Merge branch 'dyg_db' of https://github.com/wangjingyeye/PaddleOCR into dyg_db

README_ch.md

@@ -71,6 +71,8 @@ PaddleOCR旨在打造一套丰富、领先、且实用的OCR工具库，助力
 ## 《动手学OCR》电子书
 - [《动手学OCR》电子书📚](./doc/doc_ch/ocr_book.md)
 
+## 场景应用
+- PaddleOCR场景应用覆盖通用，制造、金融、交通行业的主要OCR垂类应用，在PP-OCR、PP-Structure的通用能力基础之上，以notebook的形式展示利用场景数据微调、模型优化方法、数据增广等内容，为开发者快速落地OCR应用提供示范与启发。详情可查看[README](./applications)。
 
 <a name="开源社区"></a>
 ## 开源社区

applications/PCB字符识别/PCB字符识别.md

@@ -206,7 +206,11 @@ Eval.dataset.transforms.DetResizeForTest:  尺寸
         limit_type: 'min'
 ```
 
-然后执行评估代码
+如需获取已训练模型，请扫码填写问卷，加入PaddleOCR官方交流群获取全部OCR垂类模型下载链接、《动手学OCR》电子书等全套OCR学习资料🎁
+<div align="left">
+<img src="https://ai-studio-static-online.cdn.bcebos.com/dd721099bd50478f9d5fb13d8dd00fad69c22d6848244fd3a1d3980d7fefc63e"  width = "150" height = "150" />
+</div>
+将下载或训练完成的模型放置在对应目录下即可完成模型评估。
 
 
 ```python

applications/README.md

+# 场景应用
+
+PaddleOCR场景应用覆盖通用，制造、金融、交通行业的主要OCR垂类应用，在PP-OCR、PP-Structure的通用能力基础之上，以notebook的形式展示利用场景数据微调、模型优化方法、数据增广等内容，为开发者快速落地OCR应用提供示范与启发。
+
+> 如需下载全部垂类模型，可以扫描下方二维码，关注公众号填写问卷后，加入PaddleOCR官方交流群获取20G OCR学习大礼包（内含《动手学OCR》电子书、课程回放视频、前沿论文等重磅资料）
+
+<div align="center">
+<img src="https://ai-studio-static-online.cdn.bcebos.com/dd721099bd50478f9d5fb13d8dd00fad69c22d6848244fd3a1d3980d7fefc63e"  width = "150" height = "150" />
+</div>
+
+
+> 如果您是企业开发者且未在下述场景中找到合适的方案，可以填写[OCR应用合作调研问卷](https://paddle.wjx.cn/vj/QwF7GKw.aspx)，免费与官方团队展开不同层次的合作，包括但不限于问题抽象、确定技术方案、项目答疑、共同研发等。如果您已经使用PaddleOCR落地项目，也可以填写此问卷，与飞桨平台共同宣传推广，提升企业技术品宣。期待您的提交！
+
+## 通用
+
+| 类别                                              | 亮点     | 类别       | 亮点         |
+| ------------------------------------------------- | -------- | ---------- | ------------ |
+| [高精度中文识别模型SVTR](./高精度中文识别模型.md) | 新增模型 | 手写体识别 | 新增字形支持 |
+
+## 制造
+
+| 类别                                                         | 亮点                           | 类别                                        | 亮点                 |
+| ------------------------------------------------------------ | ------------------------------ | ------------------------------------------- | -------------------- |
+| [数码管识别](./光功率计数码管字符识别/光功率计数码管字符识别.md) | 数码管数据合成、漏识别调优     | 电表识别                                    | 大分辨率图像检测调优 |
+| [液晶屏读数识别](./液晶屏读数识别.md)                        | 检测模型蒸馏、Serving部署      | [PCB文字识别](./PCB字符识别/PCB字符识别.md) | 小尺寸文本检测与识别 |
+| [包装生产日期](./包装生产日期识别.md)                        | 点阵字符合成、过曝过暗文字识别 | 液晶屏缺陷检测                              | 非文字字符识别       |
+
+## 金融
+
+| 类别                           | 亮点                     | 类别         | 亮点                  |
+| ------------------------------ | ------------------------ | ------------ | --------------------- |
+| [表单VQA](./多模态表单识别.md) | 多模态通用表单结构化提取 | 通用卡证识别 | 通用结构化提取        |
+| 增值税发票                     | 尽请期待                 | 身份证识别   | 结构化提取、图像阴影  |
+| 印章检测与识别                 | 端到端弯曲文本识别       | 合同比对     | 密集文本检测、NLP串联 |
+
+## 交通
+
+| 类别                            | 亮点                           | 类别       | 亮点     |
+| ------------------------------- | ------------------------------ | ---------- | -------- |
+| [车牌识别](./轻量级车牌识别.md) | 多角度图像、轻量模型、端侧部署 | 快递单识别 | 尽请期待 |
+| 驾驶证/行驶证识别               | 尽请期待                       |            |          |
\ No newline at end of file

applications/光功率计数码管字符识别/corpus/digital.txt

+46.39
+40.08
+89.52
+-71.93
+23.19
+-81.02
+-34.09
+05.87
+-67.80
+-51.56
+-34.58
+37.91
+56.98
+29.01
+-90.13
+35.55
+66.07
+-90.35
+-50.93
+42.42
+21.40
+-30.99
+-71.78
+25.60
+-48.69
+-72.28
+-17.55
+-99.93
+-47.35
+-64.89
+-31.28
+-90.01
+05.17
+30.91
+30.56
+-06.90
+79.05
+67.74
+-32.31
+94.22
+28.75
+51.03
+-58.96

applications/光功率计数码管字符识别/光功率计数码管字符识别.md

+# 光功率计数码管字符识别
+
+本案例将使用OCR技术自动识别光功率计显示屏文字，通过本章您可以掌握：
+
+- PaddleOCR快速使用
+- 数据合成方法
+- 数据挖掘方法
+- 基于现有数据微调
+
+## 1. 背景介绍
+
+光功率计（optical power meter ）是指用于测量绝对光功率或通过一段光纤的光功率相对损耗的仪器。在光纤系统中，测量光功率是最基本的，非常像电子学中的万用表；在光纤测量中，光功率计是重负荷常用表。
+
+<img src="https://bkimg.cdn.bcebos.com/pic/a08b87d6277f9e2f999f5e3e1c30e924b899f35a?x-bce-process=image/watermark,image_d2F0ZXIvYmFpa2U5Mg==,g_7,xp_5,yp_5/format,f_auto" width="400">
+
+目前光功率计缺少将数据直接输出的功能，需要人工读数。这一项工作单调重复，如果可以使用机器替代人工，将节约大量成本。针对上述问题，希望通过摄像头拍照->智能读数的方式高效地完成此任务。
+
+为实现智能读数，通常会采取文本检测+文本识别的方案：
+
+第一步，使用文本检测模型定位出光功率计中的数字部分；
+
+第二步，使用文本识别模型获得准确的数字和单位信息。
+
+本项目主要介绍如何完成第二步文本识别部分，包括：真实评估集的建立、训练数据的合成、基于 PP-OCRv3 和 SVTR_Tiny 两个模型进行训练，以及评估和推理。
+
+本项目难点如下：
+
+- 光功率计数码管字符数据较少，难以获取。
+- 数码管中小数点占像素较少，容易漏识别。
+
+针对以上问题， 本例选用 PP-OCRv3 和 SVTR_Tiny 两个高精度模型训练，同时提供了真实数据挖掘案例和数据合成案例。基于 PP-OCRv3 模型，在构建的真实评估集上精度从 52% 提升至 72%，SVTR_Tiny 模型精度可达到 78.9%。
+
+aistudio项目链接: [光功率计数码管字符识别](https://aistudio.baidu.com/aistudio/projectdetail/4049044?contributionType=1)
+
+## 2. PaddleOCR 快速使用
+
+PaddleOCR 旨在打造一套丰富、领先、且实用的OCR工具库，助力开发者训练出更好的模型，并应用落地。
+
+![](https://github.com/PaddlePaddle/PaddleOCR/raw/release/2.5/doc/imgs_results/ch_ppocr_mobile_v2.0/test_add_91.jpg)
+
+
+官方提供了适用于通用场景的高精轻量模型，首先使用官方提供的 PP-OCRv3 模型预测图片，验证下当前模型在光功率计场景上的效果。
+
+- 准备环境
+
+```
+python3 -m pip install -U pip
+python3 -m pip install paddleocr
+```
+
+
+- 测试效果
+
+测试图：
+
+![](https://ai-studio-static-online.cdn.bcebos.com/8dca91f016884e16ad9216d416da72ea08190f97d87b4be883f15079b7ebab9a)
+
+
+```
+paddleocr --lang=ch --det=Fase --image_dir=data
+```
+
+得到如下测试结果：
+
+```
+('.7000', 0.6885431408882141)
+```
+
+发现数字识别较准，然而对负号和小数点识别不准确。 由于PP-OCRv3的训练数据大多为通用场景数据，在特定的场景上效果可能不够好。因此需要基于场景数据进行微调。
+
+下面就主要介绍如何在光功率计（数码管）场景上微调训练。
+
+
+## 3. 开始训练
+
+### 3.1 数据准备
+
+特定的工业场景往往很难获取开源的真实数据集，光功率计也是如此。在实际工业场景中，可以通过摄像头采集的方法收集大量真实数据，本例中重点介绍数据合成方法和真实数据挖掘方法，如何利用有限的数据优化模型精度。
+
+数据集分为两个部分：合成数据，真实数据， 其中合成数据由 text_renderer 工具批量生成得到， 真实数据通过爬虫等方式在百度图片中搜索并使用 PPOCRLabel 标注得到。
+
+
+- 合成数据
+
+本例中数据合成工具使用的是 [text_renderer](https://github.com/Sanster/text_renderer)， 该工具可以合成用于文本识别训练的文本行数据:
+
+![](https://github.com/oh-my-ocr/text_renderer/raw/master/example_data/effect_layout_image/char_spacing_compact.jpg)
+
+![](https://github.com/oh-my-ocr/text_renderer/raw/master/example_data/effect_layout_image/color_image.jpg)
+
+
+```
+export https_proxy=http://172.19.57.45:3128
+git clone https://github.com/oh-my-ocr/text_renderer
+```
+
+```
+import os
+python3 setup.py develop
+python3 -m pip install -r docker/requirements.txt
+python3 main.py \
+    --config example_data/example.py \
+    --dataset img \
+    --num_processes 2 \
+    --log_period 10
+```
+
+给定字体和语料，就可以合成较为丰富样式的文本行数据。 光功率计识别场景，目标是正确识别数码管文本，因此需要收集部分数码管字体，训练语料，用于合成文本识别数据。
+
+将收集好的语料存放在 example_data 路径下：
+
+```
+ln -s ./fonts/DS* text_renderer/example_data/font/
+ln -s ./corpus/digital.txt text_renderer/example_data/text/
+```
+
+修改 text_renderer/example_data/font_list/font_list.txt ,选择需要的字体开始合成：
+
+```
+python3 main.py \
+    --config example_data/digital_example.py \
+    --dataset img \
+    --num_processes 2 \
+    --log_period 10
+```
+
+合成图片会被存在目录 text_renderer/example_data/digital/chn_data 下
+
+查看合成的数据样例：
+
+![](https://ai-studio-static-online.cdn.bcebos.com/7d5774a273f84efba5b9ce7fd3f86e9ef24b6473e046444db69fa3ca20ac0986)
+
+
+- 真实数据挖掘
+
+模型训练需要使用真实数据作为评价指标，否则很容易过拟合到简单的合成数据中。没有开源数据的情况下，可以利用部分无标注数据+标注工具获得真实数据。
+
+
+1. 数据搜集
+
+使用[爬虫工具](https://github.com/Joeclinton1/google-images-download.git)获得无标注数据
+
+2. [PPOCRLabel](https://github.com/PaddlePaddle/PaddleOCR/tree/release/2.5/PPOCRLabel) 完成半自动标注
+
+PPOCRLabel是一款适用于OCR领域的半自动化图形标注工具，内置PP-OCR模型对数据自动标注和重新识别。使用Python3和PyQT5编写，支持矩形框标注、表格标注、不规则文本标注、关键信息标注模式，导出格式可直接用于PaddleOCR检测和识别模型的训练。
+
+![](https://github.com/PaddlePaddle/PaddleOCR/raw/release/2.5/PPOCRLabel/data/gif/steps_en.gif)
+
+
+收集完数据后就可以进行分配了，验证集中一般都是真实数据，训练集中包含合成数据+真实数据。本例中标注了155张图片，其中训练集和验证集的数目为100和55。
+
+
+最终 `data` 文件夹应包含以下几部分：
+
+```
+|-data
+  |- synth_train.txt
+  |- real_train.txt
+  |- real_eval.txt
+  |- synthetic_data
+      |- word_001.png
+      |- word_002.jpg
+      |- word_003.jpg
+      | ...
+  |- real_data
+      |- word_001.png
+      |- word_002.jpg
+      |- word_003.jpg
+      | ...
+  ...
+```
+
+### 3.2 模型选择
+
+本案例提供了2种文本识别模型：PP-OCRv3 识别模型 和 SVTR_Tiny：
+
+[PP-OCRv3 识别模型](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.5/doc/doc_ch/PP-OCRv3_introduction.md)：PP-OCRv3的识别模块是基于文本识别算法SVTR优化。SVTR不再采用RNN结构，通过引入Transformers结构更加有效地挖掘文本行图像的上下文信息，从而提升文本识别能力。并进行了一系列结构改进加速模型预测。
+
+[SVTR_Tiny](https://arxiv.org/abs/2205.00159):SVTR提出了一种用于场景文本识别的单视觉模型，该模型在patch-wise image tokenization框架内，完全摒弃了序列建模，在精度具有竞争力的前提下，模型参数量更少，速度更快。
+
+以上两个策略在自建中文数据集上的精度和速度对比如下：
+
+| ID | 策略 |  模型大小 | 精度 | 预测耗时（CPU + MKLDNN)|
+|-----|-----|--------|----| --- |
+| 01 | PP-OCRv2 | 8M | 74.8% | 8.54ms |
+| 02 | SVTR_Tiny | 21M | 80.1% | 97ms |
+| 03 | SVTR_LCNet(h32) | 12M | 71.9% | 6.6ms |
+| 04 | SVTR_LCNet(h48) | 12M | 73.98% | 7.6ms |
+| 05 | + GTC | 12M | 75.8% | 7.6ms |
+| 06 | + TextConAug | 12M | 76.3% | 7.6ms |
+| 07 | + TextRotNet | 12M | 76.9% | 7.6ms |
+| 08 | + UDML | 12M | 78.4% | 7.6ms |
+| 09 | + UIM | 12M | 79.4% | 7.6ms |
+
+
+### 3.3 开始训练
+
+首先下载 PaddleOCR 代码库
+
+```
+git clone -b release/2.5 https://github.com/PaddlePaddle/PaddleOCR.git
+```
+
+PaddleOCR提供了训练脚本、评估脚本和预测脚本，本节将以 PP-OCRv3 中文识别模型为例：
+
+**Step1：下载预训练模型**
+
+首先下载 pretrain model，您可以下载训练好的模型在自定义数据上进行finetune
+
+```
+cd PaddleOCR/
+# 下载PP-OCRv3 中文预训练模型
+wget -P ./pretrain_models/ https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_rec_train.tar
+# 解压模型参数
+cd pretrain_models
+tar -xf ch_PP-OCRv3_rec_train.tar && rm -rf ch_PP-OCRv3_rec_train.tar
+```
+
+**Step2：自定义字典文件**
+
+接下来需要提供一个字典（{word_dict_name}.txt），使模型在训练时，可以将所有出现的字符映射为字典的索引。
+
+因此字典需要包含所有希望被正确识别的字符，{word_dict_name}.txt需要写成如下格式，并以 `utf-8` 编码格式保存：
+
+```
+0
+1
+2
+3
+4
+5
+6
+7
+8
+9
+-
+.
+```
+
+word_dict.txt 每行有一个单字，将字符与数字索引映射在一起，“3.14” 将被映射成 [3, 11, 1, 4]
+
+* 内置字典
+
+PaddleOCR内置了一部分字典，可以按需使用。
+
+`ppocr/utils/ppocr_keys_v1.txt` 是一个包含6623个字符的中文字典
+
+`ppocr/utils/ic15_dict.txt` 是一个包含36个字符的英文字典
+
+* 自定义字典
+
+内置字典面向通用场景，具体的工业场景中，可能需要识别特殊字符，或者只需识别某几个字符，此时自定义字典会更提升模型精度。例如在光功率计场景中，需要识别数字和单位。
+
+遍历真实数据标签中的字符，制作字典`digital_dict.txt`如下所示：
+
+```
+-
+.
+0
+1
+2
+3
+4
+5
+6
+7
+8
+9
+B
+E
+F
+H
+L
+N
+T
+W
+d
+k
+m
+n
+o
+z
+```
+
+
+
+
+**Step3：修改配置文件**
+
+为了更好的使用预训练模型，训练推荐使用[ch_PP-OCRv3_rec_distillation.yml](../../configs/rec/PP-OCRv3/ch_PP-OCRv3_rec_distillation.yml)配置文件，并参考下列说明修改配置文件：
+
+以 `ch_PP-OCRv3_rec_distillation.yml` 为例：
+```
+Global:
+  ...
+  # 添加自定义字典，如修改字典请将路径指向新字典
+  character_dict_path: ppocr/utils/dict/digital_dict.txt
+  ...
+  # 识别空格
+  use_space_char: True
+
+
+Optimizer:
+  ...
+  # 添加学习率衰减策略
+  lr:
+    name: Cosine
+    learning_rate: 0.001
+  ...
+
+...
+
+Train:
+  dataset:
+    # 数据集格式，支持LMDBDataSet以及SimpleDataSet
+    name: SimpleDataSet
+    # 数据集路径
+    data_dir: ./data/
+    # 训练集标签文件
+    label_file_list:
+    - ./train_data/digital_img/digital_train.txt  #11w
+    - ./train_data/digital_img/real_train.txt     #100
+    - ./train_data/digital_img/dbm_img/dbm.txt    #3w
+    ratio_list:
+    - 0.3
+    - 1.0
+    - 1.0
+    transforms:
+      ...
+      - RecResizeImg:
+          # 修改 image_shape 以适应长文本
+          image_shape: [3, 48, 320]
+      ...
+  loader:
+    ...
+    # 单卡训练的batch_size
+    batch_size_per_card: 256
+    ...
+
+Eval:
+  dataset:
+    # 数据集格式，支持LMDBDataSet以及SimpleDataSet
+    name: SimpleDataSet
+    # 数据集路径
+    data_dir: ./data
+    # 验证集标签文件
+    label_file_list:
+    - ./train_data/digital_img/real_val.txt
+    transforms:
+      ...
+      - RecResizeImg:
+          # 修改 image_shape 以适应长文本
+          image_shape: [3, 48, 320]
+      ...
+  loader:
+    # 单卡验证的batch_size
+    batch_size_per_card: 256
+    ...
+```
+**注意，训练/预测/评估时的配置文件请务必与训练一致。**
+
+**Step4：启动训练**
+
+*如果您安装的是cpu版本，请将配置文件中的 `use_gpu` 字段修改为false*
+
+```
+# GPU训练 支持单卡，多卡训练
+# 训练数码管数据 训练日志会自动保存为 "{save_model_dir}" 下的train.log
+
+#单卡训练（训练周期长，不建议）
+python3 tools/train.py -c configs/rec/PP-OCRv3/ch_PP-OCRv3_rec_distillation.yml -o Global.pretrained_model=./pretrain_models/ch_PP-OCRv3_rec_train/best_accuracy
+
+#多卡训练，通过--gpus参数指定卡号
+python3 -m paddle.distributed.launch --gpus '0,1,2,3'  tools/train.py -c configs/rec/PP-OCRv3/ch_PP-OCRv3_rec_distillation.yml -o Global.pretrained_model=./pretrain_models/en_PP-OCRv3_rec_train/best_accuracy
+```
+
+
+PaddleOCR支持训练和评估交替进行, 可以在 `configs/rec/PP-OCRv3/ch_PP-OCRv3_rec_distillation.yml` 中修改 `eval_batch_step` 设置评估频率，默认每500个iter评估一次。评估过程中默认将最佳acc模型，保存为 `output/ch_PP-OCRv3_rec_distill/best_accuracy` 。
+
+如果验证集很大，测试将会比较耗时，建议减少评估次数，或训练完再进行评估。
+
+### SVTR_Tiny 训练
+
+SVTR_Tiny 训练步骤与上面一致，SVTR支持的配置和模型训练权重可以参考[算法介绍文档](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.5/doc/doc_ch/algorithm_rec_svtr.md)
+
+**Step1：下载预训练模型**
+
+```
+# 下载 SVTR_Tiny 中文识别预训练模型和配置文件
+wget https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/rec_svtr_tiny_none_ctc_ch_train.tar
+# 解压模型参数
+tar -xf rec_svtr_tiny_none_ctc_ch_train.tar && rm -rf rec_svtr_tiny_none_ctc_ch_train.tar
+```
+**Step2：自定义字典文件**
+
+字典依然使用自定义的 digital_dict.txt
+
+**Step3：修改配置文件**
+
+配置文件中对应修改字典路径和数据路径
+
+**Step4：启动训练**
+
+```
+## 单卡训练
+python tools/train.py -c rec_svtr_tiny_none_ctc_ch_train/rec_svtr_tiny_6local_6global_stn_ch.yml \
+           -o Global.pretrained_model=./rec_svtr_tiny_none_ctc_ch_train/best_accuracy
+```
+
+### 3.4 验证效果
+
+如需获取已训练模型，请扫码填写问卷，加入PaddleOCR官方交流群获取全部OCR垂类模型下载链接、《动手学OCR》电子书等全套OCR学习资料🎁
+<div align="left">
+<img src="https://ai-studio-static-online.cdn.bcebos.com/dd721099bd50478f9d5fb13d8dd00fad69c22d6848244fd3a1d3980d7fefc63e"  width = "150" height = "150" />
+</div>
+将下载或训练完成的模型放置在对应目录下即可完成模型推理
+
+* 指标评估
+
+训练中模型参数默认保存在`Global.save_model_dir`目录下。在评估指标时，需要设置`Global.checkpoints`指向保存的参数文件。评估数据集可以通过 `configs/rec/PP-OCRv3/ch_PP-OCRv3_rec_distillation.yml`  修改Eval中的 `label_file_path` 设置。
+
+```
+# GPU 评估， Global.checkpoints 为待测权重
+python3 -m paddle.distributed.launch --gpus '0' tools/eval.py -c configs/rec/PP-OCRv3/ch_PP-OCRv3_rec_distillation.yml -o Global.checkpoints={path/to/weights}/best_accuracy
+```
+
+* 测试识别效果
+
+使用 PaddleOCR 训练好的模型，可以通过以下脚本进行快速预测。
+
+默认预测图片存储在 `infer_img` 里，通过 `-o Global.checkpoints` 加载训练好的参数文件：
+
+根据配置文件中设置的 `save_model_dir` 和 `save_epoch_step` 字段，会有以下几种参数被保存下来：
+
+```
+output/rec/
+├── best_accuracy.pdopt  
+├── best_accuracy.pdparams  
+├── best_accuracy.states  
+├── config.yml  
+├── iter_epoch_3.pdopt  
+├── iter_epoch_3.pdparams  
+├── iter_epoch_3.states  
+├── latest.pdopt  
+├── latest.pdparams  
+├── latest.states  
+└── train.log
+```
+
+其中 best_accuracy.* 是评估集上的最优模型；iter_epoch_x.* 是以 `save_epoch_step` 为间隔保存下来的模型；latest.* 是最后一个epoch的模型。
+
+```
+# 预测英文结果
+python3 tools/infer_rec.py -c configs/rec/PP-OCRv3/ch_PP-OCRv3_rec_distillation.yml -o Global.pretrained_model={path/to/weights}/best_accuracy  Global.infer_img=test_digital.png
+```
+
+预测图片：
+
+![](https://ai-studio-static-online.cdn.bcebos.com/8dca91f016884e16ad9216d416da72ea08190f97d87b4be883f15079b7ebab9a)
+
+
+得到输入图像的预测结果：
+
+```
+infer_img: test_digital.png
+        result: ('-70.00', 0.9998967)
+```

