def main():
parser = argparse.ArgumentParser()
parser.add_argument('-s', '--dataset', type=str, choices=DatasetBase.OPTIONS, required=True, help='name of dataset')
parser.add_argument('-b', '--backbone', type=str, choices=BackboneBase.OPTIONS, required=True, help='name of backbone model')
parser.add_argument('-c', '--checkpoint', type=str, required=True, help='path to checkpoint')
parser.add_argument('-p', '--probability_threshold', type=float, default=0.6, help='threshold of detection probability')
parser.add_argument('--image_min_side', type=float, help='default: {:g}'.format(Config.IMAGE_MIN_SIDE))
parser.add_argument('--image_max_side', type=float, help='default: {:g}'.format(Config.IMAGE_MAX_SIDE))
parser.add_argument('--anchor_ratios', type=str, help='default: "{!s}"'.format(Config.ANCHOR_RATIOS))
parser.add_argument('--anchor_sizes', type=str, help='default: "{!s}"'.format(Config.ANCHOR_SIZES))
parser.add_argument('--pooler_mode', type=str, choices=Pooler.OPTIONS, help='default: {.value:s}'.format(Config.POOLER_MODE))
parser.add_argument('--rpn_pre_nms_top_n', type=int, help='default: {:d}'.format(Config.RPN_PRE_NMS_TOP_N))
parser.add_argument('--rpn_post_nms_top_n', type=int, help='default: {:d}'.format(Config.RPN_POST_NMS_TOP_N))
args = parser.parse_args()
dataset_name = args.dataset
backbone_name = args.backbone
path_to_checkpoint = args.checkpoint
prob_thresh = args.probability_threshold
Config.setup(image_min_side=args.image_min_side, image_max_side=args.image_max_side,
anchor_ratios=args.anchor_ratios, anchor_sizes=args.anchor_sizes, pooler_mode=args.pooler_mode,
rpn_pre_nms_top_n=args.rpn_pre_nms_top_n, rpn_post_nms_top_n=args.rpn_post_nms_top_n)
print('Arguments:')
for k, v in vars(args).items():
print(f'\t{k} = {v}')
print(Config.describe())
_infer_websocket(path_to_checkpoint, dataset_name, backbone_name, prob_thresh)
def main():
parser = argparse.ArgumentParser()
parser.add_argument('input', type=str, help='path to input image')
parser.add_argument('output', type=str, help='path to output result image')
parser.add_argument('-c', '--checkpoint', type=str, required=True, help='path to checkpoint')
parser.add_argument('-s', '--dataset', type=str, choices=DatasetBase.OPTIONS, required=True, help='name of dataset')
parser.add_argument('-b', '--backbone', type=str, choices=BackboneBase.OPTIONS, required=True, help='name of backbone model')
parser.add_argument('-p', '--probability_threshold', type=float, default=0.6, help='threshold of detection probability')
parser.add_argument('--image_min_side', type=float, help='default: {:g}'.format(Config.IMAGE_MIN_SIDE))
parser.add_argument('--image_max_side', type=float, help='default: {:g}'.format(Config.IMAGE_MAX_SIDE))
parser.add_argument('--anchor_ratios', type=str, help='default: "{!s}"'.format(Config.ANCHOR_RATIOS))
parser.add_argument('--anchor_scales', type=str, help='default: "{!s}"'.format(Config.ANCHOR_SCALES))
parser.add_argument('--pooling_mode', type=str, choices=ROIWrapper.OPTIONS, help='default: {.value:s}'.format(Config.POOLING_MODE))
parser.add_argument('--rpn_pre_nms_top_n', type=int, help='default: {:d}'.format(Config.RPN_PRE_NMS_TOP_N))
parser.add_argument('--rpn_post_nms_top_n', type=int, help='default: {:d}'.format(Config.RPN_POST_NMS_TOP_N))
args = parser.parse_args()
path_to_input_image = args.input
path_to_output_image = args.output
path_to_checkpoint = args.checkpoint
dataset_name = args.dataset
backbone_name = args.backbone
