[MLOps] Deepstream 3 - 실습(Multi-stream Pose Estimation)

안녕하세요 윤도현입니다. 이번 글에서는 DeepStream을 이용하여 Multi-stream Pose Estimation app을 제작해보겠습니다.

 

관련글

[1] [MLOps] Deepstream 1 - 기본개념

[2][MLOps] Deepstream 2 - 환경세팅(with docker)

[3][MLOps] Deepstream 3 - 실습

 

 

1. 환경 설정

• Deepstream SDK 설치: 이전 글을 참조해주세요.

2. github clone

cd $DEEPSTREAM_DIR/sources/apps/sample_apps
git clone https://github.com/NVIDIA-AI-IOT/deepstream_pose_estimation.git

 

3. HRNet-pose-estimation 모델 준비

3.1 HRNet Pytorch model 준비

저는 공식 repo에서 pytorch model file을 다운받았습니다.

 

wget https://github.com/HRNet/HRNet-Human-Pose-Estimation/releases/download/v1.0/hrnet_w32-36af842e.pth

 

 

3.2 다운받은 Pytorch model을 onnx로 변환

import torch
import torch.onnx
import numpy as np

# 1. 네 HRNet PyTorch 모델 불러오기
from models.pose_hrnet import get_pose_net
from config import cfg
from config import update_config

update_config(cfg, 'experiments/hrnet_w32.yaml')
model = get_pose_net(cfg, is_train=False)

# 2. 모델에 pretrained weight 로드
checkpoint = torch.load('hrnet_w32-36af842e.pth', map_location='cpu')
model.load_state_dict(checkpoint)
model.eval()

# 3. 더미 입력 생성 (B, C, H, W)
dummy_input = torch.randn(1, 3, 256, 192)  # HRNet 기본 input size

# 4. ONNX 변환
torch.onnx.export(model, dummy_input, "hrnet_pose_estimation.onnx",
                  input_names=["input"],
                  output_names=["heatmaps"],
                  dynamic_axes={
                      'input': {0: 'batch_size'},
                      'heatmaps': {0: 'batch_size'}
                  },
                  opset_version=11)

 

 

reference

[1] Building a Real-time Redaction App Using NVIDIA DeepStream, Part 1: Training

[2] Building a Real-time Redaction App Using NVIDIA DeepStream, Part 2: Deployment

[3] https://github.com/NVIDIA-AI-IOT/deepstream_pose_estimation?tab=readme-ov-file