applications/包装生产日期识别.md

+# 一种基于PaddleOCR的产品包装生产日期识别模型
+
+- [1. 项目介绍](#1-项目介绍)
+- [2. 环境搭建](#2-环境搭建)
+- [3. 数据准备](#3-数据准备)
+- [4. 直接使用PP-OCRv3模型评估](#4-直接使用PPOCRv3模型评估)
+- [5. 基于合成数据finetune](#5-基于合成数据finetune)
+  - [5.1 Text Renderer数据合成方法](#51-TextRenderer数据合成方法)
+    - [5.1.1 下载Text Renderer代码](#511-下载TextRenderer代码)
+    - [5.1.2 准备背景图片](#512-准备背景图片)
+    - [5.1.3 准备语料](#513-准备语料)
+    - [5.1.4 下载字体](#514-下载字体)
+    - [5.1.5 运行数据合成命令](#515-运行数据合成命令)
+  - [5.2 模型训练](#52-模型训练)
+- [6. 基于真实数据finetune](#6-基于真实数据finetune)
+  - [6.1 python爬虫获取数据](#61-python爬虫获取数据)
+  - [6.2 数据挖掘](#62-数据挖掘)
+  - [6.3 模型训练](#63-模型训练)
+- [7. 基于合成+真实数据finetune](#7-基于合成+真实数据finetune)
+
+
+## 1. 项目介绍
+
+产品包装生产日期是计算机视觉图像识别技术在工业场景中的一种应用。产品包装生产日期识别技术要求能够将产品生产日期从复杂背景中提取并识别出来，在物流管理、物资管理中得到广泛应用。
+
+![](https://ai-studio-static-online.cdn.bcebos.com/d9e0533cc1df47ffa3bbe99de9e42639a3ebfa5bce834bafb1ca4574bf9db684)
+
+
+- 项目难点
+
+1. 没有训练数据
+2. 图像质量层次不齐: 角度倾斜、图片模糊、光照不足、过曝等问题严重
+
+针对以上问题， 本例选用PP-OCRv3这一开源超轻量OCR系统进行包装产品生产日期识别系统的开发。直接使用PP-OCRv3进行评估的精度为62.99%。为提升识别精度，我们首先使用数据合成工具合成了3k数据，基于这部分数据进行finetune，识别精度提升至73.66%。由于合成数据与真实数据之间的分布存在差异，为进一步提升精度，我们使用网络爬虫配合数据挖掘策略得到了1k带标签的真实数据，基于真实数据finetune的精度为71.33%。最后，我们综合使用合成数据和真实数据进行finetune，将识别精度提升至86.99%。各策略的精度提升效果如下：
+
+| 策略 | 精度|
+| :--------------- | :-------- |
+| PP-OCRv3评估 | 62.99|
+| 合成数据finetune | 73.66|
+| 真实数据finetune | 71.33|
+| 真实+合成数据finetune | 86.99|
+
+AIStudio项目链接： [一种基于PaddleOCR的包装生产日期识别方法](https://aistudio.baidu.com/aistudio/projectdetail/4287736)
+
+## 2. 环境搭建
+
+本任务基于Aistudio完成, 具体环境如下：
+
+- 操作系统: Linux
+- PaddlePaddle: 2.3
+- PaddleOCR: Release/2.5
+- text_renderer: master
+
+下载PaddlleOCR代码并安装依赖库:
+```bash
+git clone -b dygraph https://gitee.com/paddlepaddle/PaddleOCR
+
+# 安装依赖库
+cd PaddleOCR
+pip install -r PaddleOCR/requirements.txt
+```
+
+## 3. 数据准备
+
+本项目使用人工预标注的300张图像作为测试集。
+
+部分数据示例如下：
+
+![](https://ai-studio-static-online.cdn.bcebos.com/39ff30e0ab0442579712255e6a9ea6b5271169c98e624e6eb2b8781f003bfea0)
+
+
+标签文件格式如下：
+```txt
+数据路径 标签（中间以制表符分隔）
+```
+
+|数据集类型|数量|
+|---|---|
+|测试集| 300|
+
+数据集[下载链接](https://aistudio.baidu.com/aistudio/datasetdetail/149770)，下载后可以通过下方命令解压:
+
+```bash
+tar -xvf data.tar
+mv data ${PaddleOCR_root}
+```
+
+数据解压后的文件结构如下：
+
+```shell
+PaddleOCR
+├── data
+│   ├── mining_images            # 挖掘的真实数据示例
+│   ├── mining_train.list        # 挖掘的真实数据文件列表
+│   ├── render_images            # 合成数据示例
+│   ├── render_train.list        # 合成数据文件列表
+│   ├── val                      # 测试集数据
+│   └── val.list                 # 测试集数据文件列表
+|   ├── bg                       # 合成数据所需背景图像
+│   └── corpus                   # 合成数据所需语料
+```
+
+## 4. 直接使用PP-OCRv3模型评估
+
+准备好测试数据后，可以使用PaddleOCR的PP-OCRv3模型进行识别。
+
+- 下载预训练模型
+
+首先需要下载PP-OCR v3中英文识别模型文件，下载链接可以在https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.5/doc/doc_ch/ppocr_introduction.md#6 获取，下载命令:
+
+```bash
+cd ${PaddleOCR_root}
+mkdir ckpt
+wget -nc -P ckpt https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_rec_train.tar
+pushd ckpt/
+tar -xvf ch_PP-OCRv3_rec_train.tar
+popd
+```
+
+- 模型评估
+
+使用以下命令进行PP-OCRv3评估:
+
+```bash
+python tools/eval.py -c configs/rec/PP-OCRv3/ch_PP-OCRv3_rec_distillation.yml \
+                         -o Global.checkpoints=ckpt/ch_PP-OCRv3_rec_train/best_accuracy \
+                         Eval.dataset.data_dir=./data \
+                         Eval.dataset.label_file_list=["./data/val.list"]
+
+```
+
+其中各参数含义如下：
+
+```bash
+-c: 指定使用的配置文件，ch_PP-OCRv3_rec_distillation.yml对应于OCRv3识别模型。
+-o: 覆盖配置文件中参数
+Global.checkpoints: 指定评估使用的模型文件路径
+Eval.dataset.data_dir: 指定评估数据集路径
+Eval.dataset.label_file_list: 指定评估数据集文件列表
+```
+
+## 5. 基于合成数据finetune
+
+### 5.1 Text Renderer数据合成方法
+
+#### 5.1.1 下载Text Renderer代码
+
+首先从github或gitee下载Text Renderer代码，并安装相关依赖。
+
+```bash
+git clone https://gitee.com/wowowoll/text_renderer.git
+
+# 安装依赖库
+cd text_renderer
+pip install -r requirements.txt
+```
+
+使用text renderer合成数据之前需要准备好背景图片、语料以及字体库，下面将逐一介绍各个步骤。
+
+#### 5.1.2 准备背景图片
+
+观察日常生活中常见的包装生产日期图片，我们可以发现其背景相对简单。为此我们可以从网上找一下图片，截取部分图像块作为背景图像。
+
+本项目已准备了部分图像作为背景图片，在第3部分完成数据准备后,可以得到我们准备好的背景图像，示例如下：
+
+![](https://ai-studio-static-online.cdn.bcebos.com/456ae2acb27d4a94896c478812aee0bc3551c703d7bd40c9be4dc983c7b3fc8a)
+
+
+
+背景图像存放于如下位置：
+
+```shell
+PaddleOCR
+├── data
+｜   ├── bg     # 合成数据所需背景图像
+```
+
+#### 5.1.3 准备语料
+
+观察测试集生产日期图像，我们可以知道如下数据有如下特点：
+1. 由年月日组成，中间可能以“/”、“-”、“:”、“.”或者空格间隔，也可能以汉字年月日分隔
+2. 有些生产日期包含在产品批号中，此时可能包含具体时间、英文字母或数字标识
+
+基于以上两点，我们编写语料生成脚本：
+
+```python
+import random
+from random import choice
+import os
+
+cropus_num = 2000 #设置语料数量
+
+def get_cropus(f):
+    # 随机生成年份
+    year = random.randint(0, 22)
+    # 随机生成月份
+    month = random.randint(1, 12)
+    # 随机生成日期
+    day_dict = {31: [1,3,5,7,8,10,12], 30: [4,6,9,11], 28: [2]}
+    for item in day_dict:
+        if month in day_dict[item]:
+            day = random.randint(0, item)
+    # 随机生成小时
+    hours = random.randint(0, 24)
+    # 随机生成分钟
+    minute = random.randint(0, 60)
+     # 随机生成秒数
+    second = random.randint(0, 60)
+
+    # 随机生成产品标识字符
+    length = random.randint(0, 6)
+    file_id = []
+    flag = 0
+    my_dict = [i for i in range(48,58)] + [j for j in range(40, 42)] + [k for k in range(65,90)]  # 大小写字母 + 括号
+
+    for i in range(1, length):
+        if flag:
+            if i == flag+2:  #括号匹配
+                file_id.append(')')
+                flag = 0
+                continue
+        sel = choice(my_dict)
+        if sel == 41:
+            continue
+        if sel == 40:
+            if i == 1 or i > length-3:
+                continue
+            flag = i
+        my_ascii = chr(sel)
+        file_id.append(my_ascii)
+    file_id_str = ''.join(file_id)
+
+    #随机生成产品标识字符
+    file_id2 = random.randint(0, 9)
+
+    rad = random.random()
+    if rad < 0.3:
+        f.write('20{:02d}{:02d}{:02d} {}'.format(year, month, day, file_id_str))
+    elif 0.3 < rad < 0.5:
+        f.write('20{:02d}年{:02d}月{:02d}日'.format(year, month, day))
+    elif 0.5 < rad < 0.7:
+        f.write('20{:02d}/{:02d}/{:02d}'.format(year, month, day))
+    elif 0.7 < rad < 0.8:
+        f.write('20{:02d}-{:02d}-{:02d}'.format(year, month, day))
+    elif 0.8 < rad < 0.9:
+        f.write('20{:02d}.{:02d}.{:02d}'.format(year, month, day))  
+    else:
+        f.write('{:02d}:{:02d}:{:02d} {:02d}'.format(hours, minute, second, file_id2))
+
+if __name__ == "__main__":
+    file_path = '/home/aistudio/text_renderer/my_data/cropus'
+    if not os.path.exists(file_path):
+        os.makedirs(file_path)
+    file_name = os.path.join(file_path, 'books.txt')
+    f = open(file_name, 'w')
+    for i in range(cropus_num):
+        get_cropus(f)
+        if i < cropus_num-1:
+            f.write('\n')
+
+    f.close()
+```
+
+本项目已准备了部分语料，在第3部分完成数据准备后,可以得到我们准备好的语料库，默认位置如下：
+
+```shell
+PaddleOCR
+├── data
+│   └── corpus              #合成数据所需语料
+```
+
+#### 5.1.4 下载字体
+
+观察包装生产日期，我们可以发现其使用的字体为点阵体。字体可以在如下网址下载：
+https://www.fonts.net.cn/fonts-en/tag-dianzhen-1.html
+
+本项目已准备了部分字体，在第3部分完成数据准备后,可以得到我们准备好的字体，默认位置如下：
+
+```shell
+PaddleOCR
+├── data
+│   └── fonts                #合成数据所需字体
+```
+
+下载好字体后，还需要在list文件中指定字体文件存放路径，脚本如下:
+
+```bash
+cd text_renderer/my_data/
+touch fonts.list
+ls /home/aistudio/PaddleOCR/data/fonts/* > fonts.list
+```
+
+#### 5.1.5 运行数据合成命令
+
+完成数据准备后，my_data文件结构如下：
+
+```shell
+my_data/
+├── cropus
+│   └── books.txt #语料库
+├── eng.txt    #字符列表
+└── fonts.list #字体列表
+```
+
+在运行合成数据命令之前，还有两处细节需要手动修改：
+1. 将默认配置文件`text_renderer/configs/default.yaml`中第9行enable的值设为`true`，即允许合成彩色图像。否则合成的都是灰度图。
+
+```yaml
+ # color boundary is in R,G,B format
+ font_color:
++  enable: true #false
+```
+
+2. 将`text_renderer/textrenderer/renderer.py`第184行作如下修改，取消padding。否则图片两端会有一些空白。
+
+```python
+padding = random.randint(s_bbox_width // 10, s_bbox_width // 8) #修改前
+padding = 0 #修改后
+```
+
+运行数据合成命令:
+
+```bash
+cd /home/aistudio/text_renderer/
+python main.py --num_img=3000 \
+                  --fonts_list='./my_data/fonts.list' \
+                  --corpus_dir "./my_data/cropus" \
+                  --corpus_mode "list" \
+                  --bg_dir "/home/aistudio/PaddleOCR/data/bg/" \
+                  --img_width 0
+```
+
+合成好的数据默认保存在`text_renderer/output`目录下，可进入该目录查看合成的数据。
+
+
+合成数据示例如下
+![](https://ai-studio-static-online.cdn.bcebos.com/d686a48d465a43d09fbee51924fdca42ee21c50e676646da8559fb9967b94185)
+
+数据合成好后，还需要生成如下格式的训练所需的标注文件，
+```
+图像路径 标签
+```
+
+使用如下脚本即可生成标注文件：
+
+```python
+import random
+
+abspath = '/home/aistudio/text_renderer/output/default/'
+
+#标注文件生成路径
+fout = open('./render_train.list', 'w', encoding='utf-8')
+
+with open('./output/default/tmp_labels.txt','r') as f:
+    lines = f.readlines()
+    for item in lines:
+        label = item[9:]
+        filename = item[:8] + '.jpg'
+        fout.write(abspath + filename + '\t' + label)
+
+    fout.close()
+```
+
+经过以上步骤，我们便完成了包装生产日期数据合成。
+数据位于`text_renderer/output`，标注文件位于`text_renderer/render_train.list`。
+
+本项目提供了生成好的数据供大家体验,完成步骤3的数据准备后，可得数据路径位于:
+
+```shell
+PaddleOCR
+├── data
+│   ├── render_images     # 合成数据示例
+│   ├── render_train.list   #合成数据文件列表
+```
+
+###  5.2 模型训练
+
+准备好合成数据后，我们可以使用以下命令，利用合成数据进行finetune:
+```bash
+cd ${PaddleOCR_root}
+python tools/train.py -c configs/rec/PP-OCRv3/ch_PP-OCRv3_rec_distillation.yml \
+                       -o Global.pretrained_model=./ckpt/ch_PP-OCRv3_rec_train/best_accuracy \
+                       Global.epoch_num=20 \
+                       Global.eval_batch_step='[0, 20]' \
+                       Train.dataset.data_dir=./data \
+                       Train.dataset.label_file_list=['./data/render_train.list'] \
+                       Train.loader.batch_size_per_card=64 \
+                       Eval.dataset.data_dir=./data \
+                       Eval.dataset.label_file_list=["./data/val.list"] \
+                       Eval.loader.batch_size_per_card=64
+
+```
+
+其中各参数含义如下：
+
+```txt
+-c: 指定使用的配置文件，ch_PP-OCRv3_rec_distillation.yml对应于OCRv3识别模型。
+-o: 覆盖配置文件中参数
+Global.pretrained_model: 指定finetune使用的预训练模型
+Global.epoch_num: 指定训练的epoch数
+Global.eval_batch_step: 间隔多少step做一次评估
+Train.dataset.data_dir: 训练数据集路径
+Train.dataset.label_file_list: 训练集文件列表
+Train.loader.batch_size_per_card: 训练单卡batch size
+Eval.dataset.data_dir: 评估数据集路径
+Eval.dataset.label_file_list: 评估数据集文件列表
+Eval.loader.batch_size_per_card: 评估单卡batch size
+```
+
+## 6. 基于真实数据finetune
+
+
+使用合成数据finetune能提升我们模型的识别精度，但由于合成数据和真实数据之间的分布可能有一定差异，因此作用有限。为进一步提高识别精度，本节介绍如何挖掘真实数据进行模型finetune。
+
+数据挖掘的整体思路如下：
+1. 使用python爬虫从网上获取大量无标签数据
+2. 使用模型从大量无标签数据中构建出有效训练集
+
+### 6.1 python爬虫获取数据
+
+- 推荐使用[爬虫工具](https://github.com/Joeclinton1/google-images-download)获取无标签图片。
+
+图片获取后，可按如下目录格式组织：
+
+```txt
+sprider
+├── file.list
+├── data
+│   ├── 00000.jpg
+│   ├── 00001.jpg
+...
+```
+
+### 6.2 数据挖掘
+
+我们使用PaddleOCR对获取到的图片进行挖掘，具体步骤如下：
+1. 使用 PP-OCRv3检测模型+svtr-tiny识别模型，对每张图片进行预测。
+2. 使用数据挖掘策略，得到有效图片。
+3. 将有效图片对应的图像区域和标签提取出来，构建训练集。
+
+
+首先下载预训练模型，PP-OCRv3检测模型下载链接：https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_infer.tar
+
+如需获取svtr-tiny高精度中文识别预训练模型，请扫码填写问卷，加入PaddleOCR官方交流群获取全部OCR垂类模型下载链接、《动手学OCR》电子书等全套OCR学习资料🎁
+<div align="left">
+<img src="https://ai-studio-static-online.cdn.bcebos.com/dd721099bd50478f9d5fb13d8dd00fad69c22d6848244fd3a1d3980d7fefc63e"  width = "150" height = "150" />
+</div>
+
+
+完成下载后，可将模型存储于如下位置:
+
+```shell
+PaddleOCR
+├── data
+│   ├── rec_vit_sub_64_363_all/  # svtr_tiny高精度识别模型
+```
+
+```bash
+# 下载解压PP-OCRv3检测模型
+cd ${PaddleOCR_root}
+wget -nc -P ckpt https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_infer.tar
+pushd ckpt
+tar -xvf ch_PP-OCRv3_det_infer.tar
+popd ckpt
+```
+
+在使用PPOCRv3检测模型+svtr-tiny识别模型进行预测之前，有如下两处细节需要手动修改：
+1. 将`tools/infer/predict_rec.py`中第110行`imgW`修改为`320`
+
+```python
+#imgW = int((imgH * max_wh_ratio))
+imgW = 320
+```
+
+2. 将`tools/infer/predict_system.py`第169行添加如下一行，将预测分数也写入结果文件中。
+
+```python
+"scores": rec_res[idx][1],
+```
+
+模型预测命令:
+```bash
+python tools/infer/predict_system.py \
+        --image_dir="/home/aistudio/sprider/data" \
+        --det_model_dir="./ckpt/ch_PP-OCRv3_det_infer/"  \
+        --rec_model_dir="/home/aistudio/PaddleOCR/data/rec_vit_sub_64_363_all/" \
+        --rec_image_shape="3,32,320"
+```
+
+获得预测结果后，我们使用数据挖掘策略得到有效图片。具体挖掘策略如下：
+1. 预测置信度高于95%
+2. 识别结果包含字符‘20’，即年份
+3. 没有中文，或者有中文并且‘日’和'月'同时在识别结果中
+
+```python
+# 获取有效预测
+
+import json
+import re
+
+zh_pattern = re.compile(u'[\u4e00-\u9fa5]+')  #正则表达式，筛选字符是否包含中文
+
+file_path = '/home/aistudio/PaddleOCR/inference_results/system_results.txt'
+out_path = '/home/aistudio/PaddleOCR/selected_results.txt'
+f_out = open(out_path, 'w')
+
+with open(file_path, "r", encoding='utf-8') as fin:
+    lines = fin.readlines()
+
+
+for line in lines:
+    flag = False
+    # 读取文件内容
+    file_name, json_file = line.strip().split('\t')
+    preds = json.loads(json_file)
+    res = []
+    for item in preds:
+        transcription = item['transcription'] #获取识别结果
+        scores = item['scores']               #获取识别得分
+        # 挖掘策略
+        if scores > 0.95:
+            if '20' in transcription and len(transcription) > 4 and len(transcription) < 12:
+                word = transcription
+                if not(zh_pattern.search(word) and ('日' not in word or '月' not in word)):
+                    flag = True
+                    res.append(item)
+    save_pred = file_name + "\t" + json.dumps(
+        res, ensure_ascii=False) + "\n"
+    if flag ==True:
+        f_out.write(save_pred)
+
+f_out.close()
+```
+
+然后将有效预测对应的图像区域和标签提取出来，构建训练集。具体实现脚本如下：
+
+```python
+import cv2
+import json
+import numpy as np
+
+PATH = '/home/aistudio/PaddleOCR/inference_results/'  #数据原始路径
+SAVE_PATH = '/home/aistudio/mining_images/'             #裁剪后数据保存路径
+file_list = '/home/aistudio/PaddleOCR/selected_results.txt' #数据预测结果
+label_file = '/home/aistudio/mining_images/mining_train.list'  #输出真实数据训练集标签list
+
+if not os.path.exists(SAVE_PATH):
+    os.mkdir(SAVE_PATH)
+
+f_label = open(label_file, 'w')
+
+
+def get_rotate_crop_image(img, points):
+    """
+    根据检测结果points，从输入图像img中裁剪出相应的区域
+    """
+    assert len(points) == 4, "shape of points must be 4*2"
+    img_crop_width = int(
+        max(
+            np.linalg.norm(points[0] - points[1]),
+            np.linalg.norm(points[2] - points[3])))
+    img_crop_height = int(
+        max(
+            np.linalg.norm(points[0] - points[3]),
+            np.linalg.norm(points[1] - points[2])))
+    pts_std = np.float32([[0, 0], [img_crop_width, 0],
+                          [img_crop_width, img_crop_height],
+                          [0, img_crop_height]])
+    M = cv2.getPerspectiveTransform(points, pts_std)
+    # 形变或倾斜，会做透视变换，reshape成矩形
+    dst_img = cv2.warpPerspective(
+        img,
+        M, (img_crop_width, img_crop_height),
+        borderMode=cv2.BORDER_REPLICATE,
+        flags=cv2.INTER_CUBIC)
+    dst_img_height, dst_img_width = dst_img.shape[0:2]
+    if dst_img_height * 1.0 / dst_img_width >= 1.5:
+        dst_img = np.rot90(dst_img)
+    return dst_img
+
+def crop_and_get_filelist(file_list):
+    with open(file_list, "r", encoding='utf-8') as fin:
+        lines = fin.readlines()
+
+    img_num = 0
+    for line in lines:
+        img_name, json_file = line.strip().split('\t')
+        preds = json.loads(json_file)
+        for item in preds:
+            transcription = item['transcription']
+            points = item['points']
+            points = np.array(points).astype('float32')
+            #print('processing {}...'.format(img_name))
+
+            img = cv2.imread(PATH+img_name)
+            dst_img = get_rotate_crop_image(img, points)
+            h, w, c = dst_img.shape
+            newWidth = int((32. / h) * w)
+            newImg = cv2.resize(dst_img, (newWidth, 32))
+            new_img_name = '{:05d}.jpg'.format(img_num)
+            cv2.imwrite(SAVE_PATH+new_img_name, dst_img)
+            f_label.write(SAVE_PATH+new_img_name+'\t'+transcription+'\n')
+            img_num += 1
+
+
+crop_and_get_filelist(file_list)
+f_label.close()
+```
+
+### 6.3 模型训练
+
+通过数据挖掘，我们得到了真实场景数据和对应的标签。接下来使用真实数据finetune，观察精度提升效果。
+
+
+利用真实数据进行finetune:
+
+```bash
+cd ${PaddleOCR_root}
+python tools/train.py -c configs/rec/PP-OCRv3/ch_PP-OCRv3_rec_distillation.yml \
+                       -o Global.pretrained_model=./ckpt/ch_PP-OCRv3_rec_train/best_accuracy \
+                       Global.epoch_num=20 \
+                       Global.eval_batch_step='[0, 20]' \
+                       Train.dataset.data_dir=./data \
+                       Train.dataset.label_file_list=['./data/mining_train.list'] \
+                       Train.loader.batch_size_per_card=64 \
+                       Eval.dataset.data_dir=./data \
+                       Eval.dataset.label_file_list=["./data/val.list"] \
+                       Eval.loader.batch_size_per_card=64
+```
+
+各参数含义参考第6部分合成数据finetune，只需要对训练数据路径做相应的修改：
+
+```txt
+Train.dataset.data_dir: 训练数据集路径
+Train.dataset.label_file_list: 训练集文件列表
+```
+
+示例使用我们提供的真实数据进行finetune，如想换成自己的数据，只需要相应的修改`Train.dataset.data_dir`和`Train.dataset.label_file_list`参数即可。
+
+由于数据量不大，这里仅训练20个epoch即可。训练完成后，可以得到合成数据finetune后的精度为best acc=**71.33%**。
+
+由于数量比较少，精度会比合成数据finetue的略低。
+
+
+## 7. 基于合成+真实数据finetune
+
+为进一步提升模型精度，我们结合使用合成数据和挖掘到的真实数据进行finetune。
+
+利用合成+真实数据进行finetune，各参数含义参考第6部分合成数据finetune，只需要对训练数据路径做相应的修改：
+
+```txt
+Train.dataset.data_dir: 训练数据集路径
+Train.dataset.label_file_list: 训练集文件列表
+```
+
+生成训练list文件:
+```bash
+# 生成训练集文件list
+cat /home/aistudio/PaddleOCR/data/render_train.list /home/aistudio/PaddleOCR/data/mining_train.list > /home/aistudio/PaddleOCR/data/render_mining_train.list
+```
+
+启动训练:
+```bash
+cd ${PaddleOCR_root}
+python tools/train.py -c configs/rec/PP-OCRv3/ch_PP-OCRv3_rec_distillation.yml \
+                       -o Global.pretrained_model=./ckpt/ch_PP-OCRv3_rec_train/best_accuracy \
+                       Global.epoch_num=40 \
+                       Global.eval_batch_step='[0, 20]' \
+                       Train.dataset.data_dir=./data \
+                       Train.dataset.label_file_list=['./data/render_mining_train.list'] \
+                       Train.loader.batch_size_per_card=64 \
+                       Eval.dataset.data_dir=./data \
+                       Eval.dataset.label_file_list=["./data/val.list"] \
+                       Eval.loader.batch_size_per_card=64
+```
+
+示例使用我们提供的真实+合成数据进行finetune，如想换成自己的数据，只需要相应的修改Train.dataset.data_dir和Train.dataset.label_file_list参数即可。
+
+由于数据量不大，这里仅训练40个epoch即可。训练完成后，可以得到合成数据finetune后的精度为best acc=**86.99%**。
+
+可以看到，相较于原始PP-OCRv3的识别精度62.99%，使用合成数据+真实数据finetune后，识别精度能提升24%。
+
+如需获取已训练模型，可以同样扫描上方二维码下载，将下载或训练完成的模型放置在对应目录下即可完成模型推理。
+
+模型的推理部署方法可以参考repo文档： https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.5/deploy/README_ch.md

applications/多模态表单识别.md

-# 1 项目说明
+# 多模态表单识别
+- [多模态表单识别](#多模态表单识别)
+  - [1 项目说明](#1-项目说明)
+  - [2 安装说明](#2-安装说明)
+  - [3 数据准备](#3-数据准备)
+    - [3.1 下载处理好的数据集](#31-下载处理好的数据集)
+    - [3.2 转换为PaddleOCR检测和识别格式](#32-转换为paddleocr检测和识别格式)
+  - [4 OCR](#4-ocr)
+    - [4.1 文本检测](#41-文本检测)
+      - [4.1.1 方案1：预训练模型](#411-方案1预训练模型)
+      - [4.1.2 方案2：XFUND数据集+fine-tune](#412-方案2xfund数据集fine-tune)
+    - [4.2 文本识别](#42-文本识别)
+      - [4.2.1 方案1：预训练模型](#421-方案1预训练模型)
+      - [4.2.2 方案2：XFUND数据集+finetune](#422-方案2xfund数据集finetune)
+      - [4.2.3 方案3：XFUND数据集+finetune+真实通用识别数据](#423-方案3xfund数据集finetune真实通用识别数据)
+  - [5 文档视觉问答(DOC-VQA)](#5-文档视觉问答doc-vqa)
+    - [5.1 SER](#51-ser)
+      - [5.1.1 模型训练](#511-模型训练)
+      - [5.1.2 模型评估](#512-模型评估)
+      - [5.1.3 模型预测](#513-模型预测)
+    - [5.2 RE](#52-re)
+      - [5.2.1 模型训练](#521-模型训练)
+      - [5.2.2 模型评估](#522-模型评估)
+      - [5.2.3 模型预测](#523-模型预测)
+  - [6 导出Excel](#6-导出excel)
+  - [获得模型](#获得模型)
+  - [更多资源](#更多资源)
+  - [参考链接](#参考链接)
+
+## 1 项目说明
 
 计算机视觉在金融领域的应用覆盖文字识别、图像识别、视频识别等，其中文字识别（OCR）是金融领域中的核心AI能力，其应用覆盖客户服务、风险防控、运营管理等各项业务，针对的对象包括通用卡证票据识别（银行卡、身份证、营业执照等）、通用文本表格识别（印刷体、多语言、手写体等）以及一些金融特色票据凭证。通过因此如果能够在结构化信息提取时同时利用文字、页面布局等信息，便可增强不同版式下的泛化性。
 
@@ -16,39 +45,37 @@
 <center><img src='https://ai-studio-static-online.cdn.bcebos.com/9bd844b970f94e5ba0bc0c5799bd819ea9b1861bb306471fabc2d628864d418e'></center>
 <center>图1 多模态表单识别流程图</center>
 
-注：欢迎再AIStudio领取免费算力体验线上实训，项目链接: [多模态表单识别](https://aistudio.baidu.com/aistudio/projectdetail/3884375)(配备Tesla V100、A100等高级算力资源)
+注：欢迎再AIStudio领取免费算力体验线上实训，项目链接: [多模态表单识别](https://aistudio.baidu.com/aistudio/projectdetail/3884375?contributionType=1)
 
-
-
-# 2 安装说明
+## 2 安装说明
 
 
 下载PaddleOCR源码，上述AIStudio项目中已经帮大家打包好的PaddleOCR(已经修改好配置文件)，无需下载解压即可，只需安装依赖环境~
 
 
 ```python
-! unzip -q PaddleOCR.zip
+unzip -q PaddleOCR.zip
 ```
 
 
 ```python
 # 如仍需安装or安装更新，可以执行以下步骤
-# ! git clone https://github.com/PaddlePaddle/PaddleOCR.git -b dygraph
-# ! git clone https://gitee.com/PaddlePaddle/PaddleOCR
+# git clone https://github.com/PaddlePaddle/PaddleOCR.git -b dygraph
+# git clone https://gitee.com/PaddlePaddle/PaddleOCR
 ```
 
 
 ```python
 # 安装依赖包
-! pip install -U pip
-! pip install -r /home/aistudio/PaddleOCR/requirements.txt
-! pip install paddleocr
+pip install -U pip
+pip install -r /home/aistudio/PaddleOCR/requirements.txt
+pip install paddleocr
 
-! pip install yacs gnureadline paddlenlp==2.2.1
-! pip install xlsxwriter
+pip install yacs gnureadline paddlenlp==2.2.1
+pip install xlsxwriter
 ```
 
-# 3 数据准备
+## 3 数据准备
 
 这里使用[XFUN数据集](https://github.com/doc-analysis/XFUND)做为实验数据集。 XFUN数据集是微软提出的一个用于KIE任务的多语言数据集，共包含七个数据集，每个数据集包含149张训练集和50张验证集
 
@@ -59,7 +86,7 @@
 <center><img src="https://ai-studio-static-online.cdn.bcebos.com/0f84137778cd4ab6899c64109d452290e9c678ccf01744978bc9c0647adbba45" width="1000" ></center>
 <center>图2 数据集样例，左中文，右法语</center>
 
-## 3.1 下载处理好的数据集
+### 3.1 下载处理好的数据集
 
 
 处理好的XFUND中文数据集下载地址：[https://paddleocr.bj.bcebos.com/dataset/XFUND.tar](https://paddleocr.bj.bcebos.com/dataset/XFUND.tar) ,可以运行如下指令完成中文数据集下载和解压。
@@ -69,13 +96,13 @@
 
 
 ```python
-! wget https://paddleocr.bj.bcebos.com/dataset/XFUND.tar
-! tar -xf XFUND.tar
+wget https://paddleocr.bj.bcebos.com/dataset/XFUND.tar
+tar -xf XFUND.tar
 
 # XFUN其他数据集使用下面的代码进行转换
 # 代码链接：https://github.com/PaddlePaddle/PaddleOCR/blob/release%2F2.4/ppstructure/vqa/helper/trans_xfun_data.py
 # %cd PaddleOCR
-# !python3 ppstructure/vqa/tools/trans_xfun_data.py --ori_gt_path=path/to/json_path --output_path=path/to/save_path
+# python3 ppstructure/vqa/tools/trans_xfun_data.py --ori_gt_path=path/to/json_path --output_path=path/to/save_path
 # %cd ../
 ```
 
@@ -119,7 +146,7 @@
 }
 ```
 
-## 3.2 转换为PaddleOCR检测和识别格式
+### 3.2 转换为PaddleOCR检测和识别格式
 
 使用XFUND训练PaddleOCR检测和识别模型，需要将数据集格式改为训练需求的格式。
 
@@ -147,7 +174,7 @@ train_data/rec/train/word_002.jpg   用科技让复杂的世界更简单
 
 
 ```python
-! unzip -q /home/aistudio/data/data140302/XFUND_ori.zip -d /home/aistudio/data/data140302/
+unzip -q /home/aistudio/data/data140302/XFUND_ori.zip -d /home/aistudio/data/data140302/
 ```
 
 已经提供转换脚本，执行如下代码即可转换成功：
@@ -155,21 +182,20 @@ train_data/rec/train/word_002.jpg   用科技让复杂的世界更简单
 
 ```python
 %cd /home/aistudio/
-! python trans_xfund_data.py
+python trans_xfund_data.py
 ```
 
-# 4 OCR
+## 4 OCR
 
 选用飞桨OCR开发套件[PaddleOCR](https://github.com/PaddlePaddle/PaddleOCR/blob/dygraph/README_ch.md)中的PP-OCRv2模型进行文本检测和识别。PP-OCRv2在PP-OCR的基础上，进一步在5个方面重点优化，检测模型采用CML协同互学习知识蒸馏策略和CopyPaste数据增广策略；识别模型采用LCNet轻量级骨干网络、UDML 改进知识蒸馏策略和[Enhanced CTC loss](https://github.com/PaddlePaddle/PaddleOCR/blob/dygraph/doc/doc_ch/enhanced_ctc_loss.md)损失函数改进，进一步在推理速度和预测效果上取得明显提升。更多细节请参考PP-OCRv2[技术报告](https://arxiv.org/abs/2109.03144)。
 
-
-## 4.1 文本检测
+### 4.1 文本检测
 
 我们使用2种方案进行训练、评估：
 -  **PP-OCRv2中英文超轻量检测预训练模型**
 -  **XFUND数据集+fine-tune**
 
-### **4.1.1 方案1：预训练模型**
+#### 4.1.1 方案1：预训练模型
 
 **1）下载预训练模型**
 
@@ -195,8 +221,8 @@ PaddleOCR已经提供了PP-OCR系列模型，部分模型展示如下表所示
 
