def __init__(self, cfg, track_queue, action_queue, predictor_process):
self.detector = get_detector(cfg)
self.input_path = cfg.input_path
self.start_mill = cfg.start
self.duration_mill = cfg.duration
self.realtime = cfg.realtime
stream = cv2.VideoCapture(self.input_path)
assert stream.isOpened(), 'Cannot capture source'
self.fourcc = int(stream.get(cv2.CAP_PROP_FOURCC))
self.fps = stream.get(cv2.CAP_PROP_FPS)
self.frameSize = (int(stream.get(cv2.CAP_PROP_FRAME_WIDTH)),
int(stream.get(cv2.CAP_PROP_FRAME_HEIGHT)))
self.videoinfo = {
'fourcc': self.fourcc,
'fps': self.fps,
'frameSize': self.frameSize
}
stream.release()
self._stopped = mp.Value('b', False)
self.track_queue = track_queue
self.action_queue = action_queue
self.predictor_process = predictor_process
def __init__(self, cfg):
self.realtime = cfg.realtime
# Action Predictor
cfg_file_path = cfg.cfg_path
model_weight_url = cfg.weight_path
self.ava_predictor = AVAPredictor(
cfg_file_path,
model_weight_url,
cfg.detect_rate,
cfg.common_cate,
cfg.device,
)
# Object Detector
if self.ava_predictor.has_object:
object_cfg = copy.deepcopy(cfg)
object_cfg.detector = "yolo"
self.coco_det = get_detector(object_cfg)
else:
self.coco_det = None
self.track_queue = mp.Queue(maxsize=1)
self.input_queue = mp.Queue(maxsize=512)
self.output_queue = mp.Queue()
# Video Detection Loader
self.predictor_process = mp.Value("i", 0)
det_loader = VideoDetectionLoader(cfg, self.track_queue,
self.input_queue,
self.predictor_process)
det_loader.start()
self.timestamps = []
self.frame_stack = []
self.extra_stack = []
ava_cfg = self.ava_predictor.cfg
self.frame_buffer_numbers = ava_cfg.INPUT.FRAME_NUM * ava_cfg.INPUT.FRAME_SAMPLE_RATE
self.center_index = self.frame_buffer_numbers // 2
# detection interval should be 1 second like AVA,
# one reason is that our model with memory feature is trained with that.
# Since we may not be able to reach 25 fps, so the strategy here is based on the
# number of frames. The duration of these frames may be varied.
self.last_milli = -2000
self.detect_rate = cfg.detect_rate
self.interval = 1000 // self.detect_rate
self.vid_transforms = self.ava_predictor.transforms
self._stopped = mp.Value('b', False)
self._task_done = mp.Value('b', False)
self.prediction_worker = mp.Process(target=self._compute_prediction,
args=())
self.prediction_worker.start()
def pose_detection(image):
mode = ''
input_source = image
if not os.path.exists(args.outputpath):
os.makedirs(args.outputpath)
det_loader = DetectionLoader(input_source, get_detector(args), cfg, args, batchSize=args.detbatch, mode=mode,
queueSize=args.qsize)
det_worker = det_loader.start()
def __init__(self):
self.device = try_gpu()
self.cfg = update_config(
'configs/coco/resnet/256x192_res50_lr1e-3_1x.yaml')
self.detector = get_detector({'detector': yolo})
self.detector.load_model()
self.pose_net = builder.build_sppe(self.cfg.MODEL,
preset_cfg=self.cfg.DATA_PRESET)
self.pose_net.load_state_dict(
torch.load('pretrained_models/fast_res50_256x192.pth',
map_location=self.device))
pose_model.to(self.device)
def __init__(self, args, cfg):
self.args = args
self.cfg = cfg
# Load pose model
self.pose_model = builder.build_sppe(cfg.MODEL, preset_cfg=cfg.DATA_PRESET)
print(f'Loading pose model from {args.checkpoint}...')
