def __init__(self, cfg, input_shape: Dict[str, ShapeSpec]):
super().__init__()
# fmt: off
self.in_features = cfg.MODEL.FCOS.IN_FEATURES
self.fpn_strides = cfg.MODEL.FCOS.FPN_STRIDES
self.focal_loss_alpha = cfg.MODEL.FCOS.LOSS_ALPHA
self.focal_loss_gamma = cfg.MODEL.FCOS.LOSS_GAMMA
self.center_sample = cfg.MODEL.FCOS.CENTER_SAMPLE
self.strides = cfg.MODEL.FCOS.FPN_STRIDES
self.radius = cfg.MODEL.FCOS.POS_RADIUS
self.pre_nms_thresh_train = cfg.MODEL.FCOS.INFERENCE_TH_TRAIN
self.pre_nms_thresh_test = cfg.MODEL.FCOS.INFERENCE_TH_TEST
self.pre_nms_topk_train = cfg.MODEL.FCOS.PRE_NMS_TOPK_TRAIN
self.pre_nms_topk_test = cfg.MODEL.FCOS.PRE_NMS_TOPK_TEST
self.nms_thresh = cfg.MODEL.FCOS.NMS_TH
self.post_nms_topk_train = cfg.MODEL.FCOS.POST_NMS_TOPK_TRAIN
self.post_nms_topk_test = cfg.MODEL.FCOS.POST_NMS_TOPK_TEST
self.thresh_with_ctr = cfg.MODEL.FCOS.THRESH_WITH_CTR
# fmt: on
self.iou_loss = IOULoss(cfg.MODEL.FCOS.LOC_LOSS_TYPE)
# generate sizes of interest
soi = []
prev_size = -1
for s in cfg.MODEL.FCOS.SIZES_OF_INTEREST:
soi.append([prev_size, s])
prev_size = s
soi.append([prev_size, INF])
self.sizes_of_interest = soi
self.fcos_head = FCOSHead(cfg,
[input_shape[f] for f in self.in_features])
def __init__(self, cfg, input_shape: Dict[str, ShapeSpec]):
super().__init__()
# fmt: off
self.cfg = cfg
self.in_features = cfg.MODEL.DTInst.IN_FEATURES
self.fpn_strides = cfg.MODEL.DTInst.FPN_STRIDES
self.focal_loss_alpha = cfg.MODEL.DTInst.LOSS_ALPHA
self.focal_loss_gamma = cfg.MODEL.DTInst.LOSS_GAMMA
self.center_sample = cfg.MODEL.DTInst.CENTER_SAMPLE
self.strides = cfg.MODEL.DTInst.FPN_STRIDES
self.radius = cfg.MODEL.DTInst.POS_RADIUS
self.pre_nms_thresh_train = cfg.MODEL.DTInst.INFERENCE_TH_TRAIN
self.pre_nms_thresh_test = cfg.MODEL.DTInst.INFERENCE_TH_TEST
self.pre_nms_topk_train = cfg.MODEL.DTInst.PRE_NMS_TOPK_TRAIN
self.pre_nms_topk_test = cfg.MODEL.DTInst.PRE_NMS_TOPK_TEST
self.nms_thresh = cfg.MODEL.DTInst.NMS_TH
self.post_nms_topk_train = cfg.MODEL.DTInst.POST_NMS_TOPK_TRAIN
self.post_nms_topk_test = cfg.MODEL.DTInst.POST_NMS_TOPK_TEST
self.thresh_with_ctr = cfg.MODEL.DTInst.THRESH_WITH_CTR
self.mask_size = cfg.MODEL.DTInst.MASK_SIZE
self.num_codes = cfg.MODEL.DTInst.NUM_CODE
self.sparse_alpha = cfg.MODEL.DTInst.MASK_SPARSE_ALPHA
self.norm_constant = cfg.MODEL.DTInst.NORMALIZATION_CONSTANT
self.bg_constant = cfg.MODEL.DTInst.BACKGROUND_CONSTANT
self.dtm_type = cfg.MODEL.DTInst.DTM_TYPE
if self.dtm_type == 'standard':
self.dict_name = 'Centered_DTM_basis_m{}_n{}_a{:.2f}.npz'.format(
self.mask_size, self.num_codes, self.sparse_alpha)
elif self.dtm_type == 'reciprocal':
self.dict_name = 'Centered_reciprocal_DTM_basis_m{}_n{}_a{:.2f}.npz'.format(
self.mask_size, self.num_codes, self.sparse_alpha)
elif self.dtm_type == 'complement':
self.dict_name = 'Centered_complement_DTM_basis_m{}_n{}_a{:.2f}.npz'.format(
self.mask_size, self.num_codes, self.sparse_alpha)
elif self.dtm_type == 'other':
self.dict_name = 'Centered_other_DTM_basis_m{}_n{}_a{:.2f}_bg{:.2f}_norm{:.2f}.npz'.format(
self.mask_size, self.num_codes, self.sparse_alpha,
self.bg_constant, self.norm_constant)
else:
raise NotImplementedError
# fmt: on
self.iou_loss = IOULoss(cfg.MODEL.DTInst.LOC_LOSS_TYPE)
# generate sizes of interest
soi = []
prev_size = -1
for s in cfg.MODEL.DTInst.SIZES_OF_INTEREST:
soi.append([prev_size, s])
prev_size = s
soi.append([prev_size, INF])
