def __init__(self, cfg):
super().__init__()
self.device = torch.device(cfg.MODEL.DEVICE)
# fmt: off
self.num_classes = cfg.MODEL.FCOS.NUM_CLASSES
self.in_features = cfg.MODEL.FCOS.IN_FEATURES
self.fpn_strides = cfg.MODEL.FCOS.FPN_STRIDES
# Loss parameters:
self.focal_loss_alpha = cfg.MODEL.FCOS.FOCAL_LOSS_ALPHA
self.focal_loss_gamma = cfg.MODEL.FCOS.FOCAL_LOSS_GAMMA
self.iou_loss_type = cfg.MODEL.FCOS.IOU_LOSS_TYPE
self.center_sampling_radius = cfg.MODEL.FCOS.CENTER_SAMPLING_RADIUS
self.budget_loss_lambda = cfg.MODEL.FCOS.BUDGET_LOSS_LAMBDA
# Inference parameters:
self.score_threshold = cfg.MODEL.FCOS.SCORE_THRESH_TEST
self.topk_candidates = cfg.MODEL.FCOS.TOPK_CANDIDATES_TEST
self.nms_threshold = cfg.MODEL.FCOS.NMS_THRESH_TEST
self.nms_type = cfg.MODEL.NMS_TYPE
self.max_detections_per_image = cfg.TEST.DETECTIONS_PER_IMAGE
# fmt: on
self.backbone = cfg.build_backbone(
cfg, input_shape=ShapeSpec(channels=len(cfg.MODEL.PIXEL_MEAN)))
backbone_shape = self.backbone.output_shape()
feature_shapes = [backbone_shape[f] for f in self.in_features]
self.head = cfg.build_head(cfg, feature_shapes)
self.shift_generator = cfg.build_shift_generator(cfg, feature_shapes)
self.is_dynamic_head = isinstance(self.head, FCOSDynamicHead)
if self.is_dynamic_head:
self.head_complexity_buffer = None
# Matching and loss
self.shift2box_transform = Shift2BoxTransform(
weights=cfg.MODEL.FCOS.BBOX_REG_WEIGHTS)
self.object_sizes_of_interest = cfg.MODEL.FCOS.OBJECT_SIZES_OF_INTEREST
pixel_mean = torch.Tensor(cfg.MODEL.PIXEL_MEAN).to(self.device).view(
3, 1, 1)
pixel_std = torch.Tensor(cfg.MODEL.PIXEL_STD).to(self.device).view(
3, 1, 1)
self.normalizer = lambda x: (x - pixel_mean) / pixel_std
self.to(self.device)
def __init__(self, cfg):
super().__init__()
self.device = torch.device(cfg.MODEL.DEVICE)
# fmt: off
self.num_classes = cfg.MODEL.FCOS.NUM_CLASSES
self.in_features = cfg.MODEL.FCOS.IN_FEATURES
self.fpn_strides = cfg.MODEL.FCOS.FPN_STRIDES
# Loss parameters:
self.focal_loss_alpha = cfg.MODEL.FCOS.FOCAL_LOSS_ALPHA
self.focal_loss_gamma = cfg.MODEL.FCOS.FOCAL_LOSS_GAMMA
self.iou_loss_type = cfg.MODEL.FCOS.IOU_LOSS_TYPE
# Inference parameters:
self.score_threshold = cfg.MODEL.FCOS.SCORE_THRESH_TEST
self.topk_candidates = cfg.MODEL.FCOS.TOPK_CANDIDATES_TEST
self.nms_threshold = cfg.MODEL.FCOS.NMS_THRESH_TEST
self.max_detections_per_image = cfg.TEST.DETECTIONS_PER_IMAGE
# fmt: on
self.backbone = cfg.build_backbone(
cfg, input_shape=ShapeSpec(channels=len(cfg.MODEL.PIXEL_MEAN)))
backbone_shape = self.backbone.output_shape()
feature_shapes = [backbone_shape[f] for f in self.in_features]
self.head = FCOSHead(cfg, feature_shapes)
self.shift_generator = cfg.build_shift_generator(cfg, feature_shapes)
# Matching and loss
self.shift2box_transform = Shift2BoxTransform(
weights=cfg.MODEL.FCOS.BBOX_REG_WEIGHTS)
self.object_sizes_of_interest = cfg.MODEL.FCOS.OBJECT_SIZES_OF_INTEREST
self.norm_sync = cfg.MODEL.FCOS.NORM_SYNC
pixel_mean = torch.Tensor(cfg.MODEL.PIXEL_MEAN).to(self.device).view(
3, 1, 1)
pixel_std = torch.Tensor(cfg.MODEL.PIXEL_STD).to(self.device).view(
3, 1, 1)
self.normalizer = lambda x: (x - pixel_mean) / pixel_std
self.to(self.device)
self.sinkhorn = SinkhornDistance(eps=cfg.MODEL.OTA.SINKHORN_EPS,
max_iter=cfg.MODEL.OTA.SINKHORN_ITER)
self.reg_weight = cfg.MODEL.OTA.REG_WEIGHT
self.top_candidates = cfg.MODEL.OTA.TOP_CANDIDATES
def __init__(self, cfg):
super(AutoAssign, self).__init__()
self.device = torch.device(cfg.MODEL.DEVICE)
# fmt: off
self.num_classes = cfg.MODEL.FCOS.NUM_CLASSES
self.in_features = cfg.MODEL.FCOS.IN_FEATURES
self.fpn_strides = cfg.MODEL.FCOS.FPN_STRIDES
# Loss parameters:
self.focal_loss_alpha = cfg.MODEL.FCOS.FOCAL_LOSS_ALPHA
