专栏算法工具链hb_compile对onnx模型转hbm,结果异常

hb_compile对onnx模型转hbm,结果异常

已解决
穆心20082025-12-16
69
8

2025-10-20 02:15:41,572 INFO log will be stored in /workspace/hb_model_info.log

2025-10-20 02:15:41,572 INFO Start hb_model_info....

2025-10-20 02:15:41,572 INFO hb_model_info version 3.3.11

2025-10-20 02:15:41,572 INFO hbm_path: /workspace/parking_actor.hbm

[02h:15m:41s:575473567ns INFO hbrt4_log::logger] pid:164 tid:164 hbrt4_log/src/logger.rs:403: Logger of HBRT4 initialized, version = 4.1.17

[02h:15m:41s:578598358ns INFO hbrt4_loader::parsing] pid:164 tid:164 hbrt4_loader/src/parsing.rs:85: Load hbm from file; filename="parking_actor.hbm"

HBDK hbm extract desc SUCCESS

2025-10-20 02:15:41,612 INFO ************* parking_actor *************

2025-10-20 02:15:41,612 INFO ############# model deps info #############

2025-10-20 02:15:41,612 INFO builder version     : 3.3.11

2025-10-20 02:15:41,612 INFO hbdk version        : 4.1.17

2025-10-20 02:15:41,612 INFO horizon nn version  : 2.1.9

2025-10-20 02:15:41,612 INFO ############# model_parameters info #############

2025-10-20 02:15:41,612 INFO onnx_model          : /workspace/onnx/parking_actor.onnx

2025-10-20 02:15:41,612 INFO BPU march           : nash-e

2025-10-20 02:15:41,612 INFO layer_out_dump      : False

2025-10-20 02:15:41,613 INFO working dir         : /workspace/onnx/model_output

2025-10-20 02:15:41,613 INFO output_model_file_prefix: parking_actor

2025-10-20 02:15:41,613 INFO remove_node_type    : Quantize;Transpose;Dequantize;Cast;Reshape;Softmax

2025-10-20 02:15:41,613 INFO node_info           : {}

2025-10-20 02:15:41,613 INFO ############# input_parameters info #############

2025-10-20 02:15:41,613 INFO ------------------------------------------

2025-10-20 02:15:41,613 INFO ---------input info : observation ---------

2025-10-20 02:15:41,613 INFO input_name          : observation

2025-10-20 02:15:41,613 INFO input_type_rt       : featuremap

2025-10-20 02:15:41,613 INFO input_space&range   : regular

2025-10-20 02:15:41,613 INFO input_type_train    : featuremap

2025-10-20 02:15:41,613 INFO input_layout_rt     : NCHW

2025-10-20 02:15:41,613 INFO input_layout_train  : NCHW

2025-10-20 02:15:41,613 INFO norm_type           : no_preprocess

2025-10-20 02:15:41,613 INFO input_shape         : 1x126

2025-10-20 02:15:41,613 INFO mean_value          : []

2025-10-20 02:15:41,613 INFO scale_value         : []

2025-10-20 02:15:41,613 INFO std_value           : []

2025-10-20 02:15:41,613 INFO separate_batch      : False

2025-10-20 02:15:41,613 INFO ---------input info : observation end -------

2025-10-20 02:15:41,613 INFO ------------------------------------------

2025-10-20 02:15:41,613 INFO ############# calibration_parameters info #############

2025-10-20 02:15:41,613 INFO calibration_type    : skip

2025-10-20 02:15:41,613 INFO max_percentile      : None

2025-10-20 02:15:41,614 INFO optimization        : run_fast

2025-10-20 02:15:41,614 INFO per_channel         : False

2025-10-20 02:15:41,614 INFO ############# compiler_parameters info #############

2025-10-20 02:15:41,614 INFO debug               : True

2025-10-20 02:15:41,614 INFO optimize_level      : O2

2025-10-20 02:15:41,614 INFO compile_mode        : latency

2025-10-20 02:15:41,614 INFO core_num            : 1

2025-10-20 02:15:41,614 INFO balance_factor      : 100

2025-10-20 02:15:41,614 INFO input_source        : {'observation': 'ddr'}

2025-10-20 02:15:41,614 INFO hbdk3_compatible_mode: False

2025-10-20 02:15:41,614 INFO hbm_path: /workspace/parking_actor.hbm

[02h:15m:41s:614487071ns INFO hbrt4_loader::parsing] pid:164 tid:164 hbrt4_loader/src/parsing.rs:42: Load hbm header from file; filename="parking_actor.hbm"

[02h:15m:41s:614895647ns INFO hbrt4_log::logger] pid:164 tid:164 hbrt4_log/src/logger.rs:403: Logger of HBRT4 initialized, version = 4.1.17

[02h:15m:41s:614904172ns INFO hbrt4_loader::parsing] pid:164 tid:164 hbrt4_loader/src/parsing.rs:85: Load hbm from file; filename="parking_actor.hbm"

2025-10-20 02:15:41,617 INFO ############# Model input/output info #############

2025-10-20 02:15:41,617 INFO NAME        TYPE   SHAPE        DATA_TYPE

2025-10-20 02:15:41,617 INFO ----------- ------ ------------ ---------

2025-10-20 02:15:41,617 INFO observation input  [1, 126]     INT8

2025-10-20 02:15:41,617 INFO action      output [1, 1, 1, 2] INT8

root@1ac653705002:/workspace# ^C

root@1ac653705002:/workspace# hb_model_info parking_actor.onnx 

2025-10-20 02:18:54,651 INFO log will be stored in /workspace/hb_model_info.log

2025-10-20 02:18:54,651 INFO Start hb_model_info....

