BEAGLE-3P-BBONE-AI: ONNX TIDL Importer

Maciej Kaca

Part Number: BEAGLE-3P-BBONE-AI
Other Parts Discussed in Thread: AM5729

Hi,

i need to import an object detection model. This model was trained with mmdetection from this repo https://github.com/TexasInstruments/edgeai-mmdetection/

I am confused which SDK should I use. I tried with Vision-SDK and AM57X-SDK. Neither of them contain metaArchType for Onnx model. Only a Jacinto processor-sdk-rtos-jacinto7 supports this parameter. But when I use the jacinto SDK I have an error "ERROR: Wrong network binary size". So what is the workflow for importing an object detection Onnx model? Which SDK should I use?

over 3 years ago

0 kumar.desappan over 3 years ago

TI__Mastermind 22145 points

Please refer below for importing mmdetection model for Jacinto7

https://github.com/TexasInstruments/edgeai-tidl-tools

We have an example also available in the below script (ssd-lite_mobilenetv2_fpn)

github.com/.../onnxrt_ep.py

0 Maciej Kaca over 3 years ago in reply to kumar.desappan

Prodigy 30 points

You mentioned Jacinto 7. Can I use edgeai-tidl-tools to compile models for AM5729(Beagle Bone AI)? will I get .bin (model and params) files with this tool?

0 kumar.desappan over 3 years ago in reply to Maciej Kaca

TI__Mastermind 22145 points

edgeai-tidl-tools can not be used for AM5729. Please follow the documentation available in AM57X-SDK to depoly the model on AM5729

0 Maciej Kaca over 3 years ago in reply to kumar.desappan

Prodigy 30 points

As I said, I a have a problem with AM57X-SDK. it lacks parameters such as metaArchType, prostProcType, metaLayersNamesList. Which are needed for object detection model. I have an Onnx file and protxt but I'm not able to import these files with the Tidl Import Tool provided with the AM57X-SDK because the above parameters are not supported.

+1 kumar.desappan over 3 years ago in reply to Maciej Kaca

TI__Mastermind 22145 points

Only caffe based SSD OD models are supported in the AM57X. The classification and Segmentation models are supported from Caffe and TF Model formats.

Please refer to Documentation for supported models and operator types

0 Maciej Kaca over 3 years ago in reply to kumar.desappan

Prodigy 30 points

Then should I use this framework for training? https://git.ti.com/cgit/jacinto-ai/caffe-jacinto/

0 kumar.desappan over 3 years ago in reply to Maciej Kaca

TI__Mastermind 22145 points

Yes

0 Maciej Kaca over 3 years ago in reply to kumar.desappan

Prodigy 30 points

is there any current tutorial about the installation and training with caffe-jacinto? I have a lot of errors during installation, the installation instructions seems to be outdated.

I found this, https://software-dl.ti.com/processor-sdk-linux/esd/docs/06_03_00_106/linux/Foundational_Components/Machine_Learning/neo.html . Is this a good alternative for TIDL? Is it compatible with BeagleBone AI?

0 Paula Carrillo over 3 years ago in reply to Maciej Kaca

TI__Mastermind 39160 points

Hi, AM57x PSDK Linux 6.3 supports BBAI.

If HW device is still under evaluation, maybe you could consider J7 (TDA4). We have a TI EdgeAI cloud, (https://dev.ti.com/edgeaisession) where you can evaluate your onw models or from our modelzoo. We currently support TFLite, NEO-AI-DLR, and ONNX runtimes (Python and C-APIs).

We also have a starter kit board for local evaluation.

Thank you,

Paula

0 Maciej Kaca over 3 years ago in reply to kumar.desappan

Prodigy 30 points

could you please give me an email to Manu Mathew? He works with you. He is the owner of the caffe-jacinto repo. And he posted there are some building errors while trying to compile the repo. Maybe he could help me with installing this.

github.com/.../33

0 Paula Carrillo over 3 years ago in reply to Maciej Kaca

TI__Mastermind 39160 points

Hi Maciej, could you please share the errors? and also let me know if there is a way we can try to reproduce them in our side.

Also, just FYI, AM57x PSDK Linux also uses TIDL, different flavor (as we have a combination of DSP + EVE cores, instead of DSP +MMA), but TIDL nevertheless.

thank you,

Paula

0 Maciej Kaca over 3 years ago in reply to Paula Carrillo

Prodigy 30 points

I have already installed caffe-jacinto. But I have a problem while I'm trying to run the training. I have this error. I'm running caffe-jacinto on Google Colab with Tesla K80. What can I do to fix it?

F1206 12:18:32.126641 16318 permute_layer.cu:70] Check failed: error == cudaSuccess (209 vs. 0)  no kernel image is available for execution on the device
*** Check failure stack trace: ***
    @     0x7f132a8e40cd  google::LogMessage::Fail()
    @     0x7f132a8e5f33  google::LogMessage::SendToLog()
    @     0x7f132a8e3c28  google::LogMessage::Flush()
    @     0x7f132a8e6999  google::LogMessageFatal::~LogMessageFatal()
    @     0x7f132ba5dddd  caffe::PermuteLayer<>::Forward_gpu()
    @     0x7f132b45f97b  caffe::Layer<>::Forward()
    @     0x7f132b8b3d63  caffe::Net::ForwardFromTo()
    @     0x7f132b8b3ed7  caffe::Net::Forward()
    @     0x5597aad93743  test_detection()
    @     0x5597aad8ed11  main
    @     0x7f1329a10bf7  __libc_start_main
    @     0x5597aad8faaa  _start
I1206 12:18:32.433220 16336 caffe.cpp:902] This is NVCaffe 0.17.0 started at Mon Dec  6 12:18:32 2021
I1206 12:18:32.757375 16336 caffe.cpp:904] CuDNN version: 7605
I1206 12:18:32.757387 16336 caffe.cpp:905] CuBLAS version: 11300
I1206 12:18:32.757392 16336 caffe.cpp:906] CUDA version: 11010
I1206 12:18:32.757398 16336 caffe.cpp:907] CUDA driver version: 11020
I1206 12:18:32.757402 16336 caffe.cpp:908] Arguments: 
[0]: /content/caffe-jacinto/build/tools/caffe.bin
[1]: test_detection
[2]: --model=training/voc0712/JDetNet/20211206_12-13_ds_PSP_dsFac_32_hdDS8_1/test_quantize/test.prototxt
[3]: --iterations=2
[4]: --weights=/content/caffe-jacinto-models/trained/object_detection/voc0712/JDetNet/ssd512x512_ds_PSP_dsFac_32_fc_0_hdDS8_1_kerMbox_3_1stHdSameOpCh_1/sparse/voc0712_ssdJacintoNetV2_iter_104000.caffemodel
[5]: --gpu
[6]: 0
I1206 12:18:32.768014 16336 gpu_memory.cpp:105] GPUMemory::Manager initialized
I1206 12:18:32.768074 16336 gpu_memory.cpp:107] Total memory: 11996954624, Free: 11892621312, dev_info[0]: total=11996954624 free=11892621312
I1206 12:18:32.768326 16336 caffe.cpp:406] Use GPU with device ID 0
I1206 12:18:32.768502 16336 caffe.cpp:409] GPU device name: Tesla K80

Processors

Processors forum

BEAGLE-3P-BBONE-AI: ONNX TIDL Importer