Possible leakage: Need to clip (theta + margin)?

[tamamachi]

keras-arcface/metrics.py

Line 36 in 877ea00

target_logits = tf.cos(theta + self.m)

at this point the range of theta is 0 <= theta <= pi.
theta + m can be larger than pi, and in that case adding m results in negative penalty, leading to data leakage.
I think we shoud clip theta + m like

target_logits = tf.cos(K.clip(theta + self.m, 0, np.pi))