self.focal_loss_gamma = cfg.MODEL.FCOS.FOCAL_LOSS_GAMMA
self.iou_loss_type = cfg.MODEL.FCOS.IOU_LOSS_TYPE
self.reg_weight = cfg.MODEL.FCOS.REG_WEIGHT
# Inference parameters:
self.score_threshold = cfg.MODEL.FCOS.SCORE_THRESH_TEST
self.topk_candidates = cfg.MODEL.FCOS.TOPK_CANDIDATES_TEST
self.nms_threshold = cfg.MODEL.FCOS.NMS_THRESH_TEST
self.max_detections_per_image = cfg.TEST.DETECTIONS_PER_IMAGE
# fmt: on
self.backbone = cfg.build_backbone(
cfg, input_shape=ShapeSpec(channels=len(cfg.MODEL.PIXEL_MEAN)))
backbone_shape = self.backbone.output_shape()
feature_shapes = [backbone_shape[f] for f in self.in_features]
self.head = AutoAssignHead(cfg, feature_shapes)
self.shift_generator = cfg.build_shift_generator(cfg, feature_shapes)
# Matching and loss
self.shift2box_transform = Shift2BoxTransform(
weights=cfg.MODEL.FCOS.BBOX_REG_WEIGHTS)
self.mu = nn.Parameter(torch.zeros(80, 2))
self.sigma = nn.Parameter(torch.ones(80, 2))
pixel_mean = torch.Tensor(cfg.MODEL.PIXEL_MEAN).to(self.device).view(
3, 1, 1)
pixel_std = torch.Tensor(cfg.MODEL.PIXEL_STD).to(self.device).view(
3, 1, 1)
self.normalizer = lambda x: (x - pixel_mean) / pixel_std
self.to(self.device)
def __init__(self, cfg):
super().__init__()
self.device = torch.device(cfg.MODEL.DEVICE)
# fmt: off
self.num_classes = cfg.MODEL.FCOS.NUM_CLASSES
self.in_features = cfg.MODEL.FCOS.IN_FEATURES
self.fpn_strides = cfg.MODEL.FCOS.FPN_STRIDES
self.iou_smooth = cfg.MODEL.FCOS.IOU_SMOOTH
# Condinst parameters
self.mask_out_stride = cfg.MODEL.CONDINST.MASK_OUT_STRIDE
self.max_proposals = cfg.MODEL.CONDINST.MAX_PROPOSALS
self.topk_proposals_per_im = cfg.MODEL.CONDINST.TOPK_PROPOSALS_PER_IM
assert (self.max_proposals != -1) ^ (self.topk_proposals_per_im != -1),\
"MAX_PROPOSALS and TOPK_PROPOSALS_PER_IM " \
"cannot be set to -1 or enabled at the same time."
self.disable_rel_coords = cfg.MODEL.CONDINST.MASK_HEAD.DISABLE_REL_COORDS
self.mask_center_sample = cfg.MODEL.CONDINST.MASK_CENTER_SAMPLE
# Loss parameters:
self.focal_loss_alpha = cfg.MODEL.FCOS.FOCAL_LOSS_ALPHA
self.focal_loss_gamma = cfg.MODEL.FCOS.FOCAL_LOSS_GAMMA
self.iou_loss_type = cfg.MODEL.FCOS.IOU_LOSS_TYPE
self.center_sampling_radius = cfg.MODEL.FCOS.CENTER_SAMPLING_RADIUS
# Inference parameters:
self.thresh_with_centerness = cfg.MODEL.FCOS.THRESH_WITH_CENTERNESS
self.score_threshold = cfg.MODEL.FCOS.SCORE_THRESH_TEST
self.topk_candidates = cfg.MODEL.FCOS.TOPK_CANDIDATES_TEST
self.nms_threshold = cfg.MODEL.FCOS.NMS_THRESH_TEST
self.nms_type = cfg.MODEL.NMS_TYPE
self.max_detections_per_image = cfg.TEST.DETECTIONS_PER_IMAGE
self.infer_mask_threshold = cfg.MODEL.CONDINST.INFER_MASK_THRESH
# fmt: on
self.backbone = cfg.build_backbone(
cfg, input_shape=ShapeSpec(channels=len(cfg.MODEL.PIXEL_MEAN)))
backbone_shape = self.backbone.output_shape()
feature_shapes = [backbone_shape[f] for f in self.in_features]
self.dynamic_mask_head = DynamicMaskHead(cfg)
self.num_gen_params = self.dynamic_mask_head.num_gen_params
self.mask_branch = MaskBranch(cfg, backbone_shape)
self.mask_branch_out_stride = self.mask_branch.out_stride
self.mask_out_level_ind = self.fpn_strides.index(
self.mask_branch_out_stride)
self.head = CondInstHead(cfg, self.num_gen_params, feature_shapes)
self.shift_generator = cfg.build_shift_generator(cfg, feature_shapes)
# Matching and loss
self.shift2box_transform = Shift2BoxTransform(
weights=cfg.MODEL.FCOS.BBOX_REG_WEIGHTS)
self.object_sizes_of_interest = cfg.MODEL.FCOS.OBJECT_SIZES_OF_INTEREST
pixel_mean = torch.Tensor(cfg.MODEL.PIXEL_MEAN).to(
self.device).reshape(3, 1, 1)
pixel_std = torch.Tensor(cfg.MODEL.PIXEL_STD).to(self.device).reshape(
3, 1, 1)
self.normalizer = lambda x: (x - pixel_mean) / pixel_std
self.to(self.device)