def get_prediction(self, conv_feat, im_info, proposal):
p = self.p
bbox_mean = p.regress_target.mean
bbox_std = p.regress_target.std
batch_image = p.batch_image
num_class = p.num_class
class_agnostic = p.regress_target.class_agnostic
num_reg_class = 2 if class_agnostic else num_class
cls_logit, bbox_delta = self.get_output(conv_feat)
bbox_delta = X.reshape(bbox_delta,
shape=(batch_image, -1, 4 * num_reg_class),
name='bbox_delta_reshape')
bbox_xyxy = X.decode_bbox(rois=proposal,
bbox_pred=bbox_delta,
im_info=im_info,
name='decode_bbox',
bbox_mean=bbox_mean,
bbox_std=bbox_std,
class_agnostic=class_agnostic)
cls_score = X.softmax(cls_logit, axis=-1, name='bbox_cls_score')
cls_score = X.reshape(cls_score,
shape=(batch_image, -1, num_class),
name='bbox_cls_score_reshape')
return cls_score, bbox_xyxy
def get_all_proposal_with_filter(self, conv_feat, im_info, valid_ranges):
if self._proposal is not None:
return self._proposal
p = self.p
rpn_stride = p.anchor_generate.stride
anchor_scale = p.anchor_generate.scale
anchor_ratio = p.anchor_generate.ratio
pre_nms_top_n = p.proposal.pre_nms_top_n
post_nms_top_n = p.proposal.post_nms_top_n
nms_thr = p.proposal.nms_thr
min_bbox_side = p.proposal.min_bbox_side
cls_logit, bbox_delta = self.get_output(conv_feat)
# TODO: remove this reshape hell
cls_logit_reshape = X.reshape(
cls_logit,
shape=(0, -4, 2, -1, 0, 0), # (N,C,H,W) -> (N,2,C/2,H,W)
name="rpn_cls_logit_reshape_"
)
cls_score = X.softmax(
cls_logit_reshape,
axis=1,
name='rpn_cls_score'
)
cls_logit_reshape = X.reshape(
cls_score,
shape=(0, -3, 0, 0),
name='rpn_cls_score_reshape'
)
proposal = mx.sym.contrib.Proposal_v2(
cls_prob=cls_logit_reshape,
bbox_pred=bbox_delta,
im_info=im_info,
valid_ranges=valid_ranges,
name='proposal',
feature_stride=rpn_stride,
scales=tuple(anchor_scale),
ratios=tuple(anchor_ratio),
rpn_pre_nms_top_n=pre_nms_top_n,
rpn_post_nms_top_n=post_nms_top_n,
threshold=nms_thr,
rpn_min_size=min_bbox_side,
iou_loss=False,
filter_scales=True,
output_score=True
)
self._proposal = proposal
return proposal
def get_prediction(self,
rois,
roi_feat,
fpn_conv_feats,
im_info,
play=False):
'''
Args:
rois: [batch_image, image_roi, 4]
roi_feat: [batch_image * image_roi, 256, roi_size, roi_size]
fpn_conv_feats: dict of FPN features, each [batch_image, in_channels, fh, fw]
im_info: ...
Returns:
cls_score: [batch_image, image_roi, num_class]
bbox_xyxy: [batch_image, image_roi, num_class*4]
'''
p = self.p
assert not p.regress_target.class_agnostic
bbox_mean = p.regress_target.mean
bbox_std = p.regress_target.std
batch_image = p.batch_image
num_class = p.num_class
class_agnostic = p.regress_target.class_agnostic
num_reg_class = num_class
assert batch_image == 1
cls_logit, bbox_delta, tsd_cls_logit, tsd_bbox_delta, delta_c, delta_r = self.get_output(
fpn_conv_feats, roi_feat, rois, is_train=False)
rois_r = self._get_delta_r_box(delta_r, rois)
bbox_xyxy = X.decode_bbox(
rois=rois_r,
bbox_pred=X.reshape(tsd_bbox_delta,
(batch_image, -1, 4 * num_reg_class)),
im_info=im_info,
name='decode_bbox',
bbox_mean=bbox_mean,
bbox_std=bbox_std,
class_agnostic=False)
cls_score = X.reshape(X.softmax(tsd_cls_logit,
axis=-1,
name='bbox_cls_score'),
shape=(batch_image, -1, num_class),
name='bbox_cls_score_reshape')
if not play:
return cls_score, bbox_xyxy
else:
return cls_score, bbox_xyxy, rois, rois_r
def get_all_proposal(self, conv_feat, im_info):
if self._proposal is not None:
return self._proposal
p = self.p
rpn_stride = p.anchor_generate.stride
anchor_scale = p.anchor_generate.scale
anchor_ratio = p.anchor_generate.ratio
pre_nms_top_n = p.proposal.pre_nms_top_n
post_nms_top_n = p.proposal.post_nms_top_n
nms_thr = p.proposal.nms_thr
min_bbox_side = p.proposal.min_bbox_side
cls_logit, bbox_delta = self.get_output(conv_feat)
# TODO: remove this reshape hell
cls_logit_reshape = X.reshape(
cls_logit,
shape=(0, -4, 2, -1, 0, 0), # (N,C,H,W) -> (N,2,C/2,H,W)
name="rpn_cls_logit_reshape_")
cls_score = X.softmax(cls_logit_reshape, axis=1, name='rpn_cls_score')
cls_logit_reshape = X.reshape(cls_score,
shape=(0, -3, 0, 0),
