def get_refined_proposal(backbone, neck, rpn_head, roi_extractor,
bbox_head, bbox_head_2nd, bbox_head_3rd, stage):
rec_id, im_id, im_info, proposal, proposal_score = CascadeRcnn.get_rpn_test_symbol(
backbone, neck, rpn_head)
rcnn_feat = backbone.get_rcnn_feature()
rcnn_feat = neck.get_rcnn_feature(rcnn_feat)
# stage1
roi_feat = roi_extractor.get_roi_feature(rcnn_feat, proposal)
_, bbox_xyxy = bbox_head.get_prediction(roi_feat, im_info, proposal)
# stage2
proposal_2nd = bbox_xyxy
roi_feat_2nd = roi_extractor.get_roi_feature(rcnn_feat, proposal_2nd)
_, bbox_xyxy_2nd = bbox_head_2nd.get_prediction(
roi_feat_2nd, im_info, proposal_2nd)
# stage3
proposal_3rd = bbox_xyxy_2nd
roi_feat_3rd = roi_extractor.get_roi_feature(rcnn_feat, proposal_3rd)
cls_score_3rd, bbox_xyxy_3rd = bbox_head_3rd.get_prediction(
roi_feat_3rd, im_info, proposal_3rd)
# AR does not need score, just pass a dummy one
if stage == 2:
return X.group(
[rec_id, im_id, im_info, proposal_2nd, cls_score_3rd])
elif stage == 3:
return X.group(
[rec_id, im_id, im_info, proposal_3rd, cls_score_3rd])
def get_train_symbol(backbone, neck, rpn_head, bbox_roi_extractor,
mask_roi_extractor, bbox_head, mask_head,
maskiou_head):
gt_bbox = X.var("gt_bbox")
gt_poly = X.var("gt_poly")
im_info = X.var("im_info")
rpn_feat = backbone.get_rpn_feature()
rcnn_feat = backbone.get_rcnn_feature()
rpn_feat = neck.get_rpn_feature(rpn_feat)
rcnn_feat = neck.get_rcnn_feature(rcnn_feat)
rpn_head.get_anchor()
rpn_loss = rpn_head.get_loss(rpn_feat, gt_bbox, im_info)
proposal, bbox_cls, bbox_target, bbox_weight, mask_proposal, mask_target, mask_ind, mask_ratio = \
rpn_head.get_sampled_proposal(rpn_feat, gt_bbox, gt_poly, im_info)
roi_feat = bbox_roi_extractor.get_roi_feature(rcnn_feat, proposal)
mask_roi_feat = mask_roi_extractor.get_roi_feature(
rcnn_feat, mask_proposal)
bbox_loss = bbox_head.get_loss(roi_feat, bbox_cls, bbox_target,
bbox_weight)
mask_loss, mask_pred_logits = mask_head.get_loss(
mask_roi_feat, mask_target, mask_ind)
iou_loss = maskiou_head.get_loss(mask_roi_feat, mask_pred_logits,
mask_target, mask_ind, mask_ratio)
return X.group(rpn_loss + bbox_loss + mask_loss + iou_loss)
def get_test_symbol(backbone, neck, rpn_head, roi_extractor,
mask_roi_extractor, bbox_head, mask_head,
bbox_post_processor, num_branch):
rec_id, im_id, im_info, proposal, proposal_score = \
TridentFasterRcnn.get_rpn_test_symbol(backbone, neck, rpn_head, num_branch)
im_info_branches = TridentResNetV2Builder.stack_branch_symbols(
[im_info] * num_branch)
rcnn_feat = backbone.get_rcnn_feature()
rcnn_feat = neck.get_rcnn_feature(rcnn_feat)
roi_feat = roi_extractor.get_roi_feature(rcnn_feat, proposal)
cls_score, bbox_xyxy = bbox_head.get_prediction(
roi_feat, im_info_branches, proposal)
post_cls_score, post_bbox_xyxy, post_cls = bbox_post_processor.get_post_processing(
cls_score, bbox_xyxy)
mask_roi_feat = mask_roi_extractor.get_roi_feature(
rcnn_feat, post_bbox_xyxy)
mask = mask_head.get_prediction(mask_roi_feat)
# fold batch size into roi size for trident only
post_cls_score = X.reshape(post_cls_score, (-3, -2),
