一、.轻量级文本识别创新大赛简介

比赛地址：轻量级文本识别创新大赛

1.赛题介绍

近年来，深度学习技术在很多方向都取得了巨大的成功，尤其是计算机视觉领域，已经实现了广泛的落地和应用，但由于深度神经网络计算复杂度高，模型参数量大，限制了其在一些场景和设备上进行部署，特别是在移动嵌入式设备的部署。因此，对于模型大小的需求也变得越来越重要。

OCR文本识别任务作为计算机视觉领域的核心问题之一，旨在找出图像中所有的文字信息，进而确定它们的位置和内容。由于文字有不同的外观、形状、尺度和姿态，加上成像时的光照、遮挡等因素干扰，OCR文本识别一直是计算机视觉领域极具挑战性的问题之一，但是在现实世界中又具有庞大的应用基础。

本次挑战赛以 轻量级文字识别技术创新 为主题，提供大规模的中文场景文字识别数据，要求参赛选手兼顾准确率指标与模型大小，并重点考察选手的网络结构设计与训练调优能力。本次挑战赛旨在为研究者提供学术交流平台，进一步推动中文场景文字识别算法与技术的突破。

2.竞赛数据集

本次赛题数据集共包括12万张真实图片，其中10万张图片作为训练集，A榜，B榜各1万张图片作为测试集。 本次赛题数据集既有来自街景的图片（例如店铺标牌、地标等），也有来自文档、网图截取的图片。










二、环境设置

PaddleOCR https://github.com/paddlepaddle/PaddleOCR 是一款全宇宙最强的用的OCR工具库，开箱即用，速度杠杠的。

# 从gitee上下载PaddleOCR代码，也可以从GitHub链接下载
!git clone https://gitee.com/paddlepaddle/PaddleOCR.git
# 升级pip
!pip install -U pip 
# 安装依赖
%cd ~/PaddleOCR
%pip install -r requirements.txt

%cd ~/PaddleOCR/
!tree   -L 1

/home/aistudio/PaddleOCR
.
├── benchmark
├── configs
├── deploy
├── doc
├── inference
├── __init__.py
├── LICENSE
├── log
├── MANIFEST.in
├── output
├── paddleocr.py
├── ppocr
├── PPOCRLabel
├── ppstructure
├── README_ch.md
├── README.md
├── requirements.txt
├── setup.py
├── StyleText
├── test_tipc
├── tools
├── train_rec.log
└── train.sh

13 directories, 10 files

三、数据准备

据悉train数据集共10万张，解压，并划分出10000张作为测试集。

1.数据下载解压

#  解压缩数据集
%cd ~
!unzip -qa data/data92473/A榜测试数据集.zip -d data/
!unzip -qa data/data92473/训练数据集.zip -d data/

/home/aistudio

# 使用命令查看训练数据文件夹下数据量是否是10万张
!cd ~/data/训练数据集/TrainImages  &&  ls -l | grep "^-" | wc -l

100000

2. 数据集划分

# 读取数据列表文件
import pandas as pd
%cd ~
data_label=pd.read_csv('data/训练数据集/LabelTrain.txt',sep='\t')
data_label.head()

/home/aistudio
Train_000000.jpg光环
0Train_000001.jpg主营：热干面 原汤面
1Train_000002.jpg创亿
2Train_000003.jpg810
3Train_000004.jpg香花桥社区（街道）
4Train_000005.jpg藍色经典
# 对数据列表文件进行划分
%cd ~/data/训练数据集/
print(data_label.shape)
train=data_label[:90000]
val=data_label[89999:]
train.to_csv('train.txt',sep='\t',header=None,index=None)
val.to_csv('val.txt',sep='\t',header=None,index=None)

/home/aistudio/data/训练数据集
(99999, 2)

# 查看数量
print(train.shape)
print(val.shape)

(90000, 2)
(10000, 2)

