Merge pull request #877 from Jackwaterveg/benchmark

Add bentchmark

tests/benchmark/.gitignore

@@ -1,2 +0,0 @@
-old-pd_env.txt
-pd_env.txt

tests/benchmark/README.md

@@ -1,11 +1,22 @@
-# Benchmark Test
+### Prepare the environment
+Please follow the instructions shown in [here](https://github.com/PaddlePaddle/DeepSpeech/blob/develop/docs/src/install.md) to install the Deepspeech first.
 
-## Data
+### Prepare the benchmark environment
+```
+bash prepare.sh
+```
 
-* Aishell
+### Start benchmarking
+```
+bash run.sh
+```
 
-## Docker
+### Analyse the sp
+```
+bash run_analysis_sp.sh
+```
 
+### Analyse the mp
 ```
-registry.baidubce.com/paddlepaddle/paddle   2.1.1-gpu-cuda10.2-cudnn7   59d5ec1de486  
+bash run_analysis_mp.sh
 ```

tests/benchmark/analysis.py

@@ -0,0 +1,345 @@
+# copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from __future__ import print_function
+
+import argparse
+import json
+import re
+import traceback
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument(
+        "--filename", type=str, help="The name of log which need to analysis.")
+    parser.add_argument(
+        "--log_with_profiler",
+        type=str,
+        help="The path of train log with profiler")
+    parser.add_argument(
+        "--profiler_path", type=str, help="The path of profiler timeline log.")
+    parser.add_argument(
+        "--keyword", type=str, help="Keyword to specify analysis data")
+    parser.add_argument(
+        "--separator",
+        type=str,
+        default=None,
+        help="Separator of different field in log")
+    parser.add_argument(
+        '--position', type=int, default=None, help='The position of data field')
+    parser.add_argument(
+        '--range',
+        type=str,
+        default="",
+        help='The range of data field to intercept')
+    parser.add_argument(
+        '--base_batch_size', type=int, help='base_batch size on gpu')
+    parser.add_argument(
+        '--skip_steps',
+        type=int,
+        default=0,
+        help='The number of steps to be skipped')
+    parser.add_argument(
+        '--model_mode',
+        type=int,
+        default=-1,
+        help='Analysis mode, default value is -1')
+    parser.add_argument('--ips_unit', type=str, default=None, help='IPS unit')
+    parser.add_argument(
+        '--model_name',
+        type=str,
+        default=0,
+        help='training model_name, transformer_base')
+    parser.add_argument(
+        '--mission_name', type=str, default=0, help='training mission name')
+    parser.add_argument(
+        '--direction_id', type=int, default=0, help='training direction_id')
+    parser.add_argument(
+        '--run_mode',
+        type=str,
+        default="sp",
+        help='multi process or single process')
+    parser.add_argument(
+        '--index',
+        type=int,
+        default=1,
+        help='{1: speed, 2:mem, 3:profiler, 6:max_batch_size}')
+    parser.add_argument(
+        '--gpu_num', type=int, default=1, help='nums of training gpus')
+    parser.add_argument(
+        '--use_num', type=int, default=1, help='nums of used recoders')
+    args = parser.parse_args()
+    args.separator = None if args.separator == "None" else args.separator
+    return args
+
+
+def _is_number(num):
+    pattern = re.compile(r'^[-+]?[-0-9]\d*\.\d*|[-+]?\.?[0-9]\d*$')
+    result = pattern.match(num)
+    if result:
+        return True
+    else:
+        return False
+
+
+class TimeAnalyzer(object):
+    def __init__(self,
+                 filename,
+                 keyword=None,
+                 separator=None,
+                 position=None,
+                 range="-1"):
+        if filename is None:
+            raise Exception("Please specify the filename!")
+
+        if keyword is None:
+            raise Exception("Please specify the keyword!")
+
+        self.filename = filename
+        self.keyword = keyword
+        self.separator = separator
+        self.position = position
+        self.range = range
+        self.records = None
+        self._distil()
+
+    def _distil(self):
+        self.records = []
+        with open(self.filename, "r") as f_object:
+            lines = f_object.readlines()
+            for line in lines:
+                if self.keyword not in line:
+                    continue
+                try:
+                    result = None
+
+                    # Distil the string from a line.
+                    line = line.strip()
+                    line_words = line.split(
+                        self.separator) if self.separator else line.split()
+                    print("line_words", line_words)
+                    if args.position:
+                        result = line_words[self.position]
+                    else:
+                        # Distil the string following the keyword.