name="post_cls_score_fold")
post_bbox_xyxy = X.reshape(post_bbox_xyxy, (-3, -2),
name="post_bbox_xyxy_fold")
post_cls = X.reshape(post_cls, (-3, -2), name="post_cls_fold")
return X.group([
rec_id, im_id, im_info, post_cls_score, post_bbox_xyxy, post_cls,
mask
])
def get_test_symbol(backbone, neck, rpn_head, bbox_roi_extractor,
mask_roi_extractor, bbox_head, mask_head,
bbox_post_processor, maskiou_head):
rec_id, im_id, im_info, proposal, proposal_score = \
MaskScoringRcnn.get_rpn_test_symbol(backbone, neck, rpn_head)
rcnn_feat = backbone.get_rcnn_feature()
rcnn_feat = neck.get_rcnn_feature(rcnn_feat)
roi_feat = bbox_roi_extractor.get_roi_feature(rcnn_feat, proposal)
cls_score, bbox_xyxy = bbox_head.get_prediction(
roi_feat, im_info, proposal)
post_cls_score, post_bbox_xyxy, post_cls = bbox_post_processor.get_post_processing(
cls_score, bbox_xyxy)
mask_roi_feat = mask_roi_extractor.get_roi_feature(
rcnn_feat, post_bbox_xyxy)
mask = mask_head.get_prediction(mask_roi_feat)
iou_pred = maskiou_head.get_maskiou_prediction(mask, mask_roi_feat,
post_cls)
return X.group([
rec_id, im_id, im_info, post_cls_score, post_bbox_xyxy, post_cls,
mask, iou_pred
])
def get_train_symbol(backbone, neck, rpn_head, roi_extractor,
mask_roi_extractor, bbox_head, mask_head):
gt_bbox = X.var("gt_bbox")
gt_poly = X.var("gt_poly")
im_info = X.var("im_info")
rpn_cls_label = X.var("rpn_cls_label")
rpn_reg_target = X.var("rpn_reg_target")
rpn_reg_weight = X.var("rpn_reg_weight")
rpn_feat = backbone.get_rpn_feature()
rcnn_feat = backbone.get_rcnn_feature()
rpn_feat = neck.get_rpn_feature(rpn_feat)
rcnn_feat = neck.get_rcnn_feature(rcnn_feat)
rpn_loss = rpn_head.get_loss(rpn_feat, rpn_cls_label, rpn_reg_target,
rpn_reg_weight)
proposal, bbox_cls, bbox_target, bbox_weight, mask_proposal, mask_target = \
rpn_head.get_sampled_proposal(rpn_feat, gt_bbox, gt_poly, im_info)
roi_feat = roi_extractor.get_roi_feature(rcnn_feat, proposal)
mask_roi_feat = mask_roi_extractor.get_roi_feature(
rcnn_feat, mask_proposal)
bbox_loss = bbox_head.get_loss(roi_feat, bbox_cls, bbox_target,
bbox_weight)
mask_loss = mask_head.get_loss(mask_roi_feat, mask_target)
return X.group(rpn_loss + bbox_loss + mask_loss)
def get_train_symbol(backbone, neck, rpn_head, roi_extractor, bbox_head, \
bbox_head_2nd, bbox_head_3rd):
gt_bbox = X.var("gt_bbox")
im_info = X.var("im_info")
rpn_cls_label = X.var("rpn_cls_label")
rpn_reg_target = X.var("rpn_reg_target")
rpn_reg_weight = X.var("rpn_reg_weight")
rpn_feat = backbone.get_rpn_feature()
rcnn_feat = backbone.get_rcnn_feature()
rpn_feat = neck.get_rpn_feature(rpn_feat)
rcnn_feat = neck.get_rcnn_feature(rcnn_feat)
rpn_loss = rpn_head.get_loss(rpn_feat, rpn_cls_label, rpn_reg_target,
rpn_reg_weight)
# stage1
proposal, bbox_cls, bbox_target, bbox_weight = \
rpn_head.get_sampled_proposal(
rpn_feat,
gt_bbox,
im_info
)
roi_feat = roi_extractor.get_roi_feature(rcnn_feat, proposal, "1st")
bbox_loss = bbox_head.get_loss(roi_feat, bbox_cls, bbox_target,