!head ~/data/训练数据集/val.txt

Train_090000.jpg	运行时间
Train_090001.jpg	无座
Train_090002.jpg	中国进出口银行
Train_090003.jpg	被人抱紧
Train_090004.jpg	与
Train_090005.jpg	展开
Train_090006.jpg	金”
Train_090007.jpg	欢迎光临
Train_090008.jpg	在2日内到有效
Train_090009.jpg	车服务公司

!head ~/data/训练数据集/train.txt

Train_000001.jpg	主营：热干面 原汤面
Train_000002.jpg	创亿
Train_000003.jpg	810
Train_000004.jpg	香花桥社区（街道）
Train_000005.jpg	藍色经典
Train_000006.jpg	承接：国内外针 梭织订单  欢迎来料加工
Train_000007.jpg	PHOENIXTEA
Train_000008.jpg	电话：13640839583
Train_000009.jpg	B7
Train_000010.jpg	申滨路

四、配置训练参数

以PaddleOCR/configs/rec/ch_ppocr_v2.0/rec_chinese_lite_train_v2.0.yml为基准进行配置

1.配置模型网络

使用CRNN算法，backbone是MobileNetV3，损失函数是CTCLoss

Architecture:
  model_type: rec
  algorithm: CRNN
  Transform:
  Backbone:
    name: MobileNetV3
    scale: 0.5
    model_name: small
    small_stride: [1, 2, 2, 2]
  Neck:
    name: SequenceEncoder
    encoder_type: rnn
    hidden_size: 48
  Head:
    name: CTCHead
    fc_decay: 0.00001

2.配置数据

对Train.data_dir, Train.label_file_list, Eval.data_dir, Eval.label_file_list进行配置

Train:
  dataset:
    name: SimpleDataSet
    data_dir: /home/aistudio/data/训练数据集/TrainImages
    label_file_list: ["/home/aistudio/data/训练数据集/train.txt"]
...
...

Eval:
  dataset:
    name: SimpleDataSet
    data_dir: /home/aistudio/data/训练数据集/TrainImages
    label_file_list: ["/home/aistudio/data/训练数据集/val.txt"]

3. 显卡、评估设置

use_gpu、cal_metric_during_train分别是GPU、评估开关

Global:
  use_gpu: false             # true 使用GPU
  .....
  cal_metric_during_train: False   # true 打开评估

4. 多线程任务

Train.loader.num_workers:4
Eval.loader.num_workers: 4

5.完整配置

Global:
  use_gpu: true
  epoch_num: 500
  log_smooth_window: 20
  print_batch_step: 10
  save_model_dir: ./output/rec_chinese_lite_v2.0
  save_epoch_step: 3
  # evaluation is run every 5000 iterations after the 4000th iteration
  eval_batch_step: [0, 200]
  cal_metric_during_train: True
  pretrained_model:
  checkpoints: 
  save_inference_dir:
  use_visualdl: True
  infer_img: doc/imgs_words/ch/word_1.jpg
  # for data or label process
  character_dict_path: ppocr/utils/ppocr_keys_v1.txt
  max_text_length: 25
  infer_mode: False
  use_space_char: True
  # 结果保存文件
  save_res_path: ./output/rec/predicts_chinese_lite_v2.0.txt


Optimizer:
  name: Adam
  beta1: 0.9
  beta2: 0.999
  lr:
    name: Cosine
    learning_rate: 0.001
    warmup_epoch: 5
  regularizer:
    name: 'L2'
    factor: 0.00001

Architecture:
  model_type: rec
  algorithm: CRNN
  Transform:
  Backbone:
    name: MobileNetV3
    scale: 0.5
    model_name: small
    small_stride: [1, 2, 2, 2]
  Neck:
    name: SequenceEncoder
    encoder_type: rnn
    hidden_size: 48
  Head:
    name: CTCHead
    fc_decay: 0.00001

