def compute(self, y_true, y_pred):
# y.shape (batches, priors, 4 x segment_offset + n x class_label)
# TODO: negatives_for_hard?
batch_size = tf.shape(y_true)[0]
num_priors = tf.shape(y_true)[1]
num_classes = tf.shape(y_true)[2] - 4
eps = K.epsilon()
# confidence loss
conf_true = tf.reshape(y_true[:,:,4:], [-1, num_classes])
conf_pred = tf.reshape(y_pred[:,:,4:], [-1, num_classes])
conf_loss = softmax_loss(conf_true, conf_pred)
class_true = tf.argmax(conf_true, axis=1)
class_pred = tf.argmax(conf_pred, axis=1)
conf = tf.reduce_max(conf_pred, axis=1)
neg_mask_float = conf_true[:,0]
neg_mask = tf.cast(neg_mask_float, tf.bool)
pos_mask = tf.logical_not(neg_mask)
pos_mask_float = tf.cast(pos_mask, tf.float32)
num_total = tf.cast(tf.shape(conf_true)[0], tf.float32)
num_pos = tf.reduce_sum(pos_mask_float)
num_neg = num_total - num_pos
pos_conf_loss = tf.reduce_sum(conf_loss * pos_mask_float)
pos_conf_loss = pos_conf_loss / (num_pos + eps)
## take only false positives for hard negative mining
#false_pos_mask = tf.logical_and(neg_mask, tf.not_equal(class_pred, 0))
#num_false_pos = tf.reduce_sum(tf.cast(false_pos_mask, tf.float32))
#num_neg = tf.minimum(self.neg_pos_ratio * num_pos, num_false_pos)
#neg_conf_loss = tf.boolean_mask(conf_loss, false_pos_mask)
num_neg = tf.minimum(self.neg_pos_ratio * num_pos, num_neg)
neg_conf_loss = tf.boolean_mask(conf_loss, neg_mask)
neg_conf_loss = neg_conf_loss / (num_neg + eps)
vals, idxs = tf.nn.top_k(neg_conf_loss, k=tf.cast(num_neg, tf.int32))
#neg_conf_loss = tf.reduce_sum(tf.gather(neg_conf_loss, idxs))
neg_conf_loss = tf.reduce_sum(vals)
conf_loss = pos_conf_loss + neg_conf_loss
# offset loss
loc_true = tf.reshape(y_true[:,:,0:4], [-1, 4])
loc_pred = tf.reshape(y_pred[:,:,0:4], [-1, 4])
loc_loss = smooth_l1_loss(loc_true, loc_pred)
pos_loc_loss = tf.reduce_sum(loc_loss * pos_mask_float) # only for positives
loc_loss = pos_loc_loss / (num_pos + eps)
# total loss
loss = conf_loss + self.alpha * loc_loss
# metrics
precision, recall, accuracy, fmeasure = compute_metrics(class_true, class_pred, conf, top_k=100*batch_size)
return eval('{'+' '.join(['"'+n+'": '+n+',' for n in self.metric_names])+'}')
def compute(self, y_true, y_pred):
# y.shape (batches, segments, 2 x segment_label + 5 x segment_offset + 16 x inter_layer_links_label + 8 x cross_layer_links_label)
batch_size = tf.shape(y_true)[0]
eps = K.epsilon()
# segment confidence loss
seg_conf_true = tf.reshape(y_true[:, :, 0:2], [-1, 2])
seg_conf_pred = tf.reshape(y_pred[:, :, 0:2], [-1, 2])
seg_conf_loss = softmax_loss(seg_conf_true, seg_conf_pred)
seg_class_pred = tf.argmax(seg_conf_pred, axis=1)
neg_seg_mask_float = seg_conf_true[:, 0]
neg_seg_mask = tf.cast(neg_seg_mask_float, tf.bool)
pos_seg_mask = tf.logical_not(neg_seg_mask)
pos_seg_mask_float = tf.cast(pos_seg_mask, tf.float32)
num_seg = tf.cast(tf.shape(seg_conf_true)[0], tf.float32)
num_pos_seg = tf.reduce_sum(pos_seg_mask_float)
num_neg_seg = num_seg - num_pos_seg
pos_seg_conf_loss = tf.reduce_sum(seg_conf_loss * pos_seg_mask_float)
#false_pos_seg_mask = tf.logical_and(neg_seg_mask, tf.not_equal(seg_class_pred, 0))
#num_false_pos_seg = tf.reduce_sum(tf.cast(false_pos_seg_mask, tf.float32))
#num_neg_seg = tf.minimum(self.neg_pos_ratio * num_pos_seg, num_false_pos_seg)
#neg_seg_conf_loss = tf.boolean_mask(seg_conf_loss, false_pos_seg_mask)
num_neg_seg = tf.minimum(self.neg_pos_ratio * num_pos_seg, num_neg_seg)
neg_seg_conf_loss = tf.boolean_mask(seg_conf_loss, neg_seg_mask)
vals, idxs = tf.nn.top_k(neg_seg_conf_loss,
k=tf.cast(num_neg_seg, tf.int32))
neg_seg_conf_loss = tf.reduce_sum(vals)