bbox_weight)
bbox_pred = bbox_head._bbox_delta
# stage2
# though call get_sampled_proposal, bbox_head does not sample rois
proposal_2nd, bbox_cls_2nd, bbox_target_2nd, bbox_weight_2nd = \
bbox_head.get_sampled_proposal(
proposal,
bbox_pred,
gt_bbox,
im_info
)
roi_feat_2nd = roi_extractor.get_roi_feature(rcnn_feat, proposal_2nd,
"2nd")
bbox_loss_2nd = bbox_head_2nd.get_loss(roi_feat_2nd, bbox_cls_2nd,
bbox_target_2nd,
bbox_weight_2nd)
bbox_pred_2nd = bbox_head_2nd._bbox_delta
# stage3
# though call get_sampled_proposal, bbox_head does not sample rois
proposal_3rd, bbox_cls_3rd, bbox_target_3rd, bbox_weight_3rd = \
bbox_head_2nd.get_sampled_proposal(
proposal_2nd,
bbox_pred_2nd,
gt_bbox,
im_info
)
roi_feat_3rd = roi_extractor.get_roi_feature(rcnn_feat, proposal_3rd,
"3rd")
bbox_loss_3rd = bbox_head_3rd.get_loss(roi_feat_3rd, bbox_cls_3rd,
bbox_target_3rd,
bbox_weight_3rd)
return X.group(rpn_loss + bbox_loss + bbox_loss_2nd + bbox_loss_3rd)
def get_train_symbol(cls, backbone, neck, rpn_head):
rpn_feat = backbone.get_rpn_feature()
rpn_feat = neck.get_rpn_feature(rpn_feat)
rpn_loss = rpn_head.get_loss(rpn_feat, None, None)
return X.group(rpn_loss)
def get_train_symbol(cls, backbone, neck, rpn_head, roi_extractor, bbox_head, num_branch, scaleaware):
gt_bbox = X.var("gt_bbox")
im_info = X.var("im_info")
if scaleaware:
valid_ranges = X.var("valid_ranges")
rpn_cls_label = X.var("rpn_cls_label")
rpn_reg_target = X.var("rpn_reg_target")
rpn_reg_weight = X.var("rpn_reg_weight")
im_info = TridentResNetV2Builder.stack_branch_symbols([im_info] * num_branch)
gt_bbox = TridentResNetV2Builder.stack_branch_symbols([gt_bbox] * num_branch)
if scaleaware:
valid_ranges = X.reshape(valid_ranges, (-3, -2))
rpn_cls_label = X.reshape(rpn_cls_label, (-3, -2))
rpn_reg_target = X.reshape(rpn_reg_target, (-3, -2))
rpn_reg_weight = X.reshape(rpn_reg_weight, (-3, -2))
rpn_feat = backbone.get_rpn_feature()
rcnn_feat = backbone.get_rcnn_feature()
rpn_feat = neck.get_rpn_feature(rpn_feat)
rcnn_feat = neck.get_rcnn_feature(rcnn_feat)
rpn_loss = rpn_head.get_loss(rpn_feat, rpn_cls_label, rpn_reg_target, rpn_reg_weight)
if scaleaware:
proposal, bbox_cls, bbox_target, bbox_weight = rpn_head.get_sampled_proposal_with_filter(rpn_feat, gt_bbox, im_info, valid_ranges)
else:
proposal, bbox_cls, bbox_target, bbox_weight = rpn_head.get_sampled_proposal(rpn_feat, gt_bbox, im_info)
roi_feat = roi_extractor.get_roi_feature(rcnn_feat, proposal)
bbox_loss = bbox_head.get_loss(roi_feat, bbox_cls, bbox_target, bbox_weight)
return X.group(rpn_loss + bbox_loss)
def get_test_symbol(backbone, neck, rpn_head, roi_extractor,
mask_roi_extractor, bbox_head, mask_head,
bbox_post_processor):
im_info = X.var("im_info")
im_id = X.var("im_id")
rec_id = X.var("rec_id")
rpn_feat = backbone.get_rpn_feature()
rcnn_feat = backbone.get_rcnn_feature()
rpn_feat = neck.get_rpn_feature(rpn_feat)
rcnn_feat = neck.get_rcnn_feature(rcnn_feat)
proposal = rpn_head.get_all_proposal(rpn_feat, im_info)
roi_feat = roi_extractor.get_roi_feature(rcnn_feat, proposal)