Loss:
  name: CTCLoss

PostProcess:
  name: CTCLabelDecode

Metric:
  name: RecMetric
  main_indicator: acc

Train:
  dataset:
    name: SimpleDataSet
    data_dir: /home/aistudio/data/训练数据集/TrainImages
    label_file_list: ["/home/aistudio/data/训练数据集/train.txt"]
    transforms:
      - DecodeImage: # load image
          img_mode: BGR
          channel_first: False
      - RecAug: 
      - CTCLabelEncode: # Class handling label
      - RecResizeImg:
          image_shape: [3, 32, 320]
      - KeepKeys:
          keep_keys: ['image', 'label', 'length'] # dataloader will return list in this order
  loader:
    shuffle: True
    batch_size_per_card: 800
    drop_last: True
    num_workers: 4

Eval:
  dataset:
    name: SimpleDataSet
    data_dir: /home/aistudio/data/训练数据集/TrainImages
    label_file_list: ["/home/aistudio/data/训练数据集/val.txt"]
    transforms:
      - DecodeImage: # load image
          img_mode: BGR
          channel_first: False
      - CTCLabelEncode: # Class handling label
      - RecResizeImg:
          image_shape: [3, 32, 320]
      - KeepKeys:
          keep_keys: ['image', 'label', 'length'] # dataloader will return list in this order
  loader:
    shuffle: False
    drop_last: False
    batch_size_per_card: 512
    num_workers: 4

五、训练

%cd ~/PaddleOCR/
!python tools/train.py -c ./configs/rec/ch_ppocr_v2.0/rec_chinese_lite_train_v2.0.yml

1.选择合适的batch size

2.训练日志

[2021/12/30 08:35:35] root INFO: Initialize indexs of datasets:['/home/aistudio/data/训练数据集/train.txt']
[2021/12/30 08:36:09] root INFO: epoch: [72/500], iter: 7960, lr: 0.000958, loss: 8.251744, acc: 0.413749, norm_edit_dis: 0.679774, reader_cost: 0.71643 s, batch_cost: 1.08854 s, samples: 7200, ips: 661.43391
[2021/12/30 08:36:35] root INFO: epoch: [72/500], iter: 7970, lr: 0.000958, loss: 8.213228, acc: 0.414374, norm_edit_dis: 0.681107, reader_cost: 0.01124 s, batch_cost: 0.41975 s, samples: 8000, ips: 1905.90655
[2021/12/30 08:37:02] root INFO: epoch: [72/500], iter: 7980, lr: 0.000957, loss: 8.017796, acc: 0.415624, norm_edit_dis: 0.680611, reader_cost: 0.01079 s, batch_cost: 0.41746 s, samples: 8000, ips: 1916.34344
[2021/12/30 08:37:28] root INFO: epoch: [72/500], iter: 7990, lr: 0.000957, loss: 8.049696, acc: 0.421249, norm_edit_dis: 0.688969, reader_cost: 0.01128 s, batch_cost: 0.41961 s, samples: 8000, ips: 1906.51883
[2021/12/30 08:37:55] root INFO: epoch: [72/500], iter: 8000, lr: 0.000957, loss: 7.952358, acc: 0.421249, norm_edit_dis: 0.688969, reader_cost: 0.01156 s, batch_cost: 0.41822 s, samples: 8000, ips: 1912.88346
eval model:: 100%|██████████████████████████████| 20/20 [00:28<00:00,  1.17s/it]
[2021/12/30 08:38:23] root INFO: cur metric, acc: 0.6908999309100069, norm_edit_dis: 0.8983265023824453, fps: 4436.456909249088
[2021/12/30 08:38:23] root INFO: save best model is to ./output/rec_chinese_lite_v2.0/best_accuracy

3. visualdl可视化

  • 本地安装visualdl pip install visualdl
  • 下载日志至本地
  • 启动visualdl可视化 visualdl --logdir ./
  • 打开浏览器查看 http://localhost:8040/

六、模型评估

# GPU 评估， Global.checkpoints 为待测权重
%cd ~/PaddleOCR/
!python  -m paddle.distributed.launch tools/eval.py -c configs/rec/ch_ppocr_v2.0/rec_chinese_lite_train_v2.0.yml \
    -o Global.checkpoints=./output/rec_chinese_lite_v2.0/best_accuracy