 ```python
 %cd /home/aistudio/PaddleOCR/pretrain/
-! wget https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_det_distill_train.tar
-! tar -xf ch_PP-OCRv2_det_distill_train.tar && rm -rf ch_PP-OCRv2_det_distill_train.tar
+wget https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_det_distill_train.tar
+tar -xf ch_PP-OCRv2_det_distill_train.tar && rm -rf ch_PP-OCRv2_det_distill_train.tar
 % cd ..
 ```
 
@@ -226,7 +252,7 @@ Eval.dataset.label_file_list：指向验证集标注文件
 
 ```python
 %cd /home/aistudio/PaddleOCR
-! python tools/eval.py \
+python tools/eval.py \
     -c configs/det/ch_PP-OCRv2/ch_PP-OCRv2_det_distill.yml \
     -o Global.checkpoints="./pretrain_models/ch_PP-OCRv2_det_distill_train/best_accuracy"
 ```
@@ -237,9 +263,9 @@ Eval.dataset.label_file_list：指向验证集标注文件
 | -------- | -------- |
 | PP-OCRv2中英文超轻量检测预训练模型     | 77.26%     |
 
-使用文本检测预训练模型在XFUND验证集上评估，达到77%左右，充分说明ppocr提供的预训练模型有一定的泛化能力。
+使用文本检测预训练模型在XFUND验证集上评估，达到77%左右，充分说明ppocr提供的预训练模型具有泛化能力。
 
-### **4.1.2 方案2：XFUND数据集+fine-tune**
+#### 4.1.2 方案2：XFUND数据集+fine-tune
 
 PaddleOCR提供的蒸馏预训练模型包含了多个模型的参数，我们提取Student模型的参数，在XFUND数据集上进行finetune，可以参考如下代码：
 
@@ -281,7 +307,7 @@ Eval.dataset.transforms.DetResizeForTest：评估尺寸，添加如下参数
 
 
 ```python
-! CUDA_VISIBLE_DEVICES=0 python tools/train.py \
+CUDA_VISIBLE_DEVICES=0 python tools/train.py \
         -c configs/det/ch_PP-OCRv2/ch_PP-OCRv2_det_student.yml
 ```
 
@@ -290,12 +316,18 @@ Eval.dataset.transforms.DetResizeForTest：评估尺寸，添加如下参数
 <center><img src="https://ai-studio-static-online.cdn.bcebos.com/5a75137c5f924dfeb6956b5818812298cc3dc7992ac84954b4175be9adf83c77"></center>
 <center>图8 文本检测方案2-模型评估</center>
 
-使用训练好的模型进行评估，更新模型路径`Global.checkpoints`，这里为大家提供训练好的模型`./pretrain/ch_db_mv3-student1600-finetune/best_accuracy`，[模型下载地址](https://paddleocr.bj.bcebos.com/fanliku/sheet_recognition/ch_db_mv3-student1600-finetune.zip)
+使用训练好的模型进行评估，更新模型路径`Global.checkpoints`。如需获取已训练模型，请扫码填写问卷，加入PaddleOCR官方交流群获取全部OCR垂类模型下载链接、《动手学OCR》电子书等全套OCR学习资料🎁
+
+<div align="left">
+<img src="https://ai-studio-static-online.cdn.bcebos.com/dd721099bd50478f9d5fb13d8dd00fad69c22d6848244fd3a1d3980d7fefc63e"  width = "150" height = "150" />
+</div>
+
+将下载或训练完成的模型放置在对应目录下即可完成模型评估
 
 
 ```python
 %cd /home/aistudio/PaddleOCR/
-! python tools/eval.py \
+python tools/eval.py \
     -c configs/det/ch_PP-OCRv2/ch_PP-OCRv2_det_student.yml \
     -o Global.checkpoints="pretrain/ch_db_mv3-student1600-finetune/best_accuracy"
 ```
@@ -305,7 +337,7 @@ Eval.dataset.transforms.DetResizeForTest：评估尺寸，添加如下参数
 
 ```python
 %cd /home/aistudio/PaddleOCR/
-! python tools/eval.py \
+python tools/eval.py \
     -c configs/det/ch_PP-OCRv2/ch_PP-OCRv2_det_student.yml \
     -o Global.checkpoints="pretrain/ch_db_mv3-student1600/best_accuracy"
 ```
@@ -331,7 +363,7 @@ Eval.dataset.transforms.DetResizeForTest：评估尺寸，添加如下参数
 # 加载配置文件`ch_PP-OCRv2_det_student.yml`，从`pretrain/ch_db_mv3-student1600-finetune`目录下加载`best_accuracy`模型
 # inference模型保存在`./output/det_db_inference`目录下
 %cd /home/aistudio/PaddleOCR/
-! python tools/export_model.py \
+python tools/export_model.py \
     -c configs/det/ch_PP-OCRv2/ch_PP-OCRv2_det_student.yml \
     -o Global.pretrained_model="pretrain/ch_db_mv3-student1600-finetune/best_accuracy" \
     Global.save_inference_dir="./output/det_db_inference/"
@@ -374,12 +406,11 @@ use_gpu：是否使用GPU
 
 | 方案 | hmeans | 结果分析 |
 | -------- | -------- | -------- |
-| PP-OCRv2中英文超轻量检测预训练模型     | 77.26%     | ppocr提供的预训练模型有一定的泛化能力     |
+| PP-OCRv2中英文超轻量检测预训练模型     | 77.26%     | ppocr提供的预训练模型有泛化能力     |
 | XFUND数据集     | 79.27%     |     |
 | XFUND数据集+finetune     | 85.24%     | finetune会提升垂类场景效果    |
 
-
-## 4.2 文本识别
+### 4.2 文本识别
 
 我们分别使用如下3种方案进行训练、评估：
 
@@ -387,8 +418,7 @@ use_gpu：是否使用GPU
 - XFUND数据集+fine-tune
 - XFUND数据集+fine-tune+真实通用识别数据
 
-
-### **4.2.1 方案1：预训练模型**
+#### 4.2.1 方案1：预训练模型
 
 **1）下载预训练模型**
 
@@ -401,8 +431,8 @@ use_gpu：是否使用GPU
 
 ```python
 %cd /home/aistudio/PaddleOCR/pretrain/
-! wget https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_train.tar
-! tar -xf ch_PP-OCRv2_rec_train.tar && rm -rf ch_PP-OCRv2_rec_train.tar
+wget https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_train.tar
+tar -xf ch_PP-OCRv2_rec_train.tar && rm -rf ch_PP-OCRv2_rec_train.tar
 % cd ..
 ```
 
@@ -424,7 +454,7 @@ Eval.dataset.label_file_list：指向验证集标注文件
 
 ```python
 %cd /home/aistudio/PaddleOCR
-! CUDA_VISIBLE_DEVICES=0 python tools/eval.py \
+CUDA_VISIBLE_DEVICES=0 python tools/eval.py \
     -c configs/rec/ch_PP-OCRv2/ch_PP-OCRv2_rec_distillation.yml \
     -o Global.checkpoints=./pretrain/ch_PP-OCRv2_rec_train/best_accuracy
 ```
@@ -435,9 +465,9 @@ Eval.dataset.label_file_list：指向验证集标注文件
 | -------- | -------- | 
 | PP-OCRv2中英文超轻量识别预训练模型     | 67.48%     | 
 
-使用文本预训练模型在XFUND验证集上评估，acc达到67%左右，充分说明ppocr提供的预训练模型有一定的泛化能力。
+使用文本预训练模型在XFUND验证集上评估，acc达到67%左右，充分说明ppocr提供的预训练模型具有泛化能力。
 
-### **4.2.2 方案2：XFUND数据集+finetune**
+#### 4.2.2 方案2：XFUND数据集+finetune
 
 同检测模型，我们提取Student模型的参数，在XFUND数据集上进行finetune，可以参考如下代码：
 
@@ -474,11 +504,9 @@ Eval.dataset.label_file_list：指向验证集标注文件
 ```
 执行如下命令启动训练：
 
-
-
 ```python
 %cd /home/aistudio/PaddleOCR/
-! CUDA_VISIBLE_DEVICES=0 python tools/train.py \
+CUDA_VISIBLE_DEVICES=0 python tools/train.py \
         -c configs/rec/ch_PP-OCRv2/ch_PP-OCRv2_rec.yml
 ```
 
@@ -493,7 +521,7 @@ Eval.dataset.label_file_list：指向验证集标注文件
 
 ```python
 %cd /home/aistudio/PaddleOCR/
-! CUDA_VISIBLE_DEVICES=0 python tools/eval.py \
+CUDA_VISIBLE_DEVICES=0 python tools/eval.py \
     -c configs/rec/ch_PP-OCRv2/ch_PP-OCRv2_rec.yml \
     -o Global.checkpoints=./pretrain/rec_mobile_pp-OCRv2-student-finetune/best_accuracy
 ```
@@ -506,7 +534,7 @@ Eval.dataset.label_file_list：指向验证集标注文件
 
 使用XFUND数据集+finetune训练，在验证集上评估达到72%左右，说明 finetune会提升垂类场景效果。
 
-### **4.2.3 方案3：XFUND数据集+finetune+真实通用识别数据**
+#### 4.2.3 方案3：XFUND数据集+finetune+真实通用识别数据
 
 接着我们在上述`XFUND数据集+finetune`实验的基础上，添加真实通用识别数据，进一步提升识别效果。首先准备真实通用识别数据，并上传到AIStudio：
 
@@ -528,7 +556,7 @@ Train.dataset.ratio_list：动态采样
 
 ```python
 %cd /home/aistudio/PaddleOCR/
-! CUDA_VISIBLE_DEVICES=0 python tools/train.py \
+CUDA_VISIBLE_DEVICES=0 python tools/train.py \
         -c configs/rec/ch_PP-OCRv2/ch_PP-OCRv2_rec.yml
 ```
 
@@ -538,11 +566,11 @@ Train.dataset.ratio_list：动态采样
 
 <center>图16 文本识别方案3-模型评估</center>
 
-使用训练好的模型进行评估，更新模型路径`Global.checkpoints`，这里为大家提供训练好的模型`./pretrain/rec_mobile_pp-OCRv2-student-readldata/best_accuracy`，[模型下载地址](https://paddleocr.bj.bcebos.com/fanliku/sheet_recognition/rec_mobile_pp-OCRv2-student-realdata.zip)
+使用训练好的模型进行评估，更新模型路径`Global.checkpoints`。
 
 
 ```python
-! CUDA_VISIBLE_DEVICES=0 python tools/eval.py \
+CUDA_VISIBLE_DEVICES=0 python tools/eval.py \
     -c configs/rec/ch_PP-OCRv2/ch_PP-OCRv2_rec.yml \
     -o Global.checkpoints=./pretrain/rec_mobile_pp-OCRv2-student-realdata/best_accuracy
 ```
@@ -580,7 +608,7 @@ Train.dataset.ratio_list：动态采样
 
 
 ```python
-! python tools/infer/predict_system.py \
+python tools/infer/predict_system.py \
     --image_dir="./doc/vqa/input/zh_val_21.jpg" \
     --det_model_dir="./output/det_db_inference/" \
     --rec_model_dir="./output/rec_crnn_inference/" \
@@ -592,11 +620,11 @@ Train.dataset.ratio_list：动态采样
 
 | 方案 | acc  | 结果分析 |
 | -------- | -------- | -------- |
-| PP-OCRv2中英文超轻量识别预训练模型     | 67.48%    | ppocr提供的预训练模型有一定的泛化能力     |
+| PP-OCRv2中英文超轻量识别预训练模型     | 67.48%    | ppocr提供的预训练模型具有泛化能力  |
 | XFUND数据集+fine-tune    |72.33%     | finetune会提升垂类场景效果    |
 | XFUND数据集+fine-tune+真实通用识别数据    | 85.29%     | 真实通用识别数据对于性能提升很有帮助    |
 
-# 5 文档视觉问答(DOC-VQA)
+## 5 文档视觉问答(DOC-VQA)
 
 VQA指视觉问答，主要针对图像内容进行提问和回答,DOC-VQA是VQA任务中的一种，DOC-VQA主要针对文本图像的文字内容提出问题。
 
@@ -608,14 +636,13 @@ PaddleOCR中DOC-VQA系列算法基于PaddleNLP自然语言处理算法库实现L
 ```python
 %cd pretrain
 #下载SER模型
-! wget https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutXLM_xfun_zh.tar && tar -xvf ser_LayoutXLM_xfun_zh.tar
+wget https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutXLM_xfun_zh.tar && tar -xvf ser_LayoutXLM_xfun_zh.tar
 #下载RE模型
-! wget https://paddleocr.bj.bcebos.com/pplayout/re_LayoutXLM_xfun_zh.tar && tar -xvf re_LayoutXLM_xfun_zh.tar
+wget https://paddleocr.bj.bcebos.com/pplayout/re_LayoutXLM_xfun_zh.tar && tar -xvf re_LayoutXLM_xfun_zh.tar
 %cd ../
 ```
 
-
-## 5.1 SER
+### 5.1 SER
 
 SER: 语义实体识别 (Semantic Entity Recognition）, 可以完成对图像中的文本识别与分类。 
 
@@ -647,7 +674,7 @@ SER: 语义实体识别 (Semantic Entity Recognition）, 可以完成对图像
 
 ```python
 %cd /home/aistudio/PaddleOCR/
-! CUDA_VISIBLE_DEVICES=0 python tools/train.py -c configs/vqa/ser/layoutxlm.yml
+CUDA_VISIBLE_DEVICES=0 python tools/train.py -c configs/vqa/ser/layoutxlm.yml
 ```
 
 最终会打印出`precision`, `recall`, `hmean`等指标。 在`./output/ser_layoutxlm/`文件夹中会保存训练日志，最优的模型和最新epoch的模型。
@@ -664,7 +691,7 @@ SER: 语义实体识别 (Semantic Entity Recognition）, 可以完成对图像
 
 
 ```python
-! CUDA_VISIBLE_DEVICES=0 python tools/eval.py \
+CUDA_VISIBLE_DEVICES=0 python tools/eval.py \
     -c configs/vqa/ser/layoutxlm.yml \
     -o Architecture.Backbone.checkpoints=pretrain/ser_LayoutXLM_xfun_zh/
 ```
@@ -684,7 +711,7 @@ SER: 语义实体识别 (Semantic Entity Recognition）, 可以完成对图像
 
 
 ```python
-! CUDA_VISIBLE_DEVICES=0 python tools/infer_vqa_token_ser.py \
+CUDA_VISIBLE_DEVICES=0 python tools/infer_vqa_token_ser.py \
     -c configs/vqa/ser/layoutxlm.yml  \
     -o Architecture.Backbone.checkpoints=pretrain/ser_LayoutXLM_xfun_zh/ \
     Global.infer_img=doc/vqa/input/zh_val_42.jpg
@@ -704,7 +731,7 @@ plt.figure(figsize=(48,24))
 plt.imshow(img)
 ```
 
-## 5.2 RE
+### 5.2 RE
 
 基于 RE 任务，可以完成对图象中的文本内容的关系提取，如判断问题对(pair)。
 
@@ -729,7 +756,7 @@ plt.imshow(img)
 
 
 ```python
-! CUDA_VISIBLE_DEVICES=0 python3 tools/train.py -c configs/vqa/re/layoutxlm.yml
+CUDA_VISIBLE_DEVICES=0 python3 tools/train.py -c configs/vqa/re/layoutxlm.yml
 ```
 
 最终会打印出`precision`, `recall`, `hmean`等指标。 在`./output/re_layoutxlm/`文件夹中会保存训练日志，最优的模型和最新epoch的模型
@@ -744,7 +771,7 @@ plt.imshow(img)
 
 
 ```python
-! CUDA_VISIBLE_DEVICES=0 python3 tools/eval.py \
+CUDA_VISIBLE_DEVICES=0 python3 tools/eval.py \
     -c configs/vqa/re/layoutxlm.yml \
     -o Architecture.Backbone.checkpoints=pretrain/re_LayoutXLM_xfun_zh/
 ```
@@ -760,20 +787,14 @@ plt.imshow(img)
 
 <center>图26 RE-模型预测</center>
 
-    使用OCR引擎 + SER + RE串联预测
-
-使用如下命令即可完成OCR引擎 + SER + RE的串联预测, 以预训练SER和RE模型为例：
-
-
+使用如下命令即可完成OCR引擎 + SER + RE的串联预测, 以预训练SER和RE模型为例，
 
 最终会在config.Global.save_res_path字段所配置的目录下保存预测结果可视化图像以及预测结果文本文件，预测结果文本文件名为infer_results.txt。
 
 
-
-
 ```python
-%cd /home/aistudio/PaddleOCR
-! CUDA_VISIBLE_DEVICES=0 python3 tools/infer_vqa_token_ser_re.py \
+cd /home/aistudio/PaddleOCR
+CUDA_VISIBLE_DEVICES=0 python3 tools/infer_vqa_token_ser_re.py \
     -c configs/vqa/re/layoutxlm.yml \
     -o Architecture.Backbone.checkpoints=pretrain/re_LayoutXLM_xfun_zh/ \
     Global.infer_img=test_imgs/ \
@@ -787,10 +808,9 @@ plt.imshow(img)
 test_imgs/t131.jpg	{"政治面税": "群众", "性别": "男", "籍贯": "河北省邯郸市", "婚姻状况": "亏末婚口已婚口已娇", "通讯地址": "邯郸市阳光苑7号楼003", "民族": "汉族", "毕业院校": "河南工业大学", "户口性质": "口农村城镇", "户口地址": "河北省邯郸市", "联系电话": "13288888888", "健康状况": "健康", "姓名": "小六", "好高cm": "180", "出生年月": "1996年8月9日", "文化程度": "本科", "身份证号码": "458933777777777777"}
 ````
 
-
+展示预测结果
 
 ```python
-# 展示预测结果
 import cv2
 from matplotlib import pyplot as plt
 %matplotlib inline
@@ -800,7 +820,7 @@ plt.figure(figsize=(48,24))
 plt.imshow(img)
 ```
 
-# 6 导出Excel
+## 6 导出Excel
 
 <center><img src="https://ai-studio-static-online.cdn.bcebos.com/ab93d3d90d77437a81c9534b2dd1d3e39ef81e8473054fd3aeff6e837ebfb827"></center>
 <center>图27 导出Excel</center>
@@ -859,7 +879,7 @@ with open('output/re/infer_results.txt', 'r', encoding='utf-8') as fin:
 workbook.close()
 ```
 
-# 更多资源
+## 更多资源
 
 - 更多深度学习知识、产业案例、面试宝典等，请参考：[awesome-DeepLearning](https://github.com/paddlepaddle/awesome-DeepLearning)
 
@@ -869,7 +889,7 @@ workbook.close()
 
 - 飞桨框架相关资料，请参考：[飞桨深度学习平台](https://www.paddlepaddle.org.cn/?fr=paddleEdu_aistudio)
 
-# 参考链接
+## 参考链接
 
 -  LayoutXLM: Multimodal Pre-training for Multilingual Visually-rich Document Understanding, https://arxiv.org/pdf/2104.08836.pdf
 