cls_score, bbox_xyxy = bbox_head.get_prediction(
roi_feat, im_info, proposal)
post_cls_score, post_bbox_xyxy, post_cls = bbox_post_processor.get_post_processing(
cls_score, bbox_xyxy)
mask_roi_feat = mask_roi_extractor.get_roi_feature(
rcnn_feat, post_bbox_xyxy)
mask = mask_head.get_prediction(mask_roi_feat)
return X.group([
rec_id, im_id, im_info, post_cls_score, post_bbox_xyxy, post_cls,
mask
])
def get_test_symbol(cls, backbone, neck, rpn_head, roi_extractor,
bbox_head):
rec_id, im_id, im_info, proposal, proposal_score = \
FasterRcnn.get_rpn_test_symbol(backbone, neck, rpn_head)
rcnn_feat = backbone.get_rcnn_feature()
rcnn_feat = neck.get_rcnn_feature(rcnn_feat)
roi_feat = roi_extractor.get_roi_feature_test(rcnn_feat, proposal)
cls_score, bbox_xyxy = bbox_head.get_prediction(
roi_feat, im_info, proposal)
if is_deploy:
out = X.group([im_info, cls_score, bbox_xyxy])
else:
out = X.group([rec_id, im_id, im_info, cls_score, bbox_xyxy])
return out
def get_train_symbol(backbone, neck, head):
label = X.var("gt_bbox")
feat = backbone.get_rpn_feature()
feat = neck.get_rpn_feature(feat)
loss = head.get_loss(feat, label)
return X.group(loss)
def get_train_symbol(backbone, neck, head):
gt_bbox = X.var("gt_bbox")
im_info = X.var("im_info")
feat = backbone.get_rpn_feature()
feat = neck.get_rpn_feature(feat)
head.get_anchor()
loss = head.get_loss(feat, gt_bbox, im_info)
return X.group(loss)
def get_train_symbol(cls, backbone, neck, rpn_head):
rpn_cls_label = X.var("rpn_cls_label")
rpn_reg_target = X.var("rpn_reg_target")
rpn_reg_weight = X.var("rpn_reg_weight")
rpn_feat = backbone.get_rpn_feature()
rpn_feat = neck.get_rpn_feature(rpn_feat)
rpn_loss = rpn_head.get_loss(rpn_feat, rpn_cls_label, rpn_reg_target, rpn_reg_weight)
return X.group(rpn_loss)
def get_test_symbol(backbone, neck, head):
im_info = X.var("im_info")
im_id = X.var("im_id")
rec_id = X.var("rec_id")
feat = backbone.get_rpn_feature()
feat = neck.get_rpn_feature(feat)
cls_score, bbox_xyxy = head.get_prediction(feat, im_info)
return X.group([rec_id, im_id, im_info, cls_score, bbox_xyxy])
def get_test_symbol(cls, backbone, neck, rpn_head, roi_extractor,
bbox_head):
rec_id, im_id, im_info, proposal, proposal_score = \
FasterRcnn_TSD.get_rpn_test_symbol(backbone, neck, rpn_head)
rcnn_feat = backbone.get_rcnn_feature()
rcnn_feat = neck.get_rcnn_feature(rcnn_feat)
roi_feat = roi_extractor.get_roi_feature_test(rcnn_feat, proposal)
preds = bbox_head.get_prediction(proposal, roi_feat, rcnn_feat,
im_info)
return X.group([rec_id, im_id, im_info] + list(preds))
def get_test_symbol(cls, backbone, neck, rpn_head, roi_extractor,
bbox_head, num_branch):
rec_id, im_id, im_info, proposal, proposal_score = \
TridentFasterRcnn.get_rpn_test_symbol(backbone, neck, rpn_head, num_branch)
im_info_branches = TridentResNetV2Builder.stack_branch_symbols(
[im_info] * num_branch)
rcnn_feat = backbone.get_rcnn_feature()
rcnn_feat = neck.get_rcnn_feature(rcnn_feat)
roi_feat = roi_extractor.get_roi_feature(rcnn_feat, proposal)