prob_thresh = args.probability_threshold
os.makedirs(os.path.join(os.path.curdir, os.path.dirname(path_to_output_image)), exist_ok=True)
Config.setup(image_min_side=args.image_min_side, image_max_side=args.image_max_side,
anchor_ratios=args.anchor_ratios, anchor_scales=args.anchor_scales, pooling_mode=args.pooling_mode,
rpn_pre_nms_top_n=args.rpn_pre_nms_top_n, rpn_post_nms_top_n=args.rpn_post_nms_top_n)
print('Arguments:')
for k, v in vars(args).items():
print(f'\t{k} = {v}')
print(Config.describe())
_infer(path_to_input_image, path_to_output_image, path_to_checkpoint, dataset_name, backbone_name, prob_thresh)
def main():
parser = argparse.ArgumentParser()
parser.add_argument('checkpoint', type=str, help='path to evaluating checkpoint')
parser.add_argument('-s', '--dataset', type=str, choices=DatasetBase.OPTIONS, required=True, help='name of dataset')
parser.add_argument('-b', '--backbone', type=str, choices=BackboneBase.OPTIONS, required=True, help='name of backbone model')
parser.add_argument('-d', '--data_dir', type=str, default='./data', help='path to data directory')
parser.add_argument('--image_min_side', type=float, help='default: {:g}'.format(Config.IMAGE_MIN_SIDE))
parser.add_argument('--image_max_side', type=float, help='default: {:g}'.format(Config.IMAGE_MAX_SIDE))
parser.add_argument('--anchor_ratios', type=str, help='default: "{!s}"'.format(Config.ANCHOR_RATIOS))
parser.add_argument('--anchor_scales', type=str, help='default: "{!s}"'.format(Config.ANCHOR_SCALES))
parser.add_argument('--pooling_mode', type=str, choices=ROIWrapper.OPTIONS, help='default: {.value:s}'.format(Config.POOLING_MODE))
parser.add_argument('--rpn_pre_nms_top_n', type=int, help='default: {:d}'.format(Config.RPN_PRE_NMS_TOP_N))
parser.add_argument('--rpn_post_nms_top_n', type=int, help='default: {:d}'.format(Config.RPN_POST_NMS_TOP_N))
args = parser.parse_args()
path_to_checkpoint = args.checkpoint
dataset_name = args.dataset
backbone_name = args.backbone
path_to_data_dir = args.data_dir
path_to_results_dir = os.path.join(os.path.dirname(path_to_checkpoint), 'results-{:s}-{:s}-{:s}'.format(
time.strftime('%Y%m%d%H%M%S'), path_to_checkpoint.split(os.path.sep)[-1].split(os.path.curdir)[0],
str(uuid.uuid4()).split('-')[0]))
os.makedirs(path_to_results_dir)
Config.setup(image_min_side=args.image_min_side, image_max_side=args.image_max_side,
anchor_ratios=args.anchor_ratios, anchor_scales=args.anchor_scales, pooling_mode=args.pooling_mode,
rpn_pre_nms_top_n=args.rpn_pre_nms_top_n, rpn_post_nms_top_n=args.rpn_post_nms_top_n)
Log.initialize(os.path.join(path_to_results_dir, 'eval.log'))
Log.i('Arguments:')
for k, v in vars(args).items():
Log.i(f'\t{k} = {v}')
Log.i(Config.describe())
_eval(path_to_checkpoint, dataset_name, backbone_name, path_to_data_dir, path_to_results_dir)
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--image_min_side',
type=float,
help='default: {:g}'.format(Config.IMAGE_MIN_SIDE))
parser.add_argument('--image_max_side',
type=float,
help='default: {:g}'.format(Config.IMAGE_MAX_SIDE))
parser.add_argument('--anchor_ratios',
type=str,
help='default: "{!s}"'.format(
Config.ANCHOR_RATIOS))
parser.add_argument('--anchor_sizes',
type=str,
help='default: "{!s}"'.format(Config.ANCHOR_SIZES))
parser.add_argument('--pooler_mode',
type=str,
choices=Pooler.OPTIONS,
help='default: {.value:s}'.format(
Config.POOLER_MODE))
parser.add_argument('--rpn_pre_nms_top_n',
type=int,
help='default: {:d}'.format(
