From 7a33fea97a890c03f7c9c37d0f92564833641c93 Mon Sep 17 00:00:00 2001
From: 18118216380 <3224925783@qq.com>
Date: Mon, 25 Apr 2022 18:40:48 +0800
Subject: [PATCH 1/6] 4.25

---
 .../run_squad.py                              |   6 +-
 .../test/train_performance_16p.sh             | 226 ++++++++++++++++++
 .../test/train_performance_16p.sh             | 225 +++++++++++++++++
 3 files changed, 456 insertions(+), 1 deletion(-)
 create mode 100644 PyTorch/built-in/nlp/Bert-Squad_ID0470_for_PyTorch/test/train_performance_16p.sh
 create mode 100644 PyTorch/dev/cv/image_classification/Hourglass_ID1809_for_PyTorch/test/train_performance_16p.sh

diff --git a/PyTorch/built-in/nlp/Bert-Squad_ID0470_for_PyTorch/run_squad.py b/PyTorch/built-in/nlp/Bert-Squad_ID0470_for_PyTorch/run_squad.py
index c245d2cb11..85b3c8280d 100644
--- a/PyTorch/built-in/nlp/Bert-Squad_ID0470_for_PyTorch/run_squad.py
+++ b/PyTorch/built-in/nlp/Bert-Squad_ID0470_for_PyTorch/run_squad.py
@@ -916,7 +916,11 @@ def main():
             os.environ['MASTER_PORT'] = '29668'
             torch.npu.set_device("npu:%d" % args.local_rank)
             device = torch.device("npu:%d" % args.local_rank)
-            torch.distributed.init_process_group(backend='hccl', world_size=8, rank=args.local_rank)
+            if args.num_npu == 8:
+                torch.distributed.init_process_group(backend='hccl', world_size=8, rank=args.local_rank)
+            if args.num_npu == 16:
+                global_rank = int(os.getenv('RANK'))
+                torch.distributed.init_process_group(backend='hccl', world_size=16, rank=global_rank)
             n_npu = 1
         else:
             torch.cuda.set_device(args.local_rank)
diff --git a/PyTorch/built-in/nlp/Bert-Squad_ID0470_for_PyTorch/test/train_performance_16p.sh b/PyTorch/built-in/nlp/Bert-Squad_ID0470_for_PyTorch/test/train_performance_16p.sh
new file mode 100644
index 0000000000..ec2c8f2543
--- /dev/null
+++ b/PyTorch/built-in/nlp/Bert-Squad_ID0470_for_PyTorch/test/train_performance_16p.sh
@@ -0,0 +1,226 @@
+#!/bin/bash
+export PATH=/usr/local/hdf5/bin:$PATH
+export LD_LIBRARY_PATH=/usr/local/hdf5/lib:$LD_LIBRARY_PATH
+export LIBRARY_PATH=/usr/local/hdf5/lib:$LIBRARY_PATH
+export CPATH=/usr/local/hdf5/include:$CPATH
+export HDF5_DISABLE_VERSION_CHECK=1
+#当前路径,不需要修改
+cur_path=`pwd`
+
+#集合通信参数,不需要修改
+export BMMV2_ENABLE=1
+export RANK_SIZE=8
+export JOB_ID=10087
+RANK_ID_START=0
+
+
+# 数据集路径,保持为空,不需要修改
+data_path=""
+ckpt_path=""
+conf_path=""
+server_index=""
+fix_node_ip=""
+devicesnum=""
+
+
+#基础参数，需要模型审视修改
+#网络名称，同目录名称
+Network="Bert-Squad_ID0470_for_PyTorch"
+#训练epoch
+train_epochs=1
+#训练batch_size
+batch_size=32
+#训练step
+train_steps=
+#学习率
+learning_rate=2e-4
+
+
+#维测参数，precision_mode需要模型审视修改
+precision_mode="allow_fp32_to_fp16"
+#维持参数，以下不需要修改
+over_dump=False
+data_dump_flag=False
+data_dump_step="10"
+profiling=False
+
+# 帮助信息，不需要修改
+if [[ $1 == --help || $1 == -h ]];then
+    echo"usage:./train_performance_1P.sh <args>"
+    echo " "
+    echo "parameter explain:
+    --precision_mode         precision mode(allow_fp32_to_fp16/force_fp16/must_keep_origin_dtype/allow_mix_precision)
+    --over_dump		           if or not over detection, default is False
+    --data_dump_flag		     data dump flag, default is False
+    --data_dump_step		     data dump step, default is 10
+    --profiling		           if or not profiling for performance debug, default is False
+    --data_path		           source data of training
+    -h/--help		             show help message
+    "
+    exit 1
+fi
+
+#参数校验，不需要修改
+for para in $*
+do
+    if [[ $para == --precision_mode* ]];then
+        precision_mode=`echo ${para#*=}`
+    elif [[ $para == --over_dump* ]];then
+        over_dump=`echo ${para#*=}`
+        over_dump_path=${cur_path}/output/overflow_dump
+        mkdir -p ${over_dump_path}
+    elif [[ $para == --data_dump_flag* ]];then
+        data_dump_flag=`echo ${para#*=}`
+        data_dump_path=${cur_path}/output/data_dump
+        mkdir -p ${data_dump_path}
+    elif [[ $para == --data_dump_step* ]];then
+        data_dump_step=`echo ${para#*=}`
+    elif [[ $para == --profiling* ]];then
+        profiling=`echo ${para#*=}`
+        profiling_dump_path=${cur_path}/output/profiling
+        mkdir -p ${profiling_dump_path}
+    elif [[ $para == --data_path* ]];then
+        data_path=`echo ${para#*=}`
+    elif [[ $para == --ckpt_path* ]];then
+        ckpt_path=`echo ${para#*=}`
+	elif [[ $para == --fix_node_ip* ]];then
+	    fix_node_ip=`echo ${para#*=}`
+	elif [[ $para == --devicesnum* ]];then
+	    devicesnum=`echo ${para#*=}`
+    elif [[ $para == --conf_path* ]];then
+            conf_path=`echo ${para#*=}`
+    elif [[ $para == --server_index* ]];then
+            server_index=`echo ${para#*=}`
+    fi
+done
+
+one_node_ip=`find $conf_path -name "server_*0.info"|awk -F "server_" '{print $2}'|awk -F "_" '{print $1}'`
+linux_num=`find $conf_path -name "server_*.info" |wc -l`
+
+#校验是否传入data_path,不需要修改
+if [[ $data_path == "" ]];then
+    echo "[Error] para \"data_path\" must be confing"
+    exit 1
+fi
+
+#训练开始时间，不需要修改
+
+
+
+export HCCL_IF_IP=$fix_node_ip
+export MASTER_ADDR=$one_node_ip
+export MASTER_PORT=29688
+export HCCL_WHITELIST_DISABLE=1
+device_num=${#devicesnum}
+devices_num=`awk 'BEGIN{printf "%.0f\n",'${device_num}'-1}'`
+
+start_time=$(date +%s)
+NPUS=($(seq 0 $devices_num))
+rank_server=`awk 'BEGIN{printf "%.0f\n",'${device_num}'*'${server_index}'}'`
+export NPU_WORLD_SIZE=`awk 'BEGIN{printf "%.0f\n",'${device_num}'*'${linux_num}'}'`
+
+
+#进入训练脚本目录，需要模型审视修改
+cd $cur_path/../
+rank=0
+for((RANK_ID=$RANK_ID_START;RANK_ID<$((RANK_SIZE+RANK_ID_START));RANK_ID++));
+do
+    #设置环境变量，不需要修改
+    export ASCEND_DEVICE_ID=$RANK_ID
+    echo "Device ID: $ASCEND_DEVICE_ID"
+    export RANK_ID=$RANK_ID
+    export RANK=`awk 'BEGIN{printf "%.0f\n",'${rank}'+'${rank_server}'}'`
+
+
+    #创建DeviceID输出目录，不需要修改
+    if [ -d ${cur_path}/output/${ASCEND_DEVICE_ID} ];then
