Merge pull request #4204 from LDOUBLEV/fix_eval

[benchmark] add DB training benchmark

benchmark/analysis.py

@@ -0,0 +1,273 @@
+# 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 os
+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')
+    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()
+                    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:
+                    print("line is: {}; separator={}; position={}".format(line, self.separator, self.position))
+
+        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)
+            try:
+                if int(os.getenv('job_fail_flag')) == 1 or int(run_info["FINAL_RESULT"]) == 0:
+                    run_info["JOB_FAIL_FLAG"] = 1
+            except:
+                pass
+        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
+

benchmark/readme.md

@@ -0,0 +1,34 @@
+
+# PaddleOCR DB/EAST 算法训练benchmark测试
+
+PaddleOCR/benchmark目录下的文件用于获取并分析训练日志。
+训练采用icdar2015数据集，包括1000张训练图像和500张测试图像。模型配置采用resnet18_vd作为backbone，分别训练batch_size=8和batch_size=16的情况。
+
+## 运行训练benchmark
+
+benchmark/run_det.sh 中包含了三个过程：
+- 安装依赖
+- 下载数据
+- 执行训练
+- 日志分析获取IPS
+
+在执行训练部分，会执行单机单卡（默认0号卡）单机多卡训练，并分别执行batch_size=8和batch_size=16的情况。所以执行完后，每种模型会得到4个日志文件。
+
+run_det.sh 执行方式如下:
+
+```
+# cd PaddleOCR/
+bash benchmark/run_det.sh 
+```
+
+以DB为例，将得到四个日志文件，如下：
+```
+det_res18_db_v2.0_sp_bs16_fp32_1
+det_res18_db_v2.0_sp_bs8_fp32_1
+det_res18_db_v2.0_mp_bs16_fp32_1
+det_res18_db_v2.0_mp_bs8_fp32_1
+```
+
+
+
+

benchmark/run_benchmark_det.sh

@@ -20,9 +20,7 @@ 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="-c configs/det/${model_name}.yml
-               -o Train.loader.batch_size_per_card=${batch_size}
-               -o Global.epoch_num=${max_iter} "   
+    train_cmd="-c configs/det/${model_name}.yml -o Train.loader.batch_size_per_card=${batch_size} Global.epoch_num=${max_iter} "   
     case ${run_mode} in
       sp) 
         train_cmd="python3.7 tools/train.py "${train_cmd}""
@@ -47,6 +45,10 @@ function _train(){
         rm ${log_file}
         cp mylog/workerlog.0 ${log_file}
     fi
+
+    # run log analysis
+    analysis_cmd="python3.7 benchmark/analysis.py --filename ${log_file}  --mission_name ${model_name} --run_mode ${mode} --direction_id 0 --keyword 'ips:' --base_batch_size ${batch_szie} --skip_steps 1 --gpu_num ${num_gpu_devices}  --index 1  --model_mode=-1  --ips_unit=samples/sec"
+    eval $analysis_cmd
 }
 
 _set_params $@

benchmark/run_det.sh

@@ -1,26 +1,25 @@
 # 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
-cd PaddleOCR
+# 执行目录: ./PaddleOCR
 # 1 安装该模型需要的依赖 (如需开启优化策略请注明)
 python3.7 -m pip install -r requirements.txt
 # 2 拷贝该模型需要数据、预训练模型
-wget -p ./tain_data/  https://paddleocr.bj.bcebos.com/dygraph_v2.0/test/icdar2015.tar && cd train_data  && tar xf icdar2015.tar && cd ../
-wget -p ./pretrain_models/ https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/ResNet50_vd_pretrained.pdparams
+wget -c  -p ./tain_data/  https://paddleocr.bj.bcebos.com/dygraph_v2.0/test/icdar2015.tar && cd train_data  && tar xf icdar2015.tar && cd ../
+wget -c -p ./pretrain_models/ https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/ResNet50_vd_pretrained.pdparams
 # 3 批量运行（如不方便批量，1，2需放到单个模型中）
 
-model_mode_list=(det_mv3_db det_r50_vd_east)
+model_mode_list=(det_res18_db_v2.0 det_r50_vd_east)
 fp_item_list=(fp32)
-bs_list=(256 128)
+bs_list=(8 16)
 for model_mode in ${model_mode_list[@]}; do
       for fp_item in ${fp_item_list[@]}; do
           for bs_item in ${bs_list[@]}; do
             echo "index is speed, 1gpus, begin, ${model_name}"
             run_mode=sp
-            CUDA_VISIBLE_DEVICES=0 bash benchmark/run_benchmark.sh ${run_mode} ${bs_item} ${fp_item} 10 ${model_mode}     #  (5min)
+            CUDA_VISIBLE_DEVICES=0 bash benchmark/run_benchmark_det.sh ${run_mode} ${bs_item} ${fp_item} 10 ${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 benchmark/run_benchmark.sh ${run_mode} ${bs_item} ${fp_item} 10 ${model_mode} 
+            CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 bash benchmark/run_benchmark_det.sh ${run_mode} ${bs_item} ${fp_item} 10 ${model_mode} 
             sleep 60
             done
       done