applications/液晶屏读数识别.md

+# 基于PP-OCRv3的液晶屏读数识别
+
+- [1. 项目背景及意义](#1-项目背景及意义)
+- [2. 项目内容](#2-项目内容)
+- [3. 安装环境](#3-安装环境)
+- [4. 文字检测](#4-文字检测)
+  - [4.1 PP-OCRv3检测算法介绍](#41-PP-OCRv3检测算法介绍)
+  - [4.2 数据准备](#42-数据准备)
+  - [4.3 模型训练](#43-模型训练)
+    - [4.3.1 预训练模型直接评估](#431-预训练模型直接评估)
+    - [4.3.2 预训练模型直接finetune](#432-预训练模型直接finetune)
+    - [4.3.3 基于预训练模型Finetune_student模型](#433-基于预训练模型Finetune_student模型)
+    - [4.3.4 基于预训练模型Finetune_teacher模型](#434-基于预训练模型Finetune_teacher模型)
+    - [4.3.5 采用CML蒸馏进一步提升student模型精度](#435-采用CML蒸馏进一步提升student模型精度)
+    - [4.3.6 模型导出推理](#436-4.3.6-模型导出推理)
+- [5. 文字识别](#5-文字识别)
+  - [5.1 PP-OCRv3识别算法介绍](#51-PP-OCRv3识别算法介绍)
+  - [5.2 数据准备](#52-数据准备)
+  - [5.3 模型训练](#53-模型训练)
+  - [5.4 模型导出推理](#54-模型导出推理)
+- [6. 系统串联](#6-系统串联)
+  - [6.1 后处理](#61-后处理)
+- [7. PaddleServing部署](#7-PaddleServing部署)
+
+
+## 1. 项目背景及意义
+目前光学字符识别(OCR)技术在我们的生活当中被广泛使用，但是大多数模型在通用场景下的准确性还有待提高，针对于此我们借助飞桨提供的PaddleOCR套件较容易的实现了在垂类场景下的应用。
+
+该项目以国家质量基础（NQI）为准绳，充分利用大数据、云计算、物联网等高新技术，构建覆盖计量端、实验室端、数据端和硬件端的完整计量解决方案，解决传统计量校准中存在的难题，拓宽计量检测服务体系和服务领域；解决无数传接口或数传接口不统一、不公开的计量设备，以及计量设备所处的环境比较恶劣，不适合人工读取数据。通过OCR技术实现远程计量，引领计量行业向智慧计量转型和发展。
+
+## 2. 项目内容
+本项目基于PaddleOCR开源套件，以PP-OCRv3检测和识别模型为基础，针对液晶屏读数识别场景进行优化。
+
+Aistudio项目链接：[OCR液晶屏读数识别](https://aistudio.baidu.com/aistudio/projectdetail/4080130)
+
+## 3. 安装环境
+
+```python
+# 首先git官方的PaddleOCR项目，安装需要的依赖
+# 第一次运行打开该注释
+# git clone https://gitee.com/PaddlePaddle/PaddleOCR.git
+cd PaddleOCR
+pip install -r requirements.txt
+```
+
+## 4. 文字检测
+文本检测的任务是定位出输入图像中的文字区域。近年来学术界关于文本检测的研究非常丰富，一类方法将文本检测视为目标检测中的一个特定场景，基于通用目标检测算法进行改进适配，如TextBoxes[1]基于一阶段目标检测器SSD[2]算法，调整目标框使之适合极端长宽比的文本行，CTPN[3]则是基于Faster RCNN[4]架构改进而来。但是文本检测与目标检测在目标信息以及任务本身上仍存在一些区别，如文本一般长宽比较大，往往呈“条状”，文本行之间可能比较密集，弯曲文本等，因此又衍生了很多专用于文本检测的算法。本项目基于PP-OCRv3算法进行优化。
+
+### 4.1 PP-OCRv3检测算法介绍
+PP-OCRv3检测模型是对PP-OCRv2中的CML（Collaborative Mutual Learning) 协同互学习文本检测蒸馏策略进行了升级。如下图所示，CML的核心思想结合了①传统的Teacher指导Student的标准蒸馏与 ②Students网络之间的DML互学习，可以让Students网络互学习的同时，Teacher网络予以指导。PP-OCRv3分别针对教师模型和学生模型进行进一步效果优化。其中，在对教师模型优化时，提出了大感受野的PAN结构LK-PAN和引入了DML（Deep Mutual Learning）蒸馏策略；在对学生模型优化时，提出了残差注意力机制的FPN结构RSE-FPN。
+![](https://ai-studio-static-online.cdn.bcebos.com/c306b2f028364805a55494d435ab553a76cf5ae5dd3f4649a948ea9aeaeb28b8)
+
+详细优化策略描述请参考[PP-OCRv3优化策略](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.5/doc/doc_ch/PP-OCRv3_introduction.md#2)
+
+### 4.2 数据准备
+[计量设备屏幕字符检测数据集](https://aistudio.baidu.com/aistudio/datasetdetail/127845)数据来源于实际项目中各种计量设备的数显屏，以及在网上搜集的一些其他数显屏，包含训练集755张，测试集355张。
+
+```python
+# 在PaddleOCR下创建新的文件夹train_data
+mkdir train_data
+# 下载数据集并解压到指定路径下
+unzip icdar2015.zip  -d train_data
+```
+
+```python
+# 随机查看文字检测数据集图片
+from PIL import Image  
+import matplotlib.pyplot as plt
+import numpy as np
+import os
+
+
+train = './train_data/icdar2015/text_localization/test'
+# 从指定目录中选取一张图片
+def get_one_image(train):
+    plt.figure()
+    files = os.listdir(train)
+    n = len(files)
+    ind = np.random.randint(0,n)
+    img_dir = os.path.join(train,files[ind])  
+    image = Image.open(img_dir)  
+    plt.imshow(image)
+    plt.show()
+    image = image.resize([208, 208])  
+
+get_one_image(train)  
+```
+![det_png](https://ai-studio-static-online.cdn.bcebos.com/0639da09b774458096ae577e82b2c59e89ced6a00f55458f946997ab7472a4f8)
+
+### 4.3 模型训练
+
+#### 4.3.1 预训练模型直接评估
+下载我们需要的PP-OCRv3检测预训练模型，更多选择请自行选择其他的[文字检测模型](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.5/doc/doc_ch/models_list.md#1-%E6%96%87%E6%9C%AC%E6%A3%80%E6%B5%8B%E6%A8%A1%E5%9E%8B)
+
+```python
+#使用该指令下载需要的预训练模型
+wget -P ./pretrained_models/ https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_distill_train.tar
+# 解压预训练模型文件
+tar -xf ./pretrained_models/ch_PP-OCRv3_det_distill_train.tar -C pretrained_models
+```
+
+在训练之前，我们可以直接使用下面命令来评估预训练模型的效果:
+
+```python
+# 评估预训练模型
+python tools/eval.py -c configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_cml.yml -o Global.pretrained_model="./pretrained_models/ch_PP-OCRv3_det_distill_train/best_accuracy"
+```
+
+结果如下：
+
+|   | 方案                        |hmeans|
+|---|---------------------------|---|
+| 0 | PP-OCRv3中英文超轻量检测预训练模型直接预测 |47.5%|
+
+#### 4.3.2 预训练模型直接finetune
+##### 修改配置文件
+我们使用configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_cml.yml，主要修改训练轮数和学习率参相关参数，设置预训练模型路径，设置数据集路径。 另外，batch_size可根据自己机器显存大小进行调整。 具体修改如下几个地方：
+```
+epoch:100
+save_epoch_step:10
+eval_batch_step:[0, 50]
+save_model_dir: ./output/ch_PP-OCR_v3_det/
+pretrained_model: ./pretrained_models/ch_PP-OCRv3_det_distill_train/best_accuracy
+learning_rate: 0.00025
+num_workers: 0 # 如果单卡训练，建议将Train和Eval的loader部分的num_workers设置为0，否则会出现`/dev/shm insufficient`的报错
+```
+
+##### 开始训练
+使用我们上面修改的配置文件configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_cml.yml，训练命令如下：
+
+```python
+# 开始训练模型
+python tools/train.py -c configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_cml.yml -o Global.pretrained_model=./pretrained_models/ch_PP-OCRv3_det_distill_train/best_accuracy
+```
+
+评估训练好的模型：
+
+```python
+# 评估训练好的模型
+python tools/eval.py -c configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_cml.yml -o Global.pretrained_model="./output/ch_PP-OCR_v3_det/best_accuracy"
+```
+
+结果如下：
+|   | 方案                        |hmeans|
+|---|---------------------------|---|
+| 0 | PP-OCRv3中英文超轻量检测预训练模型直接预测 |47.5%|
+| 1 | PP-OCRv3中英文超轻量检测预训练模型fintune |65.2%|
+
+#### 4.3.3 基于预训练模型Finetune_student模型
+
+我们使用configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_student.yml，主要修改训练轮数和学习率参相关参数，设置预训练模型路径，设置数据集路径。 另外，batch_size可根据自己机器显存大小进行调整。 具体修改如下几个地方：
+```
+epoch:100
+save_epoch_step:10
+eval_batch_step:[0, 50]
+save_model_dir: ./output/ch_PP-OCR_v3_det_student/
+pretrained_model: ./pretrained_models/ch_PP-OCRv3_det_distill_train/student
+learning_rate: 0.00025
+num_workers: 0 # 如果单卡训练，建议将Train和Eval的loader部分的num_workers设置为0，否则会出现`/dev/shm insufficient`的报错
+```
+
+训练命令如下：
+
+```python
+python tools/train.py -c configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_student.yml -o Global.pretrained_model=./pretrained_models/ch_PP-OCRv3_det_distill_train/student
+```
+
+评估训练好的模型：
+
+```python
+# 评估训练好的模型
+python tools/eval.py -c configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_student.yml -o Global.pretrained_model="./output/ch_PP-OCR_v3_det_student/best_accuracy"
+```
+
+结果如下：
+|   | 方案                        |hmeans|
+|---|---------------------------|---|
+| 0 | PP-OCRv3中英文超轻量检测预训练模型直接预测 |47.5%|
+| 1 | PP-OCRv3中英文超轻量检测预训练模型fintune |65.2%|
+| 2 | PP-OCRv3中英文超轻量检测预训练模型fintune学生模型 |80.0%|
+
+#### 4.3.4 基于预训练模型Finetune_teacher模型
+
+首先需要从提供的预训练模型best_accuracy.pdparams中提取teacher参数，组合成适合dml训练的初始化模型，提取代码如下：
+
+```python
+cd ./pretrained_models/
+# transform teacher params in best_accuracy.pdparams into teacher_dml.paramers
+import paddle
+
+# load pretrained model
+all_params = paddle.load("ch_PP-OCRv3_det_distill_train/best_accuracy.pdparams")
+# print(all_params.keys())
+
+# keep teacher params
+t_params = {key[len("Teacher."):]: all_params[key] for key in all_params if "Teacher." in key}
+
+# print(t_params.keys())
+
+s_params = {"Student." + key: t_params[key] for key in t_params}
+s2_params = {"Student2." + key: t_params[key] for key in t_params}
+s_params = {**s_params, **s2_params}
+# print(s_params.keys())
+
+paddle.save(s_params, "ch_PP-OCRv3_det_distill_train/teacher_dml.pdparams")
+
+```
+
+我们使用configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_dml.yml，主要修改训练轮数和学习率参相关参数，设置预训练模型路径，设置数据集路径。 另外，batch_size可根据自己机器显存大小进行调整。 具体修改如下几个地方：
+```
+epoch:100
+save_epoch_step:10
+eval_batch_step:[0, 50]
+save_model_dir: ./output/ch_PP-OCR_v3_det_teacher/
+pretrained_model: ./pretrained_models/ch_PP-OCRv3_det_distill_train/teacher_dml
+learning_rate: 0.00025
+num_workers: 0 # 如果单卡训练，建议将Train和Eval的loader部分的num_workers设置为0，否则会出现`/dev/shm insufficient`的报错
+```
+
+训练命令如下：
+
+```python
+python tools/train.py -c configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_dml.yml -o Global.pretrained_model=./pretrained_models/ch_PP-OCRv3_det_distill_train/teacher_dml
+```
+
+评估训练好的模型：
+
+```python
+# 评估训练好的模型
+python tools/eval.py -c configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_dml.yml -o Global.pretrained_model="./output/ch_PP-OCR_v3_det_teacher/best_accuracy"
+```
+
+结果如下：
+|   | 方案                        |hmeans|
+|---|---------------------------|---|
+| 0 | PP-OCRv3中英文超轻量检测预训练模型直接预测 |47.5%|
+| 1 | PP-OCRv3中英文超轻量检测预训练模型fintune |65.2%|
+| 2 | PP-OCRv3中英文超轻量检测预训练模型fintune学生模型 |80.0%|
+| 3 | PP-OCRv3中英文超轻量检测预训练模型fintune教师模型 |84.8%|
+
+#### 4.3.5 采用CML蒸馏进一步提升student模型精度
+
+需要从4.3.3和4.3.4训练得到的best_accuracy.pdparams中提取各自代表student和teacher的参数，组合成适合cml训练的初始化模型，提取代码如下：
+
+```python
+# transform teacher params and student parameters into cml model
+import paddle
+
+all_params = paddle.load("./pretrained_models/ch_PP-OCRv3_det_distill_train/best_accuracy.pdparams")
+# print(all_params.keys())
+
+t_params = paddle.load("./output/ch_PP-OCR_v3_det_teacher/best_accuracy.pdparams")
+# print(t_params.keys())
+
+s_params = paddle.load("./output/ch_PP-OCR_v3_det_student/best_accuracy.pdparams")
+# print(s_params.keys())
+
+for key in all_params:
+    # teacher is OK
+    if "Teacher." in key:
+        new_key = key.replace("Teacher", "Student")
+        #print("{} >> {}\n".format(key, new_key))
+        assert all_params[key].shape == t_params[new_key].shape
+        all_params[key] = t_params[new_key]
+
+    if "Student." in key:
+        new_key = key.replace("Student.", "")
+        #print("{} >> {}\n".format(key, new_key))
+        assert all_params[key].shape == s_params[new_key].shape
+        all_params[key] = s_params[new_key]
+
+    if "Student2." in key:
+        new_key = key.replace("Student2.", "")
+        print("{} >> {}\n".format(key, new_key))
+        assert all_params[key].shape == s_params[new_key].shape
+        all_params[key] = s_params[new_key]
+
+paddle.save(all_params, "./pretrained_models/ch_PP-OCRv3_det_distill_train/teacher_cml_student.pdparams")
+```
+
+训练命令如下：
+
+```python
+python tools/train.py -c configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_cml.yml -o Global.pretrained_model=./pretrained_models/ch_PP-OCRv3_det_distill_train/teacher_cml_student Global.save_model_dir=./output/ch_PP-OCR_v3_det_finetune/
+```
+
+评估训练好的模型：
+
+```python
+# 评估训练好的模型
+python tools/eval.py -c configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_cml.yml -o Global.pretrained_model="./output/ch_PP-OCR_v3_det_finetune/best_accuracy"
+```
+
+结果如下：
+|   | 方案                        |hmeans|
+|---|---------------------------|---|
+| 0 | PP-OCRv3中英文超轻量检测预训练模型直接预测 |47.5%|
+| 1 | PP-OCRv3中英文超轻量检测预训练模型fintune |65.2%|
+| 2 | PP-OCRv3中英文超轻量检测预训练模型fintune学生模型 |80.0%|
+| 3 | PP-OCRv3中英文超轻量检测预训练模型fintune教师模型 |84.8%|
+| 4 | 基于2和3训练好的模型fintune |82.7%|
+
+如需获取已训练模型，请扫码填写问卷，加入PaddleOCR官方交流群获取全部OCR垂类模型下载链接、《动手学OCR》电子书等全套OCR学习资料🎁
+<div align="left">
+<img src="https://ai-studio-static-online.cdn.bcebos.com/dd721099bd50478f9d5fb13d8dd00fad69c22d6848244fd3a1d3980d7fefc63e"  width = "150" height = "150" />
+</div>
+将下载或训练完成的模型放置在对应目录下即可完成模型推理。
+
+####  4.3.6 模型导出推理
+训练完成后，可以将训练模型转换成inference模型。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。
+#####  4.3.6.1 模型导出
+导出命令如下：
+
+```python
+# 转化为推理模型
+python tools/export_model.py \
+-c configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_cml.yml \
+-o Global.pretrained_model=./output/ch_PP-OCR_v3_det_finetune/best_accuracy \
+-o Global.save_inference_dir="./inference/det_ppocrv3"
+
+```
+
+#####  4.3.6.2 模型推理
+导出模型后，可以使用如下命令进行推理预测：
+
+```python
+# 推理预测
+python tools/infer/predict_det.py --image_dir="train_data/icdar2015/text_localization/test/1.jpg" --det_model_dir="./inference/det_ppocrv3/Student"
+```
+
+## 5. 文字识别
+文本识别的任务是识别出图像中的文字内容，一般输入来自于文本检测得到的文本框截取出的图像文字区域。文本识别一般可以根据待识别文本形状分为规则文本识别和不规则文本识别两大类。规则文本主要指印刷字体、扫描文本等，文本大致处在水平线位置；不规则文本往往不在水平位置，存在弯曲、遮挡、模糊等问题。不规则文本场景具有很大的挑战性，也是目前文本识别领域的主要研究方向。本项目基于PP-OCRv3算法进行优化。
+
+### 5.1 PP-OCRv3识别算法介绍
+PP-OCRv3的识别模块是基于文本识别算法[SVTR](https://arxiv.org/abs/2205.00159)优化。SVTR不再采用RNN结构，通过引入Transformers结构更加有效地挖掘文本行图像的上下文信息，从而提升文本识别能力。如下图所示，PP-OCRv3采用了6个优化策略。
+![](https://ai-studio-static-online.cdn.bcebos.com/d4f5344b5b854d50be738671598a89a45689c6704c4d481fb904dd7cf72f2a1a)
+
+优化策略汇总如下：
+* SVTR_LCNet：轻量级文本识别网络
+* GTC：Attention指导CTC训练策略
+* TextConAug：挖掘文字上下文信息的数据增广策略
+* TextRotNet：自监督的预训练模型
+* UDML：联合互学习策略
+* UIM：无标注数据挖掘方案
+
+详细优化策略描述请参考[PP-OCRv3优化策略](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.5/doc/doc_ch/PP-OCRv3_introduction.md#3-%E8%AF%86%E5%88%AB%E4%BC%98%E5%8C%96)
+
+### 5.2 数据准备
+[计量设备屏幕字符识别数据集](https://aistudio.baidu.com/aistudio/datasetdetail/128714)数据来源于实际项目中各种计量设备的数显屏，以及在网上搜集的一些其他数显屏，包含训练集19912张，测试集4099张。
+
+```python
+# 解压下载的数据集到指定路径下
+unzip ic15_data.zip -d train_data
+```
+
+```python
+# 随机查看文字检测数据集图片
+from PIL import Image  
+import matplotlib.pyplot as plt
+import numpy as np
+import os
+
+train = './train_data/ic15_data/train'
+# 从指定目录中选取一张图片
+def get_one_image(train):
+    plt.figure()
+    files = os.listdir(train)
+    n = len(files)
+    ind = np.random.randint(0,n)
+    img_dir = os.path.join(train,files[ind])  
+    image = Image.open(img_dir)  
+    plt.imshow(image)
+    plt.show()
+    image = image.resize([208, 208])  
+
+get_one_image(train)
+```
+
+![rec_png](https://ai-studio-static-online.cdn.bcebos.com/3de0d475c69746d0a184029001ef07c85fd68816d66d4beaa10e6ef60030f9b4)
+
+### 5.3 模型训练
+####  下载预训练模型
+下载我们需要的PP-OCRv3识别预训练模型，更多选择请自行选择其他的[文字识别模型](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.5/doc/doc_ch/models_list.md#2-%E6%96%87%E6%9C%AC%E8%AF%86%E5%88%AB%E6%A8%A1%E5%9E%8B)
+
+```python
+# 使用该指令下载需要的预训练模型
+wget -P ./pretrained_models/ https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_rec_train.tar
+# 解压预训练模型文件
+tar -xf ./pretrained_models/ch_PP-OCRv3_rec_train.tar -C pretrained_models
+```
+
+####  修改配置文件
+我们使用configs/rec/PP-OCRv3/ch_PP-OCRv3_rec_distillation.yml，主要修改训练轮数和学习率参相关参数，设置预训练模型路径，设置数据集路径。 另外，batch_size可根据自己机器显存大小进行调整。 具体修改如下几个地方：
+```
+  epoch_num: 100 # 训练epoch数
+  save_model_dir: ./output/ch_PP-OCR_v3_rec
+  save_epoch_step: 10
+  eval_batch_step: [0, 100] # 评估间隔，每隔100step评估一次
+  cal_metric_during_train: true
+  pretrained_model: ./pretrained_models/ch_PP-OCRv3_rec_train/best_accuracy  # 预训练模型路径
+  character_dict_path: ppocr/utils/ppocr_keys_v1.txt
+  use_space_char: true  # 使用空格
+
+  lr:
+    name: Cosine # 修改学习率衰减策略为Cosine
+    learning_rate: 0.0002 # 修改fine-tune的学习率
+    warmup_epoch: 2 # 修改warmup轮数
+
+Train:
+  dataset:
+    name: SimpleDataSet
+    data_dir: ./train_data/ic15_data/ # 训练集图片路径
+    ext_op_transform_idx: 1
+    label_file_list:
+    - ./train_data/ic15_data/rec_gt_train.txt # 训练集标签
+    ratio_list:
+    - 1.0
+  loader:
+    shuffle: true
+    batch_size_per_card: 64
+    drop_last: true
+    num_workers: 4
+Eval:
+  dataset:
+    name: SimpleDataSet
+    data_dir: ./train_data/ic15_data/ # 测试集图片路径
+    label_file_list:
+    - ./train_data/ic15_data/rec_gt_test.txt # 测试集标签
+    ratio_list:
+    - 1.0
+  loader:
+    shuffle: false
+    drop_last: false
+    batch_size_per_card: 64
+    num_workers: 4
+```
+
+在训练之前，我们可以直接使用下面命令来评估预训练模型的效果:
+
+```python
+# 评估预训练模型
+python tools/eval.py -c configs/rec/PP-OCRv3/ch_PP-OCRv3_rec_distillation.yml -o Global.pretrained_model="./pretrained_models/ch_PP-OCRv3_rec_train/best_accuracy"
+```
+
+结果如下：
+|   | 方案                        |accuracy|
+|---|---------------------------|---|
+| 0 | PP-OCRv3中英文超轻量识别预训练模型直接预测 |70.4%|
+
+####  开始训练
+我们使用上面修改好的配置文件configs/rec/PP-OCRv3/ch_PP-OCRv3_rec_distillation.yml，预训练模型，数据集路径，学习率，训练轮数等都已经设置完毕后，可以使用下面命令开始训练。
+
+```python
+# 开始训练识别模型
+python tools/train.py -c configs/rec/PP-OCRv3/ch_PP-OCRv3_rec_distillation.yml
+```
+
+训练完成后，可以对训练模型中最好的进行测试，评估命令如下：
+
+```python
+# 评估finetune效果
+python tools/eval.py -c configs/rec/PP-OCRv3/ch_PP-OCRv3_rec_distillation.yml -o Global.checkpoints="./output/ch_PP-OCR_v3_rec/best_accuracy"
+```
+
+结果如下：
+|   | 方案                        |accuracy|
+|---|---------------------------|---|
+| 0 | PP-OCRv3中英文超轻量识别预训练模型直接预测 |70.4%|
+| 1 | PP-OCRv3中英文超轻量识别预训练模型finetune |82.2%|
+
+如需获取已训练模型，请扫码填写问卷，加入PaddleOCR官方交流群获取全部OCR垂类模型下载链接、《动手学OCR》电子书等全套OCR学习资料🎁
+<div align="left">
+<img src="https://ai-studio-static-online.cdn.bcebos.com/dd721099bd50478f9d5fb13d8dd00fad69c22d6848244fd3a1d3980d7fefc63e"  width = "150" height = "150" />
+</div>
+将下载或训练完成的模型放置在对应目录下即可完成模型推理。
+
+###  5.4 模型导出推理
+训练完成后，可以将训练模型转换成inference模型。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。
+####  模型导出
+导出命令如下：
+
+```python
+# 转化为推理模型
+python tools/export_model.py -c configs/rec/PP-OCRv3/ch_PP-OCRv3_rec_distillation.yml -o Global.pretrained_model="./output/ch_PP-OCR_v3_rec/best_accuracy" Global.save_inference_dir="./inference/rec_ppocrv3/"
+```
+
+#### 模型推理
+导出模型后，可以使用如下命令进行推理预测
+
+```python
+# 推理预测
+python tools/infer/predict_rec.py --image_dir="train_data/ic15_data/test/1_crop_0.jpg" --rec_model_dir="./inference/rec_ppocrv3/Student"
+```
+
+## 6. 系统串联
+我们将上面训练好的检测和识别模型进行系统串联测试，命令如下：
+
+```python
+#串联测试
+python3 tools/infer/predict_system.py --image_dir="./train_data/icdar2015/text_localization/test/142.jpg" --det_model_dir="./inference/det_ppocrv3/Student"  --rec_model_dir="./inference/rec_ppocrv3/Student"
+```
+
+测试结果保存在`./inference_results/`目录下，可以用下面代码进行可视化
+
+```python
+%cd /home/aistudio/PaddleOCR
+# 显示结果
+import matplotlib.pyplot as plt
+from PIL import Image
+img_path= "./inference_results/142.jpg"
+img = Image.open(img_path)
+plt.figure("test_img", figsize=(30,30))
+plt.imshow(img)
+plt.show()
+```
+
+![sys_res_png](https://ai-studio-static-online.cdn.bcebos.com/901ab741cb46441ebec510b37e63b9d8d1b7c95f63cc4e5e8757f35179ae6373)
+
+### 6.1 后处理
+如果需要获取key-value信息，可以基于启发式的规则，将识别结果与关键字库进行匹配；如果匹配上了，则取该字段为key, 后面一个字段为value。
+
+```python
+def postprocess(rec_res):
+    keys = ["型号", "厂家", "版本号", "检定校准分类", "计量器具编号", "烟尘流量",
+            "累积体积", "烟气温度", "动压", "静压", "时间", "试验台编号", "预测流速",
+            "全压", "烟温", "流速", "工况流量", "标杆流量", "烟尘直读嘴", "烟尘采样嘴",
+            "大气压", "计前温度", "计前压力", "干球温度", "湿球温度", "流量", "含湿量"]
+    key_value = []
+    if len(rec_res) > 1:
+        for i in range(len(rec_res) - 1):
+            rec_str, _ = rec_res[i]
+            for key in keys:
+                if rec_str in key:
+                    key_value.append([rec_str, rec_res[i + 1][0]])
+                    break
+    return key_value
+key_value = postprocess(filter_rec_res)
+```
+
+## 7. PaddleServing部署
+首先需要安装PaddleServing部署相关的环境
+
+```python
+python -m pip install paddle-serving-server-gpu
+python -m pip install paddle_serving_client
+python -m pip install paddle-serving-app
+```
+
+###  7.1 转化检测模型
+
+```python
+cd deploy/pdserving/
+python -m paddle_serving_client.convert --dirname ../../inference/det_ppocrv3/Student/  \
+                                         --model_filename inference.pdmodel          \
+                                         --params_filename inference.pdiparams       \
+                                         --serving_server ./ppocr_det_v3_serving/ \
+                                         --serving_client ./ppocr_det_v3_client/
+```
+
+### 7.2 转化识别模型
+
+```python
+python -m paddle_serving_client.convert --dirname ../../inference/rec_ppocrv3/Student \
+                                         --model_filename inference.pdmodel          \
+                                         --params_filename inference.pdiparams       \
+                                         --serving_server ./ppocr_rec_v3_serving/ \
+                                         --serving_client ./ppocr_rec_v3_client/
+```
+
+
+### 7.3 启动服务
+首先可以将后处理代码加入到web_service.py中，具体修改如下：
+```
+# 代码153行后面增加下面代码
+def _postprocess(rec_res):
+    keys = ["型号", "厂家", "版本号", "检定校准分类", "计量器具编号", "烟尘流量",
+            "累积体积", "烟气温度", "动压", "静压", "时间", "试验台编号", "预测流速",
+            "全压", "烟温", "流速", "工况流量", "标杆流量", "烟尘直读嘴", "烟尘采样嘴",
+            "大气压", "计前温度", "计前压力", "干球温度", "湿球温度", "流量", "含湿量"]
+    key_value = []
+    if len(rec_res) > 1:
+        for i in range(len(rec_res) - 1):
+            rec_str, _ = rec_res[i]
+            for key in keys:
+                if rec_str in key:
+                    key_value.append([rec_str, rec_res[i + 1][0]])
+                    break
+    return key_value
+key_value = _postprocess(rec_list)
+res = {"result": str(key_value)}
+# res = {"result": str(result_list)}
+```
+
+启动服务端
+```python
+python web_service.py 2>&1 >log.txt
+```
+
+### 7.4 发送请求
+然后再开启一个新的终端，运行下面的客户端代码
+
+```python
+python pipeline_http_client.py --image_dir ../../train_data/icdar2015/text_localization/test/142.jpg
+```
+
+可以获取到最终的key-value结果：
+```
+大气压, 100.07kPa
+干球温度, 0000℃
+计前温度, 0000℃
+湿球温度, 0000℃
+计前压力, -0000kPa
+流量, 00.0L/min
+静压, 00000kPa
+含湿量, 00.0 %
+```

applications/高精度中文识别模型.md

+# 高精度中文场景文本识别模型SVTR
+
+## 1. 简介
+
+PP-OCRv3是百度开源的超轻量级场景文本检测识别模型库，其中超轻量的场景中文识别模型SVTR_LCNet使用了SVTR算法结构。为了保证速度，SVTR_LCNet将SVTR模型的Local Blocks替换为LCNet，使用两层Global Blocks。在中文场景中，PP-OCRv3识别主要使用如下优化策略：
+- GTC：Attention指导CTC训练策略；
+- TextConAug：挖掘文字上下文信息的数据增广策略；
+- TextRotNet：自监督的预训练模型；
+- UDML：联合互学习策略；
+- UIM：无标注数据挖掘方案。
+
+其中 *UIM：无标注数据挖掘方案* 使用了高精度的SVTR中文模型进行无标注文件的刷库，该模型在PP-OCRv3识别的数据集上训练，精度对比如下表。
+
+|中文识别算法|模型|UIM|精度|
+| --- | --- | --- |--- |
+|PP-OCRv3|SVTR_LCNet| w/o |78.4%|
+|PP-OCRv3|SVTR_LCNet| w |79.4%|
+|SVTR|SVTR-Tiny|-|82.5%|
+
+aistudio项目链接: [高精度中文场景文本识别模型SVTR](https://aistudio.baidu.com/aistudio/projectdetail/4263032)
+
+## 2. SVTR中文模型使用
+
+### 环境准备
+
+
+本任务基于Aistudio完成, 具体环境如下：
+
+- 操作系统: Linux
+- PaddlePaddle: 2.3
+- PaddleOCR: dygraph
+
+下载 PaddleOCR代码
+
+```bash
+git clone -b dygraph https://github.com/PaddlePaddle/PaddleOCR
+```
+
+安装依赖库
+
+```bash
+pip install -r PaddleOCR/requirements.txt -i https://mirror.baidu.com/pypi/simple
+```
+
+### 快速使用
+
+获取SVTR中文模型文件，请扫码填写问卷，加入PaddleOCR官方交流群获取全部OCR垂类模型下载链接、《动手学OCR》电子书等全套OCR学习资料🎁
+<div align="center">
+<img src="https://ai-studio-static-online.cdn.bcebos.com/dd721099bd50478f9d5fb13d8dd00fad69c22d6848244fd3a1d3980d7fefc63e"  width = "150" height = "150" />
+</div>
+
+```bash
+# 解压模型文件
+tar xf svtr_ch_high_accuracy.tar
+```
+
+预测中文文本，以下图为例：
+![](../doc/imgs_words/ch/word_1.jpg)
+
+预测命令：
+
+```bash
+# CPU预测
+python tools/infer_rec.py -c configs/rec/rec_svtrnet_ch.yml -o Global.pretrained_model=./svtr_ch_high_accuracy/best_accuracy Global.infer_img=./doc/imgs_words/ch/word_1.jpg Global.use_gpu=False
+
+# GPU预测
+#python tools/infer_rec.py -c configs/rec/rec_svtrnet_ch.yml -o Global.pretrained_model=./svtr_ch_high_accuracy/best_accuracy Global.infer_img=./doc/imgs_words/ch/word_1.jpg Global.use_gpu=True
+```
+
+可以看到最后打印结果为
+- result: 韩国小馆    0.9853458404541016
+
+0.9853458404541016为预测置信度。
+
+### 推理模型导出与预测
+
+inference 模型（paddle.jit.save保存的模型） 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。 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。 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。
+
+运行识别模型转inference模型命令，如下：
+
+```bash
+python tools/export_model.py -c configs/rec/rec_svtrnet_ch.yml -o Global.pretrained_model=./svtr_ch_high_accuracy/best_accuracy Global.save_inference_dir=./inference/svtr_ch
+```
+
+转换成功后，在目录下有三个文件：
+```shell
+inference/svtr_ch/
+    ├── inference.pdiparams         # 识别inference模型的参数文件
+    ├── inference.pdiparams.info    # 识别inference模型的参数信息，可忽略
+    └── inference.pdmodel           # 识别inference模型的program文件
+```
+
+inference模型预测，命令如下：
+
+```bash
+# CPU预测
+python3 tools/infer/predict_rec.py --image_dir="./doc/imgs_words/ch/word_1.jpg" --rec_algorithm='SVTR' --rec_model_dir=./inference/svtr_ch/ --rec_image_shape='3, 32, 320'  --rec_char_dict_path=ppocr/utils/ppocr_keys_v1.txt --use_gpu=False
+
+# GPU预测
+#python3 tools/infer/predict_rec.py --image_dir="./doc/imgs_words/ch/word_1.jpg" --rec_algorithm='SVTR' --rec_model_dir=./inference/svtr_ch/ --rec_image_shape='3, 32, 320'  --rec_char_dict_path=ppocr/utils/ppocr_keys_v1.txt --use_gpu=True
+```
+
+**注意**
+
+- 使用SVTR算法时，需要指定--rec_algorithm='SVTR'
+- 如果使用自定义字典训练的模型，需要将--rec_char_dict_path=ppocr/utils/ppocr_keys_v1.txt修改为自定义的字典
+- --rec_image_shape='3, 32, 320' 该参数不能去掉

configs/kie/kie_unet_sdmgr.yml

@@ -17,7 +17,7 @@ Global:
   checkpoints:
   save_inference_dir:
   use_visualdl: False
-  class_path: ./train_data/wildreceipt/class_list.txt
+  class_path: &class_path ./train_data/wildreceipt/class_list.txt
   infer_img: ./train_data/wildreceipt/1.txt
   save_res_path: ./output/sdmgr_kie/predicts_kie.txt
   img_scale: [ 1024, 512 ]
@@ -72,6 +72,7 @@ Train:
           order: 'hwc'
       - KieLabelEncode: # Class handling label
           character_dict_path: ./train_data/wildreceipt/dict.txt
+          class_path: *class_path
       - KieResize:
       - ToCHWImage:
       - KeepKeys:
@@ -88,7 +89,6 @@ Eval:
     data_dir: ./train_data/wildreceipt
     label_file_list:
       - ./train_data/wildreceipt/wildreceipt_test.txt
-      # - /paddle/data/PaddleOCR/train_data/wildreceipt/1.txt
     transforms:
       - DecodeImage: # load image
           img_mode: RGB

configs/rec/rec_mtb_nrtr.yml

@@ -9,7 +9,7 @@ Global:
   eval_batch_step: [0, 2000]
   cal_metric_during_train: True
   pretrained_model:
-  checkpoints: 
+  checkpoints:
   save_inference_dir:
   use_visualdl: False
   infer_img: doc/imgs_words_en/word_10.png
@@ -49,7 +49,7 @@ Architecture:
     
