def _frac_loss(truth, pred, sort_pred):
ldict = create_loss_dict(truth, pred)
t_sigfrac = ldict['t_sigfrac']
r_energy = ldict['r_energy']
t_sigfrac = ldict['t_sigfrac']
p_sigfrac = ldict['p_sigfrac']
t_sumenergy = ldict['t_sumenergy']
t_sigfrac = sortFractions(t_sigfrac,
tf.expand_dims(r_energy,axis=2),
tf.expand_dims(ldict['r_eta'],axis=2))
if sort_pred:
p_sigfrac = sortFractions(p_sigfrac,
tf.expand_dims(r_energy,axis=2),
tf.expand_dims(ldict['r_eta'],axis=2))
r_energy = energy_weighting(r_energy, True)
t_sumenergy = tf.sqrt(ldict['t_sumenergy']+K.epsilon())
diffsq = tf.expand_dims(r_energy,axis=2)*(t_sigfrac - p_sigfrac)**2 #B x V x F
loss = tf.reduce_mean(tf.reduce_mean(diffsq,axis=1), axis=1) # / (ldict['t_n_rechits']+K.epsilon())
loss = tf.where(ldict['t_n_rechits']<1., tf.zeros_like(loss),loss)
return loss
def good_frac_sum_loss(truth, pred):
ldict = create_loss_dict(truth, pred)
t_sigfrac = ldict['t_sigfrac']
p_sigfrac = ldict['p_sigfrac']
diffsq = energy_weighting(ldict['r_energy'], True) * (tf.reduce_sum(t_sigfrac, axis=-1) - tf.reduce_sum(p_sigfrac, axis=-1))**2
diffsq = tf.reduce_mean(diffsq, axis=1)
return diffsq
def good_frac_range_loss(truth, pred):
ldict = create_loss_dict(truth, pred)
t_sigfrac = ldict['t_sigfrac']
p_sigfrac = ldict['p_sigfrac']
rangeloss = tf.where(p_sigfrac < -0.5, (p_sigfrac + 0.5)**2, tf.zeros_like(p_sigfrac))
rangeloss += tf.where(p_sigfrac > 1.5, (p_sigfrac - 1.5)**2, tf.zeros_like(p_sigfrac))
rangeloss = tf.reduce_mean(tf.reduce_mean(rangeloss, axis=2), axis=1)
return rangeloss
def n_shower_loss(truth, pred):
ldict = create_loss_dict(truth, pred)
loss = (ldict['t_showers']-ldict['r_showers'])**2
return loss