Update fp16 bsupdate fp16 bs (PaddlePaddle#728)

* update_fp16_bs, test=document_fix

* update_fp16_bs, test=document_fix

Author: hysunflower

comparision_system/mxnet_benchmark/mxnet_auto.sh

@@ -7,7 +7,7 @@
 #                                        #
 ##########################################
 
-cur_model_list=(yolov3 ResNet50_bs128_fp16)
+cur_model_list=(yolov3 ResNet50_bs208_fp16)
 
 environment(){
 export LD_LIBRARY_PATH=/home/work/418.39/lib64/:/usr/local/cuda-10.0/compat/:$LD_LIBRARY_PATH
@@ -74,7 +74,7 @@ CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 bash run_yolov3.sh train 1 ${LOG_DIR} > ${R
 
 }
 
-ResNet50_bs128_fp16(){
+ResNet50_bs208_fp16(){
 curl_model_path=${MODEL_PATH}
 cd ${curl_model_path}
 
@@ -85,10 +85,10 @@ cp ${BENCHMARK_ROOT}/static_graph/image_classification/resnet50_fp/mxnet/run_ben
 mkdir data
 ln -s ${data_path}/fusefp16_resnet50_mxnet/ILSVRC2012_mxnet_rec/ ./data
 echo "----run one card----"
-CUDA_VISIBLE_DEVICES=0 bash run_benchmark.sh  1 sp 128 1 ${LOG_DIR} > ${RES_DIR}/resnet50_fp16_1.res 2>&1 
+CUDA_VISIBLE_DEVICES=0 bash run_benchmark.sh  1 sp 208 1 ${LOG_DIR} > ${RES_DIR}/resnet50_fp16_1.res 2>&1 
 sleep 60
 echo "----run 8 card----"
-CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 bash run_benchmark.sh 1 sp 128 1 ${LOG_DIR}  > ${RES_DIR}/resnet50_fp16_8.res 2>&1 
+CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 bash run_benchmark.sh 1 sp 208 1 ${LOG_DIR}  > ${RES_DIR}/resnet50_fp16_8.res 2>&1 
 
 }
 

scripts/static_graph_models.sh

@@ -323,15 +323,15 @@ image_classification(){
     # running model cases with amp
     cp ${BENCHMARK_ROOT}/static_graph/image_classification/resnet50_fp/paddle/run_benchmark.sh ./fp16_run_benchmark.sh
     sed -i '/set\ -xe/d' fp16_run_benchmark.sh
-    model_name="ResNet50_bs128_fp16"
+    model_name="ResNet50_bs208_fp16"
     echo "index is speed, 1gpus, run_mode is sp, begin"
-    CUDA_VISIBLE_DEVICES=0 bash fp16_run_benchmark.sh 1 128 sp 800 | tee ${log_path}/${model_name}_speed_1gpus 2>&1
+    CUDA_VISIBLE_DEVICES=0 bash fp16_run_benchmark.sh 1 208 sp 800 | tee ${log_path}/${model_name}_speed_1gpus 2>&1
     sleep 60
     echo "index is speed, 8gpus, run_mode is sp, begin"
-    CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 bash fp16_run_benchmark.sh 1 128 sp 500 | tee ${log_path}/${model_name}_speed_8gpus 2>&1
+    CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 bash fp16_run_benchmark.sh 1 208 sp 500 | tee ${log_path}/${model_name}_speed_8gpus 2>&1
     sleep 60
     echo "index is speed, 8gpus, run_mode is multi_process, begin"
-    CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 bash fp16_run_benchmark.sh 1 128 mp 1000  | tee ${log_path}/${model_name}_speed_8gpus8p 2>&1
+    CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 bash fp16_run_benchmark.sh 1 208 mp 1000  | tee ${log_path}/${model_name}_speed_8gpus8p 2>&1
 }
 
 

static_graph/image_classification/resnet50_fp/paddle/run_benchmark.sh

@@ -11,7 +11,7 @@ fi
 function _set_params(){
     index=$1                         # 速度(speed)|显存占用(mem)|单卡最大支持batch_size(maxbs)                       (必填)
     base_batch_size=$2               # 单卡的batch_size，如果固定的，可以写死。                                       (必填）
-    model_name="ResNet50_bs128_fp16" # 模型名字如："SE-ResNeXt50"，如果是固定的，可以写死，如果需要其他参数可以参考bert实现（必填）
+    model_name="ResNet50_bs208_fp16" # 模型名字如："SE-ResNeXt50"，如果是固定的，可以写死，如果需要其他参数可以参考bert实现（必填）
     run_mode=${3:-"sp"}              # 单进程(sp)|多进程(mp)，默认单进程                                            （必填）
     mission_name="图像分类"          # 模型所属任务名称，具体可参考scripts/config.ini                                （必填）
     direction_id=0                   # 任务所属方向，0：CV，1：NLP，2：Rec。                                         (必填)