def cost_grad(self, final_acts, y): exp_act = safe_exp(final_acts) sum_exp = np.sum(exp_act, axis=1) return exp_act / sum_exp[:, np.newaxis] - y
def _cost(self, final_acts, y): exp_act = safe_exp(final_acts) lse_act = safe_log(np.sum(exp_act, axis=1)) return -np.mean( np.sum((y * (final_acts - lse_act[:, np.newaxis])), axis=1))
def _cost(self, final_acts, y): exp_act = safe_exp(final_acts) lse_act = safe_log(np.sum(exp_act, axis=1)) return -np.mean(np.sum((y * (final_acts - lse_act[:, np.newaxis])), axis=1))