Config.RPN_PRE_NMS_TOP_N))
parser.add_argument('--rpn_post_nms_top_n',
type=int,
help='default: {:d}'.format(
Config.RPN_POST_NMS_TOP_N))
args = parser.parse_args()
input_root = '/home/mmlab/CCTV_Server/models/detectors/FasterRCNN/frames'
output_root = input_root + '_output'
path_to_checkpoint = '/home/mmlab/CCTV_Server/models/detectors/FasterRCNN/checkpoints/obstacle/model-90000.pth'
dataset_name = 'obstacle'
backbone_name = 'resnet101'
prob_thresh = 0.6
Config.setup(image_min_side=args.image_min_side,
image_max_side=args.image_max_side,
anchor_ratios=args.anchor_ratios,
anchor_sizes=args.anchor_sizes,
pooler_mode=args.pooler_mode,
rpn_pre_nms_top_n=args.rpn_pre_nms_top_n,
rpn_post_nms_top_n=args.rpn_post_nms_top_n)
print('Arguments:')
for k, v in vars(args).items():
print(f'\t{k} = {v}')
print(Config.describe())
os.makedirs(output_root, exist_ok=True)
input_sub_dirnames = [
directory for directory in os.listdir(input_root)
if os.path.isdir(os.path.join(input_root, directory))
]
dataset_class = DatasetBase.from_name(dataset_name)
backbone = BackboneBase.from_name(backbone_name)(pretrained=False)
model = Model(backbone,
dataset_class.num_classes(),
pooler_mode=Config.POOLER_MODE,
anchor_ratios=Config.ANCHOR_RATIOS,
anchor_sizes=Config.ANCHOR_SIZES,
rpn_pre_nms_top_n=Config.RPN_PRE_NMS_TOP_N,
rpn_post_nms_top_n=Config.RPN_POST_NMS_TOP_N).cuda()
model.load(path_to_checkpoint)
for sub_dir in input_sub_dirnames:
input_sub_dirpath = os.path.join(input_root, sub_dir)
output_sub_dirpath = os.path.join(output_root, sub_dir)
filenames = [
image_basename(f) for f in os.listdir(input_sub_dirpath)
if is_image(f)
]
for filename in filenames:
path_to_input_image = image_path(input_sub_dirpath, filename,
'.jpg')
# path_to_input_image = '/faster-RCNN/frames/1_360p/1_360p_0001.jpg'
path_to_output_image = image_path(output_sub_dirpath, filename,
'.jpg')
# path_to_output_image = '/faster-RCNN/frames_output/1_360p/1_360p_0001.jpg'
os.makedirs(os.path.join(
os.path.curdir, os.path.dirname(path_to_output_image)),
exist_ok=True)
with torch.no_grad():
image = transforms.Image.open(path_to_input_image)
image_tensor, scale = dataset_class.preprocess(
image, Config.IMAGE_MIN_SIDE, Config.IMAGE_MAX_SIDE)
detection_bboxes, detection_classes, detection_probs, _, _ = \
model.eval().forward(image_tensor.unsqueeze(dim=0).cuda())
detection_bboxes /= scale
kept_indices = detection_probs > prob_thresh
detection_bboxes = detection_bboxes[kept_indices]
detection_classes = detection_classes[kept_indices]
detection_probs = detection_probs[kept_indices]
draw = ImageDraw.Draw(image)
for bbox, cls, prob in zip(detection_bboxes.tolist(),
detection_classes.tolist(),
detection_probs.tolist()):
color = random.choice([
'red', 'green', 'blue', 'yellow', 'purple', 'white'
])
bbox = BBox(left=bbox[0],
top=bbox[1],
right=bbox[2],
bottom=bbox[3])
category = dataset_class.LABEL_TO_CATEGORY_DICT[cls]
draw.rectangle(
((bbox.left, bbox.top), (bbox.right, bbox.bottom)),
outline=color)
draw.text((bbox.left, bbox.top),
text=f'{category:s} {prob:.3f}',
fill=color)
image.save(path_to_output_image)
print(f'Output image is saved to {path_to_output_image}')
def main():
parser = argparse.ArgumentParser()
'''
python infer.py -s=coco2017 -b=resnet101 -c=model-180000.pth --image_min_side=800 --image_max_side=1333 --anchor_sizes="[64, 128, 256, 512]" --rpn_post_nms_top_n=1000 field.jpg out_field.jpg
这个model-180000.pth对应的训练数据是coco2017, 因此不能使用voc2007.