configs/det/det_res18_db_v2.0.yml

@@ -0,0 +1,131 @@
+Global:
+  use_gpu: true
+  epoch_num: 1200
+  log_smooth_window: 20
+  print_batch_step: 2
+  save_model_dir: ./output/ch_db_res18/
+  save_epoch_step: 1200
+  # evaluation is run every 5000 iterations after the 4000th iteration
+  eval_batch_step: [3000, 2000]
+  cal_metric_during_train: False
+  pretrained_model: ./pretrain_models/ResNet18_vd_pretrained
+  checkpoints:
+  save_inference_dir:
+  use_visualdl: False
+  infer_img: doc/imgs_en/img_10.jpg
+  save_res_path: ./output/det_db/predicts_db.txt
+
+Architecture:
+  model_type: det
+  algorithm: DB
+  Transform:
+  Backbone:
+    name: ResNet
+    layers: 18
+    disable_se: True
+  Neck:
+    name: DBFPN
+    out_channels: 256
+  Head:
+    name: DBHead
+    k: 50
+
+Loss:
+  name: DBLoss
+  balance_loss: true
+  main_loss_type: DiceLoss
+  alpha: 5
+  beta: 10
+  ohem_ratio: 3
+
+Optimizer:
+  name: Adam
+  beta1: 0.9
+  beta2: 0.999
+  lr:
+    name: Cosine
+    learning_rate: 0.001
+    warmup_epoch: 2
+  regularizer:
+    name: 'L2'
+    factor: 0
+
+PostProcess:
+  name: DBPostProcess
+  thresh: 0.3
+  box_thresh: 0.6
+  max_candidates: 1000
+  unclip_ratio: 1.5
+
+Metric:
+  name: DetMetric
+  main_indicator: hmean
+
+Train:
+  dataset:
+    name: SimpleDataSet
+    data_dir: ./train_data/icdar2015/text_localization/
+    label_file_list:
+      - ./train_data/icdar2015/text_localization/train_icdar2015_label.txt
+    ratio_list: [1.0]
+    transforms:
+      - DecodeImage: # load image
+          img_mode: BGR
+          channel_first: False
+      - DetLabelEncode: # Class handling label
+      - IaaAugment:
+          augmenter_args:
+            - { 'type': Fliplr, 'args': { 'p': 0.5 } }
+            - { 'type': Affine, 'args': { 'rotate': [-10, 10] } }
+            - { 'type': Resize, 'args': { 'size': [0.5, 3] } }
+      - EastRandomCropData:
+          size: [960, 960]
+          max_tries: 50
+          keep_ratio: true
+      - MakeBorderMap:
+          shrink_ratio: 0.4
+          thresh_min: 0.3
+          thresh_max: 0.7
+      - MakeShrinkMap:
+          shrink_ratio: 0.4
+          min_text_size: 8
+      - NormalizeImage:
+          scale: 1./255.
+          mean: [0.485, 0.456, 0.406]
+          std: [0.229, 0.224, 0.225]
+          order: 'hwc'
+      - ToCHWImage:
+      - KeepKeys:
+          keep_keys: ['image', 'threshold_map', 'threshold_mask', 'shrink_map', 'shrink_mask'] # the order of the dataloader list
+  loader:
+    shuffle: True
+    drop_last: False
+    batch_size_per_card: 8
+    num_workers: 4
+
+Eval:
+  dataset:
+    name: SimpleDataSet
+    data_dir: ./train_data/icdar2015/text_localization/
+    label_file_list:
+      - ./train_data/icdar2015/text_localization/test_icdar2015_label.txt
+    transforms:
+      - DecodeImage: # load image
+          img_mode: BGR
+          channel_first: False
+      - DetLabelEncode: # Class handling label
+      - DetResizeForTest:
+#           image_shape: [736, 1280]
+      - NormalizeImage:
+          scale: 1./255.
+          mean: [0.485, 0.456, 0.406]
+          std: [0.229, 0.224, 0.225]
+          order: 'hwc'
+      - ToCHWImage:
+      - KeepKeys:
+          keep_keys: ['image', 'shape', 'polys', 'ignore_tags']
+  loader:
+    shuffle: False
+    drop_last: False
+    batch_size_per_card: 1 # must be 1
+    num_workers: 2

