def output_shape(self):
return {
name: ShapeSpec(
channels=self._out_feature_channels[name], stride=self._out_feature_strides[name]
)
for name in self._out_features
}
def build_backbone(cfg, input_shape=None):
if input_shape is None:
input_shape = ShapeSpec(channels=len(cfg.INPUT.PIXEL_MEAN))
if 'PLAN' in cfg.BACKBONE:
pyramidal = cfg.BACKBONE.NAME
pyramidal = BACKBONE_REGISTRY.get(pyramidal)(cfg, input_shape)
backbone = BACKBONE_REGISTRY.get('PLAN')(cfg, input_shape, pyramidal)
else:
backbone = cfg.BACKBONE.NAME
backbone = BACKBONE_REGISTRY.get(backbone)(cfg, input_shape)
assert isinstance(backbone, Backbone)
return backbone
def __init__(self, cfg, input_shape):
super().__init__(cfg)
self.in_features = cfg.META_ARCH.ROI.IN_FEATURES
self.num_classes = cfg.META_ARCH.NUM_CLASSES
in_channels = [input_shape[f].channels for f in self.in_features]
assert len(set(in_channels)) == 1, in_channels
for feat in self.in_features:
assert feat in input_shape.keys(
), f"'{feat}' is not in backbone({input_shape.keys()})"
# Matcher
iou_thres = cfg.META_ARCH.ROI.MATCHER.IOU_THRESHOLDS
iou_labels = cfg.META_ARCH.ROI.MATCHER.IOU_LABELS
allow_low_quality_matches = cfg.META_ARCH.ROI.MATCHER.LOW_QUALITY_MATCHES
self.proposal_matcher = Matcher(
iou_thres,
iou_labels,
allow_low_quality_matches=allow_low_quality_matches)
# Sampling
self.proposal_append_gt = cfg.META_ARCH.ROI.SAMPLING.PROPOSAL_APPEND_GT
self.batch_size_per_image = cfg.META_ARCH.ROI.SAMPLING.BATCH_SIZE_PER_IMAGE
self.positive_fraction = cfg.META_ARCH.ROI.SAMPLING.POSITIVE_FRACTION
# Pooling Parameters and Module
box_pooler_type = cfg.META_ARCH.ROI.BOX_POOLING.TYPE
box_pooler_resolution = cfg.META_ARCH.ROI.BOX_POOLING.RESOLUTION
box_pooler_sampling_ratio = cfg.META_ARCH.ROI.BOX_POOLING.SAMPLING_RATIO
self.box_pooler = ROIPooler(
output_size=box_pooler_resolution,
scales=tuple(1.0 / input_shape[k].stride
for k in self.in_features),
sampling_ratio=box_pooler_sampling_ratio,
pooler_type=box_pooler_type,
)
# Loss parameters
self.loss_weight = cfg.META_ARCH.ROI.BOX_LOSS.LOSS_WEIGHT
self.smooth_l1_beta = cfg.META_ARCH.ROI.BOX_LOSS.SMOOTH_L1_BETA
if isinstance(self.loss_weight, float):
self.loss_weight = {
"loss_cls": self.loss_weight,
"loss_loc": self.loss_weight
}
assert 'loss_cls' in self.loss_weight
assert 'loss_loc' in self.loss_weight
# Inference parameters
bbox_reg_weights = cfg.META_ARCH.ROI.TEST.BBOX_REG_WEIGHTS
self.box2box_transform = Box2BoxTransform(weights=bbox_reg_weights)
self.test_nms_thresh = cfg.META_ARCH.ROI.TEST.NMS_THRESH
self.score_threshhold = cfg.TEST.SCORE_THRESH
self.max_detections_per_image = cfg.TEST.DETECTIONS_PER_IMAGE
# ROI Head
self.box_head = BoxHead(
cfg,
ShapeSpec(channels=in_channels[0],
height=box_pooler_resolution,
width=box_pooler_resolution))