name='rpn_cls_score_reshape')
# TODO: ask all to add is_train filed in RPNParam
proposal = X.proposal(cls_prob=cls_logit_reshape,
bbox_pred=bbox_delta,
im_info=im_info,
name='proposal',
feature_stride=rpn_stride,
scales=tuple(anchor_scale),
ratios=tuple(anchor_ratio),
rpn_pre_nms_top_n=pre_nms_top_n,
rpn_post_nms_top_n=post_nms_top_n,
threshold=nms_thr,
rpn_min_size=min_bbox_side,
iou_loss=False)
self._proposal = proposal
return proposal
def get_all_proposal(self, conv_feat, im_info):
if self._proposal is not None:
return self._proposal
p = self.p
rpn_stride = p.anchor_generate.stride
anchor_scale = p.anchor_generate.scale
anchor_ratio = p.anchor_generate.ratio
pre_nms_top_n = p.proposal.pre_nms_top_n
post_nms_top_n = p.proposal.post_nms_top_n
nms_thr = p.proposal.nms_thr
min_bbox_side = p.proposal.min_bbox_side
cls_logit, bbox_delta = self.get_output(conv_feat)
# TODO: remove this reshape hell
cls_logit_reshape = X.reshape(
cls_logit,
shape=(0, -4, 2, -1, 0, 0), # (N,C,H,W) -> (N,2,C/2,H,W)
name="rpn_cls_logit_reshape_")
cls_score = X.softmax(cls_logit_reshape, axis=1, name='rpn_cls_score')
cls_logit_reshape = X.reshape(cls_score,
shape=(0, -3, 0, 0),
name='rpn_cls_score_reshape')
# TODO: ask all to add is_train filed in RPNParam
proposal = X.proposal(cls_prob=cls_logit_reshape,
bbox_pred=bbox_delta,
im_info=im_info,
name='proposal',
feature_stride=rpn_stride,
scales=tuple(anchor_scale),
ratios=tuple(anchor_ratio),
rpn_pre_nms_top_n=pre_nms_top_n,
rpn_post_nms_top_n=post_nms_top_n,
threshold=nms_thr,
rpn_min_size=min_bbox_side,
iou_loss=False,
output_score=True)
if p.use_symbolic_proposal is not None:
batch_size = p.batch_image
max_side = p.anchor_generate.max_side
assert max_side is not None, "symbolic proposal requires max_side of image"
from mxnext.tvm.proposal import proposal as Proposal
proposal = Proposal(cls_prob=cls_logit_reshape,
bbox_pred=bbox_delta,
im_info=im_info,
name='proposal',
feature_stride=rpn_stride,
scales=tuple(anchor_scale),
ratios=tuple(anchor_ratio),
rpn_pre_nms_top_n=pre_nms_top_n,
rpn_post_nms_top_n=post_nms_top_n,
threshold=nms_thr,
batch_size=batch_size,
max_side=max_side,
output_score=True,
variant="simpledet")
self._proposal = proposal
return proposal
def get_prediction(self, conv_feat, im_info, proposal):
p = self.p
bbox_mean = p.regress_target.mean
bbox_std = p.regress_target.std
batch_image = p.batch_image
num_class = p.num_class
class_agnostic = p.regress_target.class_agnostic
num_reg_class = 2 if class_agnostic else num_class
cls_logit, bbox_delta, cls_sec_logit, bbox_sec_delta, roi_feature = self.get_output(
conv_feat)
if p.refine_mode:
refine_cls_logit, refine_bbox_delta, refine_cls_sec_logit, refine_bbox_sec_delta = \
self.get_refine_output(roi_feature, cls_logit, bbox_delta, cls_sec_logit, bbox_sec_delta)
bbox_delta = refine_bbox_delta
bbox_sec_delta = refine_bbox_sec_delta
cls_logit = refine_cls_logit
cls_sec_logit = refine_cls_sec_logit
bbox_delta = X.reshape(bbox_delta,
shape=(batch_image, -1, 4 * num_reg_class),
name='bbox_delta_reshape')
bbox_sec_delta = X.reshape(
bbox_sec_delta,
shape=(batch_image, -1, 4 * num_reg_class),
name='bbox_sec_delta_reshape',
)
bbox_xyxy = X.decode_bbox(rois=proposal,
bbox_pred=bbox_delta,
im_info=im_info,
name='decode_bbox',
bbox_mean=bbox_mean,
bbox_std=bbox_std,
class_agnostic=class_agnostic)
bbox_sec_xyxy = X.decode_bbox(
rois=proposal,
bbox_pred=bbox_sec_delta,
im_info=im_info,
name='decode_bbox_sec',
bbox_mean=bbox_mean,
bbox_std=bbox_std,
class_agnostic=class_agnostic,
)
# NOTE: avoid name conflicts(potential bug from X.decode_bbox?)
bbox_sec_xyxy = bbox_sec_xyxy * mx.sym.ones_like(bbox_sec_xyxy)
cls_score = X.softmax(cls_logit, axis=-1, name='bbox_cls_score')
cls_score = X.reshape(cls_score,
shape=(batch_image, -1, num_class),
name='bbox_cls_score_reshape')
cls_sec_score = X.softmax(
cls_sec_logit,
axis=-1,
name='bbox_cls_sec_score',
)
cls_sec_score = X.reshape(cls_sec_score,
shape=(batch_image, -1, num_class),
name='bbox_cls_sec_score_reshape')
return cls_score, bbox_xyxy, cls_sec_score, bbox_sec_xyxy
