SK-TDA4VM: edgeai-modelzoo model don't run on SK

Michele Delledonne

Part Number: SK-TDA4VM
Other Parts Discussed in Thread: TDA4VM

Tool/software:

Hi, I've encountered a problem I can't explain. I installed the edgeai-tidl-tools package on my PC using Docker with brench rel_11_00, SOC=am68pa, and TIDL_TOOLS_TYPE=CPU (the GPU version is not compatible with my GPU device). I also installed the SD card image (version 11.00.00.08) to run on SK-TDA4VM.

The problem I'm having is that on SK, I can only run the native models of the installed image, which are located in its modelzoo folder. The models downloaded through the links provided by the edgeai-modelzoo package (branch r11.0) only run on my PC and not on SK. Models like:
od-8820_onnxrt_coco_edgeai-mmyolo-gplv3_yolov5_small-v61_lite_640x640_20240329_model_onnx
od-8850_onnxrt_coco_edgeai-mmyolo-gplv3_yolov7_tiny_lite_640x640_20230830_model_onnx
od-8870_onnxrt_coco_edgeai-mmyolo-gplv3_yolov8_nano_lite_640x640_20231118_model_onnx

For example the output that I obtain with od-8820_onnxrt_coco_edgeai-mmyolo-gplv3_yolov5_small-v61_lite_640x640_20240329_model_onnx is:

Final number of subgraphs created are : 1, - Offloaded Nodes - 283, Total Nodes - 283
APP: Init ... !!!
162152.555784 s: MEM: Init ... !!!
162152.555833 s: MEM: Initialized DMA HEAP (fd=11) !!!
162152.555974 s: MEM: Init ... Done !!!
162152.555996 s: IPC: Init ... !!!
162152.617759 s: IPC: Init ... Done !!!
REMOTE_SERVICE: Init ... !!!
REMOTE_SERVICE: Init ... Done !!!
162152.628956 s: GTC Frequency = 200 MHz
APP: Init ... Done !!!
162152.632914 s: VX_ZONE_INFO: Globally Enabled VX_ZONE_ERROR
162152.633049 s: VX_ZONE_INFO: Globally Enabled VX_ZONE_WARNING
162152.633075 s: VX_ZONE_INFO: Globally Enabled VX_ZONE_INFO
162152.642192 s: VX_ZONE_INFO: [tivxPlatformCreateTargetId:169] Added target MPU-0
162152.642426 s: VX_ZONE_INFO: [tivxPlatformCreateTargetId:169] Added target MPU-1
162152.642555 s: VX_ZONE_INFO: [tivxPlatformCreateTargetId:169] Added target MPU-2
162152.642680 s: VX_ZONE_INFO: [tivxPlatformCreateTargetId:169] Added target MPU-3
162152.642765 s: VX_ZONE_INFO: [tivxInitLocal:202] Initialization Done !!!
162152.642817 s: VX_ZONE_INFO: Globally Disabled VX_ZONE_INFO
162152.665790 s: VX_ZONE_ERROR: [ownContextSendCmd:1001] Command ack message returned failure cmd_status: -1
162152.665933 s: VX_ZONE_ERROR: [ownNodeKernelInit:704] Target kernel, TIVX_CMD_NODE_CREATE failed for node node_83
162152.665952 s: VX_ZONE_ERROR: [ownNodeKernelInit:705] Please be sure the target callbacks have been registered for this core
162152.665958 s: VX_ZONE_ERROR: [ownNodeKernelInit:706] If the target callbacks have been registered, please ensure no errors are occurring within the create callback of this kernel
162152.665966 s: VX_ZONE_ERROR: [ownGraphNodeKernelInit:793] kernel init for node 0, kernel com.ti.tidl:1:1 ... failed !!!
162152.665992 s: VX_ZONE_ERROR: [ TIDL subgraph detections ] Node kernel init failed
162152.665998 s: VX_ZONE_ERROR: [ TIDL subgraph detections ] Graph verify failed

22 days ago

0 Wen Li 21 days ago

TI__Expert 7181 points

Hi;

Thanks for the questions.

Have you compiled the model that you have downloaded on PC side first?

After compiling your model, you should copy the model's artifact (the bin files) to your EVM.

Then you can run the model inference on EVM.

The link below have the info about where the artifacts are, and how to copy them to an EVM.

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

For how to compile your downloaded model; you can edit the "model_configs.py" to add your model in it; and use the python script "onnxrt_ep.py" to do the model compiling.

They are in the following directory

"<your_tidl_docker>/edgeai-tidl-tools/examples/osrt_python/"

An example for adding a model:

#=================================================================

# add the following new model for testing

"yolox_s_pose_ti_lite": create_model_config(

source=dict(

#model_url="">software-dl.ti.com/.../yolox-s-ti-lite_39p1_57p9.onnx",

#model_url="">software-dl.ti.com/.../kd-7060_onnxrt_coco_edgeai-yolox_yolox_s_pose_ti_lite_640_20220301_model_onnx.tar.gz",

model_url="">software-dl.ti.com/.../yolox_s_pose_ti_lite_640_20220301_model.onnx",

#meta_arch_url="">software-dl.ti.com/.../yolox_s_ti_lite_metaarch.prototxt",

meta_arch_url="">software-dl.ti.com/.../yolox_s_pose_ti_lite_640_20220301_model.prototxt",

infer_shape=True,

preprocess=dict(

resize=640,

crop=640,

data_layout="NCHW",

reverse_channels=True,

resize_with_pad=[True, "corner"],

pad_color=[114, 114, 114],

session=dict(

session_name="onnxrt",

model_path=os.path.join(

models_base_path, "yolox_s_pose_ti_lite_640_20220301_model.onnx"

meta_layers_names_list=os.path.join(

models_base_path, "yolox_s_pose_ti_lite_640_20220301_model.prototxt"

meta_arch_type=6,

input_mean=[0, 0, 0],

input_scale=[0.003921568627, 0.003921568627, 0.003921568627],

input_optimization=True,

postprocess=dict(

formatter="DetectionBoxSL2BoxLS",

resize_with_pad=True,

keypoint=False,

object6dpose=False,

normalized_detections=False,

shuffle_indices=None,

squeeze_axis=None,

reshape_list=[(-1, 5), (-1, 1)],

ignore_index=None,

task_type="detection",

extra_info=dict(

od_type="SSD",

framework="MMDetection",

num_images=numImages,

num_classes=91,

label_offset_type="80to90",

label_offset=1,

#== end of adding-model

#=============================================================

Best regards

Wen Li

0 Michele Delledonne 21 days ago in reply to Wen Li

Prodigy 30 points

I'm downloading model already compiled. In edgeai-tensolab package at path edgeai-tensorlab/edgeai-modelzoo/modelartifacts/TDA4VM/8bits/ there are several links to download already compiled models.