cls_score, bbox_xyxy = bbox_head.get_prediction(
roi_feat, im_info_branches, proposal)
cls_score = X.reshape(cls_score, (-3, -2))
bbox_xyxy = X.reshape(bbox_xyxy, (-3, -2))
if is_deploy:
out = X.group([im_info, cls_score, bbox_xyxy])
else:
out = X.group([rec_id, im_id, im_info, cls_score, bbox_xyxy])
return out
def get_rpn_test_symbol(cls, backbone, neck, rpn_head):
if cls._rpn_output is not None:
return cls._rpn_output
im_info = X.var("im_info")
im_id = X.var("im_id")
rec_id = X.var("rec_id")
rpn_feat = backbone.get_rpn_feature()
rpn_feat = neck.get_rpn_feature(rpn_feat)
(proposal, proposal_score) = rpn_head.get_all_proposal(rpn_feat, im_info)
cls._rpn_output = X.group([rec_id, im_id, im_info, proposal, proposal_score])
return cls._rpn_output
def get_test_symbol(cls, backbone, neck, rpn_head, roi_extractor,
bbox_head):
rec_id, im_id, im_info, proposal, proposal_score = \
DoublePredRcnn.get_rpn_test_symbol(backbone, neck, rpn_head)
rcnn_feat = backbone.get_rcnn_feature()
rcnn_feat = neck.get_rcnn_feature(rcnn_feat)
roi_feat = roi_extractor.get_roi_feature_test(rcnn_feat, proposal)
cls_score, bbox_xyxy, cls_sec_score, bbox_sec_xyxy = bbox_head.get_prediction(
roi_feat, im_info, proposal)
cls_score_concat = mx.sym.concat(*[cls_score, cls_sec_score], axis=1)
bbox_xyxy_concat = mx.sym.concat(*[bbox_xyxy, bbox_sec_xyxy], axis=1)
return X.group(
[rec_id, im_id, im_info, cls_score_concat, bbox_xyxy_concat])
def get_rpn_test_symbol(cls, backbone, neck, rpn_head, num_branch):
if cls._rpn_output is not None:
return cls._rpn_output
im_info = X.var("im_info")
im_id = X.var("im_id")
rec_id = X.var("rec_id")
rpn_feat = backbone.get_rpn_feature()
rpn_feat = neck.get_rpn_feature(rpn_feat)
im_info_branches = TridentResNetV2Builder.stack_branch_symbols([im_info] * num_branch)
(proposal, proposal_score) = rpn_head.get_all_proposal(rpn_feat, im_info_branches)
cls._rpn_output = X.group([rec_id, im_id, im_info, proposal, proposal_score])
return cls._rpn_output
def get_train_symbol(cls, backbone, neck, rpn_head, roi_extractor, bbox_head):
gt_bbox = X.var("gt_bbox")
im_info = X.var("im_info")
rpn_feat = backbone.get_rpn_feature()
rcnn_feat = backbone.get_rcnn_feature()
rpn_feat = neck.get_rpn_feature(rpn_feat)
rcnn_feat = neck.get_rcnn_feature(rcnn_feat)
rpn_head.get_anchor()
rpn_loss = rpn_head.get_loss(rpn_feat, gt_bbox, im_info)
proposal, bbox_cls, bbox_target, bbox_weight = rpn_head.get_sampled_proposal(rpn_feat, gt_bbox, im_info)
roi_feat = roi_extractor.get_roi_feature(rcnn_feat, proposal)
bbox_loss = bbox_head.get_loss(roi_feat, bbox_cls, bbox_target, bbox_weight)
return X.group(rpn_loss + bbox_loss)
def get_train_symbol(backbone, neck, head, kd_head):
rpn_cls_label = X.var("rpn_cls_label")
rpn_reg_target = X.var("rpn_reg_target")
rpn_reg_weight = X.var("rpn_reg_weight")
teacher_label = X.var("teacher_label")
feat = backbone.get_rpn_feature()
c2, c3, c4, c5 = feat
feat_dict = {'c2': c2, 'c3': c3, 'c4': c4, 'c5': c5}
feat = neck.get_rpn_feature(feat)
loss = head.get_loss(feat, rpn_cls_label, rpn_reg_target,