self.pose_model.load_state_dict(torch.load(args.checkpoint, map_location=args.device))
self.pose_dataset = builder.retrieve_dataset(cfg.DATASET.TRAIN)
self.pose_model.to(args.device)
self.pose_model.eval()
self.det_loader = DetectionLoader(get_detector(self.args), self.cfg, self.args)
def write_coco_json(self, det_file):
from pycocotools.coco import COCO
import pathlib
_coco = COCO(self._ann_file)
image_ids = sorted(_coco.getImgIds())
det_model = get_detector(self._opt, cfg=self._detector_cfg)
dets = []
for entry in tqdm(_coco.loadImgs(image_ids)):
abs_path = os.path.join(
self._root, self._img_prefix, entry['file_name'])
det = det_model.detect_one_img(entry['id'], abs_path)
if det:
dets += det
pathlib.Path(os.path.split(det_file)[0]).mkdir(parents=True, exist_ok=True)
json.dump(dets, open(det_file, 'w'))
def loop():
n = 0
while True:
yield n
n += 1
if __name__ == "__main__":
mode, input_source = check_input()
if not os.path.exists(args.outputpath):
os.makedirs(args.outputpath)
# Load detection loader
if mode == 'webcam':
det_loader = WebCamDetectionLoader(input_source, get_detector(args),
cfg, args)
det_worker = det_loader.start()
elif mode == 'detfile':
det_loader = FileDetectionLoader(input_source, cfg, args)
det_worker = det_loader.start()
else:
det_loader = DetectionLoader(input_source,
get_detector(args),
cfg,
args,
batchSize=args.detbatch,
mode=mode,
queueSize=args.qsize)
det_worker = det_loader.start()
def run(self):
if os.path.isfile(self.video):
mode, input_source = 'video', self.video
else:
raise IOError(
'Error: --video must refer to a video file, not directory.')
if not os.path.exists(self.outputpath):
os.makedirs(self.outputpath)
det_loader = DetectionLoader(input_source,
get_detector(self),
self.cfg,
self,
batchSize=self.detbatch,
mode=mode,
queueSize=self.qsize)
det_worker = det_loader.start()
# Load pose model
pose_model = builder.build_sppe(self.cfg.MODEL,
preset_cfg=self.cfg.DATA_PRESET)
print(f'Loading pose model from {self.checkpoint}...')
pose_model.load_state_dict(
torch.load(self.checkpoint, map_location=self.device))
if self.pose_track:
tracker = Tracker(tcfg, self)
pose_model.to(self.device)
pose_model.eval()
if self.save_video:
from alphapose.utils.writer import DEFAULT_VIDEO_SAVE_OPT as video_save_opt
video_save_opt['savepath'] = self.outputpath + os.path.basename(
self.video)
video_save_opt.update(det_loader.videoinfo)
writer = DataWriter(self.cfg,
self,
save_video=True,
video_save_opt=video_save_opt,
queueSize=self.qsize).start()
else:
writer = DataWriter(self.cfg,
self,
save_video=False,
queueSize=self.qsize).start()
data_len = det_loader.length
im_names_desc = tqdm(range(data_len), dynamic_ncols=True)
batchSize = self.posebatch
try:
for i in im_names_desc:
start_time = getTime()
with torch.no_grad():
(inps, orig_img, im_name, boxes, scores, ids,
cropped_boxes) = det_loader.read()
if orig_img is None:
break
if boxes is None or boxes.nelement() == 0:
writer.save(None, None, None, None, None, orig_img,
os.path.basename(im_name))
continue
# Pose Estimation
inps = inps.to(self.device)
datalen = inps.size(0)
leftover = 0
if (datalen) % batchSize:
leftover = 1
num_batches = datalen // batchSize + leftover
hm = []
for j in range(num_batches):
inps_j = inps[j * batchSize:min((j + 1) *
batchSize, datalen)]
hm_j = pose_model(inps_j)
hm.append(hm_j)
hm = torch.cat(hm)
#hm = hm.cpu()
if self.pose_track:
boxes, scores, ids, hm, cropped_boxes = track(
tracker, self, orig_img, inps, boxes, hm,
cropped_boxes, im_name, scores)
writer.save(boxes, scores, ids, hm, cropped_boxes,
orig_img, os.path.basename(im_name))
while (writer.running()):
time.sleep(1)
print('===========================> Rendering remaining ' +
str(writer.count()) + ' images in the queue...')
