图解析，这个96x96输入，1类模型为例，应该类似于
[1, 65, 12, 12]
[1, 65, 6, 6]
[1, 65, 3, 3]
65 = 16 * 4 + labels_num
189 = 12 * 12 + 6 * 6 + 3 * 3

但是你现在给的是
[1, 64, 189]
[1, 1, 189]
[1, 4, 189]

这是flatten后的DFL分类box输出，所以MaixPy会报output node shape error, please check 2

以下是我的建议
1.用Netron重新检查当前导出的ONNX，不要机械套文档里的节点名，找到真正的三个4D检测头输出节点后再转换
2.或者只保留MaixPy mode 2能识别的两路输出：
[1, labels_num, anchor_num]
[1, 4, anchor_num]
3.使用下面的网页工具
https://wiki.sipeed.com/maixpy/doc/zh/ai_model_converter/online_converter.html

最后如果想继续使用当前三个3D输出，则需要用nn.NN自己做后处理，不能用封装好的nn.YOLO11