+        rm -rf ${cur_path}/output/${ASCEND_DEVICE_ID}
+        mkdir -p ${cur_path}/output/$ASCEND_DEVICE_ID/ckpt
+    else
+        mkdir -p ${cur_path}/output/$ASCEND_DEVICE_ID/ckpt
+    fi
+
+    
+    #执行训练脚本，以下传参不需要修改，其他需要模型审视修改
+    #--data_dir, --model_dir, --precision_mode, --over_dump, --over_dump_path，--data_dump_flag，--data_dump_step，--data_dump_path，--profiling，--profiling_dump_path
+    nohup python3.7 run_squad.py \
+          --init_checkpoint ${ckpt_path}/bert_large_pretrained_amp.pt \
+          --bert_model bert-large-uncased \
+		  --do_train \
+		  --train_file ${data_path}/train-v1.1-min.json \
+		  --train_batch_size ${batch_size} \
+		  --do_predict \
+		  --predict_batch_size ${batch_size} \
+		  --predict_file ${data_path}/dev-v1.1.json \
+		  --learning_rate ${learning_rate} \
+		  --num_train_epochs ${train_epochs} \
+		  --seed 1 \
+		  --fp16 \
+		  --max_steps 100 \
+		  --use_npu \
+		  --loss_scale 4096 \
+		  --vocab_file "data/uncased_L-24_H-1024_A-16/vocab.txt" \
+		  --do_eval \
+          --eval_script ${data_path}/evaluate-v1.1.py \
+		  --npu_id ${ASCEND_DEVICE_ID} \
+		  --do_lower_case \
+		  --output_dir ${cur_path}/../results \
+		  --config_file bert_config.json \
+		  --num_npu 16 \
+		  --local_rank=$RANK_ID \
+          --addr $one_node_ip \
+		  --json-summary ${cur_path}/output/${ASCEND_DEVICE_ID}/dllogger.json> ${cur_path}/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log 2>&1 &
+		  let rank++
+done 
+wait
+
+#训练结束时间，不需要修改
+end_time=$(date +%s)
+e2e_time=$(( $end_time - $start_time ))
+
+#结果打印，不需要修改
+echo "------------------ Final result ------------------"
+#输出性能FPS，需要模型审视修改
+step_time=`grep 'step_time : ' $cur_path/output/0/train_0.log| awk '{print$13}'| tail -n +3 |awk '{sum+=$1} END {print"",sum/NR}' | sed s/[[:space:]]//g`
+
+FPS=`awk 'BEGIN{printf "%.2f\n", '$batch_size'/'$step_time'*16}'`
+
+#打印，不需要修改
+echo "Final Performance images/sec : $FPS"
+
+#输出训练精度,需要模型审视修改
+train_accuracy=`grep 'F1 : ' $cur_path/output/0/train_0.log|awk '{print $10}'` 
+#打印，不需要修改
+echo "Final Train Accuracy : ${train_accuracy}"
+echo "E2E Training Duration sec : $e2e_time"
+
+#性能看护结果汇总
+#训练用例信息，不需要修改
+BatchSize=${batch_size}
+DeviceType=`uname -m`
+CaseName=${Network}_bs${BatchSize}_${RANK_SIZE}'p'_'perf'
+
+##获取性能数据，不需要修改
+#吞吐量
+ActualFPS=${FPS}
+#单迭代训练时长
+TrainingTime=`awk 'BEGIN{printf "%.2f\n",'${BatchSize}'*1000/'${FPS}'}'`
+
+#从train_$ASCEND_DEVICE_ID.log提取Loss到train_${CaseName}_loss.txt中，需要根据模型审视
+grep -r "step_loss :" $cur_path/output/0/train_0.log | awk '{print $19}' > $cur_path/output/0/train_${CaseName}_loss.txt
+
+#最后一个迭代loss值，不需要修改
+ActualLoss=`awk 'END {print}' $cur_path/output/0/train_${CaseName}_loss.txt`
+
+#关键信息打印到${CaseName}.log中，不需要修改
+echo "Network = ${Network}" > $cur_path/output/0/${CaseName}.log
+echo "RankSize = ${RANK_SIZE}" >> $cur_path/output/0/${CaseName}.log
+echo "BatchSize = ${BatchSize}" >> $cur_path/output/0/${CaseName}.log
+echo "DeviceType = ${DeviceType}" >> $cur_path/output/0/${CaseName}.log
+echo "CaseName = ${CaseName}" >> $cur_path/output/0/${CaseName}.log
+echo "ActualFPS = ${ActualFPS}" >> $cur_path/output/0/${CaseName}.log
+echo "TrainingTime = ${TrainingTime}" >> $cur_path/output/0/${CaseName}.log
+echo "ActualLoss = ${ActualLoss}" >> $cur_path/output/0/${CaseName}.log
+echo "E2ETrainingTime = ${e2e_time}" >> $cur_path/output/0/${CaseName}.log
+rm -rf ${data_path}/train-v1.1-min.json_bert-large-uncased_384_128_64
+export BMMV2_ENABLE=0
\ No newline at end of file
diff --git a/PyTorch/dev/cv/image_classification/Hourglass_ID1809_for_PyTorch/test/train_performance_16p.sh b/PyTorch/dev/cv/image_classification/Hourglass_ID1809_for_PyTorch/test/train_performance_16p.sh
new file mode 100644
index 0000000000..ccc9ac9928
--- /dev/null
+++ b/PyTorch/dev/cv/image_classification/Hourglass_ID1809_for_PyTorch/test/train_performance_16p.sh
@@ -0,0 +1,225 @@
+#!/bin/bash
+
+#当前路径,不需要修改
+cur_path=`pwd`
+export ASCEND_SLOG_PRINT_TO_STDOUT=0
+#export NPU_CALCULATE_DEVICE=$ASCEND_DEVICE_ID
+#集合通信参数,不需要修改
+
+export RANK_SIZE=8
+export JOB_ID=10087
+RANK_ID_START=0
+
+
+# 数据集路径,保持为空,不需要修改
+data_path=""
+conf_path=""
+server_index=""
+fix_node_ip=""
+devicesnum=""
+
+#基础参数，需要模型审视修改
+#网络名称，同目录名称
+Network="Hourglass_ID1809_for_PyTorch"
+#训练epoch
+train_epochs=1
+#训练batch_size
+batch_size=16
+#训练step
+train_steps=10
+#学习率
+learning_rate=1e-3
+
+#TF2.X独有，不需要修改
+#export NPU_LOOP_SIZE=${train_steps}
+
+#维测参数，precision_mode需要模型审视修改
+precision_mode="allow_mix_precision"
+#维持参数，以下不需要修改
+over_dump=False
+data_dump_flag=False
+data_dump_step="10"
+profiling=False
+autotune=False
+
+# 帮助信息，不需要修改
+if [[ $1 == --help || $1 == -h ]];then
+    echo"usage:./train_full_1p.sh <args>"
+    echo " "
+    echo "parameter explain:
+    --precision_mode         precision mode(allow_fp32_to_fp16/force_fp16/must_keep_origin_dtype/allow_mix_precision)
+    --over_dump		         if or not over detection, default is False
+    --data_dump_flag	     data dump flag, default is False
+    --data_dump_step		 data dump step, default is 10
+    --profiling		         if or not profiling for performance debug, default is False
+    --data_path		         source data of training
+    -h/--help		         show help message
+    "
+    exit 1
+fi
+
+#参数校验，不需要修改
+for para in $*
+do
+    if [[ $para == --precision_mode* ]];then
+        precision_mode=`echo ${para#*=}`
+	elif [[ $para == --over_dump* ]];then
+        over_dump=`echo ${para#*=}`
+        over_dump_path=${cur_path}/output/overflow_dump
+        mkdir -p ${over_dump_path}
+    elif [[ $para == --data_dump_flag* ]];then
+        data_dump_flag=`echo ${para#*=}`
+        data_dump_path=${cur_path}/output/data_dump
+        mkdir -p ${data_dump_path}
+    elif [[ $para == --data_dump_step* ]];then
+        data_dump_step=`echo ${para#*=}`
+    elif [[ $para == --profiling* ]];then
+        profiling=`echo ${para#*=}`
+        profiling_dump_path=${cur_path}/output/profiling
+        mkdir -p ${profiling_dump_path}