/home/aistudio/PaddleOCR
-----------  Configuration Arguments -----------
backend: auto
elastic_server: None
force: False
gpus: None
heter_devices: 
heter_worker_num: None
heter_workers: 
host: None
http_port: None
ips: 127.0.0.1
job_id: None
log_dir: log
np: None
nproc_per_node: None
run_mode: None
scale: 0
server_num: None
servers: 
training_script: tools/eval.py
training_script_args: ['-c', 'configs/rec/ch_ppocr_v2.0/rec_chinese_lite_train_v2.0.yml', '-o', 'Global.checkpoints=./output/rec_chinese_lite_v2.0/best_accuracy']
worker_num: None
workers: 
------------------------------------------------
WARNING 2021-12-30 08:49:47,054 launch.py:423] Not found distinct arguments and compiled with cuda or xpu. Default use collective mode
launch train in GPU mode!
INFO 2021-12-30 08:49:47,057 launch_utils.py:528] Local start 1 processes. First process distributed environment info (Only For Debug): 
    +=======================================================================================+
    |                        Distributed Envs                      Value                    |
    +---------------------------------------------------------------------------------------+
    |                       PADDLE_TRAINER_ID                        0                      |
    |                 PADDLE_CURRENT_ENDPOINT                 127.0.0.1:44352               |
    |                     PADDLE_TRAINERS_NUM                        1                      |
    |                PADDLE_TRAINER_ENDPOINTS                 127.0.0.1:44352               |
    |                     PADDLE_RANK_IN_NODE                        0                      |
    |                 PADDLE_LOCAL_DEVICE_IDS                        0                      |
    |                 PADDLE_WORLD_DEVICE_IDS                        0                      |
    |                     FLAGS_selected_gpus                        0                      |
    |             FLAGS_selected_accelerators                        0                      |
    +=======================================================================================+

