Merge remote-tracking branch 'origin/dygraph' into dygraph

benchmark/analysis.py

+# 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

+
+# 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

+#!/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=${2:-"64"}
+    fp_item=${3:-"fp32"}        # fp32|fp16
+    max_iter=${4:-"500"}       # 可选，如果需要修改代码提前中断
+    model_name=${5:-"model_name"}
+    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}
+}
+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} Global.epoch_num=${max_iter} "   
+    case ${run_mode} in
+      sp) 
+        train_cmd="python3.7 tools/train.py "${train_cmd}""
+        ;;
+      mp)
+        train_cmd="python3.7 -m paddle.distributed.launch --log_dir=./mylog --gpus=$CUDA_VISIBLE_DEVICES tools/train.py ${train_cmd}"
+        ;;
+      *) echo "choose run_mode(sp or mp)"; exit 1;
+    esac
+# 以下不用修改
+    timeout 15m ${train_cmd} > ${log_file} 2>&1
+    if [ $? -ne 0 ];then
+            echo -e "${model_name}, FAIL"
+        export job_fail_flag=1
+    else
+        echo -e "${model_name}, SUCCESS"
+        export job_fail_flag=0
+    fi
+    kill -9 `ps -ef|grep 'python3.7'|awk '{print $2}'`
+
+    if [ $run_mode = "mp" -a -d mylog ]; then
+        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 $@
+_train
+

benchmark/run_det.sh

+# 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: ./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 -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_res18_db_v2.0 det_r50_vd_east)
+fp_item_list=(fp32)
+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_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_det.sh ${run_mode} ${bs_item} ${fp_item} 10 ${model_mode} 
+            sleep 60
+            done
+      done
+done
+
+

configs/det/ch_PP-OCRv2/ch_PP-OCR_det_cml.yml

@@ -141,6 +141,7 @@ Train:
           img_mode: BGR
           channel_first: False
       - DetLabelEncode: # Class handling label
+      - CopyPaste:
       - IaaAugment:
           augmenter_args:
             - { 'type': Fliplr, 'args': { 'p': 0.5 } }

configs/det/ch_PP-OCRv2/ch_PP-OCR_det_distill.yml

@@ -68,8 +68,7 @@ Loss:
       ohem_ratio: 3
   - DistillationDBLoss:
       weight: 1.0
-      model_name_list: ["Student", "Teacher"]
-      # key: maps
+      model_name_list: ["Student"]
       name: DBLoss
       balance_loss: true
       main_loss_type: DiceLoss
@@ -116,6 +115,7 @@ Train:
           img_mode: BGR
           channel_first: False
       - DetLabelEncode: # Class handling label
+      - CopyPaste:
       - IaaAugment:
           augmenter_args:
             - { 'type': Fliplr, 'args': { 'p': 0.5 } }

configs/det/ch_PP-OCRv2/ch_PP-OCR_det_dml.yml

@@ -118,6 +118,7 @@ Train:
           img_mode: BGR
           channel_first: False
       - DetLabelEncode: # Class handling label
+      - CopyPaste:
       - IaaAugment:
           augmenter_args:
             - { 'type': Fliplr, 'args': { 'p': 0.5 } }

configs/det/det_r50_vd_east.yml

@@ -8,7 +8,7 @@ Global:
   # evaluation is run every 5000 iterations after the 4000th iteration
   eval_batch_step: [4000, 5000]
   cal_metric_during_train: False
-  pretrained_model: ./pretrain_models/ResNet50_vd_pretrained/
+  pretrained_model: ./pretrain_models/ResNet50_vd_pretrained
   checkpoints: 
   save_inference_dir:
   use_visualdl: False

configs/det/det_res18_db_v2.0.yml

+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

configs/e2e/e2e_r50_vd_pg.yml

@@ -94,7 +94,7 @@ Eval:
     label_file_list: [./train_data/total_text/test/test.txt]
     transforms:
       - DecodeImage: # load image
-          img_mode: RGB
+          img_mode: BGR
           channel_first: False
       - E2ELabelEncodeTest:
       - E2EResizeForTest:
@@ -111,4 +111,4 @@ Eval:
     shuffle: False
     drop_last: False
     batch_size_per_card: 1 # must be 1
-    num_workers: 2
\ No newline at end of file
+    num_workers: 2