ppocr/utils/profiler.py

@@ -0,0 +1,110 @@
+# copyright (c) 2021 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.
+
+import sys
+import paddle
+
+# A global variable to record the number of calling times for profiler
+# functions. It is used to specify the tracing range of training steps.
+_profiler_step_id = 0
+
+# A global variable to avoid parsing from string every time.
+_profiler_options = None
+
+
+class ProfilerOptions(object):
+    '''
+    Use a string to initialize a ProfilerOptions.
+    The string should be in the format: "key1=value1;key2=value;key3=value3".
+    For example:
+      "profile_path=model.profile"
+      "batch_range=[50, 60]; profile_path=model.profile"
+      "batch_range=[50, 60]; tracer_option=OpDetail; profile_path=model.profile"
+    ProfilerOptions supports following key-value pair:
+      batch_range      - a integer list, e.g. [100, 110].
+      state            - a string, the optional values are 'CPU', 'GPU' or 'All'. 
+      sorted_key       - a string, the optional values are 'calls', 'total',
+                         'max', 'min' or 'ave.
+      tracer_option    - a string, the optional values are 'Default', 'OpDetail',
+                         'AllOpDetail'.
+      profile_path     - a string, the path to save the serialized profile data,
+                         which can be used to generate a timeline.
+      exit_on_finished - a boolean.
+    '''
+
+    def __init__(self, options_str):
+        assert isinstance(options_str, str)
+
+        self._options = {
+            'batch_range': [10, 20],
+            'state': 'All',
+            'sorted_key': 'total',
+            'tracer_option': 'Default',
+            'profile_path': '/tmp/profile',
+            'exit_on_finished': True
+        }
+        self._parse_from_string(options_str)
+
+    def _parse_from_string(self, options_str):
+        for kv in options_str.replace(' ', '').split(';'):
+            key, value = kv.split('=')
+            if key == 'batch_range':
+                value_list = value.replace('[', '').replace(']', '').split(',')
+                value_list = list(map(int, value_list))
+                if len(value_list) >= 2 and value_list[0] >= 0 and value_list[
+                        1] > value_list[0]:
+                    self._options[key] = value_list
+            elif key == 'exit_on_finished':
+                self._options[key] = value.lower() in ("yes", "true", "t", "1")
+            elif key in [
+                    'state', 'sorted_key', 'tracer_option', 'profile_path'
+            ]:
+                self._options[key] = value
+
+    def __getitem__(self, name):
+        if self._options.get(name, None) is None:
+            raise ValueError(
+                "ProfilerOptions does not have an option named %s." % name)
+        return self._options[name]
+
+
+def add_profiler_step(options_str=None):
+    '''
+    Enable the operator-level timing using PaddlePaddle's profiler.
+    The profiler uses a independent variable to count the profiler steps.
+    One call of this function is treated as a profiler step.
+    
+    Args:
+      profiler_options - a string to initialize the ProfilerOptions.
+                         Default is None, and the profiler is disabled.
+    '''
+    if options_str is None:
+        return
+
+    global _profiler_step_id
+    global _profiler_options
+
+    if _profiler_options is None:
+        _profiler_options = ProfilerOptions(options_str)
+
+    if _profiler_step_id == _profiler_options['batch_range'][0]:
+        paddle.utils.profiler.start_profiler(
+            _profiler_options['state'], _profiler_options['tracer_option'])
+    elif _profiler_step_id == _profiler_options['batch_range'][1]:
+        paddle.utils.profiler.stop_profiler(_profiler_options['sorted_key'],
+                                            _profiler_options['profile_path'])
+        if _profiler_options['exit_on_finished']:
+            sys.exit(0)
+
+    _profiler_step_id += 1

tools/program.py

@@ -31,6 +31,7 @@ from ppocr.utils.stats import TrainingStats
 from ppocr.utils.save_load import save_model
 from ppocr.utils.utility import print_dict
 from ppocr.utils.logging import get_logger
+from ppocr.utils import profiler
 from ppocr.data import build_dataloader
 import numpy as np
 
@@ -42,6 +43,13 @@ class ArgsParser(ArgumentParser):
         self.add_argument("-c", "--config", help="configuration file to use")
         self.add_argument(
             "-o", "--opt", nargs='+', help="set configuration options")
+        self.add_argument(
+            '-p',
+            '--profiler_options',
+            type=str,
+            default=None,
+            help='The option of profiler, which should be in format \"key1=value1;key2=value2;key3=value3\".'
+        )
 
     def parse_args(self, argv=None):
         args = super(ArgsParser, self).parse_args(argv)
@@ -158,6 +166,7 @@ def train(config,
     epoch_num = config['Global']['epoch_num']
     print_batch_step = config['Global']['print_batch_step']
     eval_batch_step = config['Global']['eval_batch_step']
+    profiler_options = config['profiler_options']
 
     global_step = 0
     if 'global_step' in pre_best_model_dict:
@@ -209,6 +218,7 @@ def train(config,
         max_iter = len(train_dataloader) - 1 if platform.system(
         ) == "Windows" else len(train_dataloader)
         for idx, batch in enumerate(train_dataloader):
+            profiler.add_profiler_step(profiler_options)
             train_reader_cost += time.time() - batch_start
             if idx >= max_iter:
                 break
@@ -391,8 +401,11 @@ def eval(model,
 
 def preprocess(is_train=False):
     FLAGS = ArgsParser().parse_args()
+    profiler_options = FLAGS.profiler_options
     config = load_config(FLAGS.config)
     merge_config(FLAGS.opt)
+    profile_dic = {"profiler_options": FLAGS.profiler_options}
+    merge_config(profile_dic)
 
     if is_train:
         # save_config