mxnet fine-tune training with Memory optimization, which can adjust batch_size to 1000 while training on 8GPUs!