configs/rec/ch_PP-OCRv2/ch_PP-OCRv2_rec_enhanced_ctc_loss.yml

+Global:
+  debug: false
+  use_gpu: true
+  epoch_num: 800
+  log_smooth_window: 20
+  print_batch_step: 10
+  save_model_dir: ./output/rec_mobile_pp-OCRv2_enhanced_ctc_loss
+  save_epoch_step: 3
+  eval_batch_step: [0, 2000]
+  cal_metric_during_train: true
+  pretrained_model:
+  checkpoints:
+  save_inference_dir:
+  use_visualdl: false
+  infer_img: doc/imgs_words/ch/word_1.jpg
+  character_dict_path: ppocr/utils/ppocr_keys_v1.txt
+  character_type: ch
+  max_text_length: 25
+  infer_mode: false
+  use_space_char: true
+  distributed: true
+  save_res_path: ./output/rec/predicts_mobile_pp-OCRv2_enhanced_ctc_loss.txt
+
+
+Optimizer:
+  name: Adam
+  beta1: 0.9
+  beta2: 0.999
+  lr:
+    name: Piecewise
+    decay_epochs : [700, 800]
+    values : [0.001, 0.0001]
+    warmup_epoch: 5
+  regularizer:
+    name: L2
+    factor: 2.0e-05
+
+
+Architecture:
+  model_type: rec
+  algorithm: CRNN
+  Transform:
+  Backbone:
+    name: MobileNetV1Enhance
+    scale: 0.5
+  Neck:
+    name: SequenceEncoder
+    encoder_type: rnn
+    hidden_size: 64
+  Head:
+    name: CTCHead
+    mid_channels: 96
+    fc_decay: 0.00002
+    return_feats: true
+
+Loss:
+  name: CombinedLoss
+  loss_config_list:
+  - CTCLoss:
+      use_focal_loss: false
+      weight: 1.0
+  - CenterLoss:
+      weight: 0.05
+      num_classes: 6625
+      feat_dim: 96
+      init_center: false
+      center_file_path: "./train_center.pkl"
+  # you can also try to add ace loss on your own dataset
+  # - ACELoss:
+  #     weight: 0.1
+
+PostProcess:
+  name: CTCLabelDecode
+
+Metric:
+  name: RecMetric
+  main_indicator: acc
+
+Train:
+  dataset:
+    name: SimpleDataSet
+    data_dir: ./train_data/
+    label_file_list:
+    - ./train_data/train_list.txt
+    transforms:
+    - DecodeImage:
+        img_mode: BGR
+        channel_first: false
+    - RecAug:
+    - CTCLabelEncode:
+    - RecResizeImg:
+        image_shape: [3, 32, 320]
+    - KeepKeys:
+        keep_keys:
+        - image
+        - label
+        - length
+        - label_ace
+  loader:
+    shuffle: true
+    batch_size_per_card: 128
+    drop_last: true
+    num_workers: 8
+Eval:
+  dataset:
+    name: SimpleDataSet
+    data_dir: ./train_data
+    label_file_list:
+    - ./train_data/val_list.txt
+    transforms:
+    - DecodeImage:
+        img_mode: BGR
+        channel_first: false
+    - CTCLabelEncode:
+    - RecResizeImg:
+        image_shape: [3, 32, 320]
+    - KeepKeys:
+        keep_keys:
+        - image
+        - label
+        - length
+  loader:
+    shuffle: false
+    drop_last: false
+    batch_size_per_card: 128
+    num_workers: 8