+    elif [[ $para == --data_path* ]];then
+        data_path=`echo ${para#*=}`
+    elif [[ $para == --fix_node_ip* ]];then
+	    fix_node_ip=`echo ${para#*=}`
+	elif [[ $para == --devicesnum* ]];then
+	    devicesnum=`echo ${para#*=}`
+    elif [[ $para == --conf_path* ]];then
+            conf_path=`echo ${para#*=}`
+    elif [[ $para == --server_index* ]];then
+            server_index=`echo ${para#*=}`
+    fi
+done
+
+one_node_ip=`find $conf_path -name "server_*0.info"|awk -F "server_" '{print $2}'|awk -F "_" '{print $1}'`
+linux_num=`find $conf_path -name "server_*.info" |wc -l`
+
+#校验是否传入data_path,不需要修改
+if [[ $data_path == "" ]];then
+    echo "[Error] para \"data_path\" must be confing"
+    exit 1
+fi
+
+#训练开始时间，不需要修改
+start_time=$(date +%s)
+
+#进入训练脚本目录，需要模型审视修改
+cd $cur_path/../
+sed -i "s|'batchsize': 16|'batchsize': $batch_size|g" $cur_path/../task/pose.py
+sed -i "s|'learning_rate': 1e-3|'learning_rate': $learning_rate|g" $cur_path/../task/pose.py
+sed -i "s|'epoch_num': 200|'epoch_num': $train_epochs|g" $cur_path/../task/pose.py
+sed -i "s|'train_iters': 1000|'train_iters': $train_steps|g" $cur_path/../task/pose.py
+sed -i "s|annot_dir = 'data/MPII/annot'|annot_dir = '$data_path/data/MPII/annot'|g" $cur_path/../datat/MPII/ref.py
+sed -i "s|img_dir = 'data/MPII/images'|img_dir = '$data_path/data/MPII/images'|g" $cur_path/../datat/MPII/ref.py
+for((RANK_ID=$RANK_ID_START;RANK_ID<$((RANK_SIZE+RANK_ID_START));RANK_ID++));
+do
+    #设置环境变量，不需要修改
+    echo "Device ID: $ASCEND_DEVICE_ID"
+    export RANK_ID=$RANK_ID
+    
+    
+    
+    #创建DeviceID输出目录，不需要修改
+    if [ -d ${cur_path}/output/${ASCEND_DEVICE_ID} ];then
+        rm -rf ${cur_path}/output/${ASCEND_DEVICE_ID}
+        mkdir -p ${cur_path}/output/$ASCEND_DEVICE_ID/ckpt
+    else
+        mkdir -p ${cur_path}/output/$ASCEND_DEVICE_ID/ckpt
+    fi
+    
+    #绑核，不需要绑核的模型删除，需要绑核的模型根据实际修改
+    #cpucount=`lscpu | grep "CPU(s):" | head -n 1 | awk '{print $2}'`
+    #cpustep=`expr $cpucount / 8`
+    #echo "taskset c steps:" $cpustep
+    #let a=RANK_ID*$cpustep
+    #let b=RANK_ID+1
+    #let c=b*$cpustep-1
+	
+
+    #执行训练脚本，以下传参不需要修改，其他需要模型审视修改
+done 
+wait
+
+
+export HCCL_IF_IP=$fix_node_ip
+export MASTER_ADDR=$one_node_ip
+export MASTER_PORT=29688
+export HCCL_WHITELIST_DISABLE=1
+device_num=${#devicesnum}
+devices_num=`awk 'BEGIN{printf "%.0f\n",'${device_num}'-1}'`
+
+start_time=$(date +%s)
+NPUS=($(seq 0 $devices_num))
+rank_server=`awk 'BEGIN{printf "%.0f\n",'${device_num}'*'${server_index}'}'`
+export NPU_WORLD_SIZE=`awk 'BEGIN{printf "%.0f\n",'${device_num}'*'${linux_num}'}'`
+
+rank=0
+for i in ${NPUS[@]}
+do
+    mkdir -p  $cur_path/output/${i}/
+    export NPU_CALCULATE_DEVICE=${i}
+    export ASCEND_DEVICE_ID=${i}
+    export RANK=`awk 'BEGIN{printf "%.0f\n",'${rank}'+'${rank_server}'}'`
+    echo run process ${rank}
+
+   nohup  python3 train.py -e test_run_001   --ddp True > $cur_path/output/${i}/train_${i}.log 2>&1 &
+    let rank++
+done
+wait
+
+
+sed -i "s|'batchsize': $batch_size|'batchsize': 16|g" $cur_path/../task/pose.py
+sed -i "s|'learning_rate': $learning_rate|'learning_rate': 1e-3|g" $cur_path/../task/pose.py
+sed -i "s|'epoch_num': $train_epochs|'epoch_num': 200|g" $cur_path/../task/pose.py
+sed -i "s|'train_iters': $train_steps|'train_iters': 1000|g" $cur_path/../task/pose.py
+sed -i "s|annot_dir = '$data_path/data/MPII/annot'|annot_dir = 'data/MPII/annot'|g" $cur_path/../datat/MPII/ref.py
+sed -i "s|img_dir = '$data_path/data/MPII/images'|img_dir = 'data/MPII/images'|g" $cur_path/../datat/MPII/ref.py
+#conda deactivate
+#训练结束时间，不需要修改
+end_time=$(date +%s)
+e2e_time=$(( $end_time - $start_time ))
+
+sed -i "s|\r|\n|g" ${cur_path}/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log
+#结果打印，不需要修改
+echo "------------------ Final result ------------------"
+#输出性能FPS，需要模型审视修改
+FPS=`grep "fps:"  $cur_path/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log|awk -F "fps: " '{print $2}'|awk '{sum+=$1} END {print"",sum/NR}'|sed s/[[:space:]]//g`
+FPS=$(awk 'BEGIN{print '$FPS'*16}') 
+
+#打印，不需要修改
+echo "Final Performance images/sec : $FPS"
+
+#输出训练精度,需要模型审视修改
+#train_accuracy=`grep eval_accuracy $cur_path/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log|grep -v mlp_log|awk 'END {print $5}'| sed 's/,//g' |cut -c 1-5`
+#打印，不需要修改
+#echo "Final Train Accuracy : ${train_accuracy}"
+echo "E2E Training Duration sec : $e2e_time"
+
+#稳定性精度看护结果汇总
+#训练用例信息，不需要修改
+BatchSize=${batch_size}
+DeviceType=`uname -m`
+CaseName=${Network}_bs${BatchSize}_${RANK_SIZE}'p'_'perf'
+
+##获取性能数据
+#吞吐量，不需要修改
+ActualFPS=${FPS}
+#单迭代训练时长，不需要修改
+TrainingTime=`awk 'BEGIN{printf "%.2f\n",'${BatchSize}'*1000/'${FPS}'}'`
+
+#从train_$ASCEND_DEVICE_ID.log提取Loss到train_${CaseName}_loss.txt中，需要根据模型审视
+grep "the loss is: " $cur_path/output/$ASCEND_DEVICE_ID/train_$ASCEND_DEVICE_ID.log|awk -F "the loss is: " '{print $2}'|sed s/[[:space:]]//g > $cur_path/output/$ASCEND_DEVICE_ID/train_${CaseName}_loss.txt
+
+#最后一个迭代loss值，不需要修改
+ActualLoss=`awk 'END {print}' $cur_path/output/$ASCEND_DEVICE_ID/train_${CaseName}_loss.txt`
+
+#关键信息打印到${CaseName}.log中，不需要修改
+echo "Network = ${Network}" > $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
+echo "RankSize = ${RANK_SIZE}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
+echo "BatchSize = ${BatchSize}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
+echo "DeviceType = ${DeviceType}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
+echo "CaseName = ${CaseName}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
+echo "ActualFPS = ${ActualFPS}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
+echo "TrainingTime = ${TrainingTime}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
+#echo "TrainAccuracy = ${train_accuracy}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
+echo "ActualLoss = ${ActualLoss}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
+echo "E2ETrainingTime = ${e2e_time}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
-- 
Gitee