writer.stop()
det_loader.stop()
except KeyboardInterrupt:
det_loader.terminate()
while (writer.running()):
time.sleep(1)
print('===========================> Rendering remaining ' +
str(writer.count()) + ' images in the queue...')
writer.stop()
self.all_results = writer.results()
self._save()
action_model = joblib.load(svc_file)
else:
action_model = None
for input_source in input_sources:
if mode == 'video' and len(input_sources) > 1:
lbl_name = input_source.split(os.sep)[-2]
else:
lbl_name = 'not_categorized'
### Moved from before load pose model
# Load detection loader
if mode == 'webcam':
det_loader = WebCamDetectionLoader(input_source,
get_detector(args), cfg,
args).start()
else:
det_loader = DetectionLoader(input_source,
get_detector(args),
cfg,
args,
batchSize=args.detbatch,
mode=mode).start()
runtime_profile = {'dt': [], 'pt': [], 'pn': []}
# Init data writer
queueSize = 2 if mode == 'webcam' else args.qsize
if args.save_video and mode != 'image':
from alphapose.utils.writer import DEFAULT_VIDEO_SAVE_OPT as video_save_opt
n = 0
while True:
yield n
n += 1
if __name__ == "__main__":
# mode = 'video', input_source = './videos/blCode_action1_scene1.avi'
mode, input_source = check_input()
if not os.path.exists(args.outputpath):
os.makedirs(args.outputpath)
# Load detection loader
if mode == 'webcam':
det_loader = WebCamDetectionLoader(input_source, get_detector(args), cfg, args)
det_worker = det_loader.start()
elif mode == 'detfile':
det_loader = FileDetectionLoader(input_source, cfg, args)
det_worker = det_loader.start()
else:
# 加载yolov4检测器(将视频流放到yolo中，用于检测人物的位置)
det_loader = DetectionLoader(input_source, get_detector(args), cfg, args, batchSize=args.detbatch, mode=mode,
queueSize=args.qsize)
det_worker = det_loader.start()
# Load pose model
# 加载姿态检测模型
pose_model = builder.build_sppe(cfg.MODEL, preset_cfg=cfg.DATA_PRESET)
# 打印模型
# print(pose_model)
'===========================> If this step takes too long, you can enable the --vis_fast flag to use fast rendering (real-time).'
)
def loop():
n = 0
while True:
yield n
n += 1
if __name__ == "__main__":
mode, input_root = check_input()
output_root = args.outputpath
detector = get_detector(args)
# Load pose model
pose_model = builder.build_sppe(cfg.MODEL, preset_cfg=cfg.DATA_PRESET)
print(f'Loading pose model from {args.checkpoint}...')
pose_model.load_state_dict(
torch.load(args.checkpoint, map_location=args.device))
if len(args.gpus) > 1:
pose_model = torch.nn.DataParallel(
pose_model, device_ids=args.gpus).to(args.device)
else:
pose_model.to(args.device)
pose_model.eval()
yield n
n += 1
if __name__ == "__main__":
mode, input_source = check_input()
if not os.path.exists(args.outputpath):
os.makedirs(args.outputpath)
# Load detection loader
if mode == 'detfile':
det_loader = FileDetectionLoader(input_source, cfg, args)
det_worker = det_loader.start()
else:
det_loader = DetectionLoader(input_source, get_detector(args), cfg, args, batchSize=args.detbatch, mode=mode, queueSize=args.qsize)
det_worker = det_loader.start()
# Load pose model
pose_model = builder.build_sppe(cfg.MODEL, preset_cfg=cfg.DATA_PRESET)
print('Loading pose model from %s...' % (args.checkpoint,))
pose_model.load_state_dict(torch.load(args.checkpoint, map_location=args.device))
pose_dataset = builder.retrieve_dataset(cfg.DATASET.TRAIN)
if args.pose_track:
tracker = Tracker(tcfg, args)
if len(args.gpus) > 1:
pose_model = torch.nn.DataParallel(pose_model, device_ids=args.gpus).to(args.device)
else:
pose_model.to(args.device)
pose_model.eval()
n = 0
while True:
yield n
n += 1
if __name__ == "__main__":
mode, input_source = check_input()
if not os.path.exists(args.outputpath):
os.makedirs(args.outputpath)
