TDA4VM: Models downloaded from DL models for Edge Inference can't run

Hi Shine,

Could you please confirm you using “DL models for Edge Inference” (Documentation Link) one of the model from model-zoo repo or is it some custom compiled model ?

So that we can try to replicate this issue at our end.

Thank you,

Pratik

Hello Pratik,

I am the initiator of this problem. I'm a beginer. My TDA4 version is PROCESSOR-SDK-LINUX-SK-TDA4VM/08.04.00.10 . Using "root@tda4vm-sk:/opt/edge_ai_apps/apps_python# ./app_edgeai.py ../configs/object_detection.yaml "  to run the model in the SDK is no problem. 

So, I want to run other models.First steo Using "root@tda4vm-sk:/opt/edge_ai_apps# ./download_models.sh", download " ONR-OD-8110-yolov5-s6-ti-lite-coco-384x384":

Second step, I modified 'object_detection.yaml' under '/opt/edge_ai_apps/configs'.

Third step, root@tda4vm-sk:/opt/edge_ai_apps/apps_python# ./app_edgeai.py ../configs/object_detection.yaml , the bug is:

root@tda4vm-sk:/opt/edge_ai_apps/apps_python# ./app_edgeai.py ../configs/object_detection.yaml
Traceback (most recent call last):
File "./app_edgeai.py", line 71, in <module>
main(sys.argv)
File "./app_edgeai.py", line 45, in main
demo = EdgeAIDemo(config)
File "/opt/edge_ai_apps/apps_python/edge_ai_class.py", line 75, in __init__
model_obj = config_parser.Model(model_config)
File "/opt/edge_ai_apps/apps_python/config_parser.py", line 105, in __init__
self.mean = params['session']['input_mean']
KeyError: 'input_mean'

My problems:

1.Can't the downloaded model by using 'root@tda4vm-sk:/opt/edge_ai_apps# ./download_models.sh'  share the script(./app_edgeai.py) with the SDK model ?

2.From the generation of the.pt model to the successful operation of TDA4, is there a guiding use case for reference ?

3.Can I provide a link to this version of PROCESSOR-SDK-LINUX-SK-TDA4VM / 08.04.00.10 model transformation tool chain ?

Hi,

We have verified that, the model ONR-OD-8110-yolov5-s6-ti-lite-coco-384x384 which part of our model_zoo offering under 8.4 linux psdk is not shipped correctly.

The reason behind getting keyError : 'input_mean' is that, some features from param.yaml file for above mentioned model are missing.

I have raised internal requrest to modify this, we are expecting this can be done in short period of time.

I will update you here, once we update model artifacts for the above mentioned model on internal server so you can download it using download_models.sh script.

Answering your posted questions : 

1. Yes, downloaded model is compatible with app_edgeai.py script, by making suitable changes in config file and providing correct path to it you can perform inferencing on target.

2. You can use tidl tools to perform model compilation, please refer user guide which is part of rtos sdk offering.

3. Could you please elaborate more on this ?

Regards,

Pratik

Hi Pratik，

Not only the model ONR-OD-8110-yolov5-s6-ti-lite-coco-384x384 has problems, but also all downloaded models from download_models.sh script  may not perform inferencing. It is recommended that you check all downloaded models. The third question above has been solved.Thank you for your help, look forward to the results after modify.

Regards,

maiunlei

Hi Maiunlei,

Thanks for verifying, I am in touch with the concerned team, and once I have any update on this I will keep you posted when we update above model in model zoo.

As pointed by you, all the downloaded model are not able to perfrom inferencing on target ? Could you please point me to few example which you have tried ?

Mean while I did quick test on model : ONR-OD-8430-yolox-m-lite-mmdet-widerface-640x640 and its building fine on my target.

Also could you please verify, were there build specific issues like you pointed out in model : ONR-OD-8110-yolov5-s6-ti-lite-coco-384x384 or were there inferencing specific issue if yes, could you please elaborate them ?

--

Pratik