From cbc3d60000a23779e995be0098b0f70f504b4182 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E7=8E=8B=E5=9B=BD=E5=BA=86?= <3224925783@qq.com>
Date: Mon, 25 Apr 2022 10:48:01 +0000
Subject: [PATCH 2/6] add
 PyTorch/dev/cv/image_classification/ResNet152_ID0424_for_PyTorch/test/train_performance_16p.sh.

---
 .../test/train_performance_16p.sh             | 213 ++++++++++++++++++
 1 file changed, 213 insertions(+)
 create mode 100644 PyTorch/dev/cv/image_classification/ResNet152_ID0424_for_PyTorch/test/train_performance_16p.sh

diff --git a/PyTorch/dev/cv/image_classification/ResNet152_ID0424_for_PyTorch/test/train_performance_16p.sh b/PyTorch/dev/cv/image_classification/ResNet152_ID0424_for_PyTorch/test/train_performance_16p.sh
new file mode 100644
index 0000000000..ad5b68f395
--- /dev/null
+++ b/PyTorch/dev/cv/image_classification/ResNet152_ID0424_for_PyTorch/test/train_performance_16p.sh
@@ -0,0 +1,213 @@
+#!/bin/bash
+
+#当前路径,不需要修改
+cur_path=`pwd`
+#source env.sh
+#集合通信参数,不需要修改
+
+export RANK_SIZE=8
+export JOB_ID=10087
+RANK_ID_START=0
+
+
+# 数据集路径,保持为空,不需要修改
+data_path=""
+conf_path=""
+server_index=""
+fix_node_ip=""
+devicesnum=""
+
+#设置默认日志级别,不需要修改
+export ASCEND_GLOBAL_LOG_LEVEL=3
+
+#基础参数，需要模型审视修改
+#网络名称，同目录名称
+Network="ResNet152_ID0424_for_PyTorch"
+#训练epoch
+train_epochs=1
+#训练batch_size
+batch_size=4096
+#训练step
+#train_steps=`expr 1281167 / ${batch_size}`
+#学习率
+learning_rate=0.5
+
+#TF2.X独有，需要模型审视修改
+#export NPU_LOOP_SIZE=${train_steps}
+
+#维测参数，precision_mode需要模型审视修改
+#precision_mode="allow_mix_precision"
+#维持参数，以下不需要修改
+over_dump=False
+data_dump_flag=False
+data_dump_step="10"
+profiling=False
+
+# 帮助信息，不需要修改
+if [[ $1 == --help || $1 == -h ]];then
+    echo"usage:./train_performance_1P.sh <args>"
+    echo " "
+    echo "parameter explain:
+    --precision_mode         precision mode(allow_fp32_to_fp16/force_fp16/must_keep_origin_dtype/allow_mix_precision)
+    --over_dump		           if or not over detection, default is False
+    --data_dump_flag		     data dump flag, default is False
+    --data_dump_step		     data dump step, default is 10
+    --profiling		           if or not profiling for performance debug, default is False
+    --data_path		           source data of training
+    -h/--help		             show help message
+    "
+    exit 1
+fi
+
+#参数校验，不需要修改
+for para in $*
+do
+    if [[ $para == --precision_mode* ]];then
+        precision_mode=`echo ${para#*=}`
+    elif [[ $para == --over_dump* ]];then
+        over_dump=`echo ${para#*=}`
+        over_dump_path=${cur_path}/output/overflow_dump
+        mkdir -p ${over_dump_path}
+    elif [[ $para == --data_dump_flag* ]];then
+        data_dump_flag=`echo ${para#*=}`
+        data_dump_path=${cur_path}/output/data_dump
+        mkdir -p ${data_dump_path}
+    elif [[ $para == --data_dump_step* ]];then
+        data_dump_step=`echo ${para#*=}`
+    elif [[ $para == --profiling* ]];then
+        profiling=`echo ${para#*=}`
+        profiling_dump_path=${cur_path}/output/profiling
+        mkdir -p ${profiling_dump_path}
+    elif [[ $para == --data_path* ]];then
+        data_path=`echo ${para#*=}`
+    elif [[ $para == --fix_node_ip* ]];then
+	    fix_node_ip=`echo ${para#*=}`
+	elif [[ $para == --devicesnum* ]];then
+	    devicesnum=`echo ${para#*=}`
+    elif [[ $para == --conf_path* ]];then
+        conf_path=`echo ${para#*=}`
+    elif [[ $para == --server_index* ]];then
+        server_index=`echo ${para#*=}`
+    fi
+done
+
+one_node_ip=`find $conf_path -name "server_*0.info"|awk -F "server_" '{print $2}'|awk -F "_" '{print $1}'`
+linux_num=`find $conf_path -name "server_*.info" |wc -l`
+
+#校验是否传入data_path,不需要修改
+if [[ $data_path == "" ]];then
+    echo "[Error] para \"data_path\" must be confing"
+    exit 1
+fi
+
+export HCCL_IF_IP=$fix_node_ip
+export MASTER_ADDR=$one_node_ip
+export MASTER_PORT=29688
+export HCCL_WHITELIST_DISABLE=1
+device_num=${#devicesnum}
+devices_num=`awk 'BEGIN{printf "%.0f\n",'${device_num}'-1}'`
+
+NPUS=($(seq 0 $devices_num))
+rank_server=`awk 'BEGIN{printf "%.0f\n",'${device_num}'*'${server_index}'}'`
+export NPU_WORLD_SIZE=`awk 'BEGIN{printf "%.0f\n",'${device_num}'*'${linux_num}'}'`
+
+
+#训练开始时间，不需要修改
+start_time=$(date +%s)
+
+#进入训练脚本目录，需要模型审视修改
+cd $cur_path/..
+rank=0