# Load detection loader
if mode == 'webcam':
det_loader = WebCamDetectionLoader(input_source,
get_detector(args, cfg['DETECTOR']),
cfg, args)
det_worker = det_loader.start()
elif mode == 'detfile':
det_loader = FileDetectionLoader(input_source, cfg, args)
det_worker = det_loader.start()
else:
det_loader = DetectionLoader(input_source,
get_detector(args, cfg['DETECTOR']),
cfg,
args,
batchSize=args.detbatch,
mode=mode,
queueSize=args.qsize)
det_worker = det_loader.start()
def loop():
n = 0
while True:
yield n
n += 1
if __name__ == "__main__":
mode, input_source = check_input()
if not os.path.exists(args.outputpath):
os.makedirs(args.outputpath)
# Load detection loader
if mode == 'webcam':
det_loader = RealsenseDetectionLoader(input_source, get_detector(args),
cfg, args)
det_worker = det_loader.start()
elif mode == 'detfile':
det_loader = FileDetectionLoader(input_source, cfg, args)
det_worker = det_loader.start()
else:
det_loader = DetectionLoader(input_source,
get_detector(args),
cfg,
args,
batchSize=args.detbatch,
mode=mode,
queueSize=args.qsize)
det_worker = det_loader.start()
def start(self):
parser = argparse.ArgumentParser(description='AlphaPose Demo')
parser.add_argument(
'--cfg',
type=str,
required=False,
help='experiment configure file name',
default=
"./AlphaPose/configs/coco/resnet/256x192_res50_lr1e-3_1x.yaml")
parser.add_argument(
'--checkpoint',
type=str,
required=False,
help='checkpoint file name',
default="./AlphaPose/pretrained_models/fast_res50_256x192.pth")
parser.add_argument('--sp',
default=False,
action='store_true',
help='Use single process for pytorch')
parser.add_argument('--detector',
dest='detector',
help='detector name',
default="yolo")
parser.add_argument('--detfile',
dest='detfile',
help='detection result file',
default="")
parser.add_argument('--indir',
dest='inputpath',
help='image-directory',
default="./media/img")
parser.add_argument('--list',
dest='inputlist',
help='image-list',
default="")
parser.add_argument('--image',
dest='inputimg',
help='image-name',
default="")
parser.add_argument('--outdir',
dest='outputpath',
help='output-directory',
default="./output")
parser.add_argument('--save_img',
default=True,
action='store_true',
help='save result as image')
parser.add_argument('--vis',
default=False,
action='store_true',
help='visualize image')
parser.add_argument('--showbox',
default=False,
action='store_true',
help='visualize human bbox')
parser.add_argument('--profile',
default=False,
action='store_true',
help='add speed profiling at screen output')
parser.add_argument(
'--format',
type=str,
help=
'save in the format of cmu or coco or openpose, option: coco/cmu/open',
default="open")
parser.add_argument('--min_box_area',
type=int,
default=0,
help='min box area to filter out')
parser.add_argument('--detbatch',
type=int,
default=1,
help='detection batch size PER GPU')
parser.add_argument('--posebatch',
type=int,
default=30,
help='pose estimation maximum batch size PER GPU')
parser.add_argument(
'--eval',
dest='eval',
default=False,
action='store_true',
help=
'save the result json as coco format, using image index(int) instead of image name(str)'
)
parser.add_argument(
'--gpus',
type=str,
dest='gpus',
default="0",
help=
'choose which cuda device to use by index and input comma to use multi gpus, e.g. 0,1,2,3. (input -1 for cpu only)'
)
parser.add_argument(
'--qsize',
type=int,
dest='qsize',
default=1024,
help=
'the length of result buffer, where reducing it will lower requirement of cpu memory'
)
parser.add_argument('--flip',
default=False,
action='store_true',
help='enable flip testing')
parser.add_argument('--debug',
default=False,
action='store_true',
help='print detail information')
