def main():
"""
Main Function
"""
# Parse args and set up logging
infer_args()
if args.single_scale:
scales = [1.0]
else:
scales = [float(x) for x in args.scales.split(',')]
output_dir = os.path.join(args.ckpt_path, args.exp_name, args.split)
os.makedirs(output_dir, exist_ok=True)
save_log('eval', output_dir, date_str)
logging.info("Network Arch: %s", args.arch)
logging.info("CV split: %d", args.cv_split)
logging.info("Exp_name: %s", args.exp_name)
logging.info("Ckpt path: %s", args.ckpt_path)
logging.info("Scales : %s", ' '.join(str(e) for e in scales))
logging.info("Inference mode: %s", args.inference_mode)
# Set up network, loader, inference mode
metrics = args.dataset != 'video_folder'
test_loader = setup_loader()
runner = RunEval(output_dir,
metrics,
write_image=args.dump_images,
dataset_cls=args.dataset_cls,
inference_mode=args.inference_mode)
net = get_net()
# Fix the ASPP pool size to 105, which is the tensor size if you train with crop
# size of 840x840
if args.fixed_aspp_pool:
net.module.aspp.img_pooling = torch.nn.AvgPool2d(105)
if args.inference_mode == 'sliding':
inference = inference_sliding
elif args.inference_mode == 'pooling':
inference = inference_pool
else:
raise 'Not a valid inference mode: {}'.format(args.inference_mode)
# Run Inference!
pbar = tqdm(test_loader, desc='eval {}'.format(args.split), smoothing=1.0)
for iteration, data in enumerate(pbar):
#if iteration < 800:
# continue
if args.dataset == 'video_folder':
imgs, img_names = data
gt = None
else:
if args.inference_mode == 'pooling':
if args.pos_rfactor > 0:
base_img, gt_with_imgs, img_names, (pos_h, pos_w) = data
else:
base_img, gt_with_imgs, img_names = data
base_img = base_img[0]
imgs = gt_with_imgs[0]
gt = gt_with_imgs[1]
else:
base_img = None
if args.pos_rfactor > 0:
imgs, gt, img_names, _, (pos_h, pos_w) = data
else:
imgs, gt, img_names = data
################### Profile
# pos = (pos_h, pos_w)
# flops, params = profile(net, inputs=(imgs, pos))
# print("Flops", flops, params)
# flops, params = clever_format([flops, params], "%.3f")
# print("Flops", flops, params)
# exit()
################### Profile
runner.inf(imgs,
img_names,
gt,
inference,
net,
scales,
pbar,
base_img,
pos=(pos_h, pos_w))
if iteration > 5 and args.test_mode:
break
# Calculate final overall statistics
runner.final_dump()
def main():
"""
Main Function
"""
# Parse args and set up logging
infer_args()
if args.single_scale:
scales = [1.0]
else:
scales = [float(x) for x in args.scales.split(',')]
output_dir = os.path.join(args.ckpt_path, args.exp_name, args.split)
os.makedirs(output_dir, exist_ok=True)
save_log('eval', output_dir, date_str)
logging.info("Network Arch: %s", args.arch)
logging.info("CV split: %d", args.cv_split)
logging.info("Exp_name: %s", args.exp_name)
logging.info("Ckpt path: %s", args.ckpt_path)
logging.info("Scales : %s", ' '.join(str(e) for e in scales))
logging.info("Inference mode: %s", args.inference_mode)
# Set up network, loader, inference mode, metric设置为False则不计算MIoU
metrics = args.dataset != 'video_folder'
if args.dataset == 'kitti' and args.split == 'test':
metrics = False
test_loader = setup_loader()
runner = RunEval(output_dir,
metrics,
write_image=args.dump_images,
dataset_cls=args.dataset_cls,
inference_mode=args.inference_mode)
net = get_net()
# Fix the ASPP pool size to 105, which is the tensor size if you train with crop
# size of 840x840
if args.fixed_aspp_pool:
net.module.aspp.img_pooling = torch.nn.AvgPool2d(105)
if args.inference_mode == 'sliding': #默认是这个模式
inference = inference_sliding
elif args.inference_mode == 'pooling':
inference = inference_pool
elif args.inference_mode == 'whole':
inference = inference_whole
else:
raise 'Not a valid inference mode: {}'.format(args.inference_mode)
# Run Inference!
pbar = tqdm(test_loader, desc='eval {}'.format(args.split), smoothing=1.0)
for iteration, data in enumerate(pbar):
if args.dataset == 'video_folder':
imgs, img_names = data
gt = None
else:
if args.inference_mode == 'pooling':
base_img, gt_with_imgs, img_names = data
base_img = base_img[0]
imgs = gt_with_imgs[0]
gt = gt_with_imgs[1]
else:
base_img = None
imgs, gt, img_names = data
runner.inf(imgs, img_names, gt, inference, net, scales, pbar,
base_img) #运算模型得到分割结果
if iteration > 5 and args.test_mode:
break
# Calculate final overall statistics
runner.final_dump()
default='network.sfnet_resnet.DeepR18_SF_deeply_dsn',
help='network architecture used for inference')
parser.add_argument('--save_dir',
type=str,
default='./save',
help='path to save your results')
args = parser.parse_args()
assert_and_infer_cfg(args, train_mode=False)
cudnn.benchmark = False
torch.cuda.empty_cache()
# setup logger
date_str = str(datetime.datetime.now().strftime('%Y_%m_%d_%H_%M_%S'))
log_dir = os.path.join(args.save_dir, "log")
os.makedirs(log_dir, exist_ok=True)
save_log('log', log_dir, date_str, rank=0)
# get net
args.dataset_cls = cityscapes
net = get_net(args, criterion=None)
net = torch.nn.DataParallel(net).cuda()
logging.info('Net built.')
net, _ = restore_snapshot(net,
optimizer=None,
snapshot=args.snapshot,
restore_optimizer_bool=False)
net.eval()
logging.info('Net restored.')
# get data
data_dir = args.demo_folder