+for((RANK_ID=$RANK_ID_START;RANK_ID<1;RANK_ID++));
+do
+    #设置环境变量，不需要修改
+    ASCEND_DEVICE_ID=$RANK_ID
+    echo "Device ID: $ASCEND_DEVICE_ID"
+
+
+    #创建DeviceID输出目录，不需要修改
+    if [ -d ${cur_path}/output/${ASCEND_DEVICE_ID} ];then
+        rm -rf ${cur_path}/output/${ASCEND_DEVICE_ID}
+        mkdir -p ${cur_path}/output/$ASCEND_DEVICE_ID/ckpt
+    else
+        mkdir -p ${cur_path}/output/$ASCEND_DEVICE_ID/ckpt
+    fi
+    # 绑核，不需要的绑核的模型删除，需要的模型审视修改
+    #let a=RANK_ID*12
+    #let b=RANK_ID+1
+    #let c=b*12-1
+    
+    #执行训练脚本，以下传参不需要修改，其他需要模型审视修改
+    #--data_dir, --model_dir, --precision_mode, --over_dump, --over_dump_path，--data_dump_flag，--data_dump_step，--data_dump_path，--profiling，--profiling_dump_path
+    nohup python3 ${cur_path}/../main.py \
+	${data_path} \
+	-a resnet152 \
+    --addr=$one_node_ip \
+    --seed=49 \
+    --workers=$(nproc) \
+    --learning-rate=${learning_rate} \
+    --mom=0.9 \
+    --weight-decay=1.0e-04  \
+    --print-freq=1 \
+    --dist-url='tcp://127.0.0.1:50000' \
+    --multiprocessing-distributed \
+    --world-size=2 \
+    --rank=0 \
+    --device='npu' \
+    --dist-backend='hccl' \
+	--epochs=${train_epochs} \
+    --batch-size=${batch_size} \
+	--amp \
+    --device_list=0,1,2,3,4,5,6,7 \
+    --FusedSGD \
+	--stop-step-num=128 \
+    --loss-scale=1024  > ${cur_path}/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log 2>&1 &
+done 
+wait
+
+#训练结束时间，不需要修改
+end_time=$(date +%s)
+e2e_time=$(( $end_time - $start_time ))
+
+#结果打印，不需要修改
+echo "------------------ Final result ------------------"
+#输出性能FPS，需要模型审视修改
+fps=`grep -a 'FPS'  $cur_path/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log|awk -F " " '{print $NF}'|awk 'END {print}'`
+FPS=`awk 'BEGIN{printf "%.2f\n", '$fps'*16}'`
+#打印，不需要修改
+echo "Final Performance images/sec : $FPS"
+
+#输出训练精度,需要模型审视修改
+train_accuracy=`grep -a '* Acc@1' $cur_path/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log|awk 'END {print}'|awk -F "Acc@1" '{print $NF}'|awk -F " " '{print $1}'`
+#打印，不需要修改
+echo "Final Train Accuracy : ${train_accuracy}"
+echo "E2E Training Duration sec : $e2e_time"
+
+#性能看护结果汇总
+#训练用例信息，不需要修改
+BatchSize=${batch_size}
+DeviceType=`uname -m`
+CaseName=${Network}_bs${BatchSize}_${RANK_SIZE}'p'_'perf'
+
+##获取性能数据，不需要修改
+#吞吐量
+ActualFPS=${FPS}
+#单迭代训练时长
+TrainingTime=`awk 'BEGIN{printf "%.2f\n", '${batch_size}'*1000/'${FPS}'}'`
+
+#从train_$ASCEND_DEVICE_ID.log提取Loss到train_${CaseName}_loss.txt中，需要根据模型审视
+grep Epoch: $cur_path/output/$ASCEND_DEVICE_ID/train_$ASCEND_DEVICE_ID.log|grep -v Test|awk -F "Loss" '{print $NF}' | awk -F " " '{print $1}' >> $cur_path/output/$ASCEND_DEVICE_ID/train_${CaseName}_loss.txt
+
+#最后一个迭代loss值，不需要修改
+ActualLoss=`awk 'END {print}' $cur_path/output/$ASCEND_DEVICE_ID/train_${CaseName}_loss.txt`
+
+#关键信息打印到${CaseName}.log中，不需要修改
+echo "Network = ${Network}" > $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
+echo "RankSize = ${RANK_SIZE}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
+echo "BatchSize = ${BatchSize}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
+echo "DeviceType = ${DeviceType}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
+echo "CaseName = ${CaseName}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
+echo "ActualFPS = ${ActualFPS}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
+echo "TrainingTime = ${TrainingTime}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
+echo "ActualLoss = ${ActualLoss}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
+echo "E2ETrainingTime = ${e2e_time}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
\ No newline at end of file
-- 
Gitee


From 05d674ae75221447b23531275fcdf3e43e79bc96 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E7=8E=8B=E5=9B=BD=E5=BA=86?= <3224925783@qq.com>
Date: Mon, 25 Apr 2022 10:50:23 +0000
Subject: [PATCH 3/6] =?UTF-8?q?=E5=88=A0=E9=99=A4=E6=96=87=E4=BB=B6=20PyTo?=
 =?UTF-8?q?rch/dev/cv/image=5Fclassification/ResNet152=5FID0424=5Ffor=5FPy?=
 =?UTF-8?q?Torch/test/train=5Fperformance=5F16p.sh?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 .../test/train_performance_16p.sh             | 213 ------------------
 1 file changed, 213 deletions(-)
 delete mode 100644 PyTorch/dev/cv/image_classification/ResNet152_ID0424_for_PyTorch/test/train_performance_16p.sh

