PROCESSOR-SDK-AM62A: How can I get Yolov8 model information

Hi Ryuuichi,

It is our pleasure to help. The yolov8 model artifacts are not publicly available. You can find documentation about the model prototext here: https://github.com/TexasInstruments/edgeai-tidl-tools/blob/master/docs/tidl_fsg_od_meta_arch.md

Best regards,

Qutaiba

Hello Qutaiba-san,

Thank you for your reply.
I understood that I need to create model artifacts by myself.
Do you have checkpoint information for above yolov8 ?

Best Regards,

Hi,

Thank you for your reply.
I am checking your attached document. However I understood that you do not disclose pre-trained model.

I understand following procedure is needed to get optimized yolov8-lite model.

* Get yolo model from "https://github.com/open-mmlab/mmyolo"
* Apply patch to mmyolo
* Training model with "model-surgery" argument

Is my understanding correct ?

BR,

Hello,

I'm trying to apply patch by following below.

* git am 0001-mmyolo.commit-8c4d9dc5.-model-surgery-with-edgeai-modeloptimization.txt
(Reference link : https://github.com/TexasInstruments/edgeai-tensorlab/issues/7)

However, I could not find above patch file.
How can I get it ?

Best Regards,

Hello,

I'm trying to train the yolov8 model by following below link.
>https://github.com/TexasInstruments/edgeai-tensorlab/issues/7

I can set up envirnment to train build and create conda environment to install something packages.
After appling patch as shown below, I performed training command.

Apply patch : 
$ git am 0001-2024-Aug-2-mmyolo.commit-8c4d9dc5.-model-surgery-with-edgeai-modeloptimization.txt

Training command :
$ python3 tools/train.py configs/yolov8/yolov8_s_syncbn_fast_8xb16-500e_coco.py --model-surgery 1

As a result, I got following error.
File "xxxxxxx/anaconda3/envs/mmyolo/lib/python3.8/importlib/__init__.py", line 127, in import_module
return _bootstrap._gcd_import(name[level:], package, level)
ImportError: libtorch_cuda_cu.so: cannot open shared object file: No such file or directory

In my simple confirmation, it seems that I need to build "mmcv" itself.
However, I did not see such procedure in your github repo.

Q1. Do you have same error when you create yolov8 model ?

Best Regards,

You seem to be using python3.8. We have mostly tested with python3.10. But I am not sure if that is the reason for this error though.

Hi,

Thank you for your reply.
I could try to install python 3.10, however, error was occurred.

* conda create -n mmyolo python=3.10 pytorch==1.10.1 torchvision==0.11.2 cudatoolkit=11.3 -c pytorch -y
(Original command is below.
$ conda create -n mmyolo python=3.8 pytorch==1.10.1 torchvision==0.11.2 cudatoolkit=11.3 -c pytorch -y)

Best Regards,

Hi,

I will show current package information which I think these are related to training.

Anaconda 24.5.0
python 3.8.19
torch 1.10.1
torchvision 0.11.2
cudatoolkit 11.3.1
mmcv 2.0.1
mmdet 3.3.0
mmengine 0.10.5
mmyolo 0.6.0

Could you please comment whether these versions are same as your condition ?

Best Regards,

Hi Ryuuichi,

On option to install python 3.10 is to use a virtual environment. An example of that is explained in the edgeai-modelmaker: https://github.com/TexasInstruments/edgeai-tensorlab/tree/main/edgeai-modelmaker#step-1-option-1-with-native-ubuntu-environment-and-pyenv-recommended. Look at steps 1.3a and 1.4a. 

Best regards,

Qutaiba

Hello,

Thank you for your reply.
However, I'm a little bit confusing virtual environment.

According to original site of mmyolo for yolov8, this suggests to use "conda" as virtual environment.
mmyolo/README.md at main · open-mmlab/mmyolo · GitHub

However, according to URL which you point out, this uses "pyenv" as virtual environment.
The package which these tool prepare is different.

What you would like to say is I should create "pyenv" as base virtual environment and I should create "conda" vitual environment on this virtual environment ?

Best Regards,

>>As a result, I got following error.
>>File "xxxxxxx/anaconda3/envs/mmyolo/lib/python3.8/importlib/__init__.py", line 127, in import_module
>>return _bootstrap._gcd_import(name[level:], package, level)
>>ImportError: libtorch_cuda_cu.so: cannot open shared object file: No such file or directory

Also, the specific torch version that you are trying to install on Python 3.10 may not even be supported on that Python version (not sure)

The error that you are getting is probably because of a mismatch between Python version, cuda version and/or torch version. Can you take help from someone in your team to get the environment correct. Meanwhile, I shall request my colleague to share the list of python packages that he is using.

Hi,

Here is the list of the packages that is being used for mmyolo

python         3.10.13
torch            2.0.1+cu118
torchinfo      1.8.0
torchvision   0.15.2+cu118
mmcv          2.0.1
mmdet         3.3.0
mmengine   0.10.2
mmyolo       0.6.0

Hello,

Thank you for your reply.
I could avoid original error due to your information.
However, I got following error.

anaconda3/envs/mmyolo/lib/python3.10/site-packages/albumentations/core/composition.py", line 372, in preprocess raise ValueError(msg)
ValueError: Key img_path is not in available keys.

It seems that following thread is similar to this issue. However, according to following thread, this issue will cause  albumentations version < 1.4.11.

https://github.com/open-mmlab/mmdetection/issues/11770

According to my "conda list", albumentations version is higher than 1.4.11...

albumentations            1.4.17                   pypi_0    pypi

Have you seen same error in your condition ?

Best Regards,

Can you use this version, it should fix the issue,

pip install albumentations==1.3.1

Hi,

Thank you for your reply.

>pip install albumentations==1.3.1

By doing above command, it seems that I could avoid my issue.
I will continue to check the proceed and if I have additional question, I will let you know.

BR,

Hi,

It seems that I can train model, however I got following error log.

>mmyolo/lib/python3.10/site-packages/mmdet/evaluation/metrics/coco_metric.py - compute_metrics - 465 - The testing results of the whole dataset is empty.

Did you observe same error when you train model ?

Best Regards,

It can be possible that the checkpoint generated is not correct. Something might have gone wrong while training. Can you share the training log?

I can share total log, however, I do not want it to share public thread.
Can I send it to you directly ?

BR,

Yes, you can send it to me