rpn_reg_weight)
kd_loss = kd_head.get_loss(feat_dict, teacher_label)
return X.group(loss + (kd_loss, ))
def get_test_symbol(backbone, neck, rpn_head, roi_extractor, bbox_head):
im_info = X.var("im_info")
im_id = X.var("im_id")
rec_id = X.var("rec_id")
rpn_feat = backbone.get_rpn_feature()
rcnn_feat = backbone.get_rcnn_feature()
rpn_feat = neck.get_rpn_feature(rpn_feat)
rcnn_feat = neck.get_rcnn_feature(rcnn_feat)
proposal = rpn_head.get_all_proposal(rpn_feat, im_info)
roi_feat = roi_extractor.get_roi_feature_test(rcnn_feat, proposal)
cls_score, bbox_xyxy = bbox_head.get_prediction(
roi_feat, im_info, proposal)
return X.group([rec_id, im_id, im_info, cls_score, bbox_xyxy])
def get_test_symbol(backbone, neck, rpn_head, roi_extractor, mask_roi_extractor, bbox_head, mask_head, bbox_post_processor):
rec_id, im_id, im_info, proposal, proposal_score = \
MaskFasterRcnn.get_rpn_test_symbol(backbone, neck, rpn_head)
rcnn_feat = backbone.get_rcnn_feature()
rcnn_feat = neck.get_rcnn_feature(rcnn_feat)
roi_feat = roi_extractor.get_roi_feature(rcnn_feat, proposal)
cls_score, bbox_xyxy = bbox_head.get_prediction(roi_feat, im_info, proposal)
post_cls_score, post_bbox_xyxy, post_cls = bbox_post_processor.get_post_processing(cls_score, bbox_xyxy)
mask_roi_feat = mask_roi_extractor.get_roi_feature(rcnn_feat, post_bbox_xyxy)
mask = mask_head.get_prediction(mask_roi_feat)
# mask score only for placeholder
mask_score = mx.sym.full((1, ), -1, name='maskiou_prediction')
return X.group([rec_id, im_id, im_info, post_cls_score, post_bbox_xyxy, post_cls, mask, mask_score])
def get_test_symbol(backbone, neck, rpn_head, roi_extractor, bbox_head, \
bbox_head_2nd, bbox_head_3rd):
im_info = X.var("im_info")
im_id = X.var("im_id")
rec_id = X.var("rec_id")
rpn_feat = backbone.get_rpn_feature()
rcnn_feat = backbone.get_rcnn_feature()
rpn_feat = neck.get_rpn_feature(rpn_feat)
rcnn_feat = neck.get_rcnn_feature(rcnn_feat)
# stage1
proposal = rpn_head.get_all_proposal(rpn_feat, im_info)
roi_feat = roi_extractor.get_roi_feature(rcnn_feat, proposal, "1st")
_, bbox_xyxy = bbox_head.get_prediction(roi_feat, im_info, proposal)
# stage2
proposal_2nd = bbox_xyxy
roi_feat_2nd = roi_extractor.get_roi_feature(rcnn_feat, proposal_2nd,
"2nd")
_, bbox_xyxy_2nd = bbox_head_2nd.get_prediction(
roi_feat_2nd, im_info, proposal_2nd)
# stage3
proposal_3rd = bbox_xyxy_2nd
roi_feat_3rd = roi_extractor.get_roi_feature(rcnn_feat, proposal_3rd,
"3rd")
cls_score_3rd, bbox_xyxy_3rd = bbox_head_3rd.get_prediction(
roi_feat_3rd, im_info, proposal_3rd)
# passing feature from stage3 through stage1's weight
bbox_head.stage = "1st_3rd"
cls_score_1st_3rd, _ = bbox_head.get_prediction(roi_feat_3rd, im_info)
# passing feature from stage3 through stage2's weight
bbox_head_2nd.stage = "2nd_3rd"
cls_score_2nd_3rd, _ = bbox_head_2nd.get_prediction(
roi_feat_3rd, im_info)
# average score between [1st_3rd, 2nd_3rd, 3rd]
import mxnet as mx
cls_score_avg = mx.sym.add_n(cls_score_1st_3rd, cls_score_2nd_3rd,
cls_score_3rd) / 3