INFO 2021-12-30 08:49:47,057 launch_utils.py:532] details abouts PADDLE_TRAINER_ENDPOINTS can be found in log/endpoints.log, and detail running logs maybe found in log/workerlog.0
launch proc_id:57223 idx:0
[2021/12/30 08:49:48] root INFO: Architecture : 
[2021/12/30 08:49:48] root INFO:     Backbone : 
[2021/12/30 08:49:48] root INFO:         model_name : small
[2021/12/30 08:49:48] root INFO:         name : MobileNetV3
[2021/12/30 08:49:48] root INFO:         scale : 0.5
[2021/12/30 08:49:48] root INFO:         small_stride : [1, 2, 2, 2]
[2021/12/30 08:49:48] root INFO:     Head : 
[2021/12/30 08:49:48] root INFO:         fc_decay : 1e-05
[2021/12/30 08:49:48] root INFO:         name : CTCHead
[2021/12/30 08:49:48] root INFO:     Neck : 
[2021/12/30 08:49:48] root INFO:         encoder_type : rnn
[2021/12/30 08:49:48] root INFO:         hidden_size : 48
[2021/12/30 08:49:48] root INFO:         name : SequenceEncoder
[2021/12/30 08:49:48] root INFO:     Transform : None
[2021/12/30 08:49:48] root INFO:     algorithm : CRNN
[2021/12/30 08:49:48] root INFO:     model_type : rec
[2021/12/30 08:49:48] root INFO: Eval : 
[2021/12/30 08:49:48] root INFO:     dataset : 
[2021/12/30 08:49:48] root INFO:         data_dir : /home/aistudio/data/训练数据集/TrainImages
[2021/12/30 08:49:48] root INFO:         label_file_list : ['/home/aistudio/data/训练数据集/val.txt']
[2021/12/30 08:49:48] root INFO:         name : SimpleDataSet
[2021/12/30 08:49:48] root INFO:         transforms : 
[2021/12/30 08:49:48] root INFO:             DecodeImage : 
[2021/12/30 08:49:48] root INFO:                 channel_first : False
[2021/12/30 08:49:48] root INFO:                 img_mode : BGR
[2021/12/30 08:49:48] root INFO:             CTCLabelEncode : None
[2021/12/30 08:49:48] root INFO:             RecResizeImg : 
[2021/12/30 08:49:48] root INFO:                 image_shape : [3, 32, 320]
[2021/12/30 08:49:48] root INFO:             KeepKeys : 
[2021/12/30 08:49:48] root INFO:                 keep_keys : ['image', 'label', 'length']
[2021/12/30 08:49:48] root INFO:     loader : 
[2021/12/30 08:49:48] root INFO:         batch_size_per_card : 512
[2021/12/30 08:49:48] root INFO:         drop_last : False
[2021/12/30 08:49:48] root INFO:         num_workers : 4
[2021/12/30 08:49:48] root INFO:         shuffle : False
[2021/12/30 08:49:48] root INFO: Global : 
[2021/12/30 08:49:48] root INFO:     cal_metric_during_train : True
[2021/12/30 08:49:48] root INFO:     character_dict_path : ppocr/utils/ppocr_keys_v1.txt
[2021/12/30 08:49:48] root INFO:     checkpoints : ./output/rec_chinese_lite_v2.0/best_accuracy
[2021/12/30 08:49:48] root INFO:     debug : False
[2021/12/30 08:49:48] root INFO:     distributed : False
[2021/12/30 08:49:48] root INFO:     epoch_num : 500
[2021/12/30 08:49:48] root INFO:     eval_batch_step : [0, 200]
[2021/12/30 08:49:48] root INFO:     infer_img : doc/imgs_words/ch/word_1.jpg
[2021/12/30 08:49:48] root INFO:     infer_mode : False
[2021/12/30 08:49:48] root INFO:     log_smooth_window : 20
[2021/12/30 08:49:48] root INFO:     max_text_length : 25
[2021/12/30 08:49:48] root INFO:     pretrained_model : None
[2021/12/30 08:49:48] root INFO:     print_batch_step : 10
[2021/12/30 08:49:48] root INFO:     save_epoch_step : 3
[2021/12/30 08:49:48] root INFO:     save_inference_dir : None
[2021/12/30 08:49:48] root INFO:     save_model_dir : ./output/rec_chinese_lite_v2.0
[2021/12/30 08:49:48] root INFO:     save_res_path : ./output/rec/predicts_chinese_lite_v2.0.txt
[2021/12/30 08:49:48] root INFO:     use_gpu : True
[2021/12/30 08:49:48] root INFO:     use_space_char : True
[2021/12/30 08:49:48] root INFO:     use_visualdl : True
[2021/12/30 08:49:48] root INFO: Loss : 
[2021/12/30 08:49:48] root INFO:     name : CTCLoss
[2021/12/30 08:49:48] root INFO: Metric : 
[2021/12/30 08:49:48] root INFO:     main_indicator : acc
[2021/12/30 08:49:48] root INFO:     name : RecMetric
[2021/12/30 08:49:48] root INFO: Optimizer : 
[2021/12/30 08:49:48] root INFO:     beta1 : 0.9
[2021/12/30 08:49:48] root INFO:     beta2 : 0.999
[2021/12/30 08:49:48] root INFO:     lr : 
[2021/12/30 08:49:48] root INFO:         learning_rate : 0.001
[2021/12/30 08:49:48] root INFO:         name : Cosine
[2021/12/30 08:49:48] root INFO:         warmup_epoch : 5
[2021/12/30 08:49:48] root INFO:     name : Adam
[2021/12/30 08:49:48] root INFO:     regularizer : 
[2021/12/30 08:49:48] root INFO:         factor : 1e-05
[2021/12/30 08:49:48] root INFO:         name : L2
[2021/12/30 08:49:48] root INFO: PostProcess : 
[2021/12/30 08:49:48] root INFO:     name : CTCLabelDecode
[2021/12/30 08:49:48] root INFO: Train : 
[2021/12/30 08:49:48] root INFO:     dataset : 
[2021/12/30 08:49:48] root INFO:         data_dir : /home/aistudio/data/训练数据集/TrainImages
[2021/12/30 08:49:48] root INFO:         label_file_list : ['/home/aistudio/data/训练数据集/train.txt']
[2021/12/30 08:49:48] root INFO:         name : SimpleDataSet
[2021/12/30 08:49:48] root INFO:         transforms : 
[2021/12/30 08:49:48] root INFO:             DecodeImage : 
[2021/12/30 08:49:48] root INFO:                 channel_first : False
[2021/12/30 08:49:48] root INFO:                 img_mode : BGR
[2021/12/30 08:49:48] root INFO:             RecAug : None
[2021/12/30 08:49:48] root INFO:             CTCLabelEncode : None
[2021/12/30 08:49:48] root INFO:             RecResizeImg : 
[2021/12/30 08:49:48] root INFO:                 image_shape : [3, 32, 320]
[2021/12/30 08:49:48] root INFO:             KeepKeys : 
[2021/12/30 08:49:48] root INFO:                 keep_keys : ['image', 'label', 'length']
[2021/12/30 08:49:48] root INFO:     loader : 
[2021/12/30 08:49:48] root INFO:         batch_size_per_card : 800
[2021/12/30 08:49:48] root INFO:         drop_last : True
[2021/12/30 08:49:48] root INFO:         num_workers : 4
[2021/12/30 08:49:48] root INFO:         shuffle : True
[2021/12/30 08:49:48] root INFO: profiler_options : None
[2021/12/30 08:49:48] root INFO: train with paddle 2.2.1 and device CUDAPlace(0)
[2021/12/30 08:49:48] root INFO: Initialize indexs of datasets:['/home/aistudio/data/训练数据集/val.txt']
W1230 08:49:48.777415 57223 device_context.cc:447] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 10.1, Runtime API Version: 10.1
W1230 08:49:48.781112 57223 device_context.cc:465] device: 0, cuDNN Version: 7.6.
[2021/12/30 08:49:52] root INFO: resume from ./output/rec_chinese_lite_v2.0/best_accuracy
[2021/12/30 08:49:52] root INFO: metric in ckpt ***************
[2021/12/30 08:49:52] root INFO: acc:0.7108999289100071
[2021/12/30 08:49:52] root INFO: norm_edit_dis:0.9080091791996405
[2021/12/30 08:49:52] root INFO: fps:4172.613918329738
[2021/12/30 08:49:52] root INFO: best_epoch:74
[2021/12/30 08:49:52] root INFO: start_epoch:75

