def onThreadTopic(post):
check_tokenized(post)
# Leading post is treated specially
if post == post.thread.posts[0]:
title_tokens = post.tokenizer(post.thread.title)
r = overlap(post.token_index, title_tokens)
else:
r = overlap(post.token_index, post.thread.posts[0].token_index)
return r
def overlapping(anchors, gt_boxes, inds_inside):
"""
overlaps between the anchors and the gt boxes
:param anchors: Generated anchors
:param gt_boxes: Ground truth bounding boxes
:param inds_inside:
:return:
"""
assert keras.backend.ndim(anchors) == 2
assert keras.backend.ndim(gt_boxes) == 2
reference = common.overlap(anchors, gt_boxes)
gt_argmax_overlaps_inds = keras.backend.argmax(reference, axis=0)
argmax_overlaps_inds = keras.backend.argmax(reference, axis=1)
arranged = keras.backend.arange(0, keras.backend.shape(inds_inside)[0])
indices = keras.backend.stack(
[arranged, keras.backend.cast(argmax_overlaps_inds, "int32")], axis=0)
indices = keras.backend.transpose(indices)
max_overlaps = tensorflow_backend.gather_nd(reference, indices)
return argmax_overlaps_inds, max_overlaps, gt_argmax_overlaps_inds
def bid_override(g, bidder): bidOverrideNum = 0 bidOverrideNormNum = 0 bidOverrideNumAuction = 0 bidOverrideNormNumAuction = 0 bidOverrideRatioDistri = dict() edges = g.edges(bidder) auctionNum = len(edges) if auctionNum ==0: return [0]*9 # print(edges) for (b, a) in edges: bidderRecord = g.node[a]["bidders"] count = cm.overlap(bidder, bidderRecord) bidOverrideNum = bidOverrideNum+count bidOverrideRatioDistri[a] = count if count>0: bidOverrideNumAuction = bidOverrideNumAuction+1 if bidOverrideNum==0: return [0]*9 else: for auction in bidOverrideRatioDistri.keys(): bidOverrideRatioDistri[auction] = bidOverrideRatioDistri[auction]/bidOverrideNum bidOverrideNormNum = bidOverrideNum/bid_total(g, bidder) bidOverrideNormNumAuction = bidOverrideNumAuction/auctionNum return [bidOverrideNum, bidOverrideNormNum, bidOverrideNumAuction, bidOverrideNormNumAuction]+cm.stats(list(bidOverrideRatioDistri.values()))
def overlapDistance(post):
check_tokenized(post)
overlaps = []
for i,p in enumerate(post.thread.posts):
if p == post: break
overlaps.append(overlap(post.token_index, p.token_index))
return (i - overlaps.index(max(overlaps))) if overlaps != [] else 0
def overlapPrevious(post):
check_tokenized(post)
overlaps = []
for p in post.thread.posts:
if p == post: break
overlaps.append(overlap(post.token_index, p.token_index))
return max(overlaps) if overlaps != [] else 0.0
def sample_rois(self, all_rois, gt_boxes, gt_labels):
"""
Generate a random sample of RoIs comprising foreground and background
examples.
all_rois is (N, 4)
gt_boxes is (K, 4) with 4 coordinates
gt_labels is in one hot form
"""
num_classes = keras.backend.shape(gt_labels)[1]
gt_labels = keras.backend.argmax(gt_labels, axis=1)
# overlaps: (rois x gt_boxes)
# finds the overlapping regions between the predicted regions of interests and ground truth bounding boxes
overlaps = common.overlap(all_rois, gt_boxes)
gt_assignment = keras.backend.argmax(overlaps, axis=1)
max_overlaps = keras.backend.max(overlaps, axis=1)
# finds the ground truth labels corresponding to the ground truth boxes with greatest overlap for each predicted regions of interest
labels = keras.backend.gather(gt_labels, gt_assignment)
# Select RoIs with given parameters for fg/bg objects
keep_inds = self.get_fg_bg_rois(max_overlaps)
# Select sampled values from various arrays:
labels = keras.backend.gather(labels, keep_inds)
rois = keras.backend.gather(all_rois, keep_inds)
labels = self.set_label_background(labels)
# Compute bounding-box regression targets for an image.
gt_boxes = keras.backend.gather(
gt_boxes, keras.backend.gather(gt_assignment, keep_inds))
bbox_targets = self.get_bbox_targets(rois, gt_boxes, labels,
num_classes)
labels = keras.backend.one_hot(labels, num_classes)
return rois, labels, bbox_targets
