x86上训练好的YOLOv8模型（x86测试可用）用 https://milkv.io/zh/docs/duo/application-development/tdl-sdk/tdl-sdk-yolov8 转化过后在duoS上加载失败，请教如何将自己训练的pt模型部署到duos上,非对称模型会产生bug

[YunxiangLuo]

操作全过程参考：https://github.com/Arielfoever/milkv-duo-yolo11

参考 TPU-MLIR 文档 配置好 TPU-MLIR 工作环境，参数解析请参考 TPU-MLIR 文档。

配置好工作环境后,在与本项目同级目录下创建一个model_yolov8n目录,将模型和图片文件放入其中。

模型转换命令如下：

model_transform.py \
--model_name wangzhe \
--model_def ../best.onnx \
--input_shapes [[1,3,640,640]] \
--mean 0.0,0.0,0.0 \
--scale 0.0039216,0.0039216,0.0039216 \
--keep_aspect_ratio \
--pixel_format rgb \
--mlir wangzhe.mlir

run_calibration.py wangzhe.mlir \
--dataset ../wangzhe/ \
--input_num 100 \
-o wangzhe_cali_table

用校准表生成 int8 对称 cvimodel:

model_deploy.py \
--mlir wangzhe.mlir \
--quant_input --quant_output \
--quantize INT8 \
--calibration_table wangzhe_cali_table \
--processor cv181x \
--model wangzhe-int8-sym.cvimodel

非对称模型会产生bug

model_deploy.py \
--mlir wangzhe.mlir \
--quant_input --quant_output \
--quantize INT8 --asymmetric \
--calibration_table wangzhe_cali_table \
--processor cv181x \
--model wangzhe-int8-asym.cvimodel

使用非对称会出现bug。

root@bacc5a4054eb:/workspace/model_yolov5s/workspace# model_deploy.py \                          
--mlir wangzhe.mlir \                                                                           
--quant_input --quant_output \                                                                  
--quantize INT8 --asymmetric \                                                                  
--calibration_table wangzhe_cali_table \                                                        
--processor cv181x \                                                                            
--model wangzhe-int8-asym.cvimodel                                                              
2024/11/17 21:17:43 - INFO : TPU-MLIR 081d28a-20241114                     
2024/11/17 21:17:43 - INFO :                                                                    
  load_config Preprocess args :                                                                 
        resize_dims           : [640, 640]                                                      
        keep_aspect_ratio     : True                                                            
        keep_ratio_mode       : letterbox                                                       
        pad_value             : 0                                                               
        pad_type              : center                                                          
        input_dims            : [640, 640]                                                      
        --------------------------                                                              
        mean                  : [0.0, 0.0, 0.0]                                                 
        scale                 : [0.0039216, 0.0039216, 0.0039216]
        --------------------------
        pixel_format          : rgb                                                             
        channel_format        : nchw   
                                                                                                