eval model::   0%|          | 0/20 [00:00<?, ?it/s]
eval model::   5%|▌         | 1/20 [00:02<00:48,  2.56s/it]
eval model::  10%|█         | 2/20 [00:03<00:39,  2.19s/it]
eval model::  15%|█▌        | 3/20 [00:05<00:32,  1.94s/it]
eval model::  20%|██        | 4/20 [00:06<00:28,  1.76s/it]
eval model::  25%|██▌       | 5/20 [00:07<00:24,  1.65s/it]
eval model::  30%|███       | 6/20 [00:09<00:21,  1.57s/it]
eval model::  35%|███▌      | 7/20 [00:10<00:19,  1.51s/it]
eval model::  40%|████      | 8/20 [00:12<00:17,  1.47s/it]
eval model::  45%|████▌     | 9/20 [00:13<00:15,  1.43s/it]
eval model::  50%|█████     | 10/20 [00:14<00:14,  1.40s/it]
eval model::  55%|█████▌    | 11/20 [00:16<00:12,  1.39s/it]
eval model::  60%|██████    | 12/20 [00:17<00:10,  1.37s/it]
eval model::  65%|██████▌   | 13/20 [00:18<00:09,  1.36s/it]
eval model::  70%|███████   | 14/20 [00:20<00:08,  1.36s/it]
eval model::  75%|███████▌  | 15/20 [00:21<00:06,  1.36s/it]
eval model::  80%|████████  | 16/20 [00:22<00:05,  1.35s/it]
eval model::  85%|████████▌ | 17/20 [00:24<00:04,  1.35s/it]
eval model::  90%|█████████ | 18/20 [00:25<00:02,  1.35s/it]
eval model::  95%|█████████▌| 19/20 [00:26<00:01,  1.35s/it]
eval model:: 100%|██████████| 20/20 [00:27<00:00,  1.16s/it]
[2021/12/30 08:50:20] root INFO: metric eval ***************
[2021/12/30 08:50:20] root INFO: acc:0.7108999289100071
[2021/12/30 08:50:20] root INFO: norm_edit_dis:0.9080091791996405
[2021/12/30 08:50:20] root INFO: fps:4556.466541403698
INFO 2021-12-30 08:50:23,113 launch.py:311] Local processes completed.

