目录

使用OMG工具将ONNX模型转换为OM离线模型报错

问题现象

使用OMG工具将ONNX模型转换为OM离线模型报错,命令执行部分日志如下:

W/AI_FMK (16075): ops_kernel_ store_manager.cpp DlopenComputeLibrary(41)::"dlopen so failed: libai_npucore_itf.so: cannot open shared object file:No such file or directory"
I/AI_FMK (16075): ops_kernel_store_manager.cpp DlCloseComputeLibrary(83)::"handle is null, not need to close library" 
I/AI_FMK (16075): model_util.cpp BuildOrigin2IRGraph(229)::"Modelutil::BuildOrigin2IRGraph from file" 
I/AI_FMK (16075): proto_util.cpp ReadBytesFromBinaryFile(192)::"Read size:16697653"
W/AI_FMK (16075): parser_factory.cpp LoadcustomOpLib(58)::"dlerror: libcustom_op.so: cannot open shared object file:No such file or directory 
E/AI_FMK(16075): onnx_graph_parser.cpp ParserGraph(875):Pre-check has errors."
E/AI_INFRA (16075): model_util.cpp ParseOriginBuffer2IrGraph(475)::"ret == SUCCESS" "false, return ret." 
E/AI_INFRA (16075): model_util.cpp ParseOriginONNX2IrGraph(541)::"ParseOriginBuffer2IrGraph(inputOptions, srcbuffer, irGraph, hiai::FrameworkType::ONNX) == SUCCESS" "false, return FAIL." 
E/AI_INFRA (16075): model_util.cpp BuildOrigin2IRGraph(233)::"ret == SUCCESS" "false,return ret 
E/AI_FMK (16075): omg.cpp GenerateIRModel(113)::"Failed to generator IR graph!."
E/OMG_TOOL (16075):command_util.cpp ProcessCommand(1226)::"OMG Generate execute failed!!"
E/OMG_TOOL (16075): main.cpp main(21)::"OMG generate offline model failed. Please see the log or pre-checking report for more details."

当前目录下check_result.json文件的部分信息如下:

  {"name": "/quant1/Constant", "result": "success", "type": "Constant" },
  {"name": "/quant1/Constant_1", "result": "success", "type": "Constant" },
  {"cause": [{ "code": 1, "message": "The type: QuantizeLinear is not supported." }], "name": "/quant1/QuantizeLinear", "result": "failed", "type": "QuantizeLinear"},
  {"name": "/backbone/conv1/Cast","result": "success", "type": "Cast" },
  {"name": "/backbone/conv1/Constant", "result": "success", "type": "Constant" },
  {"name": "/backbone/conv1/Constant_1", "result": "success", "type": "Constant" },
  {"cause": [{"code": 1, "message": "The type: DequantizeLinear is not supported." }], "name": "/backbone/conv1/DequantizeLinear", "result": "failed", "type":"DequantizeLinear" },
  {"name": "/backbone/conv1/Constant_2", "result": "success", "type": "Constant" },
  {"name": "/backbone/conv1/Constant_3","result": "success", "type": "Constant"},
  {"name": "/backbone/conv1/Constant_4", "result": "success", "type": "Constant"},
  {"cause": [{ "code": 1, "message": "The type: DequantizeLinear is not supported." }], "name": "/backbone/conv1/DequantizeLinear_1", "result": "failed", "type":"DequantizeLinear"},
  {"name": "/backbone/conv1/Constant_5", "result": "success", "type": "Constant" },
  {"name": "/backbone/conv1/ConstantOfShape", "result": "success", "type": "ConstantOfShape" },
  {"name": "/backbone/conv1/Constant_6", "result": "success", "type": "Constant" },
  {"name": "/backbone/conv1/Constant_7", "result": "success", "type": "Constant" },
  {"name": "/backbone/conv1/Cast_1", "result": "success", "type": "Cast" },
  {"cause": [{ "code": 1, "message": "The type: DequantizeLinear is not supported." }], "name": "/backbone/conv1/DequantizeLinear_2", "result": "failed", "type":"DequantizeLinear" },
  {"name": "/backbone/conv1/Conv", "result": "success", "type": "Conv" },
  {"name": "/backbone/conv1/Constant_8", "result": "success", "type": "Constant" },
  {"name": "/backbone/conv1/Constant_9", "result": "success", "type": "Constant" },
  {"cause": [{ "code": 1, "message": "The type: QuantizeLinear is not supported." }], "name": "/backbone/conv1/QuantizeLinear", "result": "failed", "type":"QuantizeLinear"},
  {"name": "/backbone/act1/Cast", "result": "success", "type": "Cast" },
  {"name": "/backbone/act1/Constant", "result": "success", "type": "Constant" },
  {"name": "/backbone/act1/Constant_1", "result": "success", "type": "Constant"},
  {"cause": [{ "code": 1, "message": "The type: DequantizeLinear is not supported." }], "name": "/backbone/act1/DequantizeLinear", "result": "failed", "type":"DequantizeLinear" },
  {"name": "/backbone/act1/LeakyRelu", "result": "success", "type": "LeakyRelu" },
  {"name": "/backbone/conv1/Constant_2", "result": "success", "type": "Constant" },
  {"name": "/backbone/conv1/Constant_3","result": "success", "type": "Constant"}

背景知识

问题定位

  1. 根据以下执行日志,可得知OMG命令转换离线模型失败,更多的信息需要查看预检查报告,即check_result.json文件信息。
    E/OMG_TOOL (16075): main.cpp main(21)::"OMG generate offline model failed. Please see the log or pre-checking report for more details."
  2. 根据当前转换目录下生成的check_result.json文件内的以下报错信息可知,该ONNX模型使用了QuantizeLinear和DequantizeLinear这2个ONNX的量化算子,当前HarmonyOS 5.0的CANN Kit仅支持官网提供的量化方案:模型轻量化。
    {"cause":[{"code":1,"message":"The type: QuantizeLinear is not supported."}],"name":"/quant1/QuantizeLinear","result":"failed","type":"QuantizeLinear"}
    ···
    {"cause":[{"code":1,"message":"The type: DequantizeLinear is not supported."}],"name":"/backbone/conv1/DequantizeLinear_1","result":"failed","type":"DequantizeLinear"}

分析结论

该模型使用了HarmonyOS 5.0当前不支持的ONNX量化算子,导致模型转换时报错。

修改建议

使用HarmonyOS 5.0提供的轻量化工具对该ONNX模型进行模型轻量化优化之后,再根据官方文档进行模型转换。

目前支持ONNX的轻量化方式为无训练量化。可以参考模型轻量化示例。