2025-10-20 02:18:54,651 INFO hb_model_info version 3.3.11

2025-10-20 02:18:57,308 INFO opset version: 11

2025-10-20 02:18:57,308 INFO ############# Model input/output info #############

2025-10-20 02:18:57,308 INFO NAME        TYPE   SHAPE               DATA_TYPE

2025-10-20 02:18:57,308 INFO ----------- ------ ------------------- ---------

2025-10-20 02:18:57,308 INFO observation input  ['batch_size', 126] FLOAT32

2025-10-20 02:18:57,308 INFO action      output ['batch_size', 2]   FLOAT

 

hrt_model_exec model_info --model_file parking_actor.hbm

root@9304cd369d75:/workspace# hrt_model_exec model_info --model_file parking_actor.hbm

[UCP]: log level = 3

[UCP]: UCP version = 3.3.3

[VP]: log level = 3

[DNN]: log level = 3

[HPL]: log level = 3

[UCPT]: log level = 6

[DSP]: log level = 3

hrt_model_exec model_info --model_file parking_actor.hbm

[DNN]: 3.3.3_(4.1.17 HBRT)

Load model to DDR cost 19.807ms.

This model file has 1 model:

[parking_actor]  

---------------------------------------------------------------------

[model name]: parking_actor

 

[model desc]: {"BUILDER_VERSION": "3.3.11", "HBDK_VERSION": "4.1.17", "HBDK_RUNTIME_VERSION": null, "HORIZON_NN_VERSION": "2.1.9", "CAFFE_MODEL": null, "PROTOTXT": null, "ONNX_MODEL": "/workspace/onnx/parking_actor.onnx", "MARCH": "nash-e", "LAYER_OUT_DUMP": "False", "LOG_LEVEL": null, "WORKING_DIR": "/workspace/onnx/model_output", "MODEL_PREFIX": "parking_actor", "OUTPUT_NODES": "", "REMOVE_NODE_TYPE": "Quantize;Transpose;Dequantize;Cast;Reshape;Softmax", "REMOVE_NODE_NAME": "", "DEBUG_MODE": "", "NODE_INFO": "{}", "INPUT_NAMES": "observation", "INPUT_SPACE_AND_RANGE": "regular", "INPUT_TYPE_RT": "featuremap", "INPUT_TYPE_TRAIN": "featuremap", "INPUT_LAYOUT_TRAIN": "NCHW", "INPUT_LAYOUT_RT": "NCHW", "NORM_TYPE": "no_preprocess", "MEAN_VALUE": "[]", "SCALE_VALUE": "[]", "STD_VALUE": "[]", "INPUT_SHAPE": "1x126", "INPUT_BATCH": "", "SEPARATE_BATCH": "False", "SEPARATE_NAME": "", "CUSTOM_OP_METHOD": null, "CUSTOM_OP_DIR": null, "CUSTOM_OP_REGISTER_FILES": "", "OPTIMIZATION": "run_fast", "CALI_TYPE": "skip", "CAL_DATA_DIR": "", "PER_CHANNEL": "False", "MAX_PERCENTILE": "None", "RUN_ON_CPU": "", "RUN_ON_BPU": "", "QUANT_CONFIG": null, "ADVICE": 0, "DEBUG": "True", "OPTIMIZE_LEVEL": "O2", "COMPILE_MODE": "latency", "CORE_NUM": 1, "MAX_TIME_PER_FC": 0, "BALANCE_FACTOR": 100, "ABILITY_ENTRY": null, "INPUT_SOURCE": {"observation": "ddr"}, "hbdk3_compatible_mode": "False", "CALI_EXTRA_PARAM": {}, "EXTRA_PARAMS": {}}

input[0]: 

name: observation

valid shape: (1,126,)

aligned byte size: 128

tensor type: HB_DNN_TENSOR_TYPE_S8

quanti type: SCALE

stride: (126,1,)

scale data: 0.00787402,

 

output[0]: 

name: action

valid shape: (1,1,1,2,)

aligned byte size: 128

tensor type: HB_DNN_TENSOR_TYPE_S8

quanti type: SCALE

stride: (2,2,2,1,)

scale data: 0.0059968,

zero_point data: ,

 

 

 

算法工具链
征程6
评论3
0/1000
  • Huanghui
    Lv.5

    这个是啥?J6E+ OE3.2.0,有啥问题呢?

    2025-12-16
    0
    5
    • 穆心2008回复Huanghui:

      转完的hbm推理结果和onnx的不一样,大佬指点一下

      2025-12-17
      0
    • Vincent回复穆心2008:

      请具体描述下问题吧,

      2025-12-18
      0
    • Vincent回复穆心2008:

      这个不一样指的是完全不一样还是说有一定误差?

      2025-12-18
      0
    • Vincent回复穆心2008:

      另外,你用hb_runtime推理了onnx和hbm模型了吗 ,直接推理出来的结果余弦相似度是否很高呢? 要首先排除是不是后处理导致的

      2025-12-18
      0
    • Vincent回复穆心2008:

      还有就是我看你的这个log,是remove掉了quant和dequant,在推理hbm时候要手动量化与反量化,请问您这个步骤您做了吗

      2025-12-18
      0
  • Huanghui
    Lv.5

    你好,了解一下目前该问题的状态呢, 转完的hbm推理结果和onnx的不一样 ,如果这里的onnx是ptq的输入模型,而hbm是经过量化后的上板模型,二者之间因为存在浮转定的过程,另外校准数据和配置等也可能存在一定的偏差,从而导致存在一定的差距(不是二进制对齐的),但是这个如果在量化损失范围内是正常的。

    2025-12-26
    0
    0
  • YCJ
    Lv.4

    您好!请问这个问题现在是解决了吗?若没有解决,可否分享一下模型?这边帮您复现一下

    2026-01-22
    0
    0