 
 Loss:
-  name: NRTRLoss
+  name: CELoss
   smoothing: True
 
 PostProcess:
@@ -68,8 +68,8 @@ Train:
           img_mode: BGR
           channel_first: False
       - NRTRLabelEncode: # Class handling label
-      - NRTRRecResizeImg:
-          image_shape: [100, 32]
+      - GrayRecResizeImg:
+          image_shape: [100, 32] # W H
           resize_type: PIL # PIL or OpenCV
       - KeepKeys:
           keep_keys: ['image', 'label', 'length'] # dataloader will return list in this order
@@ -82,14 +82,14 @@ Train:
 Eval:
   dataset:
     name: LMDBDataSet
-    data_dir: ./train_data/data_lmdb_release/evaluation/
+    data_dir: ./train_data/data_lmdb_release/validation/
     transforms:
       - DecodeImage: # load image
           img_mode: BGR
           channel_first: False
       - NRTRLabelEncode: # Class handling label
-      - NRTRRecResizeImg:
-          image_shape: [100, 32]
+      - GrayRecResizeImg:
+          image_shape: [100, 32] # W H
           resize_type: PIL # PIL or OpenCV
       - KeepKeys:
           keep_keys: ['image', 'label', 'length'] # dataloader will return list in this order
@@ -97,5 +97,5 @@ Eval:
     shuffle: False
     drop_last: False
     batch_size_per_card: 256
-    num_workers: 1
+    num_workers: 4
     use_shared_memory: False

configs/rec/rec_r45_abinet.yml

+Global:
+  use_gpu: True
+  epoch_num: 10
+  log_smooth_window: 20
+  print_batch_step: 10
+  save_model_dir: ./output/rec/r45_abinet/
+  save_epoch_step: 1
+  # evaluation is run every 2000 iterations
+  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
+  max_text_length: 25
+  infer_mode: False
+  use_space_char: False
+  save_res_path: ./output/rec/predicts_abinet.txt
+
+Optimizer:
+  name: Adam
+  beta1: 0.9
+  beta2: 0.99
+  clip_norm: 20.0
+  lr:
+    name: Piecewise
+    decay_epochs: [6]
+    values: [0.0001, 0.00001] 
+  regularizer:
+    name: 'L2'
+    factor: 0.
+
+Architecture:
+  model_type: rec
+  algorithm: ABINet
+  in_channels: 3
+  Transform:
+  Backbone:
+    name: ResNet45
+  Head:
+    name: ABINetHead
+    use_lang: True
+    iter_size: 3
+    
+
+Loss:
+  name: CELoss
+  ignore_index: &ignore_index 100 # Must be greater than the number of character classes
+
+PostProcess:
+  name: ABINetLabelDecode
+
+Metric:
+  name: RecMetric
+  main_indicator: acc
+
+Train:
+  dataset:
+    name: LMDBDataSet
+    data_dir: ./train_data/data_lmdb_release/training/
+    transforms:
+      - DecodeImage: # load image
+          img_mode: RGB
+          channel_first: False
+      - ABINetRecAug:
+      - ABINetLabelEncode: # Class handling label
+          ignore_index: *ignore_index
+      - ABINetRecResizeImg:
+          image_shape: [3, 32, 128]
+      - KeepKeys:
+          keep_keys: ['image', 'label', 'length'] # dataloader will return list in this order
+  loader:
+    shuffle: True
+    batch_size_per_card: 96
+    drop_last: True
+    num_workers: 4
+
+Eval:
+  dataset:
+    name: LMDBDataSet
+    data_dir: ./train_data/data_lmdb_release/validation/
+    transforms:
+      - DecodeImage: # load image
+          img_mode: RGB
+          channel_first: False
+      - ABINetLabelEncode: # Class handling label
+          ignore_index: *ignore_index
+      - ABINetRecResizeImg:
+          image_shape: [3, 32, 128]
+      - KeepKeys:
+          keep_keys: ['image', 'label', 'length'] # dataloader will return list in this order
+  loader:
+    shuffle: False
+    drop_last: False
+    batch_size_per_card: 256
+    num_workers: 4
+    use_shared_memory: False

configs/rec/rec_svtrnet.yml

@@ -26,7 +26,7 @@ Optimizer:
   name: AdamW
   beta1: 0.9
   beta2: 0.99
-  epsilon: 0.00000008
+  epsilon: 8.e-8
   weight_decay: 0.05
   no_weight_decay_name: norm pos_embed
   one_dim_param_no_weight_decay: true
@@ -77,14 +77,13 @@ Metric:
 Train:
   dataset:
     name: LMDBDataSet
-    data_dir: ./train_data/data_lmdb_release/training/
+    data_dir: ./train_data/data_lmdb_release/training
     transforms:
       - DecodeImage: # load image
           img_mode: BGR
           channel_first: False
       - CTCLabelEncode: # Class handling label
-      - RecResizeImg:
-          character_dict_path:
+      - SVTRRecResizeImg:
           image_shape: [3, 64, 256]
           padding: False
       - KeepKeys:
@@ -98,14 +97,13 @@ Train:
 Eval:
   dataset:
     name: LMDBDataSet
-    data_dir: ./train_data/data_lmdb_release/validation/
+    data_dir: ./train_data/data_lmdb_release/validation
     transforms:
       - DecodeImage: # load image
           img_mode: BGR
           channel_first: False
       - CTCLabelEncode: # Class handling label
-      - RecResizeImg:
-          character_dict_path:
+      - SVTRRecResizeImg:
           image_shape: [3, 64, 256]
           padding: False
       - KeepKeys:

configs/rec/rec_svtrnet_ch.yml

+Global:
+  use_gpu: true
+  epoch_num: 100
+  log_smooth_window: 20
+  print_batch_step: 10
+  save_model_dir: ./output/rec/svtr_ch_all/
+  save_epoch_step: 10
+  eval_batch_step:
+  - 0
+  - 2000
+  cal_metric_during_train: true
+  pretrained_model: null
+  checkpoints: null
+  save_inference_dir: null
+  use_visualdl: false
+  infer_img: doc/imgs_words/ch/word_1.jpg
+  character_dict_path: ppocr/utils/ppocr_keys_v1.txt
+  max_text_length: 25
+  infer_mode: false
+  use_space_char: true
+  save_res_path: ./output/rec/predicts_svtr_tiny_ch_all.txt
+Optimizer:
+  name: AdamW
+  beta1: 0.9
+  beta2: 0.99
+  epsilon: 8.0e-08
+  weight_decay: 0.05
+  no_weight_decay_name: norm pos_embed
+  one_dim_param_no_weight_decay: true
+  lr:
+    name: Cosine
+    learning_rate: 0.0005
+    warmup_epoch: 2
+Architecture:
+  model_type: rec
+  algorithm: SVTR
+  Transform: null
+  Backbone:
+    name: SVTRNet
+    img_size:
+    - 32
+    - 320
+    out_char_num: 40
+    out_channels: 96
+    patch_merging: Conv
+    embed_dim:
+    - 64
+    - 128
+    - 256
+    depth:
+    - 3
+    - 6
+    - 3
+    num_heads:
+    - 2
+    - 4
+    - 8
+    mixer:
+    - Local
+    - Local
+    - Local
+    - Local
+    - Local
+    - Local
+    - Global
+    - Global
+    - Global
+    - Global
+    - Global
+    - Global
+    local_mixer:
+    - - 7
+      - 11
+    - - 7
+      - 11
+    - - 7
+      - 11
+    last_stage: true
+    prenorm: false
+  Neck:
+    name: SequenceEncoder
+    encoder_type: reshape
+  Head:
+    name: CTCHead
+Loss:
+  name: CTCLoss
+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
+    ext_op_transform_idx: 1
+    transforms:
+    - DecodeImage:
+        img_mode: BGR
+        channel_first: false
+    - RecConAug:
+        prob: 0.5
+        ext_data_num: 2
+        image_shape:
+        - 32
+        - 320
+        - 3
+    - RecAug: null
+    - CTCLabelEncode: null
+    - SVTRRecResizeImg:
+        image_shape:
+        - 3
+        - 32
+        - 320
+        padding: true
+    - KeepKeys:
+        keep_keys:
+        - image
+        - label
+        - length
+  loader:
+    shuffle: true
+    batch_size_per_card: 256
+    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: null
+    - SVTRRecResizeImg:
+        image_shape:
+        - 3
+        - 32
+        - 320
+        padding: true
+    - KeepKeys:
+        keep_keys:
+        - image
+        - label
+        - length
+  loader:
+    shuffle: false
+    drop_last: false
+    batch_size_per_card: 256
+    num_workers: 2
+profiler_options: null

configs/rec/rec_vitstr_none_ce.yml

+Global:
+  use_gpu: True
+  epoch_num: 20
+  log_smooth_window: 20
+  print_batch_step: 10
+  save_model_dir: ./output/rec/vitstr_none_ce/
+  save_epoch_step: 1
+  # evaluation is run every 2000 iterations after the 0th 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: ppocr/utils/EN_symbol_dict.txt
+  max_text_length: 25
+  infer_mode: False
+  use_space_char: False
+  save_res_path: ./output/rec/predicts_vitstr.txt
+
+
+Optimizer:
+  name: Adadelta
+  epsilon: 1.e-8
+  rho: 0.95
+  clip_norm: 5.0
+  lr:
+    learning_rate: 1.0
+
+Architecture:
+  model_type: rec
+  algorithm: ViTSTR
+  in_channels: 1
+  Transform:
+  Backbone:
+    name: ViTSTR
+    scale: tiny
+  Neck:
+    name: SequenceEncoder
+    encoder_type: reshape
+  Head:
+    name: CTCHead
+
+Loss:
+  name: CELoss
+  with_all: True
+  ignore_index: &ignore_index 0 # Must be zero or greater than the number of character classes
+
+PostProcess:
+  name: ViTSTRLabelDecode
+
+Metric:
+  name: RecMetric
+  main_indicator: acc
+
+Train:
+  dataset:
+    name: LMDBDataSet
+    data_dir: ./train_data/data_lmdb_release/training/
+    transforms:
+      - DecodeImage: # load image
+          img_mode: BGR
+          channel_first: False
+      - ViTSTRLabelEncode: # Class handling label
+          ignore_index: *ignore_index
+      - GrayRecResizeImg:
+          image_shape: [224, 224] # W H
+          resize_type: PIL # PIL or OpenCV
+          inter_type: 'Image.BICUBIC'
+          scale: false
+      - KeepKeys:
+          keep_keys: ['image', 'label', 'length'] # dataloader will return list in this order
+  loader:
+    shuffle: True
+    batch_size_per_card: 48
+    drop_last: True
+    num_workers: 8
+
+Eval:
+  dataset:
+    name: LMDBDataSet
+    data_dir: ./train_data/data_lmdb_release/validation/
+    transforms:
+      - DecodeImage: # load image
+          img_mode: BGR
+          channel_first: False
+      - ViTSTRLabelEncode: # Class handling label
+          ignore_index: *ignore_index
+      - GrayRecResizeImg:
+          image_shape: [224, 224] # W H
+          resize_type: PIL # PIL or OpenCV
+          inter_type: 'Image.BICUBIC'
+          scale: false
+      - KeepKeys:
+          keep_keys: ['image', 'label', 'length'] # dataloader will return list in this order
+  loader:
+    shuffle: False
+    drop_last: False
+    batch_size_per_card: 256
+    num_workers: 2

configs/vqa/ser/layoutlm.yml

@@ -43,7 +43,7 @@ Optimizer:
     
 PostProcess:
   name: VQASerTokenLayoutLMPostProcess
-  class_path: &class_path ppstructure/vqa/labels/labels_ser.txt
+  class_path: &class_path train_data/XFUND/class_list_xfun.txt
 
 Metric:
   name: VQASerTokenMetric
@@ -54,7 +54,7 @@ Train:
     name: SimpleDataSet
     data_dir: train_data/XFUND/zh_train/image
     label_file_list: 
-      - train_data/XFUND/zh_train/xfun_normalize_train.json
+      - train_data/XFUND/zh_train/train.json
     transforms:
       - DecodeImage: # load image
           img_mode: RGB
@@ -89,7 +89,7 @@ Eval:
     name: SimpleDataSet
     data_dir: train_data/XFUND/zh_val/image
     label_file_list:
-      - train_data/XFUND/zh_val/xfun_normalize_val.json
+      - train_data/XFUND/zh_val/val.json
     transforms:
       - DecodeImage: # load image
           img_mode: RGB

configs/vqa/ser/layoutlmv2.yml

@@ -44,7 +44,7 @@ Optimizer:
     
 PostProcess:
   name: VQASerTokenLayoutLMPostProcess
-  class_path: &class_path ppstructure/vqa/labels/labels_ser.txt
+  class_path: &class_path train_data/XFUND/class_list_xfun.txt
 
 Metric:
   name: VQASerTokenMetric
@@ -55,7 +55,7 @@ Train:
     name: SimpleDataSet
     data_dir: train_data/XFUND/zh_train/image
     label_file_list: 
-      - train_data/XFUND/zh_train/xfun_normalize_train.json
+      - train_data/XFUND/zh_train/train.json
     transforms:
       - DecodeImage: # load image
           img_mode: RGB
@@ -90,7 +90,7 @@ Eval:
     name: SimpleDataSet
     data_dir: train_data/XFUND/zh_val/image
     label_file_list:
-      - train_data/XFUND/zh_val/xfun_normalize_val.json
+      - train_data/XFUND/zh_val/val.json
     transforms:
       - DecodeImage: # load image
           img_mode: RGB

configs/vqa/ser/layoutxlm.yml

@@ -11,7 +11,7 @@ Global:
   save_inference_dir:
   use_visualdl: False
   seed: 2022
-  infer_img: doc/vqa/input/zh_val_42.jpg
+  infer_img: ppstructure/docs/vqa/input/zh_val_42.jpg
   save_res_path: ./output/ser
 
 Architecture:
@@ -54,7 +54,7 @@ Train:
     name: SimpleDataSet
     data_dir: train_data/XFUND/zh_train/image
     label_file_list: 
-      - train_data/XFUND/zh_train/xfun_normalize_train.json
+      - train_data/XFUND/zh_train/train.json
     ratio_list: [ 1.0 ]
     transforms:
       - DecodeImage: # load image
@@ -90,7 +90,7 @@ Eval:
     name: SimpleDataSet
     data_dir: train_data/XFUND/zh_val/image
     label_file_list:
-      - train_data/XFUND/zh_val/xfun_normalize_val.json
+      - train_data/XFUND/zh_val/val.json
     transforms:
       - DecodeImage: # load image
           img_mode: RGB

deploy/lite/config.txt

@@ -4,4 +4,5 @@ det_db_box_thresh  0.5
 det_db_unclip_ratio  1.6
 det_db_use_dilate 0
 det_use_polygon_score 1
-use_direction_classify  1
\ No newline at end of file
+use_direction_classify  1
+rec_image_height  32
\ No newline at end of file

deploy/lite/crnn_process.cc

@@ -19,25 +19,27 @@
 
 const std::vector<int> rec_image_shape{3, 32, 320};
 
-cv::Mat CrnnResizeImg(cv::Mat img, float wh_ratio) {
+cv::Mat CrnnResizeImg(cv::Mat img, float wh_ratio, int rec_image_height) {
   int imgC, imgH, imgW;
   imgC = rec_image_shape[0];
+  imgH = rec_image_height;
   imgW = rec_image_shape[2];
-  imgH = rec_image_shape[1];
 
-  imgW = int(32 * wh_ratio);
+  imgW = int(imgH * wh_ratio);
 
-  float ratio = static_cast<float>(img.cols) / static_cast<float>(img.rows);
+  float ratio = float(img.cols) / float(img.rows);
   int resize_w, resize_h;
+
   if (ceilf(imgH * ratio) > imgW)
     resize_w = imgW;
   else
-    resize_w = static_cast<int>(ceilf(imgH * ratio));
-  cv::Mat resize_img;
+    resize_w = int(ceilf(imgH * ratio));
+
   cv::resize(img, resize_img, cv::Size(resize_w, imgH), 0.f, 0.f,
              cv::INTER_LINEAR);
-
-  return resize_img;
+  cv::copyMakeBorder(resize_img, resize_img, 0, 0, 0,
+                     int(imgW - resize_img.cols), cv::BORDER_CONSTANT,
+                     {127, 127, 127});
 }
 
 std::vector<std::string> ReadDict(std::string path) {

deploy/lite/crnn_process.h

@@ -26,7 +26,7 @@
 #include "opencv2/imgcodecs.hpp"
 #include "opencv2/imgproc.hpp"
 
-cv::Mat CrnnResizeImg(cv::Mat img, float wh_ratio);
+cv::Mat CrnnResizeImg(cv::Mat img, float wh_ratio, int rec_image_height);
 
 std::vector<std::string> ReadDict(std::string path);
 

deploy/lite/ocr_db_crnn.cc

@@ -162,7 +162,8 @@ void RunRecModel(std::vector<std::vector<std::vector<int>>> boxes, cv::Mat img,
                  std::vector<std::string> charactor_dict,
                  std::shared_ptr<PaddlePredictor> predictor_cls,
                  int use_direction_classify,
-                 std::vector<double> *times) {
+                 std::vector<double> *times,
+                 int rec_image_height) {
   std::vector<float> mean = {0.5f, 0.5f, 0.5f};
   std::vector<float> scale = {1 / 0.5f, 1 / 0.5f, 1 / 0.5f};
 
@@ -183,7 +184,7 @@ void RunRecModel(std::vector<std::vector<std::vector<int>>> boxes, cv::Mat img,
     float wh_ratio =
         static_cast<float>(crop_img.cols) / static_cast<float>(crop_img.rows);
 
-    resize_img = CrnnResizeImg(crop_img, wh_ratio);
+    resize_img = CrnnResizeImg(crop_img, wh_ratio, rec_image_height);
     resize_img.convertTo(resize_img, CV_32FC3, 1 / 255.f);
 
     const float *dimg = reinterpret_cast<const float *>(resize_img.data);
@@ -444,7 +445,7 @@ void system(char **argv){
   //// load config from txt file
   auto Config = LoadConfigTxt(det_config_path);
   int use_direction_classify = int(Config["use_direction_classify"]);
-
+  int rec_image_height = int(Config["rec_image_height"]);
   auto charactor_dict = ReadDict(dict_path);
   charactor_dict.insert(charactor_dict.begin(), "#"); // blank char for ctc
   charactor_dict.push_back(" ");
@@ -590,12 +591,16 @@ void rec(int argc, char **argv) {
   std::string batchsize = argv[6];
   std::string img_dir = argv[7];
   std::string dict_path = argv[8];
+  std::string config_path = argv[9];
 
   if (strcmp(argv[4], "FP32") != 0 && strcmp(argv[4], "INT8") != 0) {
       std::cerr << "Only support FP32 or INT8." << std::endl;
       exit(1);
   }
 
+  auto Config = LoadConfigTxt(config_path);
+  int rec_image_height = int(Config["rec_image_height"]);
+
   std::vector<cv::String> cv_all_img_names;
   cv::glob(img_dir, cv_all_img_names);
 
@@ -630,7 +635,7 @@ void rec(int argc, char **argv) {
     std::vector<float> rec_text_score;
     std::vector<double> times;
     RunRecModel(boxes, srcimg, rec_predictor, rec_text, rec_text_score,
-                charactor_dict, cls_predictor, 0, &times);
+                charactor_dict, cls_predictor, 0, &times, rec_image_height);
   
     //// print recognized text
     for (int i = 0; i < rec_text.size(); i++) {

deploy/lite/readme.md

@@ -34,7 +34,7 @@ For the compilation process of different development environments, please refer
 ### 1.2 Prepare Paddle-Lite library
 
 There are two ways to obtain the Paddle-Lite library：
-- 1. Download directly, the download link of the Paddle-Lite library is as follows：
+- 1. [Recommended] Download directly, the download link of the Paddle-Lite library is as follows：
 
       | Platform | Paddle-Lite library download link |
       |---|---|
@@ -43,7 +43,9 @@ There are two ways to obtain the Paddle-Lite library：
 
       Note: 1. The above Paddle-Lite library is compiled from the Paddle-Lite 2.10 branch. For more information about Paddle-Lite 2.10, please refer to [link](https://github.com/PaddlePaddle/Paddle-Lite/releases/tag/v2.10).
 
-- 2. [Recommended] Compile Paddle-Lite to get the prediction library. The compilation method of Paddle-Lite is as follows：
+    **Note: It is recommended to use paddlelite>=2.10 version of the prediction library, other prediction library versions [download link](https://github.com/PaddlePaddle/Paddle-Lite/tags)**
+
+- 2. Compile Paddle-Lite to get the prediction library. The compilation method of Paddle-Lite is as follows：
 ```
 git clone https://github.com/PaddlePaddle/Paddle-Lite.git
 cd Paddle-Lite
@@ -104,21 +106,17 @@ If you directly use the model in the above table for deployment, you can skip th
 
 If the model to be deployed is not in the above table, you need to follow the steps below to obtain the optimized model.
 
-The `opt` tool can be obtained by compiling Paddle Lite.
+- Step 1: Refer to [document](https://www.paddlepaddle.org.cn/lite/v2.10/user_guides/opt/opt_python.html) to install paddlelite, which is used to convert paddle inference model to paddlelite required for running nb model
 ```
-git clone https://github.com/PaddlePaddle/Paddle-Lite.git
-cd Paddle-Lite
-git checkout release/v2.10
-./lite/tools/build.sh build_optimize_tool
+pip install paddlelite==2.10 # The paddlelite version should be the same as the prediction library version
 ```
-
-After the compilation is complete, the opt file is located under build.opt/lite/api/, You can view the operating options and usage of opt in the following ways:
-
+After installation, the following commands can view the help information
 ```
-cd build.opt/lite/api/
-./opt
+paddle_lite_opt
 ```
 
+Introduction to paddle_lite_opt parameters:
+
 |Options|Description|
 |---|---|
 |--model_dir|The path of the PaddlePaddle model to be optimized (non-combined form)|
@@ -131,6 +129,8 @@ cd build.opt/lite/api/
 
 `--model_dir` is suitable for the non-combined mode of the model to be optimized, and the inference model of PaddleOCR is the combined mode, that is, the model structure and model parameters are stored in a single file.
 
+- Step 2: Use paddle_lite_opt to convert the inference model to the mobile model format.
+
 The following takes the ultra-lightweight Chinese model of PaddleOCR as an example to introduce the use of the compiled opt file to complete the conversion of the inference model to the Paddle-Lite optimized model
 
 ```
@@ -240,6 +240,7 @@ det_db_thresh  0.3        # Used to filter the binarized image of DB prediction,
 det_db_box_thresh  0.5    # DDB post-processing filter box threshold, if there is a missing box detected, it can be reduced as appropriate
 det_db_unclip_ratio  1.6  # Indicates the compactness of the text box, the smaller the value, the closer the text box to the text
 use_direction_classify  0  # Whether to use the direction classifier, 0 means not to use, 1 means to use
+rec_image_height  32      # The height of the input image of the recognition model, the PP-OCRv3 model needs to be set to 48, and the PP-OCRv2 model needs to be set to 32
 ```
 
  5. Run Model on phone
@@ -258,8 +259,15 @@ After the above steps are completed, you can use adb to push the file to the pho
  cd /data/local/tmp/debug
  export LD_LIBRARY_PATH=${PWD}:$LD_LIBRARY_PATH
  # The use of ocr_db_crnn is:
- # ./ocr_db_crnn Detection model file Orientation classifier model file Recognition model file Test image path Dictionary file path
- ./ocr_db_crnn ch_PP-OCRv2_det_slim_opt.nb  ch_PP-OCRv2_rec_slim_opt.nb  ch_ppocr_mobile_v2.0_cls_opt.nb  ./11.jpg  ppocr_keys_v1.txt
+ # ./ocr_db_crnn Mode Detection model file Orientation classifier model file Recognition model file  Hardware  Precision  Threads Batchsize  Test image path Dictionary file path
+ ./ocr_db_crnn system ch_PP-OCRv2_det_slim_opt.nb  ch_PP-OCRv2_rec_slim_opt.nb  ch_ppocr_mobile_v2.0_cls_slim_opt.nb  arm8 INT8 10 1  ./11.jpg  config.txt  ppocr_keys_v1.txt  True
+# precision can be INT8 for quantitative model or FP32 for normal model.
+
+# Only using detection model
+./ocr_db_crnn  det ch_PP-OCRv2_det_slim_opt.nb arm8 INT8 10 1 ./11.jpg  config.txt
+
+# Only using recognition model
+./ocr_db_crnn  rec ch_PP-OCRv2_rec_slim_opt.nb arm8 INT8 10 1 word_1.jpg ppocr_keys_v1.txt config.txt
  ```
 
 If you modify the code, you need to recompile and push to the phone.
@@ -283,3 +291,7 @@ A2: Replace the .jpg test image under ./debug with the image you want to test, a
 Q3: How to package it into the mobile APP?
 
 A3: This demo aims to provide the core algorithm part that can run OCR on mobile phones. Further, PaddleOCR/deploy/android_demo is an example of encapsulating this demo into a mobile app for reference.
+
+Q4: When running the demo, an error is reported `Error: This model is not supported, because kernel for 'io_copy' is not supported by Paddle-Lite.`
+
+A4: The problem is that the installed paddlelite version does not match the downloaded prediction library version. Make sure that the paddleliteopt tool matches your prediction library version, and try to switch to the nb model again.

deploy/lite/readme_ch.md

@@ -8,7 +8,7 @@
     - [2.1 模型优化](#21-模型优化)
     - [2.2 与手机联调](#22-与手机联调)
 - [FAQ](#faq)
-  
+
 
 本教程将介绍基于[Paddle Lite](https://github.com/PaddlePaddle/Paddle-Lite) 在移动端部署PaddleOCR超轻量中文检测、识别模型的详细步骤。
 
@@ -32,7 +32,7 @@ Paddle Lite是飞桨轻量化推理引擎，为手机、IOT端提供高效推理
 ### 1.2 准备预测库
 
 预测库有两种获取方式：
-- 1. 直接下载，预测库下载链接如下：
+- 1. [推荐]直接下载，预测库下载链接如下：
 
       | 平台 | 预测库下载链接 |
       |---|---|
@@ -41,7 +41,9 @@ Paddle Lite是飞桨轻量化推理引擎，为手机、IOT端提供高效推理
 
       注：1. 上述预测库为PaddleLite 2.10分支编译得到，有关PaddleLite 2.10 详细信息可参考 [链接](https://github.com/PaddlePaddle/Paddle-Lite/releases/tag/v2.10) 。
 
-- 2. [推荐]编译Paddle-Lite得到预测库，Paddle-Lite的编译方式如下：
+**注：建议使用paddlelite>=2.10版本的预测库，其他预测库版本[下载链接](https://github.com/PaddlePaddle/Paddle-Lite/tags)**
+
+- 2. 编译Paddle-Lite得到预测库，Paddle-Lite的编译方式如下：
 ```
 git clone https://github.com/PaddlePaddle/Paddle-Lite.git
 cd Paddle-Lite
@@ -102,22 +104,16 @@ Paddle-Lite 提供了多种策略来自动优化原始的模型，其中包括
 
 如果要部署的模型不在上述表格中，则需要按照如下步骤获得优化后的模型。
 
-模型优化需要Paddle-Lite的opt可执行文件，可以通过编译Paddle-Lite源码获得，编译步骤如下：
+- 步骤1：参考[文档](https://www.paddlepaddle.org.cn/lite/v2.10/user_guides/opt/opt_python.html)安装paddlelite，用于转换paddle inference model为paddlelite运行所需的nb模型
 ```
-# 如果准备环境时已经clone了Paddle-Lite，则不用重新clone Paddle-Lite
-git clone https://github.com/PaddlePaddle/Paddle-Lite.git
-cd Paddle-Lite
-git checkout release/v2.10
-# 启动编译
-./lite/tools/build.sh build_optimize_tool
+pip install paddlelite==2.10  # paddlelite版本要与预测库版本一致
 ```
-
-编译完成后，opt文件位于`build.opt/lite/api/`下，可通过如下方式查看opt的运行选项和使用方式；
+安装完后，如下指令可以查看帮助信息
 ```
-cd build.opt/lite/api/
-./opt
+paddle_lite_opt
 ```
 
+paddle_lite_opt 参数介绍：
 |选项|说明|
 |---|---|
 |--model_dir|待优化的PaddlePaddle模型（非combined形式）的路径|
@@ -130,6 +126,8 @@ cd build.opt/lite/api/
 
 `--model_dir`适用于待优化的模型是非combined方式，PaddleOCR的inference模型是combined方式，即模型结构和模型参数使用单独一个文件存储。
 
+- 步骤2：使用paddle_lite_opt将inference模型转换成移动端模型格式。
+
 下面以PaddleOCR的超轻量中文模型为例，介绍使用编译好的opt文件完成inference模型到Paddle-Lite优化模型的转换。
 
 ```
@@ -148,7 +146,7 @@ wget  https://paddleocr.bj.bcebos.com/dygraph_v2.0/slim/ch_ppocr_mobile_v2.0_cls
 
 转换成功后，inference模型目录下会多出`.nb`结尾的文件，即是转换成功的模型文件。
 
-注意：使用paddle-lite部署时，需要使用opt工具优化后的模型。 opt 工具的输入模型是paddle保存的inference模型
+注意：使用paddle-lite部署时，需要使用opt工具优化后的模型。 opt工具的输入模型是paddle保存的inference模型
 
 <a name="2.2与手机联调"></a>
 ### 2.2 与手机联调
@@ -234,13 +232,14 @@ ppocr_keys_v1.txt   # 中文字典
 ...
 ```
 
-2.  `config.txt` 包含了检测器、分类器的超参数，如下：
+2.  `config.txt` 包含了检测器、分类器、识别器的超参数，如下：
 ```
 max_side_len  960         # 输入图像长宽大于960时，等比例缩放图像，使得图像最长边为960
 det_db_thresh  0.3        # 用于过滤DB预测的二值化图像，设置为0.-0.3对结果影响不明显
-det_db_box_thresh  0.5    # DB后处理过滤box的阈值，如果检测存在漏框情况，可酌情减小
+det_db_box_thresh  0.5    # 检测器后处理过滤box的阈值，如果检测存在漏框情况，可酌情减小
 det_db_unclip_ratio  1.6  # 表示文本框的紧致程度，越小则文本框更靠近文本
 use_direction_classify  0  # 是否使用方向分类器，0表示不使用，1表示使用
+rec_image_height  32      # 识别模型输入图像的高度，PP-OCRv3模型设置为48，PP-OCRv2模型需要设置为32
 ```
 
  5. 启动调试
@@ -259,8 +258,14 @@ use_direction_classify  0  # 是否使用方向分类器，0表示不使用，1
  cd /data/local/tmp/debug
  export LD_LIBRARY_PATH=${PWD}:$LD_LIBRARY_PATH
  # 开始使用，ocr_db_crnn可执行文件的使用方式为:
- # ./ocr_db_crnn  检测模型文件 方向分类器模型文件  识别模型文件  测试图像路径  字典文件路径
- ./ocr_db_crnn ch_PP-OCRv2_det_slim_opt.nb  ch_PP-OCRv2_rec_slim_opt.nb  ch_ppocr_mobile_v2.0_cls_slim_opt.nb  ./11.jpg  ppocr_keys_v1.txt
+ # ./ocr_db_crnn 预测模式  检测模型文件 方向分类器模型文件  识别模型文件 运行硬件 运行精度 线程数  batchsize  测试图像路径  参数配置路径  字典文件路径 是否使用benchmark参数
+ ./ocr_db_crnn system  ch_PP-OCRv2_det_slim_opt.nb  ch_PP-OCRv2_rec_slim_opt.nb  ch_ppocr_mobile_v2.0_cls_slim_opt.nb  arm8 INT8 10 1  ./11.jpg  config.txt  ppocr_keys_v1.txt  True
+
+# 仅使用文本检测模型，使用方式如下：
+./ocr_db_crnn  det ch_PP-OCRv2_det_slim_opt.nb arm8 INT8 10 1 ./11.jpg  config.txt
+
+# 仅使用文本识别模型，使用方式如下：
+./ocr_db_crnn  rec ch_PP-OCRv2_rec_slim_opt.nb arm8 INT8 10 1 word_1.jpg ppocr_keys_v1.txt config.txt
  ```
 
  如果对代码做了修改，则需要重新编译并push到手机上。
@@ -284,3 +289,7 @@ A2：替换debug下的.jpg测试图像为你想要测试的图像，adb push 到
 Q3：如何封装到手机APP中？
 