return X.group([rec_id, im_id, im_info, cls_score_avg, bbox_xyxy_3rd])
def get_test_symbol(backbone, neck, rpn_head, roi_extractor, bbox_head,
num_branch):
im_info = X.var("im_info")
im_id = X.var("im_id")
rec_id = X.var("rec_id")
im_info_branches = TridentResNetV2Builder.stack_branch_symbols(
[im_info] * num_branch)
rpn_feat = backbone.get_rpn_feature()
rcnn_feat = backbone.get_rcnn_feature()
rpn_feat = neck.get_rpn_feature(rpn_feat)
rcnn_feat = neck.get_rcnn_feature(rcnn_feat)
proposal = rpn_head.get_all_proposal(rpn_feat, im_info_branches)
roi_feat = roi_extractor.get_roi_feature(rcnn_feat, proposal)
cls_score, bbox_xyxy = bbox_head.get_prediction(
roi_feat, im_info_branches, proposal)
cls_score = X.reshape(cls_score, (-3, -2))
bbox_xyxy = X.reshape(bbox_xyxy, (-3, -2))
return X.group([rec_id, im_id, im_info, cls_score, bbox_xyxy])
def get_test_symbol(backbone, neck, rpn_head, roi_extractor, bbox_head, \
bbox_head_2nd, bbox_head_3rd):
rec_id, im_id, im_info, proposal, proposal_score = \
CascadeRcnn.get_rpn_test_symbol(backbone, neck, rpn_head)
rcnn_feat = backbone.get_rcnn_feature()
rcnn_feat = neck.get_rcnn_feature(rcnn_feat)
# stage1
roi_feat = roi_extractor.get_roi_feature(rcnn_feat, proposal)
_, bbox_xyxy = bbox_head.get_prediction(roi_feat, im_info, proposal)
# stage2
proposal_2nd = bbox_xyxy
roi_feat_2nd = roi_extractor.get_roi_feature(rcnn_feat, proposal_2nd)
_, bbox_xyxy_2nd = bbox_head_2nd.get_prediction(
roi_feat_2nd, im_info, proposal_2nd)
# stage3
proposal_3rd = bbox_xyxy_2nd
roi_feat_3rd = roi_extractor.get_roi_feature(rcnn_feat, proposal_3rd)
cls_score_3rd, bbox_xyxy_3rd = bbox_head_3rd.get_prediction(
roi_feat_3rd, im_info, proposal_3rd)
# passing feature from stage3 through stage1's weight
bbox_head.stage = "1st_3rd"
cls_score_1st_3rd, _ = bbox_head.get_prediction(roi_feat_3rd, im_info)
# passing feature from stage3 through stage2's weight
bbox_head_2nd.stage = "2nd_3rd"
cls_score_2nd_3rd, _ = bbox_head_2nd.get_prediction(
roi_feat_3rd, im_info)
# average score between [1st_3rd, 2nd_3rd, 3rd]
cls_score_avg = mx.sym.add_n(cls_score_1st_3rd, cls_score_2nd_3rd,
cls_score_3rd) / 3
return X.group([rec_id, im_id, im_info, cls_score_avg, bbox_xyxy_3rd])
def get_train_symbol(cls, backbone, neck, rpn_head, roi_extractor,
bbox_head, kd_head):
gt_bbox = X.var("gt_bbox")
im_info = X.var("im_info")
teacher_label = X.var("teacher_label")
rpn_feat = backbone.get_rpn_feature()
c2, c3, c4, c5 = rpn_feat
feat_dict = {'c2': c2, 'c3': c3, 'c4': c4, 'c5': c5}
rcnn_feat = backbone.get_rcnn_feature()
rpn_feat = neck.get_rpn_feature(rpn_feat)
rcnn_feat = neck.get_rcnn_feature(rcnn_feat)
rpn_head.get_anchor()
rpn_loss = rpn_head.get_loss(rpn_feat, gt_bbox, im_info)
proposal, bbox_cls, bbox_target, bbox_weight = rpn_head.get_sampled_proposal(
rpn_feat, gt_bbox, im_info)
roi_feat = roi_extractor.get_roi_feature(rcnn_feat, proposal)
bbox_loss = bbox_head.get_loss(roi_feat, bbox_cls, bbox_target,
bbox_weight)
kd_loss = kd_head.get_loss(feat_dict, teacher_label)
return X.group(rpn_loss + bbox_loss + (kd_loss, ))