pos_seg_conf_loss = pos_seg_conf_loss / (num_pos_seg + eps)
neg_seg_conf_loss = neg_seg_conf_loss / (num_neg_seg + eps)
seg_conf_loss = pos_seg_conf_loss + neg_seg_conf_loss
seg_conf_loss = self.lambda_segments * seg_conf_loss
# segment offset loss
seg_loc_true = tf.reshape(y_true[:, :, 2:7], [-1, 5])
seg_loc_pred = tf.reshape(y_pred[:, :, 2:7], [-1, 5])
seg_loc_loss = smooth_l1_loss(seg_loc_true, seg_loc_pred)
pos_seg_loc_loss = tf.reduce_sum(seg_loc_loss * pos_seg_mask_float)
seg_loc_loss = pos_seg_loc_loss / (num_pos_seg + eps)
seg_loc_loss = self.lambda_offsets * seg_loc_loss
# link confidence loss
inter_link_conf_true = y_true[:, :, 7:23]
cross_link_conf_true = y_true[:, self.first_map_offset:, 23:31]
link_conf_true = tf.concat([
tf.reshape(inter_link_conf_true, [-1, 2]),
tf.reshape(cross_link_conf_true, [-1, 2])
], 0)
inter_link_conf_pred = y_pred[:, :, 7:23]
cross_link_conf_pred = y_pred[:, self.first_map_offset:, 23:31]
link_conf_pred = tf.concat([
tf.reshape(inter_link_conf_pred, [-1, 2]),
tf.reshape(cross_link_conf_pred, [-1, 2])
], 0)
link_conf_loss = softmax_loss(link_conf_true, link_conf_pred)
link_class_pred = tf.argmax(link_conf_pred, axis=1)
neg_link_mask_float = link_conf_true[:, 0]
neg_link_mask = tf.cast(neg_link_mask_float, tf.bool)
pos_link_mask = tf.logical_not(neg_link_mask)
pos_link_mask_float = tf.cast(pos_link_mask, tf.float32)
num_link = tf.cast(tf.shape(link_conf_true)[0], tf.float32)
num_pos_link = tf.reduce_sum(pos_link_mask_float)
num_neg_link = num_link - num_pos_link
pos_link_conf_loss = tf.reduce_sum(link_conf_loss *
pos_link_mask_float)
#false_pos_link_mask = tf.logical_and(neg_link_mask, tf.not_equal(link_class_pred, 0))
#num_false_pos_link = tf.reduce_sum(tf.cast(false_pos_link_mask, tf.float32))
#num_neg_link = tf.minimum(self.neg_pos_ratio * num_pos_link, num_false_pos_link)
#neg_link_conf_loss = tf.boolean_mask(link_conf_loss, false_pos_link_mask)
num_neg_link = tf.minimum(self.neg_pos_ratio * num_pos_link,
num_neg_link)
neg_link_conf_loss = tf.boolean_mask(link_conf_loss, neg_link_mask)
vals, idxs = tf.nn.top_k(neg_link_conf_loss,
k=tf.cast(num_neg_link, tf.int32))
neg_link_conf_loss = tf.reduce_sum(vals)
pos_link_conf_loss = pos_link_conf_loss / (num_pos_link + eps)
neg_link_conf_loss = neg_link_conf_loss / (num_neg_link + eps)
link_conf_loss = pos_link_conf_loss + neg_link_conf_loss
link_conf_loss = self.lambda_links * link_conf_loss
# total loss
total_loss = seg_conf_loss + seg_loc_loss + link_conf_loss
seg_conf = tf.reduce_max(seg_conf_pred, axis=1)
seg_class_true = tf.argmax(seg_conf_true, axis=1)
seg_class_pred = tf.argmax(seg_conf_pred, axis=1)
seg_precision, seg_recall, seg_accuracy, seg_fmeasure = compute_metrics(
seg_class_true, seg_class_pred, seg_conf, top_k=100 * batch_size)
link_conf = tf.reduce_max(link_conf_pred, axis=1)
link_class_true = tf.argmax(link_conf_true, axis=1)
link_class_pred = tf.argmax(link_conf_pred, axis=1)
link_precision, link_recall, link_accuracy, link_fmeasure = compute_metrics(
link_class_true,
link_class_pred,
link_conf,
top_k=100 * batch_size)
# metrics
def make_fcn(t):
return lambda y_true, y_pred: t
for name in [
'seg_conf_loss',
'seg_loc_loss',
'link_conf_loss',
'num_pos_seg',
'num_neg_seg',
'pos_seg_conf_loss',
'neg_seg_conf_loss',
'pos_link_conf_loss',
'neg_link_conf_loss',
'seg_precision',
'seg_recall',
'seg_accuracy',
'seg_fmeasure',
'link_precision',
'link_recall',
'link_accuracy',
'link_fmeasure',
]:
f = make_fcn(eval(name))
f.__name__ = name
self.metrics.append(f)
return total_loss
def compute(self, y_true, y_pred):
