TDA4VM: Unable to train model after installing edgeai-mmdetection3d

xiu yin

Part Number: TDA4VM

Hi,

I am trying to train Smoke using edgeai-mmdetection3d. I installed edgeai-mmdetection3d according to the installation steps on https://github.com/TexasInstruments/edgeai-mmdetection3d/.

But there were some errors while running ./setup.sh, so I changed to run the commands line by line. And modified the installation package that would report errors.

For example, I installed torch-1.7.1+cu110, torchvision-0.8.2+cu110, numpy-1.20.3, onnx-1.9.0.

But the installation command pip3 install -e ./references/edgeailite/ will still generate an error message, indicating that the Python version is incorrect.

Then I entered in the directory ./references/edgailite/, manually installed requirements and setup.py, and the installation was successful. The following steps are similar to the installation steps on GitHub, but I installed other packages using pip.

Then I started creating data and training the model, but there was a error that appeared to indicate that Edgeailite was not successfully installed. How should I modify it?

Additionally, I have also attempted to configure edgeai-mmdetection3d using Python 3.10, but it did not work on pybind11_DIR=$(pybind11-config --cmakedir) pip3 install --no-input hub.yzuu.cf/.../tidl-j7.zip. At this step, it failed building wheel for onnx.

I don't know how to correspond to the versions of these installation packages. Please help share your latest installation steps.

Regards,

Yolanda

over 1 year ago

0 Pratik Kedar over 1 year ago

TI__Mastermind 24041 points

Hi,

I have assigned this thread to our expert.

He will look into this and get back to you.

0 xiu yin over 1 year ago in reply to Pratik Kedar

Prodigy 50 points

Okay, thank you.

0 Manu Mathew over 1 year ago in reply to xiu yin

TI__Genius 11466 points

I have added a setup.sh file and readme documentation has been updated. Please try.

0 xiu yin over 1 year ago in reply to Manu Mathew

Prodigy 50 points

Hi,

Thank you for your reply.

I installed according to the new readme documentation and setup.sh file, but still encountered some issues.

1. Installation issues. Unable to install waymo-open-dataset-tf-2-1-0-1.2.0, numba does not have version 0.54.1, and the Python 3.10 environment cannot install numpy-1.20.3 and lower version. After deleting the above packages or installing other versions, it shows that the installation was successful. Is this handling correct?

2. Training model issues. I trained SMOKE using edgeai-mdetection3d and ran the command: python tools/train.py configs/smoke/smoke_dla34_pytorch_dlaneck_gn-all_8x4_6x_kitti-mono3d.py. There was an error with the mmcv version during runtime. The version of mmcv-full-1.4.8 in the configuration file is too low. I followed the prompts to upgrade mmcv to 2.0.0rc4, but after running it again, it still reported ImportError: cannot import name 'Config' from 'mmcv'. The reason should be that the mmcv version is too high and does not support this method of importing packages. How should I modify it here?

Regards,

Yolanda

0 Manu Mathew over 1 year ago in reply to xiu yin

TI__Genius 11466 points

Can you try the installation instucrtions given in the README file: https://github.com/TexasInstruments/edgeai-mmdetection3d/blob/main/docs/en/getting_started.md

0 xiu yin over 1 year ago in reply to Manu Mathew

Prodigy 50 points

Hi,

Thank you for your reply.

Following your method, I successfully installed edgeai-mmdetection3d and trained an .pt model of SMOKE.

I need to convert it to ONNX format now. Following the instructions on https://github.com/TexasInstruments/edgeai-mmdetection3d/blob/main/docs/det3d_usage.md, I added "save_onnx_model=True" and "quantize=True" to the config file. But an error was reported, it seems that the parameter format of the model does not match.