TDA4VM: YOLOX_ti_lite training

Ramaseshan Subramanian

Part Number: TDA4VM

Hello forum

I need to train yolox_ti_lite models using colab or kaggle, what is the basic steps has to be followed in order to retrain the model in kaggle or colab. Is there any thread or references related to this kindly attach that too

Thanks

over 1 year ago

0 Pratik Kedar over 1 year ago

TI__Mastermind 24041 points

Hi,

Here are 2 options I can suggest.

1. You can Create and Train the model using std deep learning libs like Pytorch, Tensorflow etc and use our edgeai tidl tools to compile the onnx or tflite model

https://github.com/TexasInstruments/edgeai-tidl-tools

2. You can explore edgeai modelmaker repos to train/re-train(transfer learning) and compile the model here : https://github.com/TexasInstruments/edgeai-modelmaker

Check out how we have utilized this tools to create the demos here : https://github.com/TexasInstruments/edgeai-gst-apps-people-tracking

Thanks.

0 Ramaseshan Subramanian over 1 year ago in reply to Pratik Kedar

Intellectual 490 points

Hi Pratik

Thanks for the reply, is there any way to deploy yolov8 onnx model on tda4vm Or else it is possible to make it compatible to tda4vm

0 Pratik Kedar over 1 year ago in reply to Ramaseshan Subramanian

TI__Mastermind 24041 points

End - End (DSP inference) support is not available for Yolo V8, however you can try out OSRT for model inference on hardware.

Please read more about the same here :

https://github.com/TexasInstruments/edgeai-tidl-tools/tree/master

0 Ramaseshan Subramanian over 1 year ago in reply to Pratik Kedar

Intellectual 490 points

Hello Pratik

We are capable of writing DSP instrinsics for yolov8, Is there any user guide in order make yolov8 to deploy on TDA4VM

0 Pratik Kedar over 1 year ago in reply to Ramaseshan Subramanian

TI__Mastermind 24041 points

Hi,

You can use OSRT flow to get your model inferred on target SoC, the supported layers will get delegated to DSP for optimization

Pratik Kedar said:
however you can try out OSRT for model inference on hardware.

Please read more about the same here :

https://github.com/TexasInstruments/edgeai-tidl-tools/tree/master

With every release we add new set of operators and models you can look for further release to get this supported end to end.

0 Ramaseshan Subramanian over 1 year ago in reply to Pratik Kedar

Intellectual 490 points

Thanks for the reply, can I know more about calibration iteration and calibration frames what these parameters actually does while compiling models?

0 Pratik Kedar over 1 year ago in reply to Ramaseshan Subramanian

TI__Mastermind 24041 points

During quantization, to calculate the scales and offset is done by using provided calibrations frames as input for certain specified iterations.

Thanks.

0 Ramaseshan Subramanian over 1 year ago in reply to Pratik Kedar

Intellectual 490 points

Hello Pratik!
I have successfully compiled yolox_m_lite model, while making inference, we have faced some issues:

nfer : imagedet-6_onnxrt_edgeai-benchmark_model_yolox_m_onn|   0%|          || 0/100 [00:22<?, ?it/s]
Traceback (most recent call last):
  File "/home/mugu/edgeai-benchmark/edgeai_benchmark/pipelines/pipeline_runner.py", line 174, in _run_pipeline
    result = cls._run_pipeline_impl(basic_settings, pipeline_config, description)
  File "/home/mugu/edgeai-benchmark/edgeai_benchmark/pipelines/pipeline_runner.py", line 147, in _run_pipeline_impl
    accuracy_result = accuracy_pipeline(description)
  File "/home/mugu/edgeai-benchmark/edgeai_benchmark/pipelines/accuracy_pipeline.py", line 123, in __call__
    param_result = self._run(description=description)
  File "/home/mugu/edgeai-benchmark/edgeai_benchmark/pipelines/accuracy_pipeline.py", line 180, in _run
    output_list = self._infer_frames(description)
  File "/home/mugu/edgeai-benchmark/edgeai_benchmark/pipelines/accuracy_pipeline.py", line 273, in _infer_frames
    output, info_dict = postprocess(output, info_dict)
  File "/home/mugu/edgeai-benchmark/edgeai_benchmark/utils/transforms_utils.py", line 41, in __call__
    tensor, info_dict = t(tensor, info_dict)
  File "/home/mugu/edgeai-benchmark/edgeai_benchmark/postprocess/transforms.py", line 486, in __call__
    bbox_copy[..., self.dst_indices] = bbox[..., self.src_indices]
IndexError: index 5 is out of bounds for axis 1 with size 5
index 5 is out of bounds for axis 1 with size 5
TASKS                                                       | 100%|██████████|| 
TASKS                                                       | 100%|██████████|| 

