src/predict.py

import torch
from torchvision.utils import draw_bounding_boxes, draw_keypoints
from PIL import Image
import numpy as np

from .keypt_det.model import get_model, get_transforms, load_model_weight
from .roi_calib.custom.trousers import fix_trouser_orientation

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

model = get_model(num_classes=2, num_keypoints=14).to(device)
model = load_model_weight(model, device=device)
model.eval()

transforms = get_transforms()

def predict(input):
    if isinstance(input, np.ndarray):
        input = Image.fromarray(input)

    input = transforms(input).unsqueeze(0).to(device)
    with torch.no_grad():
        outputs = model(input)[0]

    image = (input[0].cpu() * 255.0).to(torch.uint8)
    _, h, w = image.shape
    max_edge = max(h, w)

    filter = outputs["scores"] >= 0.8 # Objectness threshold

    boxes = outputs["boxes"][filter]

    keypoints = outputs["keypoints"][filter].cpu().detach().numpy()
    #keypoints = np.array([fix_trouser_orientation(kypts) for kypts in keypoints])
    keypoints = torch.tensor(keypoints)

    keypoints, visibility = keypoints.split([2, 1], dim=-1)
    visibility = visibility.bool()

    connect_skeleton = [
        (0, 3),
        *[(i, i+1) for i in range(3, 13)],
        (13, 2),
        (2, 1),
        (1, 0)
    ]

    output = draw_bounding_boxes(
        image=image,
        boxes=boxes,
        colors="orange",
        width=int(max_edge * 0.0033)
    )
    
    output = draw_keypoints(
        image=output,
        keypoints=keypoints,
        visibility=visibility,
        connectivity=connect_skeleton,
        colors="blue",
        radius=int(max_edge * 0.01),
        width=int(max_edge * 0.005)
    )

    output = output.transpose(0, 2).transpose(0, 1).cpu().numpy()
    output = Image.fromarray(output)
    return output