TDA4VM: Algorithm model deployment application problem

Hi Ti,

       Regarding the above question, I would like to add the following points, please take a look together.Thank you!

       1. Tried two YOLO projects officially provided by TI, and the corresponding project links are as follows:

YOLOV5：TexasInstruments/edgeai-yolov5: YOLOv5        in PyTorch > ONNX > CoreML > TFLite (github.com)

YOLOX：TexasInstruments/edgeai-yolox: YOLOX-ti-lite models are optimized for deployment on TI edge processors (github.com)

      2. I hope to try the deployment and application of YOLOX model in TDA4. You can find the download address of the onnx file of YoloX Nano version officially provided by TI in the following link:

edgeai-yolox/yolox_nano_ti_lite_26p1_41p8.onnx.link at main · TexasInstruments/edgeai-yolox · GitHub

     3. Try to export the officially provided onnx model. If the following two options are not added to the configuration file：   

#metaArchType = 6

#metaLayersNamesList = "/projects/files/models/onnx/yolox_n.prototxt"

The export tool reports the following error:

ONNX Model (Proto) File  : /projects/files/models/onnx/yolox_nano_ti_lite_26p1_41p8.onnx

TIDL Network File      : /projects/files/models/bin/yolox_n-sim.net.bin

TIDL IO Info File      : /projects/files/models/bin/yolox_n-sim.para.bin

Current ONNX OpSet Version   : 11

 ONNX operator Gather is not suported now..  By passing

 ONNX operator Gather is not suported now..  By passing

*** WARNING : Mul with constant tensor requires input dimensions of mul layer to be present as part of the network.      

If present, this warning can be ignored. If not, please use open source runtimes offering to run this model or run shape inference on this model before executing import  ***

Unsupported slice - axis parameters, in Slice 

     4. Export the onnx model and the corresponding prototxt file in the YOLOX project by following commands:

python tools/export_onnx.py --output-name yoloxs.onnx -n yolox-s

      Enable the corresponding metaArchType and metaLayersNamesList options during conversion:

metaArchType = 6

metaLayersNamesList = "/projects/files/models/onnx/yoloxs.prototxt"

    The export tool reports the following error:

ONNX Model (Proto) File  : /projects/files/models/onnx/yoloxs.onnx

TIDL Network File      : /projects/files/models/bin/yoloxs_net.bin

TIDL IO Info File      : /projects/files/models/bin/yoloxs.para.bin

Current ONNX OpSet Version   : 11

[libprotobuf FATAL ./google/protobuf/repeated_field.h:1537] CHECK failed: (index) < (current_size_):

terminate called after throwing an instance of 'google::protobuf::FatalException'

  what():  CHECK failed: (index) < (current_size_):

fish: Job 1, 'out/tidl_model_import.out /proj…' terminated by signal SIGABRT (Abort)

     5. Now look at the YOLOv5 project and try to export the onnx file corresponding to YOLOv5s by executing the following command:

python export.py --weights yolov5s.pt --include onnx

    The following error occurs:

Traceback (most recent call last):

  File "/projects/TexasInstruments/edgeai-yolov5/export.py", line 253, in <module>

    main(opt)

  File "/projects/TexasInstruments/edgeai-yolov5/export.py", line 248, in main

    run(**vars(opt))

  File "/projects/TexasInstruments/edgeai-yolov5/export.py", line 171, in run

    model = attempt_load(weights, map_location=device)  # load FP32 model

  File "/projects/TexasInstruments/edgeai-yolov5/models/experimental.py", line 119, in attempt_load

    ckpt = torch.load(attempt_download(w), map_location=map_location)  # load

  File "/anaconda3/envs/torch-2.0.1_cu118_cp310/lib/python3.10/site-packages/torch/serialization.py", line 809, in load

    return _load(opened_zipfile, map_location, pickle_module, **pickle_load_args)

  File "/anaconda3/envs/torch-2.0.1_cu118_cp310/lib/python3.10/site-packages/torch/serialization.py", line 1172, in _load

    result = unpickler.load()

  File "/anaconda3/envs/torch-2.0.1_cu118_cp310/lib/python3.10/site-packages/torch/serialization.py", line 1165, in find_class

    return super().find_class(mod_name, name)

AttributeError: Can't get attribute 'SPPF' on <module 'models.common' from '/projects/TexasInstruments/edgeai-yolov5/models/common.py'>

 Based on the above trial and error, I have the following questions:

1. Considering that the above project is provided by TI's official Github warehouse, how to solve the above error about YOLOv5 exporting ONNX file (point 5)?

2. Considering that the above project is provided by TI's official Github warehouse, how to solve the above error of exporting the ONNX file of YOLOX into TDA4 bin file (points 3 and 4)?

Note: According to the error log, the metaArchType configuration in the configuration file may be faulty. However, from the official TI documentation:

software-dl.ti.com/.../d ocs/user_guide_html/md_tidl_fsg_meta_arch_support.html

No option is found for metaArchType to correspond to YOLOX

3. During the project, we tried a variety of different target detection models and encountered problems similar to those mentioned in the third point. However, these models (e.g. FCOS, etc.) do not belong to any of the corresponding architectures in metaArchType=1/2/3/4/5/6. Does this mean that you can only try metaArchType=1/2/3/4/5/6 one by one when converting these models?

Or, to put it another way, does the fact that these model examples (such as FCOS, etc.) encounter problems like Point 3 above in the transformation process mean that there is no solution?

    Regards，

    Kong

Hi，

    We have not received a reply on this problem for eight days. Could you please help us solve it as soon as possible?

    Thank you!

    Regards,

     Kong

I have successfully complied the model using onnxruntime-tidl. I used https://github.com/TexasInstruments/edgeai-benchmark

which is a wrapper over the underlying tidl_tools / onnxruntime-tidl. 

Sample configs for edgeai-yolox are here: https://github.com/TexasInstruments/edgeai-benchmark/blob/master/configs/detection_experimental.py#L59

See the instructions for compiling custom model to compile your own model: https://github.com/TexasInstruments/edgeai-benchmark/blob/master/docs/custom_models.md

You can also train and compile your own model using Model Composer in Edge AI Studio

https://www.ti.com/tool/EDGE-AI-STUDIO