TDA4VM: How to create prototxt file for edgeai-yolov5 pose model?

Hi Bushan,

My thought is if there is not a .protext file associated with the model it is not needed.  Model zoo models tend to be complete.  In general yolo5 models do not need a .protext file except for specific applications.  What is the intended application for this model?

Regards,

Chris

Hi Chris,

This model is yolov5-pose model which is intended to detect human pose: BBOX + 17 keypoints. We always used prototxt files for yolov5 bbox detection model. The repository I am using does not fall under model zoo.

So are you suggesting that yolov5-pose model should be imported without the prototxt in the SDK I am using? and there is no restriction from the SDK side to import the pose model? I can give it a try and respond with specifics.

Hi Brushan,

Depending on your use case a prototext file may not be needed.  I am unsure in your specific case.  Can you please include the model, any other files associated with he model you are using, import file,  inference file, input data and I will try it out and give more specific information.

Regards,

Chris 

Hi Chris,

Please find the model pytorch file, onnx file, model config yaml, coco-pose data sample, export file and tidl import config file in the following folder: https://drive.google.com/drive/folders/1DXdruZYgUE1GpWHvPKohbUij0qNiclsg?usp=drive_link

I have used the edgeai-yolov5 (branch: yolo-pose) without any modifications with the coco-pose dataset as mentioned in the repo. I am mentioning the link to repo again for your reference: https://github.com/TexasInstruments/edgeai-yolov5/tree/yolo-pose. Please let me know if you need anything else from my side.

Thanks,

Bhushan

Hi Bhushan,

Please use the following python script and it will generate the corresponding prototext file for an input ONNX model.

Run it by:

python3 ./onnx2proto.py -m model_opt.onnx

import onnx
import json
from onnx.defs import onnx_opset_version
import argparse
import os


parser = argparse.ArgumentParser()

parser.add_argument(
    "-m",
    "--model",
    required=True,
    help="ONNX Model To Execute",
)

args = parser.parse_args()

model = args.model

# Is this a valid model?
mymodel = onnx.load(model)



def onnx_to_prototxt(onnx_path, output_path):
    """
    Converts an ONNX model to a prototext file.
    """
    model = onnx.load(onnx_path)

    with open(output_path, "w") as f:
        # Print key model information
        print(f"ir_version: {json.dumps(model.ir_version)}", file=f)
        print(f"producer_name: {json.dumps(model.producer_name)}", file=f)
        print(f"producer_version: {json.dumps(model.producer_version)}", file=f)
        print(f"model_version: {json.dumps(model.model_version)}", file=f)

        # Print graph details (nodes, inputs, outputs, initializers, etc.)
        # This part requires more detailed parsing of the 'model.graph' object
        # to represent the full structure in a text format.
        # For a complete representation, you would iterate through model.graph.node,
        # model.graph.input, model.graph.output, model.graph.initializer, etc.,
        # and format their details into the output file.
        # A simple example for nodes:
        for node in model.graph.node:
            print(f"node {{", file=f)
            print(f"  name: {json.dumps(node.name)}", file=f)
            print(f"  op_type: {json.dumps(node.op_type)}", file=f)
            print(f"  input: {json.dumps(list(node.input))}", file=f)
            print(f"  output: {json.dumps(list(node.output))}", file=f)
            print(f"}}", file=f)

onnx_to_prototxt(model, model.replace(".onnx",".prototext"))

Regards,

Chris

Hi Chris, I am receiving the following error while trying to import the model using TI SDK:

TIDL Meta PipeLine (Proto) File : ../../test/testvecs/models/magna/TAD3/yolov5n_exp_pose_coco/yolov5n_exp_pose_coco.prototxt
[libprotobuf ERROR google/protobuf/text_format.cc:309] Error parsing text-format tidl_meta_arch.TIDLMetaArch: 1:11: Message type "tidl_meta_arch.TIDLMetaArch" has no field named "ir_version".
ERROR: google::protobuf::TextFormat::Parse proto file(../../test/testvecs/models/magna/TAD3/yolov5n_exp_pose_coco/yolov5n_exp_pose_coco.prototxt) FAILED !!!

I have generated the prototxt file for the same onnx model I have shared earlier. Can you also let me know if the https://drive.google.com/file/d/13O0kRU-dXarmjpntL4l5xPGivPxzslZk/view?usp=drive_link onnx model is created correctly? because there is a big difference in the size if I use nms or not while exporting to onnx. I have used the export.py in the shared folder to export onnx with nms.

Thank you for your help,

Bhushan

Hi Bhushan,

Please send me how you are compiling the model and if you are using OSRT or TIDLRT.  Also, I would recommend you move to a newer TIDL version like 11.08 as we no longer support updates to 8.5.

Regards,

Chris