AM67A: Deploy of custom model

Hi,

Its seems like you are using model analyzer tool and running model compilation and inference using jyputer notebooks.

I understand that working on web based tools where EVM is reserved for finite time sometimes frustrating and notebook hang or session disconnects issue may arrive due to limited bandwidth or some known reasons.

I would suggest to please check out our edgeai tidl tools repos usage here, where you can setup the same on your local machine and do model compilation and inference in host mode as well you can get the artifacts directly running on target evm as well.

Please help your self using this link : https://github.com/TexasInstruments/edgeai-tidl-tools/tree/09_02_09_00

OKay, I try also that but the instructions are not that clear how can I put my costume model there to compile? is a yolov8 and also I have a yolov5 

I also try to train an own model with my own dataset using this https://github.com/TexasInstruments/edgeai-tensorlab/tree/main/edgeai-modelmaker with version r9.2 but it doesn't have support for AM67x so I try with TDA4VM, and also didn't work, does It mean edgeai-modelmaker is not yet compatible with this model? I have ARM-basedapplication processors TDA4VEN,TDA4AEN and AM67x EVM

Elizabeth Morales said:

OKay, I try also that but the instructions are not that clear how can I put my costume model there to compile? is a yolov8 and also I have a yolov5 

please do dilegence of the above shared documents, you can check out section : https://github.com/TexasInstruments/edgeai-tidl-tools/blob/09_02_09_00/examples/osrt_python/README.md

Custom model eval : https://github.com/TexasInstruments/edgeai-tidl-tools/blob/09_02_09_00/docs/custom_model_evaluation.md#custom-model-evaluation

Using this repo you can compile the model for AM67A device 

Thanks 

thanks, Im following based on the examples of the model_configs.py but how do I get all this params information, I export my yolov5_model.pt to onnx but it doesn't generated me a protobox so I don't have the value of meta_layers_names_list

I test it with tflr_delegaty.py realized I was testing it wrong but then when I test it with the correct script onnxrt.py I got the following error 

AttributeError: 'InferenceSession' object has no attribute 'get_TI_benchmark_data'

Hi,

You can use edgeai-mmdetection repos that generates prototxt files at the time of model training, please refer to the standard examples 

https://github.com/TexasInstruments/edgeai-mmdetection

Once you have the model and proto txt ready use edgeai tidl tools for model compilation.

Is there any difference between that repo and this one? https://github.com/TexasInstruments/edgeai-tensorlab/tree/main/edgeai-mmdetection

And do I have to re-trained my model now with one of that examples?
I try to run the edgeai-mmdetection but gives me the following error 

train_model/lib/python3.10/site-packages/torch/distributed/elastic/agent/server/api.py", line 88, in __post_init__
assert self.local_world_size > 0
AssertionError
Done.

I modify the files of settings_import_on_pc.yaml and train_detection_main.py 
Im using python3.10 and I try with CPU and also with CUDA, but I have cuda 12.2

This script highlights standard supported models as part of model zoo. If your model doesn't fall under supported list, you can refer to the meta arch documentation and create your own prototxt files.

Here is the link : https://github.com/TexasInstruments/edgeai-tidl-tools/blob/09_02_09_00/docs/tidl_fsg_od_meta_arch.md

I also try that already and I search for this 

but all the links are broken
I try to train a new model with this https://github.com/TexasInstruments/edgeai-yolov5/blob/main/README_ultralytics.md but the libraries list there are not compatible with python3.10

Thanks for response.

Due to limited bandwidth its at the moment am not able to help on this issue, however i plan to circle back next week.

Thank for understanding.

Okay I understand, I hope you can help me again.
I train a model using this one https://github.com/TexasInstruments/edgeai-yolov5/blob/main/README_ultralytics.md and from this model then I export it to onnx and got the prototxt file.
After this I compiling with edgeai-tidl-tools/examples/osrt_python/ort onnxrt_ep.py 
version https://github.com/TexasInstruments/edgeai-tidl-tools/tree/09_02_09_00  
and then I move it to the board but in the board I get this errors now and my screen is then just green

As mentioned here will circle back again.

Pratik Kedar said:

Due to limited bandwidth its at the moment am not able to help on this issue, however i plan to circle back next week.

However in the meanwhile i would suggest to please get familiar with edgeai gst apps flow (by running std examples) post this you can try out custom generated artifacts.