self.sizes_of_interest = soi
self.DTInst_head = DTInstHead(
cfg, [input_shape[f] for f in self.in_features])
self.flag_parameters = cfg.MODEL.DTInst.FLAG_PARAMETERS
self.mask_encoding = DistanceTransformEncoding(cfg)
def __init__(self, cfg):
super(FCOSOutputs, self).__init__()
self.focal_loss_alpha = cfg.MODEL.FCOS.LOSS_ALPHA
self.focal_loss_gamma = cfg.MODEL.FCOS.LOSS_GAMMA
self.center_sample = cfg.MODEL.FCOS.CENTER_SAMPLE
self.radius = cfg.MODEL.FCOS.POS_RADIUS
self.pre_nms_thresh_train = cfg.MODEL.FCOS.INFERENCE_TH_TRAIN
self.pre_nms_topk_train = cfg.MODEL.FCOS.PRE_NMS_TOPK_TRAIN
self.post_nms_topk_train = cfg.MODEL.FCOS.POST_NMS_TOPK_TRAIN
self.loc_loss_func = IOULoss(cfg.MODEL.FCOS.LOC_LOSS_TYPE)
self.pre_nms_thresh_test = cfg.MODEL.FCOS.INFERENCE_TH_TEST
self.pre_nms_topk_test = cfg.MODEL.FCOS.PRE_NMS_TOPK_TEST
self.post_nms_topk_test = cfg.MODEL.FCOS.POST_NMS_TOPK_TEST
self.nms_thresh = cfg.MODEL.FCOS.NMS_TH
self.thresh_with_ctr = cfg.MODEL.FCOS.THRESH_WITH_CTR
self.box_quality = cfg.MODEL.FCOS.BOX_QUALITY
self.num_classes = cfg.MODEL.FCOS.NUM_CLASSES
self.strides = cfg.MODEL.FCOS.FPN_STRIDES
# whether target location should be inside mask
self.sample_in_mask = cfg.MODEL.EMBEDMASK.SAMPLE_IN_MASK
# generate sizes of interest
soi = []
prev_size = -1
for s in cfg.MODEL.FCOS.SIZES_OF_INTEREST:
soi.append([prev_size, s])
prev_size = s
soi.append([prev_size, INF])
self.sizes_of_interest = soi
self.loss_normalizer_cls = cfg.MODEL.FCOS.LOSS_NORMALIZER_CLS
assert self.loss_normalizer_cls in ("moving_fg", "fg", "all"), \
'MODEL.FCOS.CLS_LOSS_NORMALIZER can only be "moving_fg", "fg", or "all"'
# For an explanation, please refer to
# https://github.com/facebookresearch/detectron2/blob/ea8b17914fc9a5b7d82a46ccc72e7cf6272b40e4/detectron2/modeling/meta_arch/retinanet.py#L148
self.moving_num_fg = 100 # initialize with any reasonable #fg that's not too small
self.moving_num_fg_momentum = 0.9
self.loss_weight_cls = cfg.MODEL.FCOS.LOSS_WEIGHT_CLS
def __init__(self, cfg, input_shape: Dict[str, ShapeSpec]):
super().__init__()
# fmt: off
self.cfg = cfg
self.in_features = cfg.MODEL.SMInst.IN_FEATURES
self.fpn_strides = cfg.MODEL.SMInst.FPN_STRIDES
self.focal_loss_alpha = cfg.MODEL.SMInst.LOSS_ALPHA
self.focal_loss_gamma = cfg.MODEL.SMInst.LOSS_GAMMA
self.center_sample = cfg.MODEL.SMInst.CENTER_SAMPLE
self.strides = cfg.MODEL.SMInst.FPN_STRIDES
self.radius = cfg.MODEL.SMInst.POS_RADIUS
self.pre_nms_thresh_train = cfg.MODEL.SMInst.INFERENCE_TH_TRAIN
self.pre_nms_thresh_test = cfg.MODEL.SMInst.INFERENCE_TH_TEST
self.pre_nms_topk_train = cfg.MODEL.SMInst.PRE_NMS_TOPK_TRAIN
self.pre_nms_topk_test = cfg.MODEL.SMInst.PRE_NMS_TOPK_TEST
self.nms_thresh = cfg.MODEL.SMInst.NMS_TH
self.post_nms_topk_train = cfg.MODEL.SMInst.POST_NMS_TOPK_TRAIN
self.post_nms_topk_test = cfg.MODEL.SMInst.POST_NMS_TOPK_TEST
self.thresh_with_ctr = cfg.MODEL.SMInst.THRESH_WITH_CTR
# self.thresh_with_active = cfg.MODEL.SMInst.THRESH_WITH_ACTIVE
# fmt: on
self.iou_loss = IOULoss(cfg.MODEL.SMInst.LOC_LOSS_TYPE)
# generate sizes of interest
soi = []
prev_size = -1
for s in cfg.MODEL.SMInst.SIZES_OF_INTEREST:
soi.append([prev_size, s])
prev_size = s
soi.append([prev_size, INF])
self.sizes_of_interest = soi
self.SMInst_head = SMInstHead(
cfg, [input_shape[f] for f in self.in_features])
self.flag_parameters = cfg.MODEL.SMInst.FLAG_PARAMETERS
self.mask_encoding = SparseMaskEncoding(cfg)