 A3：此demo旨在提供能在手机上运行OCR的核心算法部分，PaddleOCR/deploy/android_demo是将这个demo封装到手机app的示例，供参考
+
+Q4：运行demo时遇到报错`Error: This model is not supported, because kernel for 'io_copy' is not supported by Paddle-Lite.`
+
+A4：问题是安装的paddlelite版本和下载的预测库版本不匹配，确保paddleliteopt工具和你的预测库版本匹配，重新转nb模型试试。

doc/doc_ch/algorithm_overview.md

@@ -66,6 +66,8 @@
 - [x]  [SAR](./algorithm_rec_sar.md)
 - [x]  [SEED](./algorithm_rec_seed.md)
 - [x]  [SVTR](./algorithm_rec_svtr.md)
+- [x]  [ViTSTR](./algorithm_rec_vitstr.md)
+- [x]  [ABINet](./algorithm_rec_abinet.md)
 
 参考[DTRB](https://arxiv.org/abs/1904.01906)[3]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：
 
@@ -84,6 +86,8 @@
 |SAR|Resnet31| 87.20% | rec_r31_sar | [训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.1/rec/rec_r31_sar_train.tar) |
 |SEED|Aster_Resnet| 85.35% | rec_resnet_stn_bilstm_att | [训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.1/rec/rec_resnet_stn_bilstm_att.tar) |
 |SVTR|SVTR-Tiny| 89.25% | rec_svtr_tiny_none_ctc_en | [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/rec_svtr_tiny_none_ctc_en_train.tar) |
+|ViTSTR|ViTSTR| 79.82% | rec_vitstr_none_ce | [训练模型](https://paddleocr.bj.bcebos.com/rec_vitstr_none_ce_train.tar) |
+|ABINet|Resnet45| 90.75% | rec_r45_abinet | [训练模型](https://paddleocr.bj.bcebos.com/rec_r45_abinet_train.tar) |
 
 
 <a name="2"></a>

doc/doc_ch/algorithm_rec_abinet.md

+# 场景文本识别算法-ABINet
+
+- [1. 算法简介](#1)
+- [2. 环境配置](#2)
+- [3. 模型训练、评估、预测](#3)
+    - [3.1 训练](#3-1)
+    - [3.2 评估](#3-2)
+    - [3.3 预测](#3-3)
+- [4. 推理部署](#4)
+    - [4.1 Python推理](#4-1)
+    - [4.2 C++推理](#4-2)
+    - [4.3 Serving服务化部署](#4-3)
+    - [4.4 更多推理部署](#4-4)
+- [5. FAQ](#5)
+
+<a name="1"></a>
+## 1. 算法简介
+
+论文信息：
+> [ABINet: Read Like Humans: Autonomous, Bidirectional and Iterative Language Modeling for Scene Text Recognition](https://openaccess.thecvf.com/content/CVPR2021/papers/Fang_Read_Like_Humans_Autonomous_Bidirectional_and_Iterative_Language_Modeling_for_CVPR_2021_paper.pdf)
+> Shancheng Fang and Hongtao Xie and Yuxin Wang and Zhendong Mao and Yongdong Zhang
+> CVPR, 2021
+
+
+<a name="model"></a>
+`ABINet`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，算法复现效果如下：
+
+|模型|骨干网络|配置文件|Acc|下载链接|
+| --- | --- | --- | --- | --- |
+|ABINet|ResNet45|[rec_r45_abinet.yml](../../configs/rec/rec_r45_abinet.yml)|90.75%|[预训练、训练模型](https://paddleocr.bj.bcebos.com/rec_r45_abinet_train.tar)|
+
+<a name="2"></a>
+## 2. 环境配置
+请先参考[《运行环境准备》](./environment.md)配置PaddleOCR运行环境，参考[《项目克隆》](./clone.md)克隆项目代码。
+
+
+<a name="3"></a>
+## 3. 模型训练、评估、预测
+
+<a name="3-1"></a>
+### 3.1 模型训练
+
+请参考[文本识别训练教程](./recognition.md)。PaddleOCR对代码进行了模块化，训练`ABINet`识别模型时需要**更换配置文件**为`ABINet`的[配置文件](../../configs/rec/rec_r45_abinet.yml)。
+
+#### 启动训练
+
+
+具体地，在完成数据准备后，便可以启动训练，训练命令如下：
+```shell
+#单卡训练（训练周期长，不建议）
+python3 tools/train.py -c configs/rec/rec_r45_abinet.yml
+
+#多卡训练，通过--gpus参数指定卡号
+python3 -m paddle.distributed.launch --gpus '0,1,2,3'  tools/train.py -c configs/rec/rec_r45_abinet.yml
+```
+
+<a name="3-2"></a>
+### 3.2 评估
+
+可下载已训练完成的[模型文件](#model)，使用如下命令进行评估：
+
+```shell
+# 注意将pretrained_model的路径设置为本地路径。
+python3 -m paddle.distributed.launch --gpus '0' tools/eval.py -c configs/rec/rec_r45_abinet.yml -o Global.pretrained_model=./rec_r45_abinet_train/best_accuracy
+```
+
+<a name="3-3"></a>
+### 3.3 预测
+
+使用如下命令进行单张图片预测：
+```shell
+# 注意将pretrained_model的路径设置为本地路径。
+python3 tools/infer_rec.py -c configs/rec/rec_r45_abinet.yml -o Global.infer_img='./doc/imgs_words_en/word_10.png' Global.pretrained_model=./rec_r45_abinet_train/best_accuracy
+# 预测文件夹下所有图像时，可修改infer_img为文件夹，如 Global.infer_img='./doc/imgs_words_en/'。
+```
+
+
+<a name="4"></a>
+## 4. 推理部署
+
+<a name="4-1"></a>
+### 4.1 Python推理
+首先将训练得到best模型，转换成inference model。这里以训练完成的模型为例（[模型下载地址](https://paddleocr.bj.bcebos.com/rec_r45_abinet_train.tar) )，可以使用如下命令进行转换：
+
+```shell
+# 注意将pretrained_model的路径设置为本地路径。
+python3 tools/export_model.py -c configs/rec/rec_r45_abinet.yml -o Global.pretrained_model=./rec_r45_abinet_train/best_accuracy Global.save_inference_dir=./inference/rec_r45_abinet/
+```
+**注意：**
+- 如果您是在自己的数据集上训练的模型，并且调整了字典文件，请注意修改配置文件中的`character_dict_path`是否是所需要的字典文件。
+- 如果您修改了训练时的输入大小，请修改`tools/export_model.py`文件中的对应ABINet的`infer_shape`。
+
+转换成功后，在目录下有三个文件：
+```
+/inference/rec_r45_abinet/
+    ├── inference.pdiparams         # 识别inference模型的参数文件
+    ├── inference.pdiparams.info    # 识别inference模型的参数信息，可忽略
+    └── inference.pdmodel           # 识别inference模型的program文件
+```
+
+执行如下命令进行模型推理：
+
+```shell
+python3 tools/infer/predict_rec.py --image_dir='./doc/imgs_words_en/word_10.png' --rec_model_dir='./inference/rec_r45_abinet/' --rec_algorithm='ABINet' --rec_image_shape='3,32,128' --rec_char_dict_path='./ppocr/utils/ic15_dict.txt'
+# 预测文件夹下所有图像时，可修改image_dir为文件夹，如 --image_dir='./doc/imgs_words_en/'。
+```
+
+![](../imgs_words_en/word_10.png)
+
+执行命令后，上面图像的预测结果（识别的文本和得分）会打印到屏幕上，示例如下：
+结果如下：
+```shell
+Predicts of ./doc/imgs_words_en/word_10.png:('pain', 0.9999995231628418)
+```
+
+**注意**：
+
+- 训练上述模型采用的图像分辨率是[3，32，128]，需要通过参数`rec_image_shape`设置为您训练时的识别图像形状。
+- 在推理时需要设置参数`rec_char_dict_path`指定字典，如果您修改了字典，请修改该参数为您的字典文件。
+- 如果您修改了预处理方法，需修改`tools/infer/predict_rec.py`中ABINet的预处理为您的预处理方法。
+
+
+<a name="4-2"></a>
+### 4.2 C++推理部署
+
+由于C++预处理后处理还未支持ABINet，所以暂未支持
+
+<a name="4-3"></a>
+### 4.3 Serving服务化部署
+
+暂不支持
+
+<a name="4-4"></a>
+### 4.4 更多推理部署
+
+暂不支持
+
+<a name="5"></a>
+## 5. FAQ
+
+1. MJSynth和SynthText两种数据集来自于[ABINet源repo](https://github.com/FangShancheng/ABINet) 。
+2. 我们使用ABINet作者提供的预训练模型进行finetune训练。
+
+## 引用
+
+```bibtex
+@article{Fang2021ABINet,
+  title     = {ABINet: Read Like Humans: Autonomous, Bidirectional and Iterative Language Modeling for Scene Text Recognition},
+  author    = {Shancheng Fang and Hongtao Xie and Yuxin Wang and Zhendong Mao and Yongdong Zhang},
+  booktitle = {CVPR},
+  year      = {2021},
+  url       = {https://arxiv.org/abs/2103.06495},
+  pages     = {7098-7107}
+}
+```

doc/doc_ch/algorithm_rec_nrtr.md

@@ -12,6 +12,7 @@
     - [4.3 Serving服务化部署](#4-3)
     - [4.4 更多推理部署](#4-4)
 - [5. FAQ](#5)
+- [6. 发行公告](#6)
 
 <a name="1"></a>
 ## 1. 算法简介
@@ -110,7 +111,7 @@ python3 tools/infer/predict_rec.py --image_dir='./doc/imgs_words_en/word_10.png'
 执行命令后，上面图像的预测结果（识别的文本和得分）会打印到屏幕上，示例如下：
 结果如下：
 ```shell
-Predicts of ./doc/imgs_words_en/word_10.png:('pain', 0.9265879392623901)
+Predicts of ./doc/imgs_words_en/word_10.png:('pain', 0.9465042352676392)
 ```
 
 **注意**：
@@ -140,12 +141,147 @@ Predicts of ./doc/imgs_words_en/word_10.png:('pain', 0.9265879392623901)
 
 1. `NRTR`论文中使用Beam搜索进行解码字符，但是速度较慢，这里默认未使用Beam搜索，以贪婪搜索进行解码字符。
 
+<a name="6"></a>
+## 6. 发行公告
+
+1. release/2.6更新NRTR代码结构，新版NRTR可加载旧版(release/2.5及之前)模型参数，使用下面示例代码将旧版模型参数转换为新版模型参数：
+
+```python
+
+    params = paddle.load('path/' + '.pdparams') # 旧版本参数
+    state_dict = model.state_dict() # 新版模型参数
+    new_state_dict = {}
+
+    for k1, v1 in state_dict.items():
+
+        k = k1
+        if 'encoder' in k and 'self_attn' in k and 'qkv' in k and 'weight' in k:
+
+            k_para = k[:13] + 'layers.' + k[13:]
+            q = params[k_para.replace('qkv', 'conv1')].transpose((1, 0, 2, 3))
+            k = params[k_para.replace('qkv', 'conv2')].transpose((1, 0, 2, 3))
+            v = params[k_para.replace('qkv', 'conv3')].transpose((1, 0, 2, 3))
+
+            new_state_dict[k1] = np.concatenate([q[:, :, 0, 0], k[:, :, 0, 0], v[:, :, 0, 0]], -1)
+
+        elif 'encoder' in k and 'self_attn' in k and 'qkv' in k and 'bias' in k:
+
+            k_para = k[:13] + 'layers.' + k[13:]
+            q = params[k_para.replace('qkv', 'conv1')]
+            k = params[k_para.replace('qkv', 'conv2')]
+            v = params[k_para.replace('qkv', 'conv3')]
+
+            new_state_dict[k1] = np.concatenate([q, k, v], -1)
+
+        elif 'encoder' in k and 'self_attn' in k and 'out_proj' in k:
+
+            k_para = k[:13] + 'layers.' + k[13:]
+            new_state_dict[k1] = params[k_para]
+
+        elif 'encoder' in k and 'norm3' in k:
+            k_para = k[:13] + 'layers.' + k[13:]
+            new_state_dict[k1] = params[k_para.replace('norm3', 'norm2')]
+
+        elif 'encoder' in k and 'norm1' in k:
+            k_para = k[:13] + 'layers.' + k[13:]
+            new_state_dict[k1] = params[k_para]
+
+
+        elif 'decoder' in k and 'self_attn' in k and 'qkv' in k and 'weight' in k:
+            k_para = k[:13] + 'layers.' + k[13:]
+            q = params[k_para.replace('qkv', 'conv1')].transpose((1, 0, 2, 3))
+            k = params[k_para.replace('qkv', 'conv2')].transpose((1, 0, 2, 3))
+            v = params[k_para.replace('qkv', 'conv3')].transpose((1, 0, 2, 3))
+            new_state_dict[k1] = np.concatenate([q[:, :, 0, 0], k[:, :, 0, 0], v[:, :, 0, 0]], -1)
+
+        elif 'decoder' in k and 'self_attn' in k and 'qkv' in k and 'bias' in k:
+            k_para = k[:13] + 'layers.' + k[13:]
+            q = params[k_para.replace('qkv', 'conv1')]
+            k = params[k_para.replace('qkv', 'conv2')]
+            v = params[k_para.replace('qkv', 'conv3')]
+            new_state_dict[k1] = np.concatenate([q, k, v], -1)
+
+        elif 'decoder' in k and 'self_attn' in k and 'out_proj' in k:
+
+            k_para = k[:13] + 'layers.' + k[13:]
+            new_state_dict[k1] = params[k_para]
+
+        elif 'decoder' in k and 'cross_attn' in k and 'q' in k and 'weight' in k:
+            k_para = k[:13] + 'layers.' + k[13:]
+            k_para = k_para.replace('cross_attn', 'multihead_attn')
+            q = params[k_para.replace('q', 'conv1')].transpose((1, 0, 2, 3))
+            new_state_dict[k1] = q[:, :, 0, 0]
+
+        elif 'decoder' in k and 'cross_attn' in k and 'q' in k and 'bias' in k:
+            k_para = k[:13] + 'layers.' + k[13:]
+            k_para = k_para.replace('cross_attn', 'multihead_attn')
+            q = params[k_para.replace('q', 'conv1')]
+            new_state_dict[k1] = q
+
+        elif 'decoder' in k and 'cross_attn' in k and 'kv' in k and 'weight' in k:
+            k_para = k[:13] + 'layers.' + k[13:]
+            k_para = k_para.replace('cross_attn', 'multihead_attn')
+            k = params[k_para.replace('kv', 'conv2')].transpose((1, 0, 2, 3))
+            v = params[k_para.replace('kv', 'conv3')].transpose((1, 0, 2, 3))
+            new_state_dict[k1] = np.concatenate([k[:, :, 0, 0], v[:, :, 0, 0]], -1)
+
+        elif 'decoder' in k and 'cross_attn' in k and 'kv' in k and 'bias' in k:
+            k_para = k[:13] + 'layers.' + k[13:]
+            k_para = k_para.replace('cross_attn', 'multihead_attn')
+            k = params[k_para.replace('kv', 'conv2')]
+            v = params[k_para.replace('kv', 'conv3')]
+            new_state_dict[k1] = np.concatenate([k, v], -1)
+
+        elif 'decoder' in k and 'cross_attn' in k and 'out_proj' in k:
+
+            k_para = k[:13] + 'layers.' + k[13:]
+            k_para = k_para.replace('cross_attn', 'multihead_attn')
+            new_state_dict[k1] = params[k_para]
+        elif 'decoder' in k and 'norm' in k:
+            k_para = k[:13] + 'layers.' + k[13:]
+            new_state_dict[k1] = params[k_para]
+        elif 'mlp' in k and 'weight' in k:
+            k_para = k[:13] + 'layers.' + k[13:]
+            k_para = k_para.replace('fc', 'conv')
+            k_para = k_para.replace('mlp.', '')
+            w = params[k_para].transpose((1, 0, 2, 3))
+            new_state_dict[k1] = w[:, :, 0, 0]
+        elif 'mlp' in k and 'bias' in k:
+            k_para = k[:13] + 'layers.' + k[13:]
+            k_para = k_para.replace('fc', 'conv')
+            k_para = k_para.replace('mlp.', '')
+            w = params[k_para]
+            new_state_dict[k1] = w
+
+        else:
+            new_state_dict[k1] = params[k1]
+
+        if list(new_state_dict[k1].shape) != list(v1.shape):
+            print(k1)
+
+
+    for k, v1 in state_dict.items():
+        if k not in new_state_dict.keys():
+            print(1, k)
+        elif list(new_state_dict[k].shape) != list(v1.shape):
+            print(2, k)
+
+
+
+    model.set_state_dict(new_state_dict)
+    paddle.save(model.state_dict(), 'nrtrnew_from_old_params.pdparams')
+
+```
+
+2. 新版相比与旧版，代码结构简洁，推理速度有所提高。
+
+
 ## 引用
 
 ```bibtex
 @article{Sheng2019NRTR,
   title     = {NRTR: A No-Recurrence Sequence-to-Sequence Model For Scene Text Recognition},
-  author    = {Fenfen Sheng and Zhineng Chen andBo Xu},
+  author    = {Fenfen Sheng and Zhineng Chen and Bo Xu},
   booktitle = {ICDAR},
   year      = {2019},
   url       = {http://arxiv.org/abs/1806.00926},

doc/doc_ch/algorithm_rec_svtr.md

@@ -111,7 +111,6 @@ python3 tools/export_model.py -c ./rec_svtr_tiny_none_ctc_en_train/rec_svtr_tiny
 
 **注意：**
 - 如果您是在自己的数据集上训练的模型，并且调整了字典文件，请注意修改配置文件中的`character_dict_path`是否为所正确的字典文件。
-- 如果您修改了训练时的输入大小，请修改`tools/export_model.py`文件中的对应SVTR的`infer_shape`。
 
 转换成功后，在目录下有三个文件：
 ```

doc/doc_ch/algorithm_rec_vitstr.md

+# 场景文本识别算法-ViTSTR
+
+- [1. 算法简介](#1)
+- [2. 环境配置](#2)
+- [3. 模型训练、评估、预测](#3)
+    - [3.1 训练](#3-1)
+    - [3.2 评估](#3-2)
+    - [3.3 预测](#3-3)
+- [4. 推理部署](#4)
+    - [4.1 Python推理](#4-1)
+    - [4.2 C++推理](#4-2)
+    - [4.3 Serving服务化部署](#4-3)
+    - [4.4 更多推理部署](#4-4)
+- [5. FAQ](#5)
+
+<a name="1"></a>
+## 1. 算法简介
+
+论文信息：
+> [Vision Transformer for Fast and Efficient Scene Text Recognition](https://arxiv.org/abs/2105.08582)
+> Rowel Atienza
+> ICDAR, 2021
+
+
+<a name="model"></a>
+`ViTSTR`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，算法复现效果如下：
+
+|模型|骨干网络|配置文件|Acc|下载链接|
+| --- | --- | --- | --- | --- |
+|ViTSTR|ViTSTR|[rec_vitstr_none_ce.yml](../../configs/rec/rec_vitstr_none_ce.yml)|79.82%|[训练模型](https://paddleocr.bj.bcebos.com/rec_vitstr_none_ce_train.tar)|
+
+<a name="2"></a>
+## 2. 环境配置
+请先参考[《运行环境准备》](./environment.md)配置PaddleOCR运行环境，参考[《项目克隆》](./clone.md)克隆项目代码。
+
+
+<a name="3"></a>
+## 3. 模型训练、评估、预测
+
+<a name="3-1"></a>
+### 3.1 模型训练
+
+请参考[文本识别训练教程](./recognition.md)。PaddleOCR对代码进行了模块化，训练`ViTSTR`识别模型时需要**更换配置文件**为`ViTSTR`的[配置文件](../../configs/rec/rec_vitstr_none_ce.yml)。
+
+#### 启动训练
+
+
+具体地，在完成数据准备后，便可以启动训练，训练命令如下：
+```shell
+#单卡训练（训练周期长，不建议）
+python3 tools/train.py -c configs/rec/rec_vitstr_none_ce.yml
+
+#多卡训练，通过--gpus参数指定卡号
+python3 -m paddle.distributed.launch --gpus '0,1,2,3'  tools/train.py -c configs/rec/rec_vitstr_none_ce.yml
+```
+
+<a name="3-2"></a>
+### 3.2 评估
+
+可下载已训练完成的[模型文件](#model)，使用如下命令进行评估：
+
+```shell
+# 注意将pretrained_model的路径设置为本地路径。
+python3 -m paddle.distributed.launch --gpus '0' tools/eval.py -c configs/rec/rec_vitstr_none_ce.yml -o Global.pretrained_model=./rec_vitstr_none_ce_train/best_accuracy
+```
+
+<a name="3-3"></a>
+### 3.3 预测
+
+使用如下命令进行单张图片预测：
+```shell
+# 注意将pretrained_model的路径设置为本地路径。
+python3 tools/infer_rec.py -c configs/rec/rec_vitstr_none_ce.yml -o Global.infer_img='./doc/imgs_words_en/word_10.png' Global.pretrained_model=./rec_vitstr_none_ce_train/best_accuracy
+# 预测文件夹下所有图像时，可修改infer_img为文件夹，如 Global.infer_img='./doc/imgs_words_en/'。
+```
+
+
+<a name="4"></a>
+## 4. 推理部署
+
+<a name="4-1"></a>
+### 4.1 Python推理
+首先将训练得到best模型，转换成inference model。这里以训练完成的模型为例（[模型下载地址](https://paddleocr.bj.bcebos.com/rec_vitstr_none_ce_train.tar) )，可以使用如下命令进行转换：
+
+```shell
+# 注意将pretrained_model的路径设置为本地路径。
+python3 tools/export_model.py -c configs/rec/rec_vitstr_none_ce.yml -o Global.pretrained_model=./rec_vitstr_none_ce_train/best_accuracy Global.save_inference_dir=./inference/rec_vitstr/
+```
+**注意：**
+- 如果您是在自己的数据集上训练的模型，并且调整了字典文件，请注意修改配置文件中的`character_dict_path`是否是所需要的字典文件。
+- 如果您修改了训练时的输入大小，请修改`tools/export_model.py`文件中的对应ViTSTR的`infer_shape`。
+
+转换成功后，在目录下有三个文件：
+```
+/inference/rec_vitstr/
+    ├── inference.pdiparams         # 识别inference模型的参数文件
+    ├── inference.pdiparams.info    # 识别inference模型的参数信息，可忽略
+    └── inference.pdmodel           # 识别inference模型的program文件
+```
+
+执行如下命令进行模型推理：
+
+```shell
+python3 tools/infer/predict_rec.py --image_dir='./doc/imgs_words_en/word_10.png' --rec_model_dir='./inference/rec_vitstr/' --rec_algorithm='ViTSTR' --rec_image_shape='1,224,224' --rec_char_dict_path='./ppocr/utils/EN_symbol_dict.txt'
+# 预测文件夹下所有图像时，可修改image_dir为文件夹，如 --image_dir='./doc/imgs_words_en/'。
+```
+
+![](../imgs_words_en/word_10.png)
+
+执行命令后，上面图像的预测结果（识别的文本和得分）会打印到屏幕上，示例如下：
+结果如下：
+```shell
+Predicts of ./doc/imgs_words_en/word_10.png:('pain', 0.9998350143432617)
+```
+
+**注意**：
+
+- 训练上述模型采用的图像分辨率是[1，224，224]，需要通过参数`rec_image_shape`设置为您训练时的识别图像形状。
+- 在推理时需要设置参数`rec_char_dict_path`指定字典，如果您修改了字典，请修改该参数为您的字典文件。
+- 如果您修改了预处理方法，需修改`tools/infer/predict_rec.py`中ViTSTR的预处理为您的预处理方法。
+
+
+<a name="4-2"></a>
+### 4.2 C++推理部署
+
+由于C++预处理后处理还未支持ViTSTR，所以暂未支持
+
+<a name="4-3"></a>
+### 4.3 Serving服务化部署
+
+暂不支持
+
+<a name="4-4"></a>
+### 4.4 更多推理部署
+
+暂不支持
+
+<a name="5"></a>
+## 5. FAQ
+
+1. 在`ViTSTR`论文中，使用在ImageNet1k上的预训练权重进行初始化训练，我们在训练未采用预训练权重，最终精度没有变化甚至有所提高。
+2. 我们仅仅复现了`ViTSTR`中的tiny版本，如果需要使用small、base版本，可将[ViTSTR源repo](https://github.com/roatienza/deep-text-recognition-benchmark) 中的预训练权重转为Paddle权重使用。
+
+## 引用
+
+```bibtex
+@article{Atienza2021ViTSTR,
+  title     = {Vision Transformer for Fast and Efficient Scene Text Recognition},
+  author    = {Rowel Atienza},
+  booktitle = {ICDAR},
+  year      = {2021},
+  url       = {https://arxiv.org/abs/2105.08582}
+}
+```

doc/doc_ch/application.md

+# 场景应用
+
+PaddleOCR场景应用覆盖通用，制造、金融、交通行业的主要OCR垂类应用，在PP-OCR、PP-Structure的通用能力基础之上，以notebook的形式展示利用场景数据微调、模型优化方法、数据增广等内容，为开发者快速落地OCR应用提供示范与启发。
+
+> 如需下载全部垂类模型，可以扫描下方二维码，关注公众号填写问卷后，加入PaddleOCR官方交流群获取20G OCR学习大礼包（内含《动手学OCR》电子书、课程回放视频、前沿论文等重磅资料）
+
+<div align="center">
+<img src="https://ai-studio-static-online.cdn.bcebos.com/dd721099bd50478f9d5fb13d8dd00fad69c22d6848244fd3a1d3980d7fefc63e"  width = "150" height = "150" />
+</div>
+
+
+> 如果您是企业开发者且未在下述场景中找到合适的方案，可以填写[OCR应用合作调研问卷](https://paddle.wjx.cn/vj/QwF7GKw.aspx)，免费与官方团队展开不同层次的合作，包括但不限于问题抽象、确定技术方案、项目答疑、共同研发等。如果您已经使用PaddleOCR落地项目，也可以填写此问卷，与飞桨平台共同宣传推广，提升企业技术品宣。期待您的提交！
+
+## 通用
+
+| 类别                   | 亮点     | 类别       | 亮点         |
+| ---------------------- | -------- | ---------- | ------------ |
+| 高精度中文识别模型SVTR | 新增模型 | 手写体识别 | 新增字形支持 |
+
+## 制造
+
+| 类别           | 亮点                           | 类别           | 亮点                 |
+| -------------- | ------------------------------ | -------------- | -------------------- |
+| 数码管识别     | 数码管数据合成、漏识别调优     | 电表识别       | 大分辨率图像检测调优 |
+| 液晶屏读数识别 | 检测模型蒸馏、Serving部署      | PCB文字识别    | 小尺寸文本检测与识别 |
+| 包装生产日期   | 点阵字符合成、过曝过暗文字识别 | 液晶屏缺陷检测 | 非文字形态识别       |
+
+## 金融
+
+| 类别           | 亮点                     | 类别         | 亮点                  |
+| -------------- | ------------------------ | ------------ | --------------------- |
+| 表单VQA        | 多模态通用表单结构化提取 | 通用卡证识别 | 通用结构化提取        |
+| 增值税发票     | 尽请期待                 | 身份证识别   | 结构化提取、图像阴影  |
+| 印章检测与识别 | 端到端弯曲文本识别       | 合同比对     | 密集文本检测、NLP串联 |
+
+## 交通
+
+| 类别              | 亮点                           | 类别       | 亮点     |
+| ----------------- | ------------------------------ | ---------- | -------- |
+| 车牌识别          | 多角度图像、轻量模型、端侧部署 | 快递单识别 | 尽请期待 |
+| 驾驶证/行驶证识别 | 尽请期待                       |            |          |
\ No newline at end of file

doc/doc_en/algorithm_overview_en.md

@@ -65,6 +65,8 @@ Supported text recognition algorithms (Click the link to get the tutorial):
 - [x]  [SAR](./algorithm_rec_sar_en.md)
 - [x]  [SEED](./algorithm_rec_seed_en.md)
 - [x]  [SVTR](./algorithm_rec_svtr_en.md)
+- [x]  [ViTSTR](./algorithm_rec_vitstr_en.md)
+- [x]  [ABINet](./algorithm_rec_abinet_en.md)
 
 Refer to [DTRB](https://arxiv.org/abs/1904.01906), the training and evaluation result of these above text recognition (using MJSynth and SynthText for training, evaluate on IIIT, SVT, IC03, IC13, IC15, SVTP, CUTE) is as follow:
 
@@ -83,6 +85,8 @@ Refer to [DTRB](https://arxiv.org/abs/1904.01906), the training and evaluation r
 |SAR|Resnet31| 87.20% | rec_r31_sar | [trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.1/rec/rec_r31_sar_train.tar) |
 |SEED|Aster_Resnet| 85.35% | rec_resnet_stn_bilstm_att | [trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.1/rec/rec_resnet_stn_bilstm_att.tar) |
 |SVTR|SVTR-Tiny| 89.25% | rec_svtr_tiny_none_ctc_en | [trained model](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/rec_svtr_tiny_none_ctc_en_train.tar) |
+|ViTSTR|ViTSTR| 79.82% | rec_vitstr_none_ce | [trained model](https://paddleocr.bj.bcebos.com/rec_vitstr_none_none_train.tar) |
+|ABINet|Resnet45| 90.75% | rec_r45_abinet | [trained model](https://paddleocr.bj.bcebos.com/rec_r45_abinet_train.tar) |
 
