ViTPose/easy_ViTPose/inference.py

import argparse
import json
import os
import time

from PIL import Image
import cv2
import numpy as np
import torch
import tqdm

from easy_ViTPose.vit_utils.inference import NumpyEncoder, VideoReader
from easy_ViTPose.inference import VitInference
from easy_ViTPose.vit_utils.visualization import joints_dict

try:
    import onnxruntime  # noqa: F401
    has_onnx = True
except ModuleNotFoundError:
    has_onnx = False


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument('--input', type=str, required=True,
                        help='path to image / video or webcam ID (=cv2)')
    parser.add_argument('--output-path', type=str, default='',
                        help='output path, if the path provided is a directory '
                        'output files are "input_name +_result{extension}".')
    parser.add_argument('--model', type=str, required=True,
                        help='checkpoint path of the model')
    parser.add_argument('--yolo', type=str, required=False, default=None,
                        help='checkpoint path of the yolo model')
    parser.add_argument('--dataset', type=str, required=False, default=None,
                        help='Name of the dataset. If None it"s extracted from the file name. \
                              ["coco", "coco_25", "wholebody", "mpii", "ap10k", "apt36k", "aic"]')
    parser.add_argument('--det-class', type=str, required=False, default=None,
                        help='["human", "cat", "dog", "horse", "sheep", \
                               "cow", "elephant", "bear", "zebra", "giraffe", "animals"]')
    parser.add_argument('--model-name', type=str, required=False, choices=['s', 'b', 'l', 'h'],
                        help='[s: ViT-S, b: ViT-B, l: ViT-L, h: ViT-H]')
    parser.add_argument('--yolo-size', type=int, required=False, default=320,
                        help='YOLOv8 image size during inference')
    parser.add_argument('--conf-threshold', type=float, required=False, default=0.5,
                        help='Minimum confidence for keypoints to be drawn. [0, 1] range')
    parser.add_argument('--rotate', type=int, choices=[0, 90, 180, 270],
                        required=False, default=0,
                        help='Rotate the image of [90, 180, 270] degress counterclockwise')
    parser.add_argument('--yolo-step', type=int,
                        required=False, default=1,
                        help='The tracker can be used to predict the bboxes instead of yolo for performance, '
                             'this flag specifies how often yolo is applied (e.g. 1 applies yolo every frame). '
                             'This does not have any effect when is_video is False')
    parser.add_argument('--single-pose', default=False, action='store_true',
                        help='Do not use SORT tracker because single pose is expected in the video')
    parser.add_argument('--show', default=False, action='store_true',
                        help='preview result during inference')
    parser.add_argument('--show-yolo', default=False, action='store_true',
                        help='draw yolo results')
    parser.add_argument('--show-raw-yolo', default=False, action='store_true',
                        help='draw yolo result before that SORT is applied for tracking'
                        ' (only valid during video inference)')
    parser.add_argument('--save-img', default=False, action='store_true',
                        help='save image results')
    parser.add_argument('--save-json', default=False, action='store_true',
                        help='save json results')
    args = parser.parse_args()

    use_mps = hasattr(torch.backends, 'mps') and torch.backends.mps.is_available()
    use_cuda = torch.cuda.is_available()

    # Load Yolo
    yolo = args.yolo
    if yolo is None:
        yolo = 'easy_ViTPose/' + ('yolov8s' + ('.onnx' if has_onnx and not (use_mps or use_cuda) else '.pt'))
    input_path = args.input

    # Load the image / video reader
    try:  # Check if is webcam
        int(input_path)
        is_video = True
    except ValueError:
        assert os.path.isfile(input_path), 'The input file does not exist'
        is_video = input_path[input_path.rfind('.') + 1:].lower() in ['mp4', 'mov']

    ext = '.mp4' if is_video else '.png'
    assert not (args.save_img or args.save_json) or args.output_path, \
        'Specify an output path if using save-img or save-json flags'
    output_path = args.output_path
    if output_path:
        if os.path.isdir(output_path):
            og_ext = input_path[input_path.rfind('.'):]
            save_name_img = os.path.basename(input_path).replace(og_ext, f"_result{ext}")
            save_name_json = os.path.basename(input_path).replace(og_ext, "_result.json")
            output_path_img = os.path.join(output_path, save_name_img)
            output_path_json = os.path.join(output_path, save_name_json)
        else:
            output_path_img = output_path + f'{ext}'
            output_path_json = output_path + '.json'

    wait = 0
    total_frames = 1
    if is_video:
        reader = VideoReader(input_path, args.rotate)
        cap = cv2.VideoCapture(input_path)  # type: ignore
        total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
        cap.release()
        wait = 15
        if args.save_img:
            cap = cv2.VideoCapture(input_path)  # type: ignore
            fps = cap.get(cv2.CAP_PROP_FPS)
            ret, frame = cap.read()
            cap.release()
            assert ret
            assert fps > 0
            output_size = frame.shape[:2][::-1]

            # Check if we have X264 otherwise use default MJPG
            try:
                temp_video = cv2.VideoWriter('/tmp/checkcodec.mp4',
                                             cv2.VideoWriter_fourcc(*'h264'), 30, (32, 32))
                opened = temp_video.isOpened()
            except Exception:
                opened = False
            codec = 'h264' if opened else 'MJPG'
            out_writer = cv2.VideoWriter(output_path_img,
                                         cv2.VideoWriter_fourcc(*codec),  # More efficient codec
                                         fps, output_size)  # type: ignore
    else:
        reader = [np.array(Image.open(input_path).rotate(args.rotate))]  # type: ignore

    # Initialize model
    model = VitInference(args.model, yolo, args.model_name,
                         args.det_class, args.dataset,
                         args.yolo_size, is_video=is_video,
                         single_pose=args.single_pose,
                         yolo_step=args.yolo_step)  # type: ignore
    print(f">>> Model loaded: {args.model}")

    print(f'>>> Running inference on {input_path}')
    keypoints = []
    fps = []
    tot_time = 0.
    for (ith, img) in tqdm.tqdm(enumerate(reader), total=total_frames):
        t0 = time.time()

        # Run inference
        frame_keypoints = model.inference(img)
        keypoints.append(frame_keypoints)

        delta = time.time() - t0
        tot_time += delta
        fps.append(delta)

        # Draw the poses and save the output img
        if args.show or args.save_img:
            # Draw result and transform to BGR
            img = model.draw(args.show_yolo, args.show_raw_yolo, args.conf_threshold)[..., ::-1]

            if args.save_img:
                # TODO: If exists add (1), (2), ...
                if is_video:
                    out_writer.write(img)
                else:
                    print('>>> Saving output image')
                    cv2.imwrite(output_path_img, img)

            if args.show:
                cv2.imshow('preview', img)
                cv2.waitKey(wait)

    if is_video:
        tot_poses = sum(len(k) for k in keypoints)
        print(f'>>> Mean inference FPS: {1 / np.mean(fps):.2f}')
        print(f'>>> Total poses predicted: {tot_poses} mean per frame: '
              f'{(tot_poses / (ith + 1)):.2f}')
        print(f'>>> Mean FPS per pose: {(tot_poses / tot_time):.2f}')

    if args.save_json:
        print('>>> Saving output json')
        with open(output_path_json, 'w') as f:
            out = {'keypoints': keypoints,
                   'skeleton': joints_dict()[model.dataset]['keypoints']}
            json.dump(out, f, cls=NumpyEncoder)

    if is_video and args.save_img:
        out_writer.release()
    cv2.destroyAllWindows()