diff --git a/PyTorch/dev/cv/image_classification/ResNet152_ID0424_for_PyTorch/test/train_performance_16p.sh b/PyTorch/dev/cv/image_classification/ResNet152_ID0424_for_PyTorch/test/train_performance_16p.sh
deleted file mode 100644
index ad5b68f395..0000000000
--- a/PyTorch/dev/cv/image_classification/ResNet152_ID0424_for_PyTorch/test/train_performance_16p.sh
+++ /dev/null
@@ -1,213 +0,0 @@
-#!/bin/bash
-
-#当前路径,不需要修改
-cur_path=`pwd`
-#source env.sh
-#集合通信参数,不需要修改
-
-export RANK_SIZE=8
-export JOB_ID=10087
-RANK_ID_START=0
-
-
-# 数据集路径,保持为空,不需要修改
-data_path=""
-conf_path=""
-server_index=""
-fix_node_ip=""
-devicesnum=""
-
-#设置默认日志级别,不需要修改
-export ASCEND_GLOBAL_LOG_LEVEL=3
-
-#基础参数，需要模型审视修改
-#网络名称，同目录名称
-Network="ResNet152_ID0424_for_PyTorch"
-#训练epoch
-train_epochs=1
-#训练batch_size
-batch_size=4096
-#训练step
-#train_steps=`expr 1281167 / ${batch_size}`
-#学习率
-learning_rate=0.5
-
-#TF2.X独有，需要模型审视修改
-#export NPU_LOOP_SIZE=${train_steps}
-
-#维测参数，precision_mode需要模型审视修改
-#precision_mode="allow_mix_precision"
-#维持参数，以下不需要修改
-over_dump=False
-data_dump_flag=False
-data_dump_step="10"
-profiling=False
-
-# 帮助信息，不需要修改
-if [[ $1 == --help || $1 == -h ]];then
-    echo"usage:./train_performance_1P.sh <args>"
-    echo " "
-    echo "parameter explain:
-    --precision_mode         precision mode(allow_fp32_to_fp16/force_fp16/must_keep_origin_dtype/allow_mix_precision)
-    --over_dump		           if or not over detection, default is False
-    --data_dump_flag		     data dump flag, default is False
-    --data_dump_step		     data dump step, default is 10
-    --profiling		           if or not profiling for performance debug, default is False
-    --data_path		           source data of training
-    -h/--help		             show help message
-    "
-    exit 1
-fi
-
-#参数校验，不需要修改
-for para in $*
-do
-    if [[ $para == --precision_mode* ]];then
-        precision_mode=`echo ${para#*=}`
-    elif [[ $para == --over_dump* ]];then
-        over_dump=`echo ${para#*=}`
-        over_dump_path=${cur_path}/output/overflow_dump
-        mkdir -p ${over_dump_path}
-    elif [[ $para == --data_dump_flag* ]];then
-        data_dump_flag=`echo ${para#*=}`
-        data_dump_path=${cur_path}/output/data_dump
-        mkdir -p ${data_dump_path}
-    elif [[ $para == --data_dump_step* ]];then
-        data_dump_step=`echo ${para#*=}`
-    elif [[ $para == --profiling* ]];then
-        profiling=`echo ${para#*=}`
-        profiling_dump_path=${cur_path}/output/profiling
-        mkdir -p ${profiling_dump_path}
-    elif [[ $para == --data_path* ]];then
-        data_path=`echo ${para#*=}`
-    elif [[ $para == --fix_node_ip* ]];then
-	    fix_node_ip=`echo ${para#*=}`
-	elif [[ $para == --devicesnum* ]];then
-	    devicesnum=`echo ${para#*=}`
-    elif [[ $para == --conf_path* ]];then
-        conf_path=`echo ${para#*=}`
-    elif [[ $para == --server_index* ]];then
-        server_index=`echo ${para#*=}`
-    fi
-done
-
-one_node_ip=`find $conf_path -name "server_*0.info"|awk -F "server_" '{print $2}'|awk -F "_" '{print $1}'`
-linux_num=`find $conf_path -name "server_*.info" |wc -l`
-
-#校验是否传入data_path,不需要修改
-if [[ $data_path == "" ]];then
-    echo "[Error] para \"data_path\" must be confing"
-    exit 1
-fi
-
-export HCCL_IF_IP=$fix_node_ip
-export MASTER_ADDR=$one_node_ip
-export MASTER_PORT=29688
-export HCCL_WHITELIST_DISABLE=1
-device_num=${#devicesnum}
-devices_num=`awk 'BEGIN{printf "%.0f\n",'${device_num}'-1}'`
-
-NPUS=($(seq 0 $devices_num))
-rank_server=`awk 'BEGIN{printf "%.0f\n",'${device_num}'*'${server_index}'}'`
-export NPU_WORLD_SIZE=`awk 'BEGIN{printf "%.0f\n",'${device_num}'*'${linux_num}'}'`
-
-
-#训练开始时间，不需要修改
-start_time=$(date +%s)
-
-#进入训练脚本目录，需要模型审视修改
-cd $cur_path/..
-rank=0
-for((RANK_ID=$RANK_ID_START;RANK_ID<1;RANK_ID++));
-do
-    #设置环境变量，不需要修改
-    ASCEND_DEVICE_ID=$RANK_ID
-    echo "Device ID: $ASCEND_DEVICE_ID"
-
-
-    #创建DeviceID输出目录，不需要修改
-    if [ -d ${cur_path}/output/${ASCEND_DEVICE_ID} ];then
-        rm -rf ${cur_path}/output/${ASCEND_DEVICE_ID}
-        mkdir -p ${cur_path}/output/$ASCEND_DEVICE_ID/ckpt
-    else
-        mkdir -p ${cur_path}/output/$ASCEND_DEVICE_ID/ckpt
-    fi
-    # 绑核，不需要的绑核的模型删除，需要的模型审视修改
-    #let a=RANK_ID*12
-    #let b=RANK_ID+1
-    #let c=b*12-1
-    
-    #执行训练脚本，以下传参不需要修改，其他需要模型审视修改
-    #--data_dir, --model_dir, --precision_mode, --over_dump, --over_dump_path，--data_dump_flag，--data_dump_step，--data_dump_path，--profiling，--profiling_dump_path
-    nohup python3 ${cur_path}/../main.py \
-	${data_path} \
-	-a resnet152 \
-    --addr=$one_node_ip \
-    --seed=49 \
-    --workers=$(nproc) \
-    --learning-rate=${learning_rate} \
-    --mom=0.9 \
-    --weight-decay=1.0e-04  \
-    --print-freq=1 \
-    --dist-url='tcp://127.0.0.1:50000' \
-    --multiprocessing-distributed \
-    --world-size=2 \
-    --rank=0 \
-    --device='npu' \
-    --dist-backend='hccl' \
-	--epochs=${train_epochs} \
-    --batch-size=${batch_size} \
-	--amp \
-    --device_list=0,1,2,3,4,5,6,7 \
-    --FusedSGD \
-	--stop-step-num=128 \
-    --loss-scale=1024  > ${cur_path}/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log 2>&1 &
-done 
-wait
-
-#训练结束时间，不需要修改
-end_time=$(date +%s)
-e2e_time=$(( $end_time - $start_time ))
-
-#结果打印，不需要修改
-echo "------------------ Final result ------------------"
-#输出性能FPS，需要模型审视修改
-fps=`grep -a 'FPS'  $cur_path/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log|awk -F " " '{print $NF}'|awk 'END {print}'`
-FPS=`awk 'BEGIN{printf "%.2f\n", '$fps'*16}'`
-#打印，不需要修改
-echo "Final Performance images/sec : $FPS"
-
-#输出训练精度,需要模型审视修改
-train_accuracy=`grep -a '* Acc@1' $cur_path/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log|awk 'END {print}'|awk -F "Acc@1" '{print $NF}'|awk -F " " '{print $1}'`
-#打印，不需要修改
-echo "Final Train Accuracy : ${train_accuracy}"
-echo "E2E Training Duration sec : $e2e_time"
-
-#性能看护结果汇总
-#训练用例信息，不需要修改
-BatchSize=${batch_size}
-DeviceType=`uname -m`
-CaseName=${Network}_bs${BatchSize}_${RANK_SIZE}'p'_'perf'
-
-##获取性能数据，不需要修改
-#吞吐量
-ActualFPS=${FPS}
-#单迭代训练时长
-TrainingTime=`awk 'BEGIN{printf "%.2f\n", '${batch_size}'*1000/'${FPS}'}'`
-
-#从train_$ASCEND_DEVICE_ID.log提取Loss到train_${CaseName}_loss.txt中，需要根据模型审视
-grep Epoch: $cur_path/output/$ASCEND_DEVICE_ID/train_$ASCEND_DEVICE_ID.log|grep -v Test|awk -F "Loss" '{print $NF}' | awk -F " " '{print $1}' >> $cur_path/output/$ASCEND_DEVICE_ID/train_${CaseName}_loss.txt
-
-#最后一个迭代loss值，不需要修改
-ActualLoss=`awk 'END {print}' $cur_path/output/$ASCEND_DEVICE_ID/train_${CaseName}_loss.txt`
-
-#关键信息打印到${CaseName}.log中，不需要修改
-echo "Network = ${Network}" > $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
-echo "RankSize = ${RANK_SIZE}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
-echo "BatchSize = ${BatchSize}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
-echo "DeviceType = ${DeviceType}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
-echo "CaseName = ${CaseName}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
-echo "ActualFPS = ${ActualFPS}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
-echo "TrainingTime = ${TrainingTime}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
-echo "ActualLoss = ${ActualLoss}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
-echo "E2ETrainingTime = ${e2e_time}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
\ No newline at end of file
-- 
Gitee