七、结果预测

预测脚本使用预测训练好的模型，并将结果保存成txt格式，可以直接送到比赛提交入口测评，文件默认保存在output/rec/predicts_chinese_lite_v2.0.txt

!python tools/infer_rec.py -c configs/rec/ch_ppocr_v2.0/rec_chinese_lite_train_v2.0.yml \
    -o Global.infer_img="/home/aistudio/data/A榜测试数据集/TestAImages" \
    Global.pretrained_model="./output/rec_chinese_lite_v2.0/latest.pdparams"

预测日志

[2021/12/30 08:52:55] root INFO: 	 result: 门	0.5737782
[2021/12/30 08:52:55] root INFO: infer_img: /home/aistudio/data/A榜测试数据集/TestAImages/TestA_000704.jpg
[2021/12/30 08:52:55] root INFO: 	 result: 红豆米营美码	0.5796438
[2021/12/30 08:52:55] root INFO: infer_img: /home/aistudio/data/A榜测试数据集/TestAImages/TestA_000705.jpg
[2021/12/30 08:52:55] root INFO: 	 result: 少年成长加油站	0.95518243
[2021/12/30 08:52:55] root INFO: infer_img: /home/aistudio/data/A榜测试数据集/TestAImages/TestA_000706.jpg
[2021/12/30 08:52:55] root INFO: 	 result: 展板奖牌铜牌活动策划	0.6614779
[2021/12/30 08:52:55] root INFO: infer_img: /home/aistudio/data/A榜测试数据集/TestAImages/TestA_000707.jpg
[2021/12/30 08:52:55] root INFO: 	 result: USG铁工代室	0.61983764
[2021/12/30 08:52:55] root INFO: infer_img: /home/aistudio/data/A榜测试数据集/TestAImages/TestA_000708.jpg
[2021/12/30 08:52:55] root INFO: 	 result: 家部	0.10427441
[2021/12/30 08:52:55] root INFO: infer_img: /home/aistudio/data/A榜测试数据集/TestAImages/TestA_000709.jpg
...
...

八、基于预测引擎的预测

1.模型大小限制

约束性条件1：模型总大小不超过10MB（以.pdmodel和.pdiparams文件非压缩状态磁盘占用空间之和为准）；

2.解决办法

训练过程中保存的模型是checkpoints模型，保存的只有模型的参数，多用于恢复训练等。实际上，此处的约束条件限制的是inference 模型的大小。inference 模型一般是模型训练，把模型结构和模型参数保存在文件中的固化模型，多用于预测部署场景。与checkpoints模型相比，inference 模型会额外保存模型的结构信息，在预测部署、加速推理上性能优越，灵活方便，适合于实际系统集成，模型大小也会小一些。

# 静态模型导出
%cd ~/PaddleOCR/
!python tools/export_model.py -c configs/rec/ch_ppocr_v2.0/rec_chinese_lite_train_v2.0.yml -o Global.pretrained_model=./output/rec_chinese_lite_v2.0/latest.pdparams  Global.save_inference_dir=./inference/rec_inference/