+                        for i in range(len(line_words) - 1):
+                            if line_words[i] == self.keyword:
+                                result = line_words[i + 1]
+                                break
+
+                    # Distil the result from the picked string.
+                    if not self.range:
+                        result = result[0:]
+                    elif _is_number(self.range):
+                        result = result[0:int(self.range)]
+                    else:
+                        result = result[int(self.range.split(":")[0]):int(
+                            self.range.split(":")[1])]
+                    self.records.append(float(result))
+                except Exception as exc:
+                    pass
+                    #print("line is: {}; separator={}; position={}".format(line, self.separator, self.position))
+        self.records.sort()
+        self.records = self.records[:args.use_num]
+        print("records", self.records)
+        print("Extract {} records: separator={}; position={}".format(
+            len(self.records), self.separator, self.position))
+
+    def _get_fps(self,
+                 mode,
+                 batch_size,
+                 gpu_num,
+                 avg_of_records,
+                 run_mode,
+                 unit=None):
+        if mode == -1 and run_mode == 'sp':
+            assert unit, "Please set the unit when mode is -1."
+            fps = gpu_num * avg_of_records
+        elif mode == -1 and run_mode == 'mp':
+            assert unit, "Please set the unit when mode is -1."
+            fps = gpu_num * avg_of_records  #temporarily, not used now
+            print("------------this is mp")
+        elif mode == 0:
+            # s/step -> samples/s
+            fps = (batch_size * gpu_num) / avg_of_records
+            unit = "samples/s"
+        elif mode == 1:
+            # steps/s -> steps/s
+            fps = avg_of_records
+            unit = "steps/s"
+        elif mode == 2:
+            # s/step -> steps/s
+            fps = 1 / avg_of_records
+            unit = "steps/s"
+        elif mode == 3:
+            # steps/s -> samples/s
+            fps = batch_size * gpu_num * avg_of_records
+            unit = "samples/s"
+        elif mode == 4:
+            # s/epoch -> s/epoch
+            fps = avg_of_records
+            unit = "s/epoch"
+        else:
+            ValueError("Unsupported analysis mode.")
+
+        return fps, unit
+
+    def analysis(self,
+                 batch_size,
+                 gpu_num=1,
+                 skip_steps=0,
+                 mode=-1,
+                 run_mode='sp',
+                 unit=None):
+        if batch_size <= 0:
+            print("base_batch_size should larger than 0.")
+            return 0, ''
+
+        if len(
+                self.records
+        ) <= skip_steps:  # to address the condition which item of log equals to skip_steps
+            print("no records")
+            return 0, ''
+
+        sum_of_records = 0
+        sum_of_records_skipped = 0
+        skip_min = self.records[skip_steps]
+        skip_max = self.records[skip_steps]
+
+        count = len(self.records)
+        for i in range(count):
+            sum_of_records += self.records[i]
+            if i >= skip_steps:
+                sum_of_records_skipped += self.records[i]
+                if self.records[i] < skip_min:
+                    skip_min = self.records[i]
+                if self.records[i] > skip_max:
+                    skip_max = self.records[i]
+
+        avg_of_records = sum_of_records / float(count)
+        avg_of_records_skipped = sum_of_records_skipped / float(count -
+                                                                skip_steps)
+
+        fps, fps_unit = self._get_fps(mode, batch_size, gpu_num, avg_of_records,
+                                      run_mode, unit)
+        fps_skipped, _ = self._get_fps(mode, batch_size, gpu_num,
+                                       avg_of_records_skipped, run_mode, unit)
+        if mode == -1:
+            print("average ips of %d steps, skip 0 step:" % count)
+            print("\tAvg: %.3f %s" % (avg_of_records, fps_unit))
+            print("\tFPS: %.3f %s" % (fps, fps_unit))
+            if skip_steps > 0:
+                print("average ips of %d steps, skip %d steps:" %
+                      (count, skip_steps))
+                print("\tAvg: %.3f %s" % (avg_of_records_skipped, fps_unit))
+                print("\tMin: %.3f %s" % (skip_min, fps_unit))
+                print("\tMax: %.3f %s" % (skip_max, fps_unit))
+                print("\tFPS: %.3f %s" % (fps_skipped, fps_unit))
+        elif mode == 1 or mode == 3:
+            print("average latency of %d steps, skip 0 step:" % count)
+            print("\tAvg: %.3f steps/s" % avg_of_records)
+            print("\tFPS: %.3f %s" % (fps, fps_unit))
