TDA4VM: 使用tidl_model_import.out进行模型转换报错

SSS

Part Number: TDA4VM

不太理解哪里出了问题，请帮忙解答一下,THANKS

错误信息如下:

Current ONNX OpSet Version   : 9
ONNX operator Div is not suported now.. By passing
Could not find Indata for data ID 16
INFO : Couldn't open inData file:   , Skipping Range Collection for Quantization

------------------ Network Compiler Traces -----------------------------
Wrong Network structure: Higher ( 128 > 127) number of co-ex Layers
Wrong Network structure: Higher ( 129 > 127) number of co-ex Layers
Error : Error Code = <ERR_WRONG_NETWORK_STRUCTURE>
Could not open /home/swl/ws/tidl/squeezenet1.1/perfSimInfo.bin
ERROR: TIDL_E_QUANT_STATS_NOT_AVAILABLE] tidl_quant_stats_tool.out fails to collect dynamic range. Please look into quant stats log. This model will get fault on target.
WARNING: [TIDL_E_DATAFLOW_INFO_NULL] ti_cnnperfsim.out fails to allocate memory in MSMC. Please look into perfsim log. This model can only be used on PC emulation, it will get fault on target.
****************************************************
**          1 WARNINGS          1 ERRORS          **
****************************************************

3 年多前

0 Shine 3 年多前

TI__Guru**** 357097 points

请问用的是哪个版本的SDK? 转换命令是怎么写的？

0 SSS 3 年多前回复 Shine

ti-processor-sdk-rtos-j721e-evm-07_01_00_11+tidl_j7_01_03_00_11

命令:./out/tidl_model_import.out ~/ws/tidl/tidl_import_onnx.txt

tidl_import_onnx.txt:

modelType          = 2
inputNetFile      = "/home/sss/ws/tidl/superpoint_v1.onnx"
outputNetFile      = "/home/sss/ws/tidl/tidl_net_squeezenet1.1.bin"
outputParamsFile   = "/home/sss/ws/tidl/tidl_io_squeezenet1.1_"
numParamBits = 8
numFeatureBits = 8
inWidth = 224
inHeight = 224
inNumChannels = 1
perfSimConfig = "/home/sss/ti-processor-sdk-rtos-j721e-evm-07_01_00_11/tidl_j7_01_03_00_11/ti_dl/test/testvecs/config/import/device_config.cfg"

0 Shine 3 年多前回复 SSS

TI__Guru**** 357097 points

我把您的问题升级到e2e论坛，请关注下面帖子的回复。
https://e2e.ti.com/support/processors-group/processors/f/processors-forum/1050325/tda4vm-error-when-using-tidl_model_import-out-to-convert-model

+1 Shine 3 年多前回复 Shine

TI__Guru**** 357097 points

请看下面e2e工程师的回复。

The customer needs to specify "inData" as part of import config file - they can refer to any of the examples in the following documentation:

https://software-dl.ti.com/jacinto7/esd/processor-sdk-rtos-jacinto7/07_01_00_11/exports/docs/tidl_j7_01_03_00_11/ti_dl/docs/user_guide_html/md_tidl_model_import.html

Also, the DIV operation is not supported by TIDL, so model as it is won't run on standalone TIDL. Here's the list of TIDL supported layers:

https://software-dl.ti.com/jacinto7/esd/processor-sdk-rtos-jacinto7/07_01_00_11/exports/docs/tidl_j7_01_03_00_11/ti_dl/docs/user_guide_html/md_tidl_layers_info.html

In case the customer can modify the network and replace Div layer by some TIDL supported layers, they should be able to import the network.

Otherwise, we also have open source runtimes based offering which can be used even if some layer is not supported by TIDL. Documentation for the same:

https://software-dl.ti.com/jacinto7/esd/processor-sdk-rtos-jacinto7/07_01_00_11/exports/docs/tidl_j7_01_03_00_11/ti_dl/docs/user_guide_html/usergroup5.html

I would recommend migrating to the latest version as it supports ONNX runtime as well in addition to TVM and tflite runtime.

If customer can migrate to SDK 8.0, they can refer to the following github repo which works independent of the need to install the SDK : https://github.com/TexasInstruments/edgeai-tidl-tools

0 SSS 3 年多前回复 Shine

Thanks

处理器

处理器论坛

TDA4VM: 使用tidl_model_import.out进行模型转换报错