[Running]: tpuc-opt wangzhe.mlir --processor-assign="chip=cv181x mode=INT8 num_device=1 num_core=1 addr_mode=auto" --import-calibration-table="file=wangzhe_cali_table asymmetric=True" --processor-top-optimize --convert-top-to-tpu=" asymmetric=True doWinograd=False ignore_f16_overflow=False q_group_size=0 matmul_perchannel=False gelu_mode=normal" --canonicalize --weight-fold -o wang
zhe_cv181x_int8_asym_tpu.mlir                                                                   
<unknown>:0: error: illegal min and max: (8.494865e+00:-8.494865e+00)          
tpuc-opt: /llvm-project/mlir/include/mlir/IR/StorageUniquerSupport.h:180: static ConcreteT mlir::detail::StorageUserBase<mlir::quant::CalibratedQuantizedType, mlir::quant::QuantizedType, mlir::quant::detail::CalibratedQuantizedTypeStorage, mlir::detail::TypeUniquer>::get(mlir::MLIRContext *, Args...) [ConcreteT = mlir::quant::CalibratedQuantizedType, BaseT = mlir::quant::QuantizedType, StorageT = mlir::quant::detail::CalibratedQuantizedTypeStorage, UniquerT = mlir::detail::TypeUniquer, Traits = <>, Args = <mlir::Type, double, double>]: Assertion `succeeded(ConcreteT::verify(getDefaultDiagnosticEmitFn(ctx), args...))' failed.
PLEASE submit a bug report to https://github.com/llvm/llvm-project/issues/ and include the crash backtrace.
Stack dump:
0.      Program arguments: tpuc-opt wangzhe.mlir --init "--processor-assign=chip=cv181x mode=INT8 num_device=1 num_core=1 addr_mode=auto" "--import-calibration-table=file=wangzhe_cali_table asymmetric=True" --processor-top-optimize "--convert-top-to-tpu= asymmetric=True doWinograd=False ignore_f16_overflow=False q_group_size=0 matmul_perchannel=False gelu_mode=normal" --canonicalize --weight-fold --deinit --mlir-print-debuginfo -o wangzhe_cv181x_int8_asym_tpu.mlir
 #0 0x00005fde6aca1e87 (/workspace/tpu-mlir/install/bin/tpuc-opt+0x86ae87)
 #1 0x00005fde6ac9fbae (/workspace/tpu-mlir/install/bin/tpuc-opt+0x868bae)
 #2 0x00005fde6aca280a (/workspace/tpu-mlir/install/bin/tpuc-opt+0x86b80a)
 #3 0x000077eeb6842520 (/lib/x86_64-linux-gnu/libc.so.6+0x42520)
 #4 0x000077eeb68969fc pthread_kill (/lib/x86_64-linux-gnu/libc.so.6+0x969fc)
 #5 0x000077eeb6842476 gsignal (/lib/x86_64-linux-gnu/libc.so.6+0x42476)
 #6 0x000077eeb68287f3 abort (/lib/x86_64-linux-gnu/libc.so.6+0x287f3)
 #7 0x000077eeb682871b (/lib/x86_64-linux-gnu/libc.so.6+0x2871b)
 #8 0x000077eeb6839e96 (/lib/x86_64-linux-gnu/libc.so.6+0x39e96)
 #9 0x00005fde6c3ad806 (/workspace/tpu-mlir/install/bin/tpuc-opt+0x1f76806)
#10 0x00005fde6c3ad714 (/workspace/tpu-mlir/install/bin/tpuc-opt+0x1f76714)
#11 0x00005fde6c160bc2 (/workspace/tpu-mlir/install/bin/tpuc-opt+0x1d29bc2)
#12 0x00005fde6adf3c5e (/workspace/tpu-mlir/install/bin/tpuc-opt+0x9bcc5e)
#13 0x00005fde6c15f1f7 (/workspace/tpu-mlir/install/bin/tpuc-opt+0x1d281f7)
#14 0x00005fde6c273ad4 (/workspace/tpu-mlir/install/bin/tpuc-opt+0x1e3cad4)
#15 0x00005fde6c274101 (/workspace/tpu-mlir/install/bin/tpuc-opt+0x1e3d101)
#16 0x00005fde6c2765a8 (/workspace/tpu-mlir/install/bin/tpuc-opt+0x1e3f5a8)
#17 0x00005fde6ac9353b (/workspace/tpu-mlir/install/bin/tpuc-opt+0x85c53b)
#18 0x00005fde6ac92904 (/workspace/tpu-mlir/install/bin/tpuc-opt+0x85b904)
#19 0x00005fde6c48a458 (/workspace/tpu-mlir/install/bin/tpuc-opt+0x2053458)
#20 0x00005fde6ac8cc0a (/workspace/tpu-mlir/install/bin/tpuc-opt+0x855c0a)
#21 0x00005fde6ac8d0d4 (/workspace/tpu-mlir/install/bin/tpuc-opt+0x8560d4)
#22 0x00005fde6ac8bb1a (/workspace/tpu-mlir/install/bin/tpuc-opt+0x854b1a)
#23 0x000077eeb6829d90 (/lib/x86_64-linux-gnu/libc.so.6+0x29d90)
#24 0x000077eeb6829e40 __libc_start_main (/lib/x86_64-linux-gnu/libc.so.6+0x29e40)
#25 0x00005fde6ac8af25 (/workspace/tpu-mlir/install/bin/tpuc-opt+0x853f25)
Aborted (core dumped)
Traceback (most recent call last):
  File "/workspace/tpu-mlir/python/tools/model_deploy.py", line 479, in <module>
    lowering_patterns = tool.lowering()
  File "/workspace/tpu-mlir/python/tools/model_deploy.py", line 170, in lowering
    patterns = mlir_lowering(self.mlir_file,
  File "/workspace/tpu-mlir/python/utils/mlir_shell.py", line 202, in mlir_lowering
    _os_system(cmd, mute=mute,log_level=log_level)
  File "/workspace/tpu-mlir/python/utils/mlir_shell.py", line 62, in _os_system
    raise RuntimeError("[!Error]: {}".format(cmd_str))
RuntimeError: [!Error]: tpuc-opt wangzhe.mlir --processor-assign="chip=cv181x mode=INT8 num_device=1 num_core=1 addr_mode=auto" --import-calibration-table="file=wangzhe_cali_table asymmetric=True" --processor-top-optimize --convert-top-to-tpu=" asymmetric=True doWinograd=False ignore_f16_overflow=False q_group_size=0 matmul_perchannel=False gelu_mode=normal" --canonicalize --weight-fold -o wangzhe_cv181x_int8_asym_tpu.mlir