inference.py

from transformers import (
    VitPoseForPoseEstimation,
    AutoProcessor,
    RTDetrForObjectDetection,
)
from PIL import Image
import torch

# load models
det_proc = AutoProcessor.from_pretrained("PekingU/rtdetr_r50vd_coco_o365")
det_model = RTDetrForObjectDetection.from_pretrained("PekingU/rtdetr_r50vd_coco_o365").eval()

pose_proc = AutoProcessor.from_pretrained("usyd-community/vitpose-base-simple")
pose_model = VitPoseForPoseEstimation.from_pretrained("usyd-community/vitpose-base-simple").eval()

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
det_model.to(device)
pose_model.to(device)

# Hugging Face will call this function
def predict(inputs: dict) -> dict:
    """
    inputs: {"image": PIL.Image}
    returns: {"poses": ...}
    """
    image = inputs["image"]

    # detect people
    det_inputs = det_proc(images=image, return_tensors="pt").to(device)
    det_outputs = det_model(**det_inputs)
    results = det_proc.post_process_object_detection(
        det_outputs,
        threshold=0.5,
        target_sizes=[(image.height, image.width)]
    )
    # keep only "person" class (label 0)
    person_boxes = results[0]["boxes"][results[0]["labels"] == 0]

    # run pose estimation
    pose_inputs = pose_proc(image, boxes=[person_boxes], return_tensors="pt").to(device)
    with torch.no_grad():
        pose_outputs = pose_model(**pose_inputs)
    poses = pose_proc.post_process_pose_estimation(pose_outputs, boxes=[person_boxes])

    return {"poses": poses[0]}