 
 <a name="2"></a>

doc/doc_en/algorithm_rec_abinet_en.md

+# ABINet
+
+- [1. Introduction](#1)
+- [2. Environment](#2)
+- [3. Model Training / Evaluation / Prediction](#3)
+    - [3.1 Training](#3-1)
+    - [3.2 Evaluation](#3-2)
+    - [3.3 Prediction](#3-3)
+- [4. Inference and Deployment](#4)
+    - [4.1 Python Inference](#4-1)
+    - [4.2 C++ Inference](#4-2)
+    - [4.3 Serving](#4-3)
+    - [4.4 More](#4-4)
+- [5. FAQ](#5)
+
+<a name="1"></a>
+## 1. Introduction
+
+Paper:
+> [ABINet: Read Like Humans: Autonomous, Bidirectional and Iterative Language Modeling for Scene Text Recognition](https://openaccess.thecvf.com/content/CVPR2021/papers/Fang_Read_Like_Humans_Autonomous_Bidirectional_and_Iterative_Language_Modeling_for_CVPR_2021_paper.pdf)
+> Shancheng Fang and Hongtao Xie and Yuxin Wang and Zhendong Mao and Yongdong Zhang
+> CVPR, 2021
+
+Using MJSynth and SynthText two text recognition datasets for training, and evaluating on IIIT, SVT, IC03, IC13, IC15, SVTP, CUTE datasets, the algorithm reproduction effect is as follows:
+
+|Model|Backbone|config|Acc|Download link|
+| --- | --- | --- | --- | --- |
+|ABINet|ResNet45|[rec_r45_abinet.yml](../../configs/rec/rec_r45_abinet.yml)|90.75%|[pretrained & trained model](https://paddleocr.bj.bcebos.com/rec_r45_abinet_train.tar)|
+
+<a name="2"></a>
+## 2. Environment
+Please refer to ["Environment Preparation"](./environment_en.md) to configure the PaddleOCR environment, and refer to ["Project Clone"](./clone_en.md) to clone the project code.
+
+
+<a name="3"></a>
+## 3. Model Training / Evaluation / Prediction
+
+Please refer to [Text Recognition Tutorial](./recognition_en.md). PaddleOCR modularizes the code, and training different recognition models only requires **changing the configuration file**.
+
+Training:
+
+Specifically, after the data preparation is completed, the training can be started. The training command is as follows:
+
+```
+#Single GPU training (long training period, not recommended)
+python3 tools/train.py -c configs/rec/rec_r45_abinet.yml
+
+#Multi GPU training, specify the gpu number through the --gpus parameter
+python3 -m paddle.distributed.launch --gpus '0,1,2,3'  tools/train.py -c configs/rec/rec_r45_abinet.yml
+```
+
+Evaluation:
+
+```
+# GPU evaluation
+python3 -m paddle.distributed.launch --gpus '0' tools/eval.py -c configs/rec/rec_r45_abinet.yml -o Global.pretrained_model={path/to/weights}/best_accuracy
+```
+
+Prediction:
+
+```
+# The configuration file used for prediction must match the training
+python3 tools/infer_rec.py -c configs/rec/rec_r45_abinet.yml -o Global.infer_img='./doc/imgs_words_en/word_10.png' Global.pretrained_model=./rec_r45_abinet_train/best_accuracy
+```
+
+<a name="4"></a>
+## 4. Inference and Deployment
+
+<a name="4-1"></a>
+### 4.1 Python Inference
+First, the model saved during the ABINet text recognition training process is converted into an inference model. ( [Model download link](https://paddleocr.bj.bcebos.com/rec_r45_abinet_train.tar)) ), you can use the following command to convert:
+
+```
+python3 tools/export_model.py -c configs/rec/rec_r45_abinet.yml -o Global.pretrained_model=./rec_r45_abinet_train/best_accuracy  Global.save_inference_dir=./inference/rec_r45_abinet
+```
+
+**Note:**
+- If you are training the model on your own dataset and have modified the dictionary file, please pay attention to modify the `character_dict_path` in the configuration file to the modified dictionary file.
+- If you modified the input size during training, please modify the `infer_shape` corresponding to ABINet in the `tools/export_model.py` file.
+
+After the conversion is successful, there are three files in the directory:
+```
+/inference/rec_r45_abinet/
+    ├── inference.pdiparams
+    ├── inference.pdiparams.info
+    └── inference.pdmodel
+```
+
+
+For ABINet text recognition model inference, the following commands can be executed:
+
+```
+python3 tools/infer/predict_rec.py --image_dir='./doc/imgs_words_en/word_10.png' --rec_model_dir='./inference/rec_r45_abinet/' --rec_algorithm='ABINet' --rec_image_shape='3,32,128' --rec_char_dict_path='./ppocr/utils/ic15_dict.txt'
+```
+
+![](../imgs_words_en/word_10.png)
+
+After executing the command, the prediction result (recognized text and score) of the image above is printed to the screen, an example is as follows:
+The result is as follows:
+```shell
+Predicts of ./doc/imgs_words_en/word_10.png:('pain', 0.9999995231628418)
+```
+
+<a name="4-2"></a>
+### 4.2 C++ Inference
+
+Not supported
+
+<a name="4-3"></a>
+### 4.3 Serving
+
+Not supported
+
+<a name="4-4"></a>
+### 4.4 More
+
+Not supported
+
+<a name="5"></a>
+## 5. FAQ
+
+1. Note that the MJSynth and SynthText datasets come from [ABINet repo](https://github.com/FangShancheng/ABINet).
+2. We use the pre-trained model provided by the ABINet authors for finetune training.
+
+## Citation
+
+```bibtex
+@article{Fang2021ABINet,
+  title     = {ABINet: Read Like Humans: Autonomous, Bidirectional and Iterative Language Modeling for Scene Text Recognition},
+  author    = {Shancheng Fang and Hongtao Xie and Yuxin Wang and Zhendong Mao and Yongdong Zhang},
+  booktitle = {CVPR},
+  year      = {2021},
+  url       = {https://arxiv.org/abs/2103.06495},
+  pages     = {7098-7107}
+}
+```

doc/doc_en/algorithm_rec_nrtr_en.md

@@ -12,6 +12,7 @@
     - [4.3 Serving](#4-3)
     - [4.4 More](#4-4)
 - [5. FAQ](#5)
+- [6. Release Note](#6)
 
 <a name="1"></a>
 ## 1. Introduction
@@ -25,7 +26,7 @@ Using MJSynth and SynthText two text recognition datasets for training, and eval
 
 |Model|Backbone|config|Acc|Download link|
 | --- | --- | --- | --- | --- |
-|NRTR|MTB|[rec_mtb_nrtr.yml](../../configs/rec/rec_mtb_nrtr.yml)|84.21%|[train model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/rec_mtb_nrtr_train.tar)|
+|NRTR|MTB|[rec_mtb_nrtr.yml](../../configs/rec/rec_mtb_nrtr.yml)|84.21%|[trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/rec_mtb_nrtr_train.tar)|
 
 <a name="2"></a>
 ## 2. Environment
@@ -98,7 +99,7 @@ python3 tools/infer/predict_rec.py --image_dir='./doc/imgs_words_en/word_10.png'
 After executing the command, the prediction result (recognized text and score) of the image above is printed to the screen, an example is as follows:
 The result is as follows:
 ```shell
-Predicts of ./doc/imgs_words_en/word_10.png:('pain', 0.9265879392623901)
+Predicts of ./doc/imgs_words_en/word_10.png:('pain', 0.9465042352676392)
 ```
 
 <a name="4-2"></a>
@@ -121,12 +122,146 @@ Not supported
 
 1. In the `NRTR` paper, Beam search is used to decode characters, but the speed is slow. Beam search is not used by default here, and greedy search is used to decode characters.
 
+<a name="6"></a>
+## 6. Release Note
+
+1. The release/2.6 version updates the NRTR code structure. The new version of NRTR can load the model parameters of the old version (release/2.5 and before), and you may use the following code to convert the old version model parameters to the new version model parameters:
+
+```python
+
+    params = paddle.load('path/' + '.pdparams') # the old version parameters
+    state_dict = model.state_dict() # the new version model parameters
+    new_state_dict = {}
+
+    for k1, v1 in state_dict.items():
+
+        k = k1
+        if 'encoder' in k and 'self_attn' in k and 'qkv' in k and 'weight' in k:
+
+            k_para = k[:13] + 'layers.' + k[13:]
+            q = params[k_para.replace('qkv', 'conv1')].transpose((1, 0, 2, 3))
+            k = params[k_para.replace('qkv', 'conv2')].transpose((1, 0, 2, 3))
+            v = params[k_para.replace('qkv', 'conv3')].transpose((1, 0, 2, 3))
+
+            new_state_dict[k1] = np.concatenate([q[:, :, 0, 0], k[:, :, 0, 0], v[:, :, 0, 0]], -1)
+
+        elif 'encoder' in k and 'self_attn' in k and 'qkv' in k and 'bias' in k:
+
+            k_para = k[:13] + 'layers.' + k[13:]
+            q = params[k_para.replace('qkv', 'conv1')]
+            k = params[k_para.replace('qkv', 'conv2')]
+            v = params[k_para.replace('qkv', 'conv3')]
+
+            new_state_dict[k1] = np.concatenate([q, k, v], -1)
+
+        elif 'encoder' in k and 'self_attn' in k and 'out_proj' in k:
+
+            k_para = k[:13] + 'layers.' + k[13:]
+            new_state_dict[k1] = params[k_para]
+
+        elif 'encoder' in k and 'norm3' in k:
+            k_para = k[:13] + 'layers.' + k[13:]
+            new_state_dict[k1] = params[k_para.replace('norm3', 'norm2')]
+
+        elif 'encoder' in k and 'norm1' in k:
+            k_para = k[:13] + 'layers.' + k[13:]
+            new_state_dict[k1] = params[k_para]
+
+
+        elif 'decoder' in k and 'self_attn' in k and 'qkv' in k and 'weight' in k:
+            k_para = k[:13] + 'layers.' + k[13:]
+            q = params[k_para.replace('qkv', 'conv1')].transpose((1, 0, 2, 3))
+            k = params[k_para.replace('qkv', 'conv2')].transpose((1, 0, 2, 3))
+            v = params[k_para.replace('qkv', 'conv3')].transpose((1, 0, 2, 3))
+            new_state_dict[k1] = np.concatenate([q[:, :, 0, 0], k[:, :, 0, 0], v[:, :, 0, 0]], -1)
+
+        elif 'decoder' in k and 'self_attn' in k and 'qkv' in k and 'bias' in k:
+            k_para = k[:13] + 'layers.' + k[13:]
+            q = params[k_para.replace('qkv', 'conv1')]
+            k = params[k_para.replace('qkv', 'conv2')]
+            v = params[k_para.replace('qkv', 'conv3')]
+            new_state_dict[k1] = np.concatenate([q, k, v], -1)
+
+        elif 'decoder' in k and 'self_attn' in k and 'out_proj' in k:
+
+            k_para = k[:13] + 'layers.' + k[13:]
+            new_state_dict[k1] = params[k_para]
+
+        elif 'decoder' in k and 'cross_attn' in k and 'q' in k and 'weight' in k:
+            k_para = k[:13] + 'layers.' + k[13:]
+            k_para = k_para.replace('cross_attn', 'multihead_attn')
+            q = params[k_para.replace('q', 'conv1')].transpose((1, 0, 2, 3))
+            new_state_dict[k1] = q[:, :, 0, 0]
+
+        elif 'decoder' in k and 'cross_attn' in k and 'q' in k and 'bias' in k:
+            k_para = k[:13] + 'layers.' + k[13:]
+            k_para = k_para.replace('cross_attn', 'multihead_attn')
+            q = params[k_para.replace('q', 'conv1')]
+            new_state_dict[k1] = q
+
+        elif 'decoder' in k and 'cross_attn' in k and 'kv' in k and 'weight' in k:
+            k_para = k[:13] + 'layers.' + k[13:]
+            k_para = k_para.replace('cross_attn', 'multihead_attn')
+            k = params[k_para.replace('kv', 'conv2')].transpose((1, 0, 2, 3))
+            v = params[k_para.replace('kv', 'conv3')].transpose((1, 0, 2, 3))
+            new_state_dict[k1] = np.concatenate([k[:, :, 0, 0], v[:, :, 0, 0]], -1)
+
+        elif 'decoder' in k and 'cross_attn' in k and 'kv' in k and 'bias' in k:
+            k_para = k[:13] + 'layers.' + k[13:]
+            k_para = k_para.replace('cross_attn', 'multihead_attn')
+            k = params[k_para.replace('kv', 'conv2')]
+            v = params[k_para.replace('kv', 'conv3')]
+            new_state_dict[k1] = np.concatenate([k, v], -1)
+
+        elif 'decoder' in k and 'cross_attn' in k and 'out_proj' in k:
+
+            k_para = k[:13] + 'layers.' + k[13:]
+            k_para = k_para.replace('cross_attn', 'multihead_attn')
+            new_state_dict[k1] = params[k_para]
+        elif 'decoder' in k and 'norm' in k:
+            k_para = k[:13] + 'layers.' + k[13:]
+            new_state_dict[k1] = params[k_para]
+        elif 'mlp' in k and 'weight' in k:
+            k_para = k[:13] + 'layers.' + k[13:]
+            k_para = k_para.replace('fc', 'conv')
+            k_para = k_para.replace('mlp.', '')
+            w = params[k_para].transpose((1, 0, 2, 3))
+            new_state_dict[k1] = w[:, :, 0, 0]
+        elif 'mlp' in k and 'bias' in k:
+            k_para = k[:13] + 'layers.' + k[13:]
+            k_para = k_para.replace('fc', 'conv')
+            k_para = k_para.replace('mlp.', '')
+            w = params[k_para]
+            new_state_dict[k1] = w
+
+        else:
+            new_state_dict[k1] = params[k1]
+
+        if list(new_state_dict[k1].shape) != list(v1.shape):
+            print(k1)
+
+
+    for k, v1 in state_dict.items():
+        if k not in new_state_dict.keys():
+            print(1, k)
+        elif list(new_state_dict[k].shape) != list(v1.shape):
+            print(2, k)
+
+
+
+    model.set_state_dict(new_state_dict)
+    paddle.save(model.state_dict(), 'nrtrnew_from_old_params.pdparams')
+
+```
+
+2. The new version has a clean code structure and improved inference speed compared with the old version.
+
 ## Citation
 
 ```bibtex
 @article{Sheng2019NRTR,
   title     = {NRTR: A No-Recurrence Sequence-to-Sequence Model For Scene Text Recognition},
-  author    = {Fenfen Sheng and Zhineng Chen andBo Xu},
+  author    = {Fenfen Sheng and Zhineng Chen and Bo Xu},
   booktitle = {ICDAR},
   year      = {2019},
   url       = {http://arxiv.org/abs/1806.00926},

doc/doc_en/algorithm_rec_svtr_en.md

@@ -88,7 +88,6 @@ python3 tools/export_model.py -c configs/rec/rec_svtrnet.yml -o Global.pretraine
 
 **Note:**
 - If you are training the model on your own dataset and have modified the dictionary file, please pay attention to modify the `character_dict_path` in the configuration file to the modified dictionary file.
-- If you modified the input size during training, please modify the `infer_shape` corresponding to SVTR in the `tools/export_model.py` file.
 
 After the conversion is successful, there are three files in the directory:
 ```

doc/doc_en/algorithm_rec_vitstr_en.md

+# ViTSTR
+
+- [1. Introduction](#1)
+- [2. Environment](#2)
+- [3. Model Training / Evaluation / Prediction](#3)
+    - [3.1 Training](#3-1)
+    - [3.2 Evaluation](#3-2)
+    - [3.3 Prediction](#3-3)
+- [4. Inference and Deployment](#4)
+    - [4.1 Python Inference](#4-1)
+    - [4.2 C++ Inference](#4-2)
+    - [4.3 Serving](#4-3)
+    - [4.4 More](#4-4)
+- [5. FAQ](#5)
+
+<a name="1"></a>
+## 1. Introduction
+
+Paper:
+> [Vision Transformer for Fast and Efficient Scene Text Recognition](https://arxiv.org/abs/2105.08582)
+> Rowel Atienza
+> ICDAR, 2021
+
+Using MJSynth and SynthText two text recognition datasets for training, and evaluating on IIIT, SVT, IC03, IC13, IC15, SVTP, CUTE datasets, the algorithm reproduction effect is as follows:
+
+|Model|Backbone|config|Acc|Download link|
+| --- | --- | --- | --- | --- |
+|ViTSTR|ViTSTR|[rec_vitstr_none_ce.yml](../../configs/rec/rec_vitstr_none_ce.yml)|79.82%|[trained model](https://paddleocr.bj.bcebos.com/rec_vitstr_none_none_train.tar)|
+
+<a name="2"></a>
+## 2. Environment
+Please refer to ["Environment Preparation"](./environment_en.md) to configure the PaddleOCR environment, and refer to ["Project Clone"](./clone_en.md) to clone the project code.
+
+
+<a name="3"></a>
+## 3. Model Training / Evaluation / Prediction
+
+Please refer to [Text Recognition Tutorial](./recognition_en.md). PaddleOCR modularizes the code, and training different recognition models only requires **changing the configuration file**.
+
+Training:
+
+Specifically, after the data preparation is completed, the training can be started. The training command is as follows:
+
+```
+#Single GPU training (long training period, not recommended)
+python3 tools/train.py -c configs/rec/rec_vitstr_none_ce.yml
+
+#Multi GPU training, specify the gpu number through the --gpus parameter
+python3 -m paddle.distributed.launch --gpus '0,1,2,3'  tools/train.py -c configs/rec/rec_vitstr_none_ce.yml
+```
+
+Evaluation:
+
+```
+# GPU evaluation
+python3 -m paddle.distributed.launch --gpus '0' tools/eval.py -c configs/rec/rec_vitstr_none_ce.yml -o Global.pretrained_model={path/to/weights}/best_accuracy
+```
+
+Prediction:
+
+```
+# The configuration file used for prediction must match the training
+python3 tools/infer_rec.py -c configs/rec/rec_vitstr_none_ce.yml -o Global.infer_img='./doc/imgs_words_en/word_10.png' Global.pretrained_model=./rec_vitstr_none_ce_train/best_accuracy
+```
+
+<a name="4"></a>
+## 4. Inference and Deployment
+
+<a name="4-1"></a>
+### 4.1 Python Inference
+First, the model saved during the ViTSTR text recognition training process is converted into an inference model. ( [Model download link](https://paddleocr.bj.bcebos.com/rec_vitstr_none_none_train.tar)) ), you can use the following command to convert:
+
+```
+python3 tools/export_model.py -c configs/rec/rec_vitstr_none_ce.yml -o Global.pretrained_model=./rec_vitstr_none_ce_train/best_accuracy  Global.save_inference_dir=./inference/rec_vitstr
+```
+
+**Note:**
+- If you are training the model on your own dataset and have modified the dictionary file, please pay attention to modify the `character_dict_path` in the configuration file to the modified dictionary file.
+- If you modified the input size during training, please modify the `infer_shape` corresponding to ViTSTR in the `tools/export_model.py` file.
+
+After the conversion is successful, there are three files in the directory:
+```
+/inference/rec_vitstr/
+    ├── inference.pdiparams
+    ├── inference.pdiparams.info
+    └── inference.pdmodel
+```
+
+
+For ViTSTR text recognition model inference, the following commands can be executed:
+
+```
+python3 tools/infer/predict_rec.py --image_dir='./doc/imgs_words_en/word_10.png' --rec_model_dir='./inference/rec_vitstr/' --rec_algorithm='ViTSTR' --rec_image_shape='1,224,224' --rec_char_dict_path='./ppocr/utils/EN_symbol_dict.txt'
+```
+
+![](../imgs_words_en/word_10.png)
+
+After executing the command, the prediction result (recognized text and score) of the image above is printed to the screen, an example is as follows:
+The result is as follows:
+```shell
+Predicts of ./doc/imgs_words_en/word_10.png:('pain', 0.9998350143432617)
+```
+
+<a name="4-2"></a>
+### 4.2 C++ Inference
+
+Not supported
+
+<a name="4-3"></a>
+### 4.3 Serving
+
+Not supported
+
+<a name="4-4"></a>
+### 4.4 More
+
+Not supported
+
+<a name="5"></a>
+## 5. FAQ
+
+1. In the `ViTSTR` paper, using pre-trained weights on ImageNet1k for initial training, we did not use pre-trained weights in training, and the final accuracy did not change or even improved.
+
+## Citation
+
+```bibtex
+@article{Atienza2021ViTSTR,
+  title     = {Vision Transformer for Fast and Efficient Scene Text Recognition},
+  author    = {Rowel Atienza},
+  booktitle = {ICDAR},
+  year      = {2021},
+  url       = {https://arxiv.org/abs/2105.08582}
+}
+```

ppocr/data/imaug/__init__.py

@@ -22,8 +22,11 @@ from .make_shrink_map import MakeShrinkMap
 from .random_crop_data import EastRandomCropData, RandomCropImgMask
 from .make_pse_gt import MakePseGt
 
+
+
 from .rec_img_aug import BaseDataAugmentation, RecAug, RecConAug, RecResizeImg, ClsResizeImg, \
-    SRNRecResizeImg, NRTRRecResizeImg, SARRecResizeImg, PRENResizeImg
+                         SRNRecResizeImg, GrayRecResizeImg, SARRecResizeImg, PRENResizeImg, \
+                         ABINetRecResizeImg, SVTRRecResizeImg, ABINetRecAug
 from .ssl_img_aug import SSLRotateResize
 from .randaugment import RandAugment
 from .copy_paste import CopyPaste

ppocr/data/imaug/abinet_aug.py

+# copyright (c) 2020 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.
+"""
+This code is refer from:
+https://github.com/FangShancheng/ABINet/blob/main/transforms.py
+"""
+import math
+import numbers
+import random
+
+import cv2
+import numpy as np
+from paddle.vision.transforms import Compose, ColorJitter
+
+
+def sample_asym(magnitude, size=None):
+    return np.random.beta(1, 4, size) * magnitude
+
+
+def sample_sym(magnitude, size=None):
+    return (np.random.beta(4, 4, size=size) - 0.5) * 2 * magnitude
+
+
+def sample_uniform(low, high, size=None):
+    return np.random.uniform(low, high, size=size)
+
+
+def get_interpolation(type='random'):
+    if type == 'random':
+        choice = [
+            cv2.INTER_NEAREST, cv2.INTER_LINEAR, cv2.INTER_CUBIC, cv2.INTER_AREA
+        ]
+        interpolation = choice[random.randint(0, len(choice) - 1)]
+    elif type == 'nearest':
+        interpolation = cv2.INTER_NEAREST
+    elif type == 'linear':
+        interpolation = cv2.INTER_LINEAR
+    elif type == 'cubic':
+        interpolation = cv2.INTER_CUBIC
+    elif type == 'area':
+        interpolation = cv2.INTER_AREA
+    else:
+        raise TypeError(
+            'Interpolation types only nearest, linear, cubic, area are supported!'
+        )
+    return interpolation
+
+
+class CVRandomRotation(object):
+    def __init__(self, degrees=15):
+        assert isinstance(degrees,
+                          numbers.Number), "degree should be a single number."
+        assert degrees >= 0, "degree must be positive."
+        self.degrees = degrees
+
+    @staticmethod
+    def get_params(degrees):
+        return sample_sym(degrees)
+
+    def __call__(self, img):
+        angle = self.get_params(self.degrees)
+        src_h, src_w = img.shape[:2]
+        M = cv2.getRotationMatrix2D(
+            center=(src_w / 2, src_h / 2), angle=angle, scale=1.0)
+        abs_cos, abs_sin = abs(M[0, 0]), abs(M[0, 1])
+        dst_w = int(src_h * abs_sin + src_w * abs_cos)
+        dst_h = int(src_h * abs_cos + src_w * abs_sin)
+        M[0, 2] += (dst_w - src_w) / 2
+        M[1, 2] += (dst_h - src_h) / 2
+
+        flags = get_interpolation()
+        return cv2.warpAffine(
+            img,
+            M, (dst_w, dst_h),
+            flags=flags,
+            borderMode=cv2.BORDER_REPLICATE)
+
+
+class CVRandomAffine(object):
+    def __init__(self, degrees, translate=None, scale=None, shear=None):
+        assert isinstance(degrees,
+                          numbers.Number), "degree should be a single number."
+        assert degrees >= 0, "degree must be positive."
+        self.degrees = degrees
+
+        if translate is not None:
+            assert isinstance(translate, (tuple, list)) and len(translate) == 2, \
+                "translate should be a list or tuple and it must be of length 2."
+            for t in translate:
+                if not (0.0 <= t <= 1.0):
+                    raise ValueError(
+                        "translation values should be between 0 and 1")
+        self.translate = translate
+
+        if scale is not None:
+            assert isinstance(scale, (tuple, list)) and len(scale) == 2, \
+                "scale should be a list or tuple and it must be of length 2."
+            for s in scale:
+                if s <= 0:
+                    raise ValueError("scale values should be positive")
+        self.scale = scale
+
+        if shear is not None:
+            if isinstance(shear, numbers.Number):
+                if shear < 0:
+                    raise ValueError(
+                        "If shear is a single number, it must be positive.")
+                self.shear = [shear]
+            else:
+                assert isinstance(shear, (tuple, list)) and (len(shear) == 2), \
+                    "shear should be a list or tuple and it must be of length 2."
+                self.shear = shear
+        else:
+            self.shear = shear
+
+    def _get_inverse_affine_matrix(self, center, angle, translate, scale,
+                                   shear):
+        # https://github.com/pytorch/vision/blob/v0.4.0/torchvision/transforms/functional.py#L717
+        from numpy import sin, cos, tan
+
+        if isinstance(shear, numbers.Number):
+            shear = [shear, 0]
+
+        if not isinstance(shear, (tuple, list)) and len(shear) == 2:
+            raise ValueError(
+                "Shear should be a single value or a tuple/list containing " +
+                "two values. Got {}".format(shear))
+
+        rot = math.radians(angle)
+        sx, sy = [math.radians(s) for s in shear]
+
+        cx, cy = center
+        tx, ty = translate
+
+        # RSS without scaling
+        a = cos(rot - sy) / cos(sy)
+        b = -cos(rot - sy) * tan(sx) / cos(sy) - sin(rot)
+        c = sin(rot - sy) / cos(sy)
+        d = -sin(rot - sy) * tan(sx) / cos(sy) + cos(rot)
+
+        # Inverted rotation matrix with scale and shear
+        # det([[a, b], [c, d]]) == 1, since det(rotation) = 1 and det(shear) = 1
+        M = [d, -b, 0, -c, a, 0]
+        M = [x / scale for x in M]
+
+        # Apply inverse of translation and of center translation: RSS^-1 * C^-1 * T^-1
+        M[2] += M[0] * (-cx - tx) + M[1] * (-cy - ty)
+        M[5] += M[3] * (-cx - tx) + M[4] * (-cy - ty)
+
+        # Apply center translation: C * RSS^-1 * C^-1 * T^-1
+        M[2] += cx
+        M[5] += cy
+        return M
+
+    @staticmethod
+    def get_params(degrees, translate, scale_ranges, shears, height):
+        angle = sample_sym(degrees)
+        if translate is not None:
+            max_dx = translate[0] * height
+            max_dy = translate[1] * height
+            translations = (np.round(sample_sym(max_dx)),
+                            np.round(sample_sym(max_dy)))
+        else:
+            translations = (0, 0)
+
+        if scale_ranges is not None:
+            scale = sample_uniform(scale_ranges[0], scale_ranges[1])
+        else:
+            scale = 1.0
+
+        if shears is not None:
+            if len(shears) == 1:
+                shear = [sample_sym(shears[0]), 0.]
+            elif len(shears) == 2:
+                shear = [sample_sym(shears[0]), sample_sym(shears[1])]
+        else:
+            shear = 0.0
+
+        return angle, translations, scale, shear
+
+    def __call__(self, img):
+        src_h, src_w = img.shape[:2]
+        angle, translate, scale, shear = self.get_params(
+            self.degrees, self.translate, self.scale, self.shear, src_h)
+
+        M = self._get_inverse_affine_matrix((src_w / 2, src_h / 2), angle,
+                                            (0, 0), scale, shear)
+        M = np.array(M).reshape(2, 3)
+
+        startpoints = [(0, 0), (src_w - 1, 0), (src_w - 1, src_h - 1),
+                       (0, src_h - 1)]
+        project = lambda x, y, a, b, c: int(a * x + b * y + c)
+        endpoints = [(project(x, y, *M[0]), project(x, y, *M[1]))
+                     for x, y in startpoints]
+
+        rect = cv2.minAreaRect(np.array(endpoints))
+        bbox = cv2.boxPoints(rect).astype(dtype=np.int)
+        max_x, max_y = bbox[:, 0].max(), bbox[:, 1].max()
+        min_x, min_y = bbox[:, 0].min(), bbox[:, 1].min()
+
+        dst_w = int(max_x - min_x)
+        dst_h = int(max_y - min_y)
+        M[0, 2] += (dst_w - src_w) / 2
+        M[1, 2] += (dst_h - src_h) / 2
+
+        # add translate
+        dst_w += int(abs(translate[0]))
+        dst_h += int(abs(translate[1]))
+        if translate[0] < 0: M[0, 2] += abs(translate[0])
+        if translate[1] < 0: M[1, 2] += abs(translate[1])
+
+        flags = get_interpolation()
+        return cv2.warpAffine(
+            img,
+            M, (dst_w, dst_h),
+            flags=flags,
+            borderMode=cv2.BORDER_REPLICATE)
+
+
+class CVRandomPerspective(object):
+    def __init__(self, distortion=0.5):
+        self.distortion = distortion
+
+    def get_params(self, width, height, distortion):
+        offset_h = sample_asym(
+            distortion * height / 2, size=4).astype(dtype=np.int)
+        offset_w = sample_asym(
+            distortion * width / 2, size=4).astype(dtype=np.int)
+        topleft = (offset_w[0], offset_h[0])
+        topright = (width - 1 - offset_w[1], offset_h[1])
+        botright = (width - 1 - offset_w[2], height - 1 - offset_h[2])
+        botleft = (offset_w[3], height - 1 - offset_h[3])
+
+        startpoints = [(0, 0), (width - 1, 0), (width - 1, height - 1),
+                       (0, height - 1)]
+        endpoints = [topleft, topright, botright, botleft]
+        return np.array(
+            startpoints, dtype=np.float32), np.array(
+                endpoints, dtype=np.float32)
+
+    def __call__(self, img):
+        height, width = img.shape[:2]
+        startpoints, endpoints = self.get_params(width, height, self.distortion)
+        M = cv2.getPerspectiveTransform(startpoints, endpoints)
+
+        # TODO: more robust way to crop image
+        rect = cv2.minAreaRect(endpoints)
+        bbox = cv2.boxPoints(rect).astype(dtype=np.int)
+        max_x, max_y = bbox[:, 0].max(), bbox[:, 1].max()
+        min_x, min_y = bbox[:, 0].min(), bbox[:, 1].min()
+        min_x, min_y = max(min_x, 0), max(min_y, 0)
+
+        flags = get_interpolation()
+        img = cv2.warpPerspective(
+            img,
+            M, (max_x, max_y),
+            flags=flags,
+            borderMode=cv2.BORDER_REPLICATE)
+        img = img[min_y:, min_x:]
+        return img
+
+
+class CVRescale(object):
+    def __init__(self, factor=4, base_size=(128, 512)):
+        """ Define image scales using gaussian pyramid and rescale image to target scale.
+        
+        Args:
+            factor: the decayed factor from base size, factor=4 keeps target scale by default.
+            base_size: base size the build the bottom layer of pyramid
+        """
+        if isinstance(factor, numbers.Number):
+            self.factor = round(sample_uniform(0, factor))
+        elif isinstance(factor, (tuple, list)) and len(factor) == 2:
+            self.factor = round(sample_uniform(factor[0], factor[1]))
+        else:
+            raise Exception('factor must be number or list with length 2')
+        # assert factor is valid
+        self.base_h, self.base_w = base_size[:2]
+
+    def __call__(self, img):
+        if self.factor == 0: return img
+        src_h, src_w = img.shape[:2]
+        cur_w, cur_h = self.base_w, self.base_h
+        scale_img = cv2.resize(
+            img, (cur_w, cur_h), interpolation=get_interpolation())
+        for _ in range(self.factor):
+            scale_img = cv2.pyrDown(scale_img)
+        scale_img = cv2.resize(
+            scale_img, (src_w, src_h), interpolation=get_interpolation())
+        return scale_img
+
+
+class CVGaussianNoise(object):
+    def __init__(self, mean=0, var=20):
+        self.mean = mean
+        if isinstance(var, numbers.Number):
+            self.var = max(int(sample_asym(var)), 1)
+        elif isinstance(var, (tuple, list)) and len(var) == 2:
+            self.var = int(sample_uniform(var[0], var[1]))
+        else:
+            raise Exception('degree must be number or list with length 2')
+
+    def __call__(self, img):
+        noise = np.random.normal(self.mean, self.var**0.5, img.shape)
+        img = np.clip(img + noise, 0, 255).astype(np.uint8)
+        return img
+
+
+class CVMotionBlur(object):
+    def __init__(self, degrees=12, angle=90):
+        if isinstance(degrees, numbers.Number):
+            self.degree = max(int(sample_asym(degrees)), 1)
+        elif isinstance(degrees, (tuple, list)) and len(degrees) == 2:
+            self.degree = int(sample_uniform(degrees[0], degrees[1]))
+        else:
+            raise Exception('degree must be number or list with length 2')
+        self.angle = sample_uniform(-angle, angle)
+
+    def __call__(self, img):
+        M = cv2.getRotationMatrix2D((self.degree // 2, self.degree // 2),
+                                    self.angle, 1)
+        motion_blur_kernel = np.zeros((self.degree, self.degree))
+        motion_blur_kernel[self.degree // 2, :] = 1
+        motion_blur_kernel = cv2.warpAffine(motion_blur_kernel, M,
+                                            (self.degree, self.degree))
+        motion_blur_kernel = motion_blur_kernel / self.degree
+        img = cv2.filter2D(img, -1, motion_blur_kernel)
+        img = np.clip(img, 0, 255).astype(np.uint8)
+        return img
+
+
+class CVGeometry(object):
+    def __init__(self,
+                 degrees=15,
+                 translate=(0.3, 0.3),
+                 scale=(0.5, 2.),
+                 shear=(45, 15),
+                 distortion=0.5,
+                 p=0.5):
+        self.p = p
+        type_p = random.random()
+        if type_p < 0.33:
+            self.transforms = CVRandomRotation(degrees=degrees)
+        elif type_p < 0.66:
+            self.transforms = CVRandomAffine(
+                degrees=degrees, translate=translate, scale=scale, shear=shear)
+        else:
+            self.transforms = CVRandomPerspective(distortion=distortion)
+
+    def __call__(self, img):
+        if random.random() < self.p:
+            return self.transforms(img)
+        else:
+            return img
+
+
+class CVDeterioration(object):
+    def __init__(self, var, degrees, factor, p=0.5):
+        self.p = p
+        transforms = []
+        if var is not None:
+            transforms.append(CVGaussianNoise(var=var))
+        if degrees is not None:
+            transforms.append(CVMotionBlur(degrees=degrees))
+        if factor is not None:
+            transforms.append(CVRescale(factor=factor))
+
+        random.shuffle(transforms)
+        transforms = Compose(transforms)
+        self.transforms = transforms
+
+    def __call__(self, img):
+        if random.random() < self.p:
+
+            return self.transforms(img)
+        else:
+            return img
+
+
+class CVColorJitter(object):
+    def __init__(self,
+                 brightness=0.5,
+                 contrast=0.5,
+                 saturation=0.5,
+                 hue=0.1,
+                 p=0.5):
+        self.p = p
+        self.transforms = ColorJitter(
+            brightness=brightness,
+            contrast=contrast,
+            saturation=saturation,
+            hue=hue)
+
+    def __call__(self, img):
+        if random.random() < self.p: return self.transforms(img)
+        else: return img

