export CUDA_VISIBLE_DEVICES=0,1,2,3 python -m paddle.distributed.launch \ --selected_gpus=0,1,2,3 \ --log_dir ./mylog \ train.py --use_data_parallel True