"""----------------------------- Video options -----------------------------"""
parser.add_argument('--video',
dest='video',
help='video-name',
default="")
parser.add_argument('--webcam',
dest='webcam',
type=int,
help='webcam number',
default=-1)
parser.add_argument('--save_video',
dest='save_video',
help='whether to save rendered video',
default=False,
action='store_true')
parser.add_argument('--vis_fast',
dest='vis_fast',
help='use fast rendering',
action='store_true',
default=False)
"""----------------------------- Tracking options -----------------------------"""
parser.add_argument('--pose_flow',
dest='pose_flow',
help='track humans in video with PoseFlow',
action='store_true',
default=False)
parser.add_argument('--pose_track',
dest='pose_track',
help='track humans in video with reid',
action='store_true',
default=True)
args = parser.parse_args()
cfg = update_config(args.cfg)
if platform.system() == 'Windows':
args.sp = True
args.gpus = [int(i) for i in args.gpus.split(',')
] if torch.cuda.device_count() >= 1 else [-1]
args.device = torch.device(
"cuda:" + str(args.gpus[0]) if args.gpus[0] >= 0 else "cpu")
args.detbatch = args.detbatch * len(args.gpus)
args.posebatch = args.posebatch * len(args.gpus)
args.tracking = args.pose_track or args.pose_flow or args.detector == 'tracker'
if not args.sp:
torch.multiprocessing.set_start_method('forkserver', force=True)
torch.multiprocessing.set_sharing_strategy('file_system')
def check_input():
# for wecam
if args.webcam != -1:
args.detbatch = 1
return 'webcam', int(args.webcam)
# for video
if len(args.video):
if os.path.isfile(args.video):
videofile = args.video
return 'video', videofile
else:
raise IOError(
'Error: --video must refer to a video file, not directory.'
)
# for detection results
if len(args.detfile):
if os.path.isfile(args.detfile):
detfile = args.detfile
return 'detfile', detfile
else:
raise IOError(
'Error: --detfile must refer to a detection json file, not directory.'
)
# for images
if len(args.inputpath) or len(args.inputlist) or len(
args.inputimg):
inputpath = args.inputpath
inputlist = args.inputlist
inputimg = args.inputimg
if len(inputlist):
im_names = open(inputlist, 'r').readlines()
elif len(inputpath) and inputpath != '/':
for root, dirs, files in os.walk(inputpath):
im_names = files
im_names = natsort.natsorted(im_names)
elif len(inputimg):
args.inputpath = os.path.split(inputimg)[0]
im_names = [os.path.split(inputimg)[1]]
return 'image', im_names
else:
raise NotImplementedError
def print_finish_info():
print('===========================> Finish Model Running.')
if (args.save_img or args.save_video) and not args.vis_fast:
print(
'===========================> Rendering remaining images in the queue...'
)
print(
'===========================> If this step takes too long, you can enable the --vis_fast flag to use fast rendering (real-time).'
)
def loop():
n = 0
while True:
yield n
n += 1
# dirList = os.listdir(args.inputpath)
# inDir = args.inputpath
# outDir = args.outputpath
# for i in dirList :
mode, input_source = check_input()
if not os.path.exists(args.outputpath):
os.makedirs(args.outputpath)
# Load detection loader
if mode == 'webcam':
det_loader = WebCamDetectionLoader(input_source,
get_detector(args), cfg, args)
det_worker = det_loader.start()
elif mode == 'detfile':
det_loader = FileDetectionLoader(input_source, cfg, args)
det_worker = det_loader.start()
else:
det_loader = DetectionLoader(input_source,
get_detector(args),
cfg,
args,
batchSize=args.detbatch,
mode=mode,
queueSize=args.qsize)
det_worker = det_loader.start()
# Load pose model
pose_model = builder.build_sppe(cfg.MODEL, preset_cfg=cfg.DATA_PRESET)
print(f'Loading pose model from {args.checkpoint}...')