+            if skip_steps > 0:
+                print("average latency of %d steps, skip %d steps:" %
+                      (count, skip_steps))
+                print("\tAvg: %.3f steps/s" % avg_of_records_skipped)
+                print("\tMin: %.3f steps/s" % skip_min)
+                print("\tMax: %.3f steps/s" % skip_max)
+                print("\tFPS: %.3f %s" % (fps_skipped, fps_unit))
+        elif mode == 0 or mode == 2:
+            print("average latency of %d steps, skip 0 step:" % count)
+            print("\tAvg: %.3f s/step" % avg_of_records)
+            print("\tFPS: %.3f %s" % (fps, fps_unit))
+            if skip_steps > 0:
+                print("average latency of %d steps, skip %d steps:" %
+                      (count, skip_steps))
+                print("\tAvg: %.3f s/step" % avg_of_records_skipped)
+                print("\tMin: %.3f s/step" % skip_min)
+                print("\tMax: %.3f s/step" % skip_max)
+                print("\tFPS: %.3f %s" % (fps_skipped, fps_unit))
+
+        return round(fps_skipped, 3), fps_unit
+
+
+if __name__ == "__main__":
+    args = parse_args()
+    run_info = dict()
+    run_info["log_file"] = args.filename
+    run_info["model_name"] = args.model_name
+    run_info["mission_name"] = args.mission_name
+    run_info["direction_id"] = args.direction_id
+    run_info["run_mode"] = args.run_mode
+    run_info["index"] = args.index
+    run_info["gpu_num"] = args.gpu_num
+    run_info["FINAL_RESULT"] = 0
+    run_info["JOB_FAIL_FLAG"] = 0
+
+    try:
+        if args.index == 1:
+            if args.gpu_num == 1:
+                run_info["log_with_profiler"] = args.log_with_profiler
+                run_info["profiler_path"] = args.profiler_path
+            analyzer = TimeAnalyzer(args.filename, args.keyword, args.separator,
+                                    args.position, args.range)
+            run_info["FINAL_RESULT"], run_info["UNIT"] = analyzer.analysis(
+                batch_size=args.base_batch_size,
+                gpu_num=args.gpu_num,
+                skip_steps=args.skip_steps,
+                mode=args.model_mode,
+                run_mode=args.run_mode,
+                unit=args.ips_unit)
+    #     if int(os.getenv('job_fail_flag')) == 1 or int(run_info["FINAL_RESULT"]) == 0:
+    #         run_info["JOB_FAIL_FLAG"] = 1
+        elif args.index == 3:
+            run_info["FINAL_RESULT"] = {}
+            records_fo_total = TimeAnalyzer(args.filename, 'Framework overhead',
+                                            None, 3, '').records
+            records_fo_ratio = TimeAnalyzer(args.filename, 'Framework overhead',
+                                            None, 5).records
+            records_ct_total = TimeAnalyzer(args.filename, 'Computation time',
+                                            None, 3, '').records
+            records_gm_total = TimeAnalyzer(args.filename,
+                                            'GpuMemcpy                Calls',
+                                            None, 4, '').records
+            records_gm_ratio = TimeAnalyzer(args.filename,
+                                            'GpuMemcpy                Calls',
+                                            None, 6).records
+            records_gmas_total = TimeAnalyzer(args.filename,
+                                              'GpuMemcpyAsync         Calls',
+                                              None, 4, '').records
+            records_gms_total = TimeAnalyzer(args.filename,
+                                             'GpuMemcpySync          Calls',
+                                             None, 4, '').records
+            run_info["FINAL_RESULT"]["Framework_Total"] = records_fo_total[
+                0] if records_fo_total else 0
+            run_info["FINAL_RESULT"]["Framework_Ratio"] = records_fo_ratio[
+                0] if records_fo_ratio else 0
+            run_info["FINAL_RESULT"][
+                "ComputationTime_Total"] = records_ct_total[
+                    0] if records_ct_total else 0
+            run_info["FINAL_RESULT"]["GpuMemcpy_Total"] = records_gm_total[
+                0] if records_gm_total else 0
+            run_info["FINAL_RESULT"]["GpuMemcpy_Ratio"] = records_gm_ratio[
+                0] if records_gm_ratio else 0
+            run_info["FINAL_RESULT"][
+                "GpuMemcpyAsync_Total"] = records_gmas_total[
+                    0] if records_gmas_total else 0
+            run_info["FINAL_RESULT"]["GpuMemcpySync_Total"] = records_gms_total[
+                0] if records_gms_total else 0
+        else:
+            print("Not support!")
+    except Exception:
+        traceback.print_exc()
+    print("{}".format(json.dumps(run_info))
+          )  # it's required, for the log file path  insert to the database

