Fix ruotianluo#26.

The code was using the learning_rate from optimizer.pth after starting self critical training.

train.py

@@ -89,9 +89,9 @@ def train(opt):
                 frac = (epoch - opt.learning_rate_decay_start) // opt.learning_rate_decay_every
                 decay_factor = opt.learning_rate_decay_rate  ** frac
                 opt.current_lr = opt.learning_rate * decay_factor
-                utils.set_lr(optimizer, opt.current_lr) # set the decayed rate
             else:
                 opt.current_lr = opt.learning_rate
+            utils.set_lr(optimizer, opt.current_lr)
             # Assign the scheduled sampling prob
             if epoch > opt.scheduled_sampling_start and opt.scheduled_sampling_start >= 0:
                 frac = (epoch - opt.scheduled_sampling_start) // opt.scheduled_sampling_increase_every