'''
#指定映射参数 > COOC模型以及验证
# parser.add_argument('-s', '--dataset', type=str, default='coco2017',choices=DatasetBase.OPTIONS, help='name of dataset')
# parser.add_argument('-b', '--backbone', type=str, default='resnet101',choices=BackboneBase.OPTIONS, help='name of backbone model')
# parser.add_argument('-c', '--checkpoint', type=str, default='model-180000.pth', help='path to checkpoint')
#>VOC2007
parser.add_argument('-s',
'--dataset',
type=str,
default='voc2007',
choices=DatasetBase.OPTIONS,
help='name of dataset')
parser.add_argument('-b',
'--backbone',
type=str,
default='resnet101',
choices=BackboneBase.OPTIONS,
help='name of backbone model')
model_path = 'outputs/checkpoints-20190702190950-voc2007-resnet101-f8049269/model-22500.pth'
parser.add_argument('-c',
'--checkpoint',
type=str,
default=model_path,
help='path to checkpoint')
parser.add_argument('-p',
'--probability_threshold',
type=float,
default=0.6,
help='threshold of detection probability')
parser.add_argument('--image_min_side',
default=600,
type=float,
help='default: {:g}'.format(Config.IMAGE_MIN_SIDE))
parser.add_argument('--image_max_side',
default=1000,
type=float,
help='default: {:g}'.format(Config.IMAGE_MAX_SIDE))
parser.add_argument('--anchor_ratios',
type=str,
help='default: "{!s}"'.format(
Config.ANCHOR_RATIOS))
parser.add_argument('--anchor_sizes',
type=str,
default="[128, 256, 512]",
help='default: "{!s}"'.format(Config.ANCHOR_SIZES))
parser.add_argument('--pooler_mode',
type=str,
choices=Pooler.OPTIONS,
help='default: {.value:s}'.format(
Config.POOLER_MODE))
parser.add_argument('--rpn_pre_nms_top_n',
type=int,
help='default: {:d}'.format(
Config.RPN_PRE_NMS_TOP_N))
parser.add_argument('--rpn_post_nms_top_n',
type=int,
default=1000,
help='default: {:d}'.format(
Config.RPN_POST_NMS_TOP_N))
#最后两个是位置参数，没有指定映射参数
#parser.add_argument('input', type=str, default='field.jpg',help='path to input image')
#parser.add_argument('output', type=str, default='out_filed.jpg',help='path to output result image')
args = parser.parse_args()
path_to_input_image = '1.jpg'
path_to_output_image = 'out_coco.jpg'
dataset_name = args.dataset
backbone_name = args.backbone
path_to_checkpoint = args.checkpoint
prob_thresh = args.probability_threshold
os.makedirs(os.path.join(os.path.curdir,
os.path.dirname(path_to_output_image)),
exist_ok=True)
#Config是由类变量和类方法构成的属性配置器
Config.setup(image_min_side=args.image_min_side,
image_max_side=args.image_max_side,
anchor_ratios=args.anchor_ratios,
anchor_sizes=args.anchor_sizes,
pooler_mode=args.pooler_mode,
rpn_pre_nms_top_n=args.rpn_pre_nms_top_n,
rpn_post_nms_top_n=args.rpn_post_nms_top_n)
print('Arguments:')
for k, v in vars(args).items():
print(f'\t{k} = {v}')
print(Config.describe())
_infer(path_to_input_image, path_to_output_image, path_to_checkpoint,
dataset_name, backbone_name, prob_thresh)