ppocr/data/imaug/label_ops.py

@@ -157,37 +157,6 @@ class BaseRecLabelEncode(object):
         return text_list
 
 
-class NRTRLabelEncode(BaseRecLabelEncode):
-    """ Convert between text-label and text-index """
-
-    def __init__(self,
-                 max_text_length,
-                 character_dict_path=None,
-                 use_space_char=False,
-                 **kwargs):
-
-        super(NRTRLabelEncode, self).__init__(
-            max_text_length, character_dict_path, use_space_char)
-
-    def __call__(self, data):
-        text = data['label']
-        text = self.encode(text)
-        if text is None:
-            return None
-        if len(text) >= self.max_text_len - 1:
-            return None
-        data['length'] = np.array(len(text))
-        text.insert(0, 2)
-        text.append(3)
-        text = text + [0] * (self.max_text_len - len(text))
-        data['label'] = np.array(text)
-        return data
-
-    def add_special_char(self, dict_character):
-        dict_character = ['blank', '<unk>', '<s>', '</s>'] + dict_character
-        return dict_character
-
-
 class CTCLabelEncode(BaseRecLabelEncode):
     """ Convert between text-label and text-index """
 
@@ -290,15 +259,26 @@ class E2ELabelEncodeTrain(object):
 
 
 class KieLabelEncode(object):
-    def __init__(self, character_dict_path, norm=10, directed=False, **kwargs):
+    def __init__(self,
+                 character_dict_path,
+                 class_path,
+                 norm=10,
+                 directed=False,
+                 **kwargs):
         super(KieLabelEncode, self).__init__()
         self.dict = dict({'': 0})
+        self.label2classid_map = dict()
         with open(character_dict_path, 'r', encoding='utf-8') as fr:
             idx = 1
             for line in fr:
                 char = line.strip()
                 self.dict[char] = idx
                 idx += 1
+        with open(class_path, "r") as fin:
+            lines = fin.readlines()
+            for idx, line in enumerate(lines):
+                line = line.strip("\n")
+                self.label2classid_map[line] = idx
         self.norm = norm
         self.directed = directed
 
@@ -439,7 +419,7 @@ class KieLabelEncode(object):
             text_ind = [self.dict[c] for c in text if c in self.dict]
             text_inds.append(text_ind)
             if 'label' in ann.keys():
-                labels.append(ann['label'])
+                labels.append(self.label2classid_map[ann['label']])
             elif 'key_cls' in ann.keys():
                 labels.append(ann['key_cls'])
             else:
@@ -907,15 +887,16 @@ class VQATokenLabelEncode(object):
         for info in ocr_info:
             if train_re:
                 # for re
-                if len(info["text"]) == 0:
+                if len(info["transcription"]) == 0:
                     empty_entity.add(info["id"])
                     continue
                 id2label[info["id"]] = info["label"]
                 relations.extend([tuple(sorted(l)) for l in info["linking"]])
             # smooth_box
+            info["bbox"] = self.trans_poly_to_bbox(info["points"])
             bbox = self._smooth_box(info["bbox"], height, width)
 
-            text = info["text"]
+            text = info["transcription"]
             encode_res = self.tokenizer.encode(
                 text, pad_to_max_seq_len=False, return_attention_mask=True)
 
@@ -931,7 +912,7 @@ class VQATokenLabelEncode(object):
                 label = info['label']
                 gt_label = self._parse_label(label, encode_res)
 
-            # construct entities for re
+# construct entities for re
             if train_re:
                 if gt_label[0] != self.label2id_map["O"]:
                     entity_id_to_index_map[info["id"]] = len(entities)
@@ -975,29 +956,29 @@ class VQATokenLabelEncode(object):
             data['entity_id_to_index_map'] = entity_id_to_index_map
         return data
 
-    def _load_ocr_info(self, data):
-        def trans_poly_to_bbox(poly):
-            x1 = np.min([p[0] for p in poly])
-            x2 = np.max([p[0] for p in poly])
-            y1 = np.min([p[1] for p in poly])
-            y2 = np.max([p[1] for p in poly])
-            return [x1, y1, x2, y2]
+    def trans_poly_to_bbox(self, poly):
+        x1 = np.min([p[0] for p in poly])
+        x2 = np.max([p[0] for p in poly])
+        y1 = np.min([p[1] for p in poly])
+        y2 = np.max([p[1] for p in poly])
+        return [x1, y1, x2, y2]
 
+    def _load_ocr_info(self, data):
         if self.infer_mode:
             ocr_result = self.ocr_engine.ocr(data['image'], cls=False)
             ocr_info = []
             for res in ocr_result:
                 ocr_info.append({
-                    "text": res[1][0],
-                    "bbox": trans_poly_to_bbox(res[0]),
-                    "poly": res[0],
+                    "transcription": res[1][0],
+                    "bbox": self.trans_poly_to_bbox(res[0]),
+                    "points": res[0],
                 })
             return ocr_info
         else:
             info = data['label']
             # read text info
             info_dict = json.loads(info)
-            return info_dict["ocr_info"]
+            return info_dict
 
     def _smooth_box(self, bbox, height, width):
         bbox[0] = int(bbox[0] * 1000.0 / width)
@@ -1008,7 +989,7 @@ class VQATokenLabelEncode(object):
 
     def _parse_label(self, label, encode_res):
         gt_label = []
-        if label.lower() == "other":
+        if label.lower() in ["other", "others", "ignore"]:
             gt_label.extend([0] * len(encode_res["input_ids"]))
         else:
             gt_label.append(self.label2id_map[("b-" + label).upper()])
@@ -1046,3 +1027,99 @@ class MultiLabelEncode(BaseRecLabelEncode):
         data_out['label_sar'] = sar['label']
         data_out['length'] = ctc['length']
         return data_out
+
+
+class NRTRLabelEncode(BaseRecLabelEncode):
+    """ Convert between text-label and text-index """
+
+    def __init__(self,
+                 max_text_length,
+                 character_dict_path=None,
+                 use_space_char=False,
+                 **kwargs):
+
+        super(NRTRLabelEncode, self).__init__(
+            max_text_length, character_dict_path, use_space_char)
+
+    def __call__(self, data):
+        text = data['label']
+        text = self.encode(text)
+        if text is None:
+            return None
+        if len(text) >= self.max_text_len - 1:
+            return None
+        data['length'] = np.array(len(text))
+        text.insert(0, 2)
+        text.append(3)
+        text = text + [0] * (self.max_text_len - len(text))
+        data['label'] = np.array(text)
+        return data
+
+    def add_special_char(self, dict_character):
+        dict_character = ['blank', '<unk>', '<s>', '</s>'] + dict_character
+        return dict_character
+
+
+class ViTSTRLabelEncode(BaseRecLabelEncode):
+    """ Convert between text-label and text-index """
+
+    def __init__(self,
+                 max_text_length,
+                 character_dict_path=None,
+                 use_space_char=False,
+                 ignore_index=0,
+                 **kwargs):
+
+        super(ViTSTRLabelEncode, self).__init__(
+            max_text_length, character_dict_path, use_space_char)
+        self.ignore_index = ignore_index
+
+    def __call__(self, data):
+        text = data['label']
+        text = self.encode(text)
+        if text is None:
+            return None
+        if len(text) >= self.max_text_len:
+            return None
+        data['length'] = np.array(len(text))
+        text.insert(0, self.ignore_index)
+        text.append(1)
+        text = text + [self.ignore_index] * (self.max_text_len + 2 - len(text))
+        data['label'] = np.array(text)
+        return data
+
+    def add_special_char(self, dict_character):
+        dict_character = ['<s>', '</s>'] + dict_character
+        return dict_character
+
+
+class ABINetLabelEncode(BaseRecLabelEncode):
+    """ Convert between text-label and text-index """
+
+    def __init__(self,
+                 max_text_length,
+                 character_dict_path=None,
+                 use_space_char=False,
+                 ignore_index=100,
+                 **kwargs):
+
+        super(ABINetLabelEncode, self).__init__(
+            max_text_length, character_dict_path, use_space_char)
+        self.ignore_index = ignore_index
+
+    def __call__(self, data):
+        text = data['label']
+        text = self.encode(text)
+        if text is None:
+            return None
+        if len(text) >= self.max_text_len:
+            return None
+        data['length'] = np.array(len(text))
+        text.append(0)
+        text = text + [self.ignore_index] * (self.max_text_len + 1 - len(text))
+        data['label'] = np.array(text)
+        return data
+
+    def add_special_char(self, dict_character):
+        dict_character = ['</s>'] + dict_character
+        return dict_character

ppocr/data/imaug/operators.py

@@ -67,39 +67,6 @@ class DecodeImage(object):
         return data
 
 
-class NRTRDecodeImage(object):
-    """ decode image """
-
-    def __init__(self, img_mode='RGB', channel_first=False, **kwargs):
-        self.img_mode = img_mode
-        self.channel_first = channel_first
-
-    def __call__(self, data):
-        img = data['image']
-        if six.PY2:
-            assert type(img) is str and len(
-                img) > 0, "invalid input 'img' in DecodeImage"
-        else:
-            assert type(img) is bytes and len(
-                img) > 0, "invalid input 'img' in DecodeImage"
-        img = np.frombuffer(img, dtype='uint8')
-
-        img = cv2.imdecode(img, 1)
-
-        if img is None:
-            return None
-        if self.img_mode == 'GRAY':
-            img = cv2.cvtColor(img, cv2.COLOR_GRAY2BGR)
-        elif self.img_mode == 'RGB':
-            assert img.shape[2] == 3, 'invalid shape of image[%s]' % (img.shape)
-            img = img[:, :, ::-1]
-        img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
-        if self.channel_first:
-            img = img.transpose((2, 0, 1))
-        data['image'] = img
-        return data
-
-
 class NormalizeImage(object):
     """ normalize image such as substract mean, divide std
     """

ppocr/data/imaug/rec_img_aug.py

@@ -19,6 +19,8 @@ import random
 import copy
 from PIL import Image
 from .text_image_aug import tia_perspective, tia_stretch, tia_distort
+from .abinet_aug import CVGeometry, CVDeterioration, CVColorJitter
+from paddle.vision.transforms import Compose
 
 
 class RecAug(object):
@@ -94,6 +96,36 @@ class BaseDataAugmentation(object):
         return data
 
 
+class ABINetRecAug(object):
+    def __init__(self,
+                 geometry_p=0.5,
+                 deterioration_p=0.25,
+                 colorjitter_p=0.25,
+                 **kwargs):
+        self.transforms = Compose([
+            CVGeometry(
+                degrees=45,
+                translate=(0.0, 0.0),
+                scale=(0.5, 2.),
+                shear=(45, 15),
+                distortion=0.5,
+                p=geometry_p), CVDeterioration(
+                    var=20, degrees=6, factor=4, p=deterioration_p),
+            CVColorJitter(
+                brightness=0.5,
+                contrast=0.5,
+                saturation=0.5,
+                hue=0.1,
+                p=colorjitter_p)
+        ])
+
+    def __call__(self, data):
+        img = data['image']
+        img = self.transforms(img)
+        data['image'] = img
+        return data
+
+
 class RecConAug(object):
     def __init__(self,
                  prob=0.5,
@@ -148,46 +180,6 @@ class ClsResizeImg(object):
         return data
 
 
-class NRTRRecResizeImg(object):
-    def __init__(self, image_shape, resize_type, padding=False, **kwargs):
-        self.image_shape = image_shape
-        self.resize_type = resize_type
-        self.padding = padding
-
-    def __call__(self, data):
-        img = data['image']
-        img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
-        image_shape = self.image_shape
-        if self.padding:
-            imgC, imgH, imgW = image_shape
-            # todo: change to 0 and modified image shape
-            h = img.shape[0]
-            w = img.shape[1]
-            ratio = w / float(h)
-            if math.ceil(imgH * ratio) > imgW:
-                resized_w = imgW
-            else:
-                resized_w = int(math.ceil(imgH * ratio))
-            resized_image = cv2.resize(img, (resized_w, imgH))
-            norm_img = np.expand_dims(resized_image, -1)
-            norm_img = norm_img.transpose((2, 0, 1))
-            resized_image = norm_img.astype(np.float32) / 128. - 1.
-            padding_im = np.zeros((imgC, imgH, imgW), dtype=np.float32)
-            padding_im[:, :, 0:resized_w] = resized_image
-            data['image'] = padding_im
-            return data
-        if self.resize_type == 'PIL':
-            image_pil = Image.fromarray(np.uint8(img))
-            img = image_pil.resize(self.image_shape, Image.ANTIALIAS)
-            img = np.array(img)
-        if self.resize_type == 'OpenCV':
-            img = cv2.resize(img, self.image_shape)
-        norm_img = np.expand_dims(img, -1)
-        norm_img = norm_img.transpose((2, 0, 1))
-        data['image'] = norm_img.astype(np.float32) / 128. - 1.
-        return data
-
-
 class RecResizeImg(object):
     def __init__(self,
                  image_shape,
@@ -268,6 +260,84 @@ class PRENResizeImg(object):
         return data
 
 
+class GrayRecResizeImg(object):
+    def __init__(self,
+                 image_shape,
+                 resize_type,
+                 inter_type='Image.ANTIALIAS',
+                 scale=True,
+                 padding=False,
+                 **kwargs):
+        self.image_shape = image_shape
+        self.resize_type = resize_type
+        self.padding = padding
+        self.inter_type = eval(inter_type)
+        self.scale = scale
+
+    def __call__(self, data):
+        img = data['image']
+        img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
+        image_shape = self.image_shape
+        if self.padding:
+            imgC, imgH, imgW = image_shape
+            # todo: change to 0 and modified image shape
+            h = img.shape[0]
+            w = img.shape[1]
+            ratio = w / float(h)
+            if math.ceil(imgH * ratio) > imgW:
+                resized_w = imgW
+            else:
+                resized_w = int(math.ceil(imgH * ratio))
+            resized_image = cv2.resize(img, (resized_w, imgH))
+            norm_img = np.expand_dims(resized_image, -1)
+            norm_img = norm_img.transpose((2, 0, 1))
+            resized_image = norm_img.astype(np.float32) / 128. - 1.
+            padding_im = np.zeros((imgC, imgH, imgW), dtype=np.float32)
+            padding_im[:, :, 0:resized_w] = resized_image
+            data['image'] = padding_im
+            return data
+        if self.resize_type == 'PIL':
+            image_pil = Image.fromarray(np.uint8(img))
+            img = image_pil.resize(self.image_shape, self.inter_type)
+            img = np.array(img)
+        if self.resize_type == 'OpenCV':
+            img = cv2.resize(img, self.image_shape)
+        norm_img = np.expand_dims(img, -1)
+        norm_img = norm_img.transpose((2, 0, 1))
+        if self.scale:
+            data['image'] = norm_img.astype(np.float32) / 128. - 1.
+        else:
+            data['image'] = norm_img.astype(np.float32) / 255.
+        return data
+
+
+class ABINetRecResizeImg(object):
+    def __init__(self, image_shape, **kwargs):
+        self.image_shape = image_shape
+
+    def __call__(self, data):
+        img = data['image']
+        norm_img, valid_ratio = resize_norm_img_abinet(img, self.image_shape)
+        data['image'] = norm_img
+        data['valid_ratio'] = valid_ratio
+        return data
+
+
+class SVTRRecResizeImg(object):
+    def __init__(self, image_shape, padding=True, **kwargs):
+        self.image_shape = image_shape
+        self.padding = padding
+
+    def __call__(self, data):
+        img = data['image']
+
+        norm_img, valid_ratio = resize_norm_img(img, self.image_shape,
+                                                self.padding)
+        data['image'] = norm_img
+        data['valid_ratio'] = valid_ratio
+        return data
+
+
 def resize_norm_img_sar(img, image_shape, width_downsample_ratio=0.25):
     imgC, imgH, imgW_min, imgW_max = image_shape
     h = img.shape[0]
@@ -386,6 +456,26 @@ def resize_norm_img_srn(img, image_shape):
     return np.reshape(img_black, (c, row, col)).astype(np.float32)
 
 
+def resize_norm_img_abinet(img, image_shape):
+    imgC, imgH, imgW = image_shape
+
+    resized_image = cv2.resize(
+        img, (imgW, imgH), interpolation=cv2.INTER_LINEAR)
+    resized_w = imgW
+    resized_image = resized_image.astype('float32')
+    resized_image = resized_image / 255.
+
+    mean = np.array([0.485, 0.456, 0.406])
+    std = np.array([0.229, 0.224, 0.225])
+    resized_image = (
+        resized_image - mean[None, None, ...]) / std[None, None, ...]
+    resized_image = resized_image.transpose((2, 0, 1))
+    resized_image = resized_image.astype('float32')
+
+    valid_ratio = min(1.0, float(resized_w / imgW))
+    return resized_image, valid_ratio
+
+
 def srn_other_inputs(image_shape, num_heads, max_text_length):
 
     imgC, imgH, imgW = image_shape

ppocr/data/imaug/vqa/__init__.py

@@ -13,7 +13,12 @@
 # limitations under the License.
 
 from .token import VQATokenPad, VQASerTokenChunk, VQAReTokenChunk, VQAReTokenRelation
+from .augment import DistortBBox
 
 __all__ = [
-    'VQATokenPad', 'VQASerTokenChunk', 'VQAReTokenChunk', 'VQAReTokenRelation'
+    'VQATokenPad',
+    'VQASerTokenChunk',
+    'VQAReTokenChunk',
+    'VQAReTokenRelation',
+    'DistortBBox',
 ]

ppocr/data/imaug/vqa/augment.py

+# copyright (c) 2022 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 os
+import sys
+import numpy as np
+import random
+
+
+class DistortBBox:
+    def __init__(self, prob=0.5, max_scale=1, **kwargs):
+        """Random distort bbox
+        """
+        self.prob = prob
+        self.max_scale = max_scale
+
+    def __call__(self, data):
+        if random.random() > self.prob:
+            return data
+        bbox = np.array(data['bbox'])
+        rnd_scale = (np.random.rand(*bbox.shape) - 0.5) * 2 * self.max_scale
+        bbox = np.round(bbox + rnd_scale).astype(bbox.dtype)
+        data['bbox'] = np.clip(data['bbox'], 0, 1000)
+        data['bbox'] = bbox.tolist()
+        sys.stdout.flush()
+        return data

ppocr/losses/__init__.py

@@ -30,7 +30,7 @@ from .det_fce_loss import FCELoss
 from .rec_ctc_loss import CTCLoss
 from .rec_att_loss import AttentionLoss
 from .rec_srn_loss import SRNLoss
-from .rec_nrtr_loss import NRTRLoss
+from .rec_ce_loss import CELoss
 from .rec_sar_loss import SARLoss
 from .rec_aster_loss import AsterLoss
 from .rec_pren_loss import PRENLoss
@@ -60,7 +60,7 @@ def build_loss(config):
     support_dict = [
         'DBLoss', 'PSELoss', 'EASTLoss', 'SASTLoss', 'FCELoss', 'CTCLoss',
         'ClsLoss', 'AttentionLoss', 'SRNLoss', 'PGLoss', 'CombinedLoss',
-        'NRTRLoss', 'TableAttentionLoss', 'SARLoss', 'AsterLoss', 'SDMGRLoss',
+        'CELoss', 'TableAttentionLoss', 'SARLoss', 'AsterLoss', 'SDMGRLoss',
         'VQASerTokenLayoutLMLoss', 'LossFromOutput', 'PRENLoss', 'MultiLoss'
     ]
     config = copy.deepcopy(config)

ppocr/losses/rec_ce_loss.py

+import paddle
+from paddle import nn
+import paddle.nn.functional as F
+
+
+class CELoss(nn.Layer):
+    def __init__(self,
+                 smoothing=False,
+                 with_all=False,
+                 ignore_index=-1,
+                 **kwargs):
+        super(CELoss, self).__init__()
+        if ignore_index >= 0:
+            self.loss_func = nn.CrossEntropyLoss(
+                reduction='mean', ignore_index=ignore_index)
+        else:
+            self.loss_func = nn.CrossEntropyLoss(reduction='mean')
+        self.smoothing = smoothing
+        self.with_all = with_all
+
+    def forward(self, pred, batch):
+
+        if isinstance(pred, dict):  # for ABINet
+            loss = {}
+            loss_sum = []
+            for name, logits in pred.items():
+                if isinstance(logits, list):
+                    logit_num = len(logits)
+                    all_tgt = paddle.concat([batch[1]] * logit_num, 0)
+                    all_logits = paddle.concat(logits, 0)
+                    flt_logtis = all_logits.reshape([-1, all_logits.shape[2]])
+                    flt_tgt = all_tgt.reshape([-1])
+                else:
+                    flt_logtis = logits.reshape([-1, logits.shape[2]])
+                    flt_tgt = batch[1].reshape([-1])
+                loss[name + '_loss'] = self.loss_func(flt_logtis, flt_tgt)
+                loss_sum.append(loss[name + '_loss'])
+            loss['loss'] = sum(loss_sum)
+            return loss
+        else:
+            if self.with_all:  # for ViTSTR
+                tgt = batch[1]
+                pred = pred.reshape([-1, pred.shape[2]])
+                tgt = tgt.reshape([-1])
+                loss = self.loss_func(pred, tgt)
+                return {'loss': loss}
+            else:  # for NRTR
+                max_len = batch[2].max()
+                tgt = batch[1][:, 1:2 + max_len]
+                pred = pred.reshape([-1, pred.shape[2]])
+                tgt = tgt.reshape([-1])
+                if self.smoothing:
+                    eps = 0.1
+                    n_class = pred.shape[1]
+                    one_hot = F.one_hot(tgt, pred.shape[1])
+                    one_hot = one_hot * (1 - eps) + (1 - one_hot) * eps / (
+                        n_class - 1)
+                    log_prb = F.log_softmax(pred, axis=1)
+                    non_pad_mask = paddle.not_equal(
+                        tgt, paddle.zeros(
+                            tgt.shape, dtype=tgt.dtype))
+                    loss = -(one_hot * log_prb).sum(axis=1)
+                    loss = loss.masked_select(non_pad_mask).mean()
+                else:
+                    loss = self.loss_func(pred, tgt)
+                return {'loss': loss}

ppocr/losses/rec_nrtr_loss.py

-import paddle
-from paddle import nn
-import paddle.nn.functional as F
-
-
-class NRTRLoss(nn.Layer):
-    def __init__(self, smoothing=True, **kwargs):
-        super(NRTRLoss, self).__init__()
-        self.loss_func = nn.CrossEntropyLoss(reduction='mean', ignore_index=0)
-        self.smoothing = smoothing
-
-    def