From 635ba7ffd5a2b57b78298b3af359c9d8fa1ab2e0 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E7=8E=8B=E5=9B=BD=E5=BA=86?= <3224925783@qq.com>
Date: Mon, 25 Apr 2022 12:23:30 +0000
Subject: [PATCH 4/6] add
 ResNet152_ID0424_for_PyTorch/test/train_performance_16p.sh.

---
 .../test/train_performance_16p.sh             | 213 ++++++++++++++++++
 1 file changed, 213 insertions(+)
 create mode 100644 PyTorch/dev/cv/image_classification/ResNet152_ID0424_for_PyTorch/test/train_performance_16p.sh

diff --git a/PyTorch/dev/cv/image_classification/ResNet152_ID0424_for_PyTorch/test/train_performance_16p.sh b/PyTorch/dev/cv/image_classification/ResNet152_ID0424_for_PyTorch/test/train_performance_16p.sh
new file mode 100644
index 0000000000..6debce121d
--- /dev/null
+++ b/PyTorch/dev/cv/image_classification/ResNet152_ID0424_for_PyTorch/test/train_performance_16p.sh
@@ -0,0 +1,213 @@
+#!/bin/bash
+
+#当前路径,不需要修改
+cur_path=`pwd`
+#source env.sh
+#集合通信参数,不需要修改
+
+export RANK_SIZE=8
+export JOB_ID=10087
+RANK_ID_START=0
+
+
+# 数据集路径,保持为空,不需要修改
+data_path=""
+conf_path=""
+server_index=""
+fix_node_ip=""
+devicesnum=""
+
+#设置默认日志级别,不需要修改
+export ASCEND_GLOBAL_LOG_LEVEL=3
+
+#基础参数，需要模型审视修改
+#网络名称，同目录名称
+Network="ResNet152_ID0424_for_PyTorch"
+#训练epoch
+train_epochs=1
+#训练batch_size
+batch_size=4096
+#训练step
+#train_steps=`expr 1281167 / ${batch_size}`
+#学习率
+learning_rate=0.5
+
+#TF2.X独有，需要模型审视修改
+#export NPU_LOOP_SIZE=${train_steps}
+
+#维测参数，precision_mode需要模型审视修改
+#precision_mode="allow_mix_precision"
+#维持参数，以下不需要修改
+over_dump=False
+data_dump_flag=False
+data_dump_step="10"
+profiling=False
+
+# 帮助信息，不需要修改
+if [[ $1 == --help || $1 == -h ]];then
+    echo"usage:./train_performance_1P.sh <args>"
+    echo " "
+    echo "parameter explain:
+    --precision_mode         precision mode(allow_fp32_to_fp16/force_fp16/must_keep_origin_dtype/allow_mix_precision)
+    --over_dump		           if or not over detection, default is False
+    --data_dump_flag		     data dump flag, default is False
+    --data_dump_step		     data dump step, default is 10
+    --profiling		           if or not profiling for performance debug, default is False
+    --data_path		           source data of training
+    -h/--help		             show help message
+    "
+    exit 1
+fi
+
+#参数校验，不需要修改
+for para in $*
+do
+    if [[ $para == --precision_mode* ]];then
+        precision_mode=`echo ${para#*=}`
+    elif [[ $para == --over_dump* ]];then
+        over_dump=`echo ${para#*=}`
+        over_dump_path=${cur_path}/output/overflow_dump
+        mkdir -p ${over_dump_path}
+    elif [[ $para == --data_dump_flag* ]];then
+        data_dump_flag=`echo ${para#*=}`
+        data_dump_path=${cur_path}/output/data_dump
+        mkdir -p ${data_dump_path}
+    elif [[ $para == --data_dump_step* ]];then
+        data_dump_step=`echo ${para#*=}`
+    elif [[ $para == --profiling* ]];then
+        profiling=`echo ${para#*=}`
+        profiling_dump_path=${cur_path}/output/profiling
+        mkdir -p ${profiling_dump_path}
+    elif [[ $para == --data_path* ]];then
+        data_path=`echo ${para#*=}`
+    elif [[ $para == --fix_node_ip* ]];then
+	    fix_node_ip=`echo ${para#*=}`
+	elif [[ $para == --devicesnum* ]];then
+	    devicesnum=`echo ${para#*=}`
+    elif [[ $para == --conf_path* ]];then
+        conf_path=`echo ${para#*=}`
+    elif [[ $para == --server_index* ]];then
+        server_index=`echo ${para#*=}`
+    fi
+done
+
+one_node_ip=`find $conf_path -name "server_*0.info"|awk -F "server_" '{print $2}'|awk -F "_" '{print $1}'`
+linux_num=`find $conf_path -name "server_*.info" |wc -l`
+
+#校验是否传入data_path,不需要修改
+if [[ $data_path == "" ]];then
+    echo "[Error] para \"data_path\" must be confing"
+    exit 1
+fi
+
+export HCCL_IF_IP=$fix_node_ip
+export MASTER_ADDR=$one_node_ip
+export MASTER_PORT=29688
+export HCCL_WHITELIST_DISABLE=1
+device_num=${#devicesnum}
+devices_num=`awk 'BEGIN{printf "%.0f\n",'${device_num}'-1}'`
+
+NPUS=($(seq 0 $devices_num))
+rank_server=`awk 'BEGIN{printf "%.0f\n",'${device_num}'*'${server_index}'}'`
+export NPU_WORLD_SIZE=`awk 'BEGIN{printf "%.0f\n",'${device_num}'*'${linux_num}'}'`
+
+
+#训练开始时间，不需要修改
+start_time=$(date +%s)
+
+#进入训练脚本目录，需要模型审视修改
+cd $cur_path/..
+rank=0
+for((RANK_ID=$RANK_ID_START;RANK_ID<1;RANK_ID++));
+do
+    #设置环境变量，不需要修改
+    ASCEND_DEVICE_ID=$RANK_ID
+    echo "Device ID: $ASCEND_DEVICE_ID"
+
+
+    #创建DeviceID输出目录，不需要修改
+    if [ -d ${cur_path}/output/${ASCEND_DEVICE_ID} ];then
+        rm -rf ${cur_path}/output/${ASCEND_DEVICE_ID}
+        mkdir -p ${cur_path}/output/$ASCEND_DEVICE_ID/ckpt
+    else
+        mkdir -p ${cur_path}/output/$ASCEND_DEVICE_ID/ckpt
+    fi
+    # 绑核，不需要的绑核的模型删除，需要的模型审视修改
+    #let a=RANK_ID*12
+    #let b=RANK_ID+1
+    #let c=b*12-1
+    
+    #执行训练脚本，以下传参不需要修改，其他需要模型审视修改
+    #--data_dir, --model_dir, --precision_mode, --over_dump, --over_dump_path，--data_dump_flag，--data_dump_step，--data_dump_path，--profiling，--profiling_dump_path
+    nohup python3 ${cur_path}/../main.py \
+	${data_path} \
+	-a resnet152 \
+    --addr=$one_node_ip \
+    --seed=49 \
+    --workers=$(nproc) \
+    --learning-rate=${learning_rate} \
+    --mom=0.9 \
+    --weight-decay=1.0e-04  \
+    --print-freq=1 \
+    --dist-url='tcp://127.0.0.1:50000' \
+    --multiprocessing-distributed \
+    --world-size=2 \
+    --rank=${server_index} \
+    --device='npu' \
+    --dist-backend='hccl' \
+	--epochs=${train_epochs} \
+    --batch-size=${batch_size} \
+	--amp \
+    --device_list=0,1,2,3,4,5,6,7 \
+    --FusedSGD \
+	--stop-step-num=128 \
+    --loss-scale=1024  > ${cur_path}/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log 2>&1 &
+done 
+wait
+
+#训练结束时间，不需要修改
+end_time=$(date +%s)
+e2e_time=$(( $end_time - $start_time ))
+
+#结果打印，不需要修改
+echo "------------------ Final result ------------------"
+#输出性能FPS，需要模型审视修改
+fps=`grep -a 'FPS'  $cur_path/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log|awk -F " " '{print $NF}'|awk 'END {print}'`
+FPS=`awk 'BEGIN{printf "%.2f\n", '$fps'*16}'`
+#打印，不需要修改
+echo "Final Performance images/sec : $FPS"
+
+#输出训练精度,需要模型审视修改
+train_accuracy=`grep -a '* Acc@1' $cur_path/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log|awk 'END {print}'|awk -F "Acc@1" '{print $NF}'|awk -F " " '{print $1}'`
+#打印，不需要修改
+echo "Final Train Accuracy : ${train_accuracy}"
+echo "E2E Training Duration sec : $e2e_time"
+
+#性能看护结果汇总
+#训练用例信息，不需要修改
+BatchSize=${batch_size}
+DeviceType=`uname -m`
+CaseName=${Network}_bs${BatchSize}_${RANK_SIZE}'p'_'perf'
+
+##获取性能数据，不需要修改
+#吞吐量
+ActualFPS=${FPS}
+#单迭代训练时长
+TrainingTime=`awk 'BEGIN{printf "%.2f\n", '${batch_size}'*1000/'${FPS}'}'`
+
+#从train_$ASCEND_DEVICE_ID.log提取Loss到train_${CaseName}_loss.txt中，需要根据模型审视
+grep Epoch: $cur_path/output/$ASCEND_DEVICE_ID/train_$ASCEND_DEVICE_ID.log|grep -v Test|awk -F "Loss" '{print $NF}' | awk -F " " '{print $1}' >> $cur_path/output/$ASCEND_DEVICE_ID/train_${CaseName}_loss.txt
+
+#最后一个迭代loss值，不需要修改
+ActualLoss=`awk 'END {print}' $cur_path/output/$ASCEND_DEVICE_ID/train_${CaseName}_loss.txt`
+
+#关键信息打印到${CaseName}.log中，不需要修改
+echo "Network = ${Network}" > $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
+echo "RankSize = ${RANK_SIZE}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
+echo "BatchSize = ${BatchSize}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
+echo "DeviceType = ${DeviceType}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
+echo "CaseName = ${CaseName}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
+echo "ActualFPS = ${ActualFPS}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
+echo "TrainingTime = ${TrainingTime}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
+echo "ActualLoss = ${ActualLoss}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
+echo "E2ETrainingTime = ${e2e_time}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log
\ No newline at end of file
-- 
Gitee