/home/aistudio/PaddleOCR
W1230 13:19:16.513371   531 device_context.cc:447] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 11.0, Runtime API Version: 10.1
W1230 13:19:16.517957   531 device_context.cc:465] device: 0, cuDNN Version: 7.6.
[2021/12/30 13:19:19] root INFO: load pretrain successful from ./output/rec_chinese_lite_v2.0/latest
[2021/12/30 13:19:21] root INFO: inference model is saved to ./inference/rec_inference/inference

%cd ~/PaddleOCR/
!du -sh ./inference/rec_inference/

/home/aistudio/PaddleOCR
5.2M	./inference/rec_inference/
  • 可以看到，当前训练使用的CRNN算法导出inference后，仅有5.2M。
  • 导出的inference模型也可以用来预测，预测逻辑如下代码所示。
# 使用导出静态模型预测
%cd ~/PaddleOCR/
。


```python
# 使用导出静态模型预测
%cd ~/PaddleOCR/
!python3.7 tools/infer/predict_rec.py  --rec_model_dir=./inference/rec_inference/  --image_dir="/home/aistudio/data/A榜测试数据集/TestAImages"

预测日志

[2021/12/30 13:20:08] root INFO: Predicts of /home/aistudio/data/A榜测试数据集/TestAImages/TestA_000001.jpg:('MJ', 0.2357887)
[2021/12/30 13:20:08] root INFO: Predicts of /home/aistudio/data/A榜测试数据集/TestAImages/TestA_000002.jpg:('中门', 0.7167614)
[2021/12/30 13:20:08] root INFO: Predicts of /home/aistudio/data/A榜测试数据集/TestAImages/TestA_000003.jpg:('黄焖鸡米饭', 0.7325407)
[2021/12/30 13:20:08] root INFO: Predicts of /home/aistudio/data/A榜测试数据集/TestAImages/TestA_000004.jpg:('加行', 0.06699998)
[2021/12/30 13:20:08] root INFO: Predicts of /home/aistudio/data/A榜测试数据集/TestAImages/TestA_000005.jpg:('学商烤面航', 0.40579563)
[2021/12/30 13:20:08] root INFO: Predicts of /home/aistudio/data/A榜测试数据集/TestAImages/TestA_000006.jpg:('绿村装机 滋光彩机 CP口出国', 0.38243735)
[2021/12/30 13:20:08] root INFO: Predicts of /home/aistudio/data/A榜测试数据集/TestAImages/TestA_000007.jpg:('有酸锁 四好吃', 0.38957664)
[2021/12/30 13:20:08] root INFO: Predicts of /home/aistudio/data/A榜测试数据集/TestAImages/TestA_000008.jpg:('婚汽中海', 0.36037388)
[2021/12/30 13:20:08] root INFO: Predicts of /home/aistudio/data/A榜测试数据集/TestAImages/TestA_000009.jpg:('L', 0.25453746)
[2021/12/30 13:20:08] root INFO: Predicts of /home/aistudio/data/A榜测试数据集/TestAImages/TestA_000010.jpg:('清女装', 0.79736567)
[2021/12/30 13:20:08] root INFO: Predicts of /home/aistudio/data/A榜测试数据集/TestAImages/TestA_000011.jpg:('幼小数学视食', 0.50577885)
...
...

九、提交

预测结果保存到配置文件指定的 output/rec/predicts_chinese_lite_v2.0.txt文件，可直接提交即可。

# 深度学习 # 计算机视觉 # 人工智能 # paddlepaddle
Logo
百度飞桨AI Studio社区

学大模型，用大模型上飞桨星河社区！每天8点V100G算力免费领！免费领取ERNIE 4.0 100w Token >>>

更多推荐

cover

利用Amazon Bedrock畅玩Claude 3等多种领先模型，抢占AI高地(体验倒计时4小时)

avatar 百度飞桨星河社区
cover

LangChain的RAG实践

avatar 百度飞桨星河社区
cover

RAG综述汇总

avatar 百度飞桨星河社区