configs/rec/rec_resnet_stn_bilstm_att.yml

+Global:
+  use_gpu: True
+  epoch_num: 400
+  log_smooth_window: 20
+  print_batch_step: 10
+  save_model_dir: ./output/rec/seed
+  save_epoch_step: 3
+  # evaluation is run every 5000 iterations after the 4000th iteration
+  eval_batch_step: [0, 2000]
+  cal_metric_during_train: True
+  pretrained_model:
+  checkpoints:
+  save_inference_dir:
+  use_visualdl: False
+  infer_img: doc/imgs_words_en/word_10.png
+  # for data or label process
+  character_dict_path: 
+  character_type: EN_symbol
+  max_text_length: 100
+  infer_mode: False
+  use_space_char: False
+  save_res_path: ./output/rec/predicts_seed.txt
+
+
+Optimizer:
+  name: Adadelta
+  weight_deacy: 0.0
+  momentum: 0.9
+  lr:
+    name: Piecewise
+    decay_epochs: [4,5,8]
+    values: [1.0, 0.1, 0.01]
+  regularizer:
+    name: 'L2'
+    factor: 2.0e-05
+
+
+Architecture:
+  model_type: rec
+  algorithm: SEED
+  Transform:
+    name: STN_ON
+    tps_inputsize: [32, 64]
+    tps_outputsize: [32, 100]
+    num_control_points: 20
+    tps_margins: [0.05,0.05]
+    stn_activation: none
+  Backbone:
+    name: ResNet_ASTER
+  Head:
+    name: AsterHead  # AttentionHead
+    sDim: 512
+    attDim: 512
+    max_len_labels: 100
+
+Loss:
+  name: AsterLoss
+
+PostProcess:
+  name: SEEDLabelDecode
+
+Metric:
+  name: RecMetric
+  main_indicator: acc
+  is_filter: True
+
+Train:
+  dataset:
+    name: LMDBDataSet
+    data_dir: ./train_data/data_lmdb_release/training/
+    transforms:
+      - Fasttext:
+          path: "./cc.en.300.bin"
+      - DecodeImage: # load image
+          img_mode: BGR
+          channel_first: False
+      - SEEDLabelEncode: # Class handling label
+      - RecResizeImg:
+          character_type: en
+          image_shape: [3, 64, 256]
+          padding: False
+      - KeepKeys:
+          keep_keys: ['image', 'label', 'length', 'fast_label'] # dataloader will return list in this order
+  loader:
+    shuffle: True
+    batch_size_per_card: 256
+    drop_last: True
+    num_workers: 6
+
+Eval:
+  dataset:
+    name: LMDBDataSet
+    data_dir: ./train_data/data_lmdb_release/evaluation/
+    transforms:
+      - DecodeImage: # load image
+          img_mode: BGR
+          channel_first: False
+      - SEEDLabelEncode: # Class handling label
+      - RecResizeImg:
+          character_type: en
+          image_shape: [3, 64, 256]
+          padding: False
+      - KeepKeys:
+          keep_keys: ['image', 'label', 'length'] # dataloader will return list in this order
+  loader:
+    shuffle: False
+    drop_last: True
+    batch_size_per_card: 256
+    num_workers: 4

deploy/cpp_infer/src/ocr_rec.cpp

@@ -112,12 +112,16 @@ void CRNNRecognizer::LoadModel(const std::string &model_dir) {
           1 << 20, 10, 3,
           precision,
           false, false);
+
       std::map<std::string, std::vector<int>> min_input_shape = {
-          {"x", {1, 3, 32, 10}}};
+          {"x", {1, 3, 32, 10}},
+          {"lstm_0.tmp_0", {10, 1, 96}}};
       std::map<std::string, std::vector<int>> max_input_shape = {
-          {"x", {1, 3, 32, 2000}}};
+          {"x", {1, 3, 32, 2000}},
+          {"lstm_0.tmp_0", {1000, 1, 96}}};
       std::map<std::string, std::vector<int>> opt_input_shape = {
-          {"x", {1, 3, 32, 320}}};
+          {"x", {1, 3, 32, 320}},
+          {"lstm_0.tmp_0", {25, 1, 96}}};
 
       config.SetTRTDynamicShapeInfo(min_input_shape, max_input_shape,
                                     opt_input_shape);
@@ -139,7 +143,7 @@ void CRNNRecognizer::LoadModel(const std::string &model_dir) {
   config.SwitchIrOptim(true);
 
   config.EnableMemoryOptim();
-  config.DisableGlogInfo();
+//   config.DisableGlogInfo();
 
   this->predictor_ = CreatePredictor(config);
 }

deploy/slim/prune/sensitivity_anal.py

@@ -110,25 +110,42 @@ def main(config, device, logger, vdl_writer):
         logger.info("metric['hmean']: {}".format(metric['hmean']))
         return metric['hmean']
 