tests/benchmark/prepare.sh

@@ -0,0 +1,7 @@
+source ../tools/venv/bin/activate
+
+#Enter the example dir
+pushd ../examples/aishell/s1
+
+#Prepare the data
+bash run.sh --stage 0 --stop_stage 0

tests/benchmark/run.sh

@@ -0,0 +1,34 @@
+
+# 提供可稳定复现性能的脚本，默认在标准docker环境内py37执行： paddlepaddle/paddle:latest-gpu-cuda10.1-cudnn7  paddle=2.1.2  py=37
+# 执行目录：需说明
+CUR_DIR=${PWD}
+source ../tools/venv/bin/activate
+#cd **
+pushd ../examples/aishell/s1
+# 1 安装该模型需要的依赖 (如需开启优化策略请注明)
+# 2 拷贝该模型需要数据、预训练模型
+
+
+source path.sh
+fp_item_list=(fp32)
+bs_item=(16)
+config_path=conf/conformer.yaml
+seed=0
+output=exp/conformer
+profiler_options=None
+for fp_item in ${fp_item_list[@]}; do
+    for batch_size in ${bs_item[@]}
+        do
+        rm exp -rf
+        echo "index is speed, 8gpus, run_mode is multi_process, begin, conformer"
+        run_mode=mp
+        ngpu=8
+        CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 bash ${CUR_DIR}/run_benchmark.sh ${run_mode} ${config_path} ${output} ${seed} ${ngpu} ${profiler_options} ${batch_size} ${fp_item} ${CUR_DIR}
+        rm exp -rf
+        echo "index is speed, 1gpus, begin, conformer"
+        run_mode=sp
+        ngpu=1
+        CUDA_VISIBLE_DEVICES=7 bash ${CUR_DIR}/run_benchmark.sh ${run_mode} ${config_path} ${output} ${seed} ${ngpu} ${profiler_options} ${batch_size} ${fp_item} ${CUR_DIR}
+    done
+done
+

tests/benchmark/run_all.sh

@@ -1,49 +0,0 @@
-#!/bin/bash
-
-CUR_DIR=${PWD}
-ROOT_DIR=../../
-
-# 提供可稳定复现性能的脚本，默认在标准docker环境内py37执行：
-# collect env info
-bash ${ROOT_DIR}/utils/pd_env_collect.sh
-#cat pd_env.txt
-
-
-# 1 安装该模型需要的依赖 (如需开启优化策略请注明)
-#pushd ${ROOT_DIR}/tools; make; popd
-#source ${ROOT_DIR}/tools/venv/bin/activate
-#pushd ${ROOT_DIR}; bash setup.sh; popd
-
-
-# 2 拷贝该模型需要数据、预训练模型
-
-# 执行目录：需说明
-#pushd ${ROOT_DIR}/examples/aishell/s1
-pushd ${ROOT_DIR}/examples/tiny/s1
-
-mkdir -p exp/log
-. path.sh
-#bash local/data.sh &> exp/log/data.log
-
-# 3 批量运行（如不方便批量，1，2需放到单个模型中）
-
-model_mode_list=(conformer transformer)
-fp_item_list=(fp32)
-bs_item_list=(32 64 96)
-for model_mode in ${model_mode_list[@]}; do
-      for fp_item in ${fp_item_list[@]}; do
-          for bs_item in ${bs_item_list[@]}
-            do
-            echo "index is speed, 1gpus, begin, ${model_name}"
-            run_mode=sp
-            CUDA_VISIBLE_DEVICES=0 bash ${CUR_DIR}/run_benchmark.sh ${run_mode} ${bs_item} ${fp_item} 500 ${model_mode}     #  (5min)
-            sleep 60
-            echo "index is speed, 8gpus, run_mode is multi_process, begin, ${model_name}"
-            run_mode=mp
-            CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 bash ${CUR_DIR}/run_benchmark.sh ${run_mode} ${bs_item} ${fp_item} 500 ${model_mode}
-            sleep 60
-            done
-      done
-done
-
-popd # aishell/s1