pose_model.load_state_dict(
torch.load(args.checkpoint, map_location=args.device))
pose_dataset = builder.retrieve_dataset(cfg.DATASET.TRAIN)
if args.pose_track:
tracker = Tracker(tcfg, args)
if len(args.gpus) > 1:
pose_model = torch.nn.DataParallel(pose_model,
device_ids=args.gpus).to(
args.device)
else:
pose_model.to(args.device)
pose_model.eval()
runtime_profile = {'dt': [], 'pt': [], 'pn': []}
# Init data writer
queueSize = 2 if mode == 'webcam' else args.qsize
if args.save_video and mode != 'image':
from alphapose.utils.writer import DEFAULT_VIDEO_SAVE_OPT as video_save_opt
if mode == 'video':
video_save_opt['savepath'] = os.path.join(
args.outputpath,
'AlphaPose_' + os.path.basename(input_source))
else:
video_save_opt['savepath'] = os.path.join(
args.outputpath,
'AlphaPose_webcam' + str(input_source) + '.mp4')
video_save_opt.update(det_loader.videoinfo)
writer = DataWriter(cfg,
args,
save_video=True,
video_save_opt=video_save_opt,
queueSize=queueSize).start()
else:
writer = DataWriter(cfg,
args,
save_video=False,
queueSize=queueSize).start()
if mode == 'webcam':
print('Starting webcam demo, press Ctrl + C to terminate...')
sys.stdout.flush()
im_names_desc = tqdm(loop())
else:
data_len = det_loader.length
im_names_desc = tqdm(range(data_len), dynamic_ncols=True)
batchSize = args.posebatch
if args.flip:
batchSize = int(batchSize / 2)
try:
self.percentage[2] = '관절정보 분석중'
for i in range(len(im_names_desc)):
start_time = getTime()
# print(start_time)
self.percentage[0] += 1
#
with torch.no_grad():
(inps, orig_img, im_name, boxes, scores, ids,
cropped_boxes) = det_loader.read()
if orig_img is None:
break
if boxes is None or boxes.nelement() == 0:
writer.save(None, None, None, None, None, orig_img,
im_name)
continue
if args.profile:
ckpt_time, det_time = getTime(start_time)
runtime_profile['dt'].append(det_time)
# Pose Estimation
inps = inps.to(args.device)
datalen = inps.size(0)
leftover = 0
if (datalen) % batchSize:
leftover = 1
num_batches = datalen // batchSize + leftover
hm = []
for j in range(num_batches):
inps_j = inps[j * batchSize:min((j + 1) *
batchSize, datalen)]
if args.flip:
inps_j = torch.cat((inps_j, flip(inps_j)))
hm_j = pose_model(inps_j)
if args.flip:
hm_j_flip = flip_heatmap(hm_j[int(len(hm_j) / 2):],
pose_dataset.joint_pairs,
shift=True)
hm_j = (hm_j[0:int(len(hm_j) / 2)] + hm_j_flip) / 2
hm.append(hm_j)
hm = torch.cat(hm)
if args.profile:
ckpt_time, pose_time = getTime(ckpt_time)
runtime_profile['pt'].append(pose_time)
if args.pose_track:
boxes, scores, ids, hm, cropped_boxes = track(
tracker, args, orig_img, inps, boxes, hm,
cropped_boxes, im_name, scores)
hm = hm.cpu()
writer.save(boxes, scores, ids, hm, cropped_boxes,
orig_img, im_name)
if args.profile:
ckpt_time, post_time = getTime(ckpt_time)
runtime_profile['pn'].append(post_time)
if args.profile:
# TQDM
im_names_desc.set_description(
'det time: {dt:.4f} | pose time: {pt:.4f} | post processing: {pn:.4f}'
.format(dt=np.mean(runtime_profile['dt']),
pt=np.mean(runtime_profile['pt']),
pn=np.mean(runtime_profile['pn'])))
print_finish_info()
print("마무리 작업중...")
while (writer.running()):
time.sleep(1)
print('===========================> Rendering remaining ' +
str(writer.count()) + ' images in the queue...')
