• State Resolved
  • Locked Locked
  • Replies 2 replies
  • Subscribers 98 subscribers
  • Views 500 views
  • Users 0 members are here
Support feedback
Options
Options
Related

This thread has been locked.

If you have a related question, please click the "Ask a related question" button in the top right corner. The newly created question will be automatically linked to this question.

Original question:

TDA4VM: TVM/Neo-AI-DLR::getting an error , while Inferring with TVM runtime for host(PC)

TDA4VM: TVM/Neo-AI-DLR::Inference steps

Vasanthkumar Mattihalli
Prodigy 80 points

Part Number: TDA4VM

Dear TI Team 

Is it possible to share the example shows -how to use "tidl_denylist" option in TIDLCompiler function of test_tidl_j7.py module. 

Will there be any difference in the generated deliverable/artifacts when SW is compiled with and without configuration of tidl_denylist. if yes, which module in generated artifiacts depicts the changes ?

Kindly share the detail steps/procedure to infer the TVM compiled model on target(j7-EVM) using TVM inference approch 

Note

Versions used :-

SDK: ti-processor-sdk-rtos-j721e-evm-07_01_00_11

TVM version: REL.TIDL.J7.01.03.00.11

OS : Ubuntu 18.04

Thanks :-) 

  • +1
    TI__Expert 4105 points

    You can use tvm_dlr tvm_compilation_tflite_example.py as an example.  Run the compilation, open the model artifacts, tempDir/relay_graph.import.txt and see which layers are offloaded to TIDL.  For this tflite inceptionnet v3 example, all layers are offloaded to TIDL.

    Say, for example, you want the last layer, "nn.softmax", not offloaded to TIDL, you can use the deny_list option in TIDLCompiler.  Modify tvm_compilation_tflite_example.py, add the following line to TIDLCompiler arguments:

    deny_list="nn.softmax",

    Run the compilation again, and see the difference in tempDir/relay_graph.import.txt.

    deny_list is mainly provided as a debugging option, it may not bring value in real application.

    -Yuan

  • 0 in reply to Yuan Zhao
    Prodigy 80 points

    Dear Yuan , 

    Thanks for the information.