# y.shape (batches, priors, 4 x segment_offset + n x class_label)
# TODO: negatives_for_hard?
# mask based on y_true or y_pred?
batch_size = tf.shape(y_true)[0]
num_priors = tf.shape(y_true)[1]
num_classes = tf.shape(y_true)[2] - 4
eps = K.epsilon()
# confidence loss
conf_true = tf.reshape(y_true[:, :, 4:], [-1, num_classes])
conf_pred = tf.reshape(y_pred[:, :, 4:], [-1, num_classes])
conf_loss = softmax_loss(conf_true, conf_pred)
class_true = tf.argmax(conf_true, axis=1)
class_pred = tf.argmax(conf_pred, axis=1)
conf = tf.reduce_max(conf_pred, axis=1)
neg_mask_float = conf_true[:, 0]
neg_mask = tf.cast(neg_mask_float, tf.bool)
pos_mask = tf.logical_not(neg_mask)
pos_mask_float = tf.cast(pos_mask, tf.float32)
num_total = tf.cast(tf.shape(conf_true)[0], tf.float32)
num_pos = tf.reduce_sum(pos_mask_float)
num_neg = num_total - num_pos
pos_conf_loss = tf.reduce_sum(conf_loss * pos_mask_float)
## take only false positives for hard negative mining
#false_pos_mask = tf.logical_and(neg_mask, tf.not_equal(class_pred, 0))
#num_false_pos = tf.reduce_sum(tf.cast(false_pos_mask, tf.float32))
#num_neg = tf.minimum(self.neg_pos_ratio * num_pos, num_false_pos)
#neg_conf_loss = tf.boolean_mask(conf_loss, false_pos_mask)
num_neg = tf.minimum(self.neg_pos_ratio * num_pos, num_neg)
neg_conf_loss = tf.boolean_mask(conf_loss, neg_mask)
vals, idxs = tf.nn.top_k(neg_conf_loss, k=tf.cast(num_neg, tf.int32))
#neg_conf_loss = tf.reduce_sum(tf.gather(neg_conf_loss, idxs))
neg_conf_loss = tf.reduce_sum(vals)
conf_loss = (pos_conf_loss + neg_conf_loss) / (num_pos + num_neg + eps)
# offset loss
loc_true = tf.reshape(y_true[:, :, 0:4], [-1, 4])
loc_pred = tf.reshape(y_pred[:, :, 0:4], [-1, 4])
loc_loss = smooth_l1_loss(loc_true, loc_pred)
pos_loc_loss = tf.reduce_sum(loc_loss *
pos_mask_float) # only for positives
loc_loss = pos_loc_loss / (num_pos + eps)
# total loss
total_loss = conf_loss + self.alpha * loc_loss
# metrics
pos_conf_loss = pos_conf_loss / (num_pos + eps)
neg_conf_loss = neg_conf_loss / (num_neg + eps)
pos_loc_loss = pos_loc_loss / (num_pos + eps)
precision, recall, accuracy, fmeasure = compute_metrics(class_true,
class_pred,
conf,
top_k=100 *
batch_size)
def make_fcn(t):
return lambda y_true, y_pred: t
for name in [
'num_pos',
'num_neg',
'pos_conf_loss',
'neg_conf_loss',
'pos_loc_loss',
'precision',
'recall',
'accuracy',
'fmeasure',
]:
f = make_fcn(eval(name))
f.__name__ = name
self.metrics.append(f)
return total_loss