tests/benchmark/run_analysis_mp.sh

@@ -0,0 +1,12 @@
+python analysis.py \
+    --filename "recoder_mp_bs16_fp32_ngpu8.txt" \
+    --keyword "ips[sent./sec]:" \
+    --base_batch_size 16 \
+    --model_name "Conformer" \
+    --mission_name "eight gpu" \
+    --run_mode "mp" \
+    --ips_unit "sent./sec" \
+    --gpu_num 8 \
+    --use_num 480 \
+    --separator " " \
+

tests/benchmark/run_analysis_sp.sh

@@ -0,0 +1,12 @@
+python analysis.py \
+    --filename "recoder_sp_bs16_fp32_ngpu1.txt" \
+    --keyword "ips[sent./sec]:" \
+    --base_batch_size 16 \
+    --model_name "Conformer" \
+    --mission_name "one gpu" \
+    --run_mode "sp" \
+    --ips_unit "sent./sec" \
+    --gpu_num 1 \
+    --use_num 60 \
+    --separator " " \
+

tests/benchmark/run_benchmark.sh

@@ -1,39 +1,49 @@
-#!/bin/bash
+#!/usr/bin/env bash
-
 set -xe
-
 # 运行示例：CUDA_VISIBLE_DEVICES=0 bash run_benchmark.sh ${run_mode} ${bs_item} ${fp_item} 500 ${model_mode}
 # 参数说明
 function _set_params(){
+
     run_mode=${1:-"sp"}          # 单卡sp|多卡mp
-    batch_size=${2:-"64"}
+    config_path=${2:-"conf/conformer.yaml"}
-    fp_item=${3:-"fp32"}        # fp32|fp16
+    output=${3:-"exp/conformer"}
-    max_iter=${4:-"500"}       # 可选，如果需要修改代码提前中断
+    seed=${4:-"0"}
-    model_name=${5:-"model_name"}
+    ngpu=${5:-"1"}
+    profiler_options=${6:-"None"}
+    batch_size=${7:-"32"}
+    fp_item=${8:-"fp32"}
+    TRAIN_LOG_DIR=${9:-$(pwd)}
+
+    benchmark_max_step=0
+
     run_log_path=${TRAIN_LOG_DIR:-$(pwd)}  # TRAIN_LOG_DIR 后续QA设置该参数
 
 #   以下不用修改
     device=${CUDA_VISIBLE_DEVICES//,/ }
     arr=(${device})
     num_gpu_devices=${#arr[*]}
-    log_file=${run_log_path}/${model_name}_${run_mode}_bs${batch_size}_${fp_item}_${num_gpu_devices}
+    log_file=${run_log_path}/recoder_${run_mode}_bs${batch_size}_${fp_item}_ngpu${ngpu}.txt
 }
 
 function _train(){
     echo "Train on ${num_gpu_devices} GPUs"
     echo "current CUDA_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES, gpus=$num_gpu_devices, batch_size=$batch_size"
+    train_cmd="--config=${config_path}
+               --output=${output}
+               --seed=${seed}
+               --nproc=${ngpu}
+               --profiler-options "${profiler_options}"
+               --benchmark-batch-size ${batch_size}
+               --benchmark-max-step ${benchmark_max_step} "
 
-    train_cmd="--benchmark-batch-size ${batch_size}
+    echo "run_mode "${run_mode}
-               --benchmark-max-step ${max_iter}
-               conf/${model_name}.yaml ${model_name}"
 
     case ${run_mode} in
-    sp) train_cmd="bash local/train.sh "${train_cmd}"" ;;
+    sp) train_cmd="python3 -u ${BIN_DIR}/train.py "${train_cmd} ;;
-    mp)
+    mp) train_cmd="python3 -u ${BIN_DIR}/train.py "${train_cmd} ;;
-        train_cmd="bash local/train.sh "${train_cmd}"" ;;
     *) echo "choose run_mode(sp or mp)"; exit 1;
     esac
-
+    echo ${train_cmd}
 # 以下不用修改
     timeout 15m ${train_cmd} > ${log_file} 2>&1
     if [ $? -ne 0 ];then
@@ -43,7 +53,6 @@ function _train(){
         echo -e "${model_name}, SUCCESS"
         export job_fail_flag=0
     fi
-
     trap 'for pid in $(jobs -pr); do kill -KILL $pid; done' INT QUIT TERM
 
     if [ $run_mode = "mp" -a -d mylog ]; then
@@ -54,4 +63,3 @@ function _train(){
 
 _set_params $@
 _train
-