def loss(logit_fh, logit_line, gt_fh, gt_line, user_stroke, scope='loss'):
with tf.name_scope(scope) as _:
# stitching face heat map (l2)
gt_fh = slice_tensor(gt_fh, logit_fh)
fh_loss = tf.losses.mean_squared_error(gt_fh, logit_fh)
real_fh_loss = tf.losses.absolute_difference(gt_fh, logit_fh)
# regression loss (l2)
img_shape = tf.shape(logit_fh)
N = img_shape[0]
H = img_shape[1]
W = img_shape[2]
C = img_shape[3]
full_one = tf.ones([N, H, W, C], tf.float32)
stroke_mask = full_one - user_stroke
logit_line = logit_line * stroke_mask
line_loss = tf.losses.mean_squared_error(gt_line,
logit_line,
weights=stroke_mask)
real_line_loss = tf.losses.absolute_difference(gt_line,
logit_line,
weights=stroke_mask)
# total loss
total_loss = fh_loss + line_loss
return total_loss, fh_loss, real_fh_loss, line_loss, real_line_loss
def loss(logit_fh, logit_curve, fh, curve, user_stroke, scope='loss'):
with tf.name_scope(scope) as _:
# stitching face heat map (l2)
gt_fh = slice_tensor(fh, logit_fh)
fh_loss = tf.losses.mean_squared_error(gt_fh, logit_fh)
real_fh_loss = tf.losses.absolute_difference(gt_fh, logit_fh)
# curve loss (l2)
curve_mask = slice_tensor(curve, logit_curve)
img_shape = tf.shape(logit_fh)
N = img_shape[0]
H = img_shape[1]
W = img_shape[2]
C = img_shape[3]
full_one = tf.ones([N, H, W, C], tf.float32)
full_zero = tf.zeros([N, H, W, C], tf.float32)
stroke_mask = full_one - user_stroke
diff_mask = stroke_mask - curve_mask
fg_pred = logit_curve * curve_mask
bg_pred = logit_curve * diff_mask
curve_fg_sum = tf.reduce_sum(tf.pow(curve_mask - fg_pred, 2),
axis=[1, 2, 3]) # offset curve loss sum
curve_bg_sum = tf.reduce_sum(tf.pow(full_zero - bg_pred, 2),
axis=[1, 2,
3]) # orthogonal curve loss sum
nb_stroke_pixels = tf.reduce_sum(stroke_mask,
axis=[1, 2,
3]) # number of stroke pixels
curve_loss = (curve_fg_sum +
curve_bg_sum) / nb_stroke_pixels # mean loss
c_loss = tf.reduce_mean(curve_loss)
real_curve_fg_sum = tf.reduce_sum(tf.abs(curve_mask - fg_pred),
axis=[1, 2, 3])
real_curve_bg_sum = tf.reduce_sum(tf.abs(full_zero - bg_pred),
axis=[1, 2, 3])
real_curve_loss = (real_curve_fg_sum +
real_curve_bg_sum) / nb_stroke_pixels
real_c_loss = tf.reduce_mean(real_curve_loss)
total_loss = fh_loss + c_loss
return total_loss, fh_loss, c_loss, real_fh_loss, real_c_loss