packaging artifacts to ./work_dirs/modelpackage/TDA4VM/8bits please wait...
SUCCESS:20240225-231436: finished packaging - imagedet-6_onnxrt_edgeai-benchmark_model_yolox_m_onnx
SUCCESS:20240225-231437: finished packaging - imageseg-3_onnxrt_edgeai-benchmark_model_model_bes_onnx
-------------------------------------------------------------------
===================================================================
settings: {'include_files': None, 'pipeline_type': 'accuracy', 'num_frames': 100, 'calibration_frames': 25, 'calibration_iterations': 25, 'configs_path': './configs', 'models_path': '../edgeai-modelzoo/models', 'modelartifacts_path': './work_dirs/modelartifacts/', 'modelpackage_path': './work_dirs/modelpackage/', 'datasets_path': './dependencies/datasets', 'target_device': None, 'target_machine': 'pc', 'run_suffix': None, 'parallel_devices': None, 'parallel_processes': 1, 'tensor_bits': 8, 'runtime_options': None, 'run_import': True, 'run_inference': True, 'run_missing': True, 'detection_threshold': 0.3, 'detection_top_k': 200, 'detection_nms_threshold': None, 'detection_keep_top_k': None, 'save_output': False, 'num_output_frames': 50, 'model_selection': None, 'model_shortlist': None, 'model_exclusion': None, 'task_selection': None, 'runtime_selection': None, 'session_type_dict': {'onnx': 'onnxrt', 'tflite': 'tflitert', 'mxnet': 'tvmdlr'}, 'dataset_type_dict': {'imagenet': 'imagenetv2c'}, 'dataset_selection': None, 'dataset_loading': True, 'config_range': None, 'enable_logging': True, 'verbose': False, 'capture_log': False, 'experimental_models': False, 'rewrite_results': False, 'with_udp': True, 'flip_test': False, 'model_transformation_dict': None, 'report_perfsim': False, 'tidl_offload': True, 'input_optimization': False, 'run_dir_tree_depth': None, 'target_device_preset': True, 'fast_calibration_factor': None, 'skip_pattern': '_package', 'settings_file': 'settings_import_on_pc.yaml', 'basic_keys': ['include_files', 'pipeline_type', 'num_frames', 'calibration_frames', 'calibration_iterations', 'configs_path', 'models_path', 'modelartifacts_path', 'modelpackage_path', 'datasets_path', 'target_device', 'target_machine', 'run_suffix', 'parallel_devices', 'parallel_processes', 'tensor_bits', 'runtime_options', 'run_import', 'run_inference', 'run_missing', 'detection_threshold', 'detection_top_k', 'detection_nms_threshold', 'detection_keep_top_k', 'save_output', 'num_output_frames', 'model_selection', 'model_shortlist', 'model_exclusion', 'task_selection', 'runtime_selection', 'session_type_dict', 'dataset_type_dict', 'dataset_selection', 'dataset_loading', 'config_range', 'enable_logging', 'verbose', 'capture_log', 'experimental_models', 'rewrite_results', 'with_udp', 'flip_test', 'model_transformation_dict', 'report_perfsim', 'tidl_offload', 'input_optimization', 'run_dir_tree_depth', 'target_device_preset', 'fast_calibration_factor', 'skip_pattern', 'settings_file'], 'dataset_cache': None}
results found for 2 models
Report generated at ./work_dirs/modelartifacts/

Processors

Processors forum

TDA4VM: YOLOX_ti_lite training