SK-AM62A-LP: edgeai model maker custom model compilation

Mickael LE STEPHAN

Part Number: SK-AM62A-LP

Tool/software:

Hi,

I have trained and compiled a model with edgeai-modelmaker with success. I use r9.2 banch

My model is based on yolox_s_lite 640x640 and trained with a custom dataset

The model is stored here : /home/ubuntu/TI/edgeai-tensorlab/edgeai-modelmaker/data/projects/parking_ploum_aug_5k_V1/run/20240911-175207/training

The dataset path : /home/ubuntu/TI/edgeai-tensorlab/edgeai-modelmaker/data/projects/parking_ploum_aug_5k/dataset

The dataset tree :

.
├── annotations
├── images
├── train -> images
└── val -> images

The parameter i want to change for compilation in settings_base.yaml i suppose (i want the highest precision as possible) :

calibration_iterations : 10

calibration_frames : 10

tensor_bits : 16

detection_threshold : 0.05
detection_top_k : 500

If i want to re-compile a model and change parameters, i suppose i need to use the notebook : tutorial_detection.ipynb

what i need to change on the file to use it with my parameters please ?

I suppose this :

What is the value split , num_frame ...

Then in Session runtime_options use

session_name = constants.SESSION_NAME_ONNXRT

Then on the next lines

model_path=f'{settings.models_path}/vision/detection/coco/mlperf/ssd_mobilenet_v1_coco_20180128.tflite'), 

with my path ?

Thank you for your help !

4 months ago

0 Qutaiba Saleh 4 months ago

TI__Intellectual 2830 points

Hi Mickael,

One option to recompile with different compilation settings is to change the setting as you specified in the settings_base.yaml and re-run the edgeai-modelmaker. The drawback (if it is) of this option is that the model will be retrained first and then the new version of the model is compiled.

Another option is to run the compilation directly from edgeai-benchmark. In this method, only the compilation part will be executed. You can run the edgeai-benchmark directly using run_custom_pc.sh Follow the details in this document: https://github.com/TexasInstruments/edgeai-tensorlab/blob/main/edgeai-benchmark/docs/custom_models.md. Using this method, you can setup the required inputs/outputs/parameters with no need to change the python script.

A third option is to use edgeai-tidl-tools. See this documentation: https://github.com/TexasInstruments/edgeai-tidl-tools/tree/master/examples/osrt_python. For the yolox_s_lite, you can edit this file https://github.com/TexasInstruments/edgeai-tidl-tools/blob/master/examples/osrt_python/model_configs.py to add your model config and then run this to start compilation (simple edits is needed here too) https://github.com/TexasInstruments/edgeai-tidl-tools/blob/master/examples/osrt_python/ort/onnxrt_ep.py.

Best regards,

Qutaiba

0 Mickael LE STEPHAN 4 months ago in reply to Qutaiba Saleh

Intellectual 380 points

Hi Qutaiba,

I think the second option is perfect for me, but can you just help me how to change this parameters for my case

def get_imagedet_dataset_loaders(settings, download=False):
    # this example uses the datasets.COCODetection data loader
    dataset_calib_cfg = dict(
        path=f'{settings.datasets_path}/coco',
        split='val2017',
        shuffle=True,
        num_frames=min(settings.calibration_frames,5000))

    # dataset parameters for actual inference
    dataset_val_cfg = dict(
        path=f'{settings.datasets_path}/coco',
        split='val2017',
        shuffle=False, # can be set to True as well, if needed
        num_frames=min(settings.num_frames,5000))

    # you are free to use any other data loaded provided in datasets folder or write your own instead of this
    calib_dataset = datasets.COCODetection(**dataset_calib_cfg, download=download)
    val_dataset = datasets.COCODetection(**dataset_val_cfg, download=download)
    return calib_dataset, val_dataset

And for the topk, tensor bit settings, is the settings_base.yaml ?

Regards

+1 Qutaiba Saleh 4 months ago in reply to Mickael LE STEPHAN

TI__Intellectual 2830 points

Hi Mickael,

To refer to your dataset you can use something like this

dataset_calib_cfg = dict(

        path='/home/ubuntu/TI/edgeai-tensorlab/edgeai-modelmaker/data/projects/parking_ploum_aug_5k/dataset',
        split='val',
        shuffle=True,
        num_frames=min(settings.calibration_frames,5000))
        
.
.
.
calib_dataset = datasets.COCODetection(**dataset_calib_cfg, download=False)

Make sure that you have a folder 'val' with images under /home/ubuntu/TI/edgeai-tensorlab/edgeai-modelmaker/data/projects/parking_ploum_aug_5k/dataset/images

Mickael LE STEPHAN said:
And for the topk, tensor bit settings, is the settings_base.yaml ?

Yes.

Best regards,

Qutaiba

Processors

Processors forum

SK-AM62A-LP: edgeai model maker custom model compilation