SK-TDA4VM: Executing yolox model on TI SOC failed.

Hi,

Thank you for the question. Could you provide the following information so we could debug further?

Could you clarify what version of edgeai-tidl-tools you are seeing this behavior?

Abdulghani Altaweel said:

I have used a model from yolox-ti-edgeai, and I have also tried models from the model zoo

From my understanding you are running models that we offer out of the box, from edgeai-modelzoo and edgeai-yolox which are contained in edgeai-tensorlab (https://github.com/TexasInstruments/edgeai-tensorlab/tree/main). Can you specify what models you are seeing this issue with?

Best,

Asha

Hi,

Thanks for the response. The version of edge-tidl-tools is 09_02_06_00. The model is own yolox model from https://github.com/TexasInstruments/edgeai-yolox. 

Hi Abdulghani,

I do apologize for the delay in getting back to you. Is this still a blocking issue for you currently? 

When you mean own yolex model, could you specify more? Is this one of the pretrained models (if so can you specify the name)?

From my understanding, the inference on target is failing. Have you checked the inference on host emulation (running on x86 PC)? Does this result in any errors?

Can you provide the compilation logs and the inference logs with debug_level = 2 as a txt file for further debug?

Best,

Asha

Hello,

• Actually I got the same error when trying to compile the model edgeai-yolox-nano-ti-lite-onnx from model zoo and integrate it to the edge_ai_apps on beagleboard-ai64.
• I did not try doing inference on host pc.

Note:

- I have used edgeai-tidl-tools (with another version as mentioned above) and edgeai-benchmark to compile the model this time.
- The model compile successfully but it does not execute on target with the example. (The example could run but no bounding boxes appear on screen)

thanks.

 33668.604213 s:  VX_ZONE_ERROR:[ownContextSendCmd:815] Command ack message returned failure cmd_status: -1
 33668.604464 s:  VX_ZONE_ERROR:[ownContextSendCmd:851] tivxEventWait() failed.
 33668.604536 s:  VX_ZONE_ERROR:[ownNodeKernelInit:538] Target kernel, TIVX_CMD_NODE_CREATE failed for node TIDLNode
 33668.604597 s:  VX_ZONE_ERROR:[ownNodeKernelInit:539] Please be sure the target callbacks have been registered for this core
 33668.604656 s:  VX_ZONE_ERROR:[ownNodeKernelInit:540] If the target callbacks have been registered, please ensure no errors are occurring within the create callback of this kernel
 33668.604719 s:  VX_ZONE_ERROR:[ownGraphNodeKernelInit:583] kernel init for node 0, kernel com.ti.tidl ... failed !!!
 33668.604797 s:  VX_ZONE_ERROR:[vxVerifyGraph:2055] Node kernel init failed
 33668.604868 s:  VX_ZONE_ERROR:[vxVerifyGraph:2109] Graph verify failed

hello,

Also when compiling the model I got the warning:

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.

Hi,

Copying your question from the other thread:

And the model cannot run on target with the edge_ai_apps. According to this post TDA4VM: TIDL ONNX import error with 32-bit numParamBits the problem is solved by setting the parameter numParamBits = 32.

Where can I set this parameter in the edgeai-tidl-tools (version 08_02_00_05) 

Running on EVM does not support 32-bit. You can use the tensor_bits parameter in edgeai-tidl-tools to modify this (either 8 or 16 bit to run on target). Also, it is recommended to run on host emulation first to determine functionality.

Best,

Asha

Hello
thanks for response. The tensor_bit parameter is set already to 8 by default. Despite the warning I could run the model on target.
The problem was solved by setting the num_class of coco dataset from 80 to 90.

Hi Abdulghani,

Good to know that you were able to resolve the issue. Can I close this ticket?

Best,

Asha

Hello,

When we could consider the warning as obsolete, then yes.

Thanks again for help