-    params_sensitive = pruner.sensitive(
-        eval_func=eval_fn,
-        sen_file="./sen.pickle",
-        skip_vars=[
-            "conv2d_57.w_0", "conv2d_transpose_2.w_0", "conv2d_transpose_3.w_0"
-        ])
-
-    logger.info(
-        "The sensitivity analysis results of model parameters saved in sen.pickle"
-    )
-    # calculate pruned params's ratio
-    params_sensitive = pruner._get_ratios_by_loss(params_sensitive, loss=0.02)
-    for key in params_sensitive.keys():
-        logger.info("{}, {}".format(key, params_sensitive[key]))
-
-    #params_sensitive = {}
-    #for param in model.parameters():
-    #    if 'transpose' not in param.name and 'linear' not in param.name:
-    #        params_sensitive[param.name] = 0.1  
+    run_sensitive_analysis = False
+    """
+    run_sensitive_analysis=True: 
+        Automatically compute the sensitivities of convolutions in a model. 
+        The sensitivity of a convolution is the losses of accuracy on test dataset in 
+        differenct pruned ratios. The sensitivities can be used to get a group of best 
+        ratios with some condition.
+    
+    run_sensitive_analysis=False: 
+        Set prune trim ratio to a fixed value, such as 10%. The larger the value, 
+        the more convolution weights will be cropped.
+
+    """
+
+    if run_sensitive_analysis:
+        params_sensitive = pruner.sensitive(
+            eval_func=eval_fn,
+            sen_file="./deploy/slim/prune/sen.pickle",
+            skip_vars=[
+                "conv2d_57.w_0", "conv2d_transpose_2.w_0",
+                "conv2d_transpose_3.w_0"
+            ])
+        logger.info(
+            "The sensitivity analysis results of model parameters saved in sen.pickle"
+        )
+        # calculate pruned params's ratio
+        params_sensitive = pruner._get_ratios_by_loss(
+            params_sensitive, loss=0.02)
+        for key in params_sensitive.keys():
+            logger.info("{}, {}".format(key, params_sensitive[key]))
+    else:
+        params_sensitive = {}
+        for param in model.parameters():
+            if 'transpose' not in param.name and 'linear' not in param.name:
+                # set prune ratio as 10%. The larger the value, the more convolution weights will be cropped
+                params_sensitive[param.name] = 0.1
 
     plan = pruner.prune_vars(params_sensitive, [0])
 

doc/doc_ch/algorithm_overview.md

@@ -50,6 +50,7 @@ PaddleOCR基于动态图开源的文本识别算法列表：
 - [x]  SRN([paper](https://arxiv.org/abs/2003.12294))
 - [x]  NRTR([paper](https://arxiv.org/abs/1806.00926v2))
 - [x]  SAR([paper](https://arxiv.org/abs/1811.00751v2))
+- [x]  SEED([paper](https://arxiv.org/pdf/2005.10977.pdf))
 
 参考[DTRB](https://arxiv.org/abs/1904.01906) 文字识别训练和评估流程，使用MJSynth和SynthText两个文字识别数据集训练，在IIIT, SVT, IC03, IC13, IC15, SVTP, CUTE数据集上进行评估，算法效果如下：
 