From 995181a02460f002b56ab80f950e6cf5f6bcc93b Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E7=8E=8B=E5=9B=BD=E5=BA=86?= <3224925783@qq.com>
Date: Mon, 25 Apr 2022 12:49:04 +0000
Subject: [PATCH 5/6] update
 ResNet152_ID0424_for_PyTorch/test/train_performance_16p.sh.

---
 .../ResNet152_ID0424_for_PyTorch/test/train_performance_16p.sh | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/PyTorch/dev/cv/image_classification/ResNet152_ID0424_for_PyTorch/test/train_performance_16p.sh b/PyTorch/dev/cv/image_classification/ResNet152_ID0424_for_PyTorch/test/train_performance_16p.sh
index 6debce121d..c02d307296 100644
--- a/PyTorch/dev/cv/image_classification/ResNet152_ID0424_for_PyTorch/test/train_performance_16p.sh
+++ b/PyTorch/dev/cv/image_classification/ResNet152_ID0424_for_PyTorch/test/train_performance_16p.sh
@@ -172,8 +172,7 @@ e2e_time=$(( $end_time - $start_time ))
 #结果打印，不需要修改
 echo "------------------ Final result ------------------"
 #输出性能FPS，需要模型审视修改
-fps=`grep -a 'FPS'  $cur_path/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log|awk -F " " '{print $NF}'|awk 'END {print}'`
-FPS=`awk 'BEGIN{printf "%.2f\n", '$fps'*16}'`
+FPS=`grep -a 'FPS'  $cur_path/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log|awk -F " " '{print $NF}'|awk 'END {print}'`
 #打印，不需要修改
 echo "Final Performance images/sec : $FPS"
 
-- 
Gitee


From c6e681ce4dc2abbf0a8bed49c62c13e15ba52ee9 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E7=8E=8B=E5=9B=BD=E5=BA=86?= <3224925783@qq.com>
Date: Tue, 26 Apr 2022 01:16:23 +0000
Subject: [PATCH 6/6] update
 ResNet152_ID0424_for_PyTorch/test/train_performance_16p.sh.

---
 .../ResNet152_ID0424_for_PyTorch/test/train_performance_16p.sh | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/PyTorch/dev/cv/image_classification/ResNet152_ID0424_for_PyTorch/test/train_performance_16p.sh b/PyTorch/dev/cv/image_classification/ResNet152_ID0424_for_PyTorch/test/train_performance_16p.sh
index c02d307296..e713468687 100644
--- a/PyTorch/dev/cv/image_classification/ResNet152_ID0424_for_PyTorch/test/train_performance_16p.sh
+++ b/PyTorch/dev/cv/image_classification/ResNet152_ID0424_for_PyTorch/test/train_performance_16p.sh
@@ -172,7 +172,8 @@ e2e_time=$(( $end_time - $start_time ))
 #结果打印，不需要修改
 echo "------------------ Final result ------------------"
 #输出性能FPS，需要模型审视修改
-FPS=`grep -a 'FPS'  $cur_path/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log|awk -F " " '{print $NF}'|awk 'END {print}'`
+fps=`grep -a 'FPS'  $cur_path/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log|awk -F " " '{print $NF}'|awk 'END {print}'`
+FPS=`awk 'BEGIN{printf "%.2f\n",'${fps}'*2}'`
 #打印，不需要修改
 echo "Final Performance images/sec : $FPS"
 
-- 
Gitee