writer.stop()
det_loader.stop()
print("작업종료")
except Exception as e:
print(repr(e))
print(
'An error as above occurs when processing the images, please check it'
)
pass
except KeyboardInterrupt:
print_finish_info()
# Thread won't be killed when press Ctrl+C
if args.sp:
det_loader.terminate()
while (writer.running()):
time.sleep(1)
print('===========================> Rendering remaining ' +
str(writer.count()) + ' images in the queue...')
writer.stop()
else:
# subprocesses are killed, manually clear queues
det_loader.terminate()
writer.terminate()
writer.clear_queues()
det_loader.clear_queues()
def predict(self, image, img_name):
args = self.args
# Load detection loader
det_loader = DetectionLoader(self.input_source, [img_name], [image],
get_detector(args),
self.cfg,
args,
batchSize=args.detbatch,
mode=self.mode).start()
# Init data writer
queueSize = args.qsize
self.writer = DataWriter(self.cfg,
args,
save_video=False,
queueSize=queueSize).start()
runtime_profile = {'dt': [], 'pt': [], 'pn': []}
data_len = det_loader.length
im_names_desc = tqdm(range(data_len), dynamic_ncols=True)
batchSize = args.posebatch
if args.flip:
batchSize = int(batchSize / 2)
try:
for i in im_names_desc:
start_time = getTime()
with torch.no_grad():
(inps, orig_img, im_name, boxes, scores, ids,
cropped_boxes) = det_loader.read()
if orig_img is None:
break
if boxes is None or boxes.nelement() == 0:
self.writer.save(None, None, None, None, None,
orig_img, os.path.basename(im_name))
continue
if args.profile:
ckpt_time, det_time = getTime(start_time)
runtime_profile['dt'].append(det_time)
# Pose Estimation
inps = inps.to(args.device)
datalen = inps.size(0)
leftover = 0
if (datalen) % batchSize:
leftover = 1
num_batches = datalen // batchSize + leftover
hm = []
for j in range(num_batches):
inps_j = inps[j * batchSize:min((j + 1) *
batchSize, datalen)]
if args.flip:
inps_j = torch.cat((inps_j, flip(inps_j)))
hm_j = self.pose_model(inps_j)
if args.flip:
hm_j_flip = flip_heatmap(hm_j[int(len(hm_j) / 2):],
det_loader.joint_pairs,
shift=True)
hm_j = (hm_j[0:int(len(hm_j) / 2)] + hm_j_flip) / 2
hm.append(hm_j)
hm = torch.cat(hm)
if args.profile:
ckpt_time, pose_time = getTime(ckpt_time)
runtime_profile['pt'].append(pose_time)
hm = hm.cpu()
self.writer.save(boxes, scores, ids, hm, cropped_boxes,
orig_img, os.path.basename(im_name))
if args.profile:
ckpt_time, post_time = getTime(ckpt_time)
runtime_profile['pn'].append(post_time)
if args.profile:
# TQDM
im_names_desc.set_description(
'det time: {dt:.4f} | pose time: {pt:.4f} | post processing: {pn:.4f}'
.format(dt=np.mean(runtime_profile['dt']),
pt=np.mean(runtime_profile['pt']),
pn=np.mean(runtime_profile['pn'])))
while (self.writer.running()):
time.sleep(1)
print('===========================> Rendering remaining ' +
str(self.writer.count()) + ' images in the queue...')
self.writer.stop()
det_loader.stop()
except KeyboardInterrupt:
self.print_finish_info(args)
# Thread won't be killed when press Ctrl+C
if args.sp:
det_loader.terminate()
while (self.writer.running()):
time.sleep(1)
print('===========================> Rendering remaining ' +
str(self.writer.count()) + ' images in the queue...')
self.writer.stop()
else:
# subprocesses are killed, manually clear queues
self.writer.commit()
self.writer.clear_queues()
# det_loader.clear_queues()
final_result = self.writer.results()
return write_json(final_result,
args.outputpath,
form=args.format,
for_eval=args.eval)
