Did you use FP16 to train your model? And was the performance in the paper trained by FP32?
Did you use FP16 to train your model? And was the performance in the paper trained by FP32?