def forward(self, inputs, truth_boxes=None, truth_labels=None, truth_instances=None):
cfg = self.cfg
mode = self.mode
features = self.data_parallel(self.feature_net, inputs)
# rpn proposals -------------------------------------------
self.rpn_logits_flat, self.rpn_deltas_flat = self.data_parallel(self.rpn_head, features)
self.rpn_window = make_rpn_windows(cfg, features)
self.rpn_proposals = rpn_nms(cfg, mode, inputs, self.rpn_window, self.rpn_logits_flat, self.rpn_deltas_flat)
if mode in ['train', 'valid']:
self.rpn_labels, self.rpn_label_assigns, self.rpn_label_weights, self.rpn_targets, self.rpn_target_weights = \
make_rpn_target(cfg, mode, inputs, self.rpn_window, truth_boxes, truth_labels)
self.rpn_proposals, self.rcnn_labels, self.rcnn_assigns, self.rcnn_targets = \
make_rcnn_target(cfg, mode, inputs, self.rpn_proposals, truth_boxes, truth_labels) # TODO the new rpn_proposals is sampled, but why?
# rcnn proposals ------------------------------------------------
self.rcnn_proposals = self.rpn_proposals
if len(self.rpn_proposals) > 0:
rcnn_crops = self.rcnn_crop(features, self.rpn_proposals)
self.rcnn_logits, self.rcnn_deltas = self.data_parallel(self.rcnn_head, rcnn_crops)
self.rcnn_proposals = rcnn_nms(cfg, mode, inputs, self.rpn_proposals, self.rcnn_logits, self.rcnn_deltas)
if mode in ['train', 'valid']:
self.rcnn_proposals, self.mask_labels, self.mask_assigns, self.mask_instances, = \
make_mask_target(cfg, mode, inputs, self.rcnn_proposals, truth_boxes, truth_labels, truth_instances)
# segmentation -------------------------------------------
self.detections = self.rcnn_proposals
self.masks = make_empty_masks(cfg, mode, inputs)
if len(self.detections) > 0:
mask_crops = self.mask_crop(features, self.detections)
self.mask_logits = self.data_parallel(self.mask_head, mask_crops)
self.masks = mask_nms(cfg, mode, inputs, self.detections, self.mask_logits) # <todo> better nms for mask
def get_detections(self, inputs):
self.rcnn_proposals = self.rpn_proposals
self.detections = self.rcnn_proposals
if len(self.rpn_proposals) > 0:
self.rcnn_proposals = rcnn_nms(
self.cfg, self.mode, inputs, self.rpn_proposals, self.rcnn_logits, self.rcnn_deltas)
self.detections = self.rcnn_proposals
return self.detections