@@ -66,5 +67,5 @@ PaddleOCR基于动态图开源的文本识别算法列表：
 |SRN|Resnet50_vd_fpn| 88.52% | rec_r50fpn_vd_none_srn | [下载链接](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/rec_r50_vd_srn_train.tar) |
 |NRTR|NRTR_MTB| 84.3% | rec_mtb_nrtr | [下载链接](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/rec_mtb_nrtr_train.tar) |
 |SAR|Resnet31| 87.2% | rec_r31_sar | [下载链接](https://paddleocr.bj.bcebos.com/dygraph_v2.1/rec/rec_r31_sar_train.tar) |
-
+|SEED| Aster_Resnet | 85.2% | rec_resnet_stn_bilstm_att | [下载链接](https://paddleocr.bj.bcebos.com/dygraph_v2.1/rec/rec_resnet_stn_bilstm_att.tar)|
 PaddleOCR文本识别算法的训练和使用请参考文档教程中[模型训练/评估中的文本识别部分](./recognition.md)。

doc/doc_ch/enhanced_ctc_loss.md

+# Enhanced CTC Loss
+
+在OCR识别中， CRNN是一种在工业界广泛使用的文字识别算法。 在训练阶段，其采用CTCLoss来计算网络损失； 在推理阶段，其采用CTCDecode来获得解码结果。虽然CRNN算法在实际业务中被证明能够获得很好的识别效果， 然而用户对识别准确率的要求却是无止境的，如何进一步提升文字识别的准确率呢？ 本文以CTCLoss为切人点，分别从难例挖掘、 多任务学习、 Metric Learning 3个不同的角度探索了CTCLoss的改进融合方案，提出了EnhancedCTCLoss，其包括如下3个组成部分： Focal-CTC Loss，A-CTC Loss， C-CTC Loss。
+
+## 1. Focal-CTC Loss
+Focal Loss 出自论文《Focal Loss for Dense Object Detection》, 该loss最先提出的时候主要是为了解决one-stage目标检测中正负样本比例严重失衡的问题。该损失函数降低了大量简单负样本在训练中所占的权重，也可理解为一种困难样本挖掘。
+其损失函数形式如下：
+<div align="center"> 
+<img src="./focal_loss_formula.png" width = "600" /> 
+</div>
+ 
+其中,  y' 是经过激活函数的输出，取值在0-1之间。其在原始的交叉熵损失的基础上加了一个调制系数（1 – y’)^ &gamma;和平衡因子&alpha;。 当&alpha; = 1，y=1时，其损失函数与交叉熵损失的对比如下图所示:   
+<div align="center"> 
+<img src="./focal_loss_image.png" width = "600" /> 
+</div>
+
+从上图可以看到, 当&gamma;> 0时，调整系数（1-y’）^&gamma; 赋予易分类样本损失一个更小的权重，使得网络更关注于困难的、错分的样本。 调整因子&gamma;用于调节简单样本权重降低的速率，当&gamma;为0时即为交叉熵损失函数，当&gamma;增加时，调整因子的影响也会随之增大。实验发现&gamma;为2是最优。平衡因子&alpha;用来平衡正负样本本身的比例不均，文中&alpha;取0.25。
+
+对于经典的CTC算法，假设某个特征序列（f<sub>1</sub>, f<sub>2</sub>, ......f<sub>t</sub>), 经过CTC解码之后结果等于label的概率为y’, 则CTC解码结果不为label的概率即为（1-y’)；不难发现 CTCLoss值和y’有如下关系：
+<div align="center"> 
+<img src="./equation_ctcloss.png" width = "250" /> 
+</div>
+
+结合Focal Loss的思想，赋予困难样本较大的权重，简单样本较小的权重，可以使网络更加聚焦于对困难样本的挖掘，进一步提升识别的准确率，由此我们提出了Focal-CTC Loss； 其定义如下所示：
+<div align="center"> 
+<img src="./equation_focal_ctc.png" width = "500" /> 
+</div>
+
+实验中，&gamma;取值为2, &alpha;= 1, 具体实现见:  [rec_ctc_loss.py](../../ppocr/losses/rec_ctc_loss.py)
+
+## 2. A-CTC Loss
+A-CTC Loss是CTC Loss + ACE Loss的简称。 其中ACE Loss出自论文< Aggregation Cross-Entropy for Sequence Recognition>.  ACE Loss相比于CTCLoss，主要有如下两点优势: 
++ ACE Loss能够解决2-D文本的识别问题;  CTCLoss只能够处理1-D文本
++ ACE Loss 在时间复杂度和空间复杂度上优于CTC loss
+
+前人总结的OCR识别算法的优劣如下图所示：
+<div align="center">
+<img src="./rec_algo_compare.png" width = "1000" /> 
+</div>
+ 
+虽然ACELoss确实如上图所说，可以处理2D预测，在内存占用及推理速度方面具备优势，但在实践过程中，我们发现单独使用ACE Loss,  识别效果并不如CTCLoss.  因此，我们尝试将CTCLoss和ACELoss进行组合，同时以CTCLoss为主，将ACELoss 定位为一个辅助监督loss。 这一尝试收到了效果，在我们内部的实验数据集上，相比单独使用CTCLoss，识别准确率可以提升1%左右。
+A_CTC Loss定义如下:  
+<div align="center">
+<img src="./equation_a_ctc.png" width = "300" /> 
+</div>
+
+实验中，λ = 0.1.  ACE loss实现代码见:  [ace_loss.py](../../ppocr/losses/ace_loss.py)
+
+## 3. C-CTC Loss
+C-CTC Loss是CTC Loss + Center Loss的简称。 其中Center Loss出自论文 < A Discriminative Feature Learning Approach for Deep Face Recognition>.  最早用于人脸识别任务，用于增大累间距离，减小类内距离,  是Metric Learning领域一种较早的、也比较常用的一种算法。 
+在中文OCR识别任务中，通过对badcase分析， 我们发现中文识别的一大难点是相似字符多，容易误识。 由此我们想到是否可以借鉴Metric Learing的想法， 增大相似字符的类间距，从而提高识别准确率。然而，MetricLearning主要用于图像识别领域，训练数据的标签为一个固定的值；而对于OCR识别来说，其本质上是一个序列识别任务，特征和label之间并不具有显式的对齐关系，因此两者如何结合依然是一个值得探索的方向。
+通过尝试Arcmargin, Cosmargin等方法， 我们最终发现Centerloss 有助于进一步提升识别的准确率。C_CTC Loss定义如下：
+<div align="center">
+<img src="./equation_c_ctc.png" width = "300" /> 
+</div>
+
+实验中，我们设置λ=0.25. center_loss实现代码见:  [center_loss.py](../../ppocr/losses/center_loss.py)
+
+值得一提的是， 在C-CTC Loss中，选择随机初始化Center并不能够带来明显的提升. 我们的Center初始化方法如下：
++ 基于原始的CTCLoss， 训练得到一个网络N
++ 挑选出训练集中，识别完全正确的部分, 组成集合G
++ 将G中的每个样本送入网络，进行前向计算， 提取最后一个FC层的输入（即feature）及其经过argmax计算的结果（即index）之间的对应关系
++ 将相同index的feature进行聚合，计算平均值，得到各自字符的初始center. 
+
+以配置文件`configs/rec/ch_PP-OCRv2/ch_PP-OCRv2_rec.yml`为例， center提取命令如下所示:
+```
+python tools/export_center.py -c configs/rec/ch_PP-OCRv2/ch_PP-OCRv2_rec.yml -o  Global.pretrained_model: "./output/rec_mobile_pp-OCRv2/best_accuracy"
+```
+运行完后，会在PaddleOCR主目录下生成`train_center.pkl`.
+
+## 4. 实验
+对于上述的三种方案，我们基于百度内部数据集进行了训练、评测，实验情况如下表所示：
+|algorithm| Focal_CTC | A_CTC | C-CTC |
+|:------| :------| ------: | :------: |
+|gain| +0.3% | +0.7% | +1.7% | 
+
+基于上述实验结论，我们在PP-OCRv2中，采用了C-CTC的策略。 值得一提的是，由于PP-OCRv2 处理的是6625个中文字符的识别任务，字符集比较大，形似字较多，所以在该任务上C-CTC 方案带来的提升较大。 但如果换做其他OCR识别任务，结论可能会有所不同。大家可以尝试Focal-CTC，A-CTC, C-CTC以及组合方案EnhancedCTC，相信会带来不同程度的提升效果。
+统一的融合方案见如下文件：  [rec_enhanced_ctc_loss.py](../../ppocr/losses/rec_enhanced_ctc_loss.py)