Recall decreases when training with a large rotation angle

[cxmscb]

Thanks for the sharing. I would like to train a kp2d model with a large rotation angle, e.g., pi.
However, I find the recall decreases in such training setting, although the total loss also decreases.

Train [ E 0, T 0, R 0.0839, R_Avg 0.0839, L 4.6522, L_Avg 4.6522]
Train [ E 0, T 0, R 0.0287, R_Avg 0.0791, L 1.9301, L_Avg 2.5305]
Train [ E 0, T 0, R 0.0056, R_Avg 0.0472, L 1.4714, L_Avg 2.0956]
Train [ E 0, T 0, R 0.0037, R_Avg 0.0333, L 1.3760, L_Avg 1.8747]
Train [ E 0, T 0, R 0.0040, R_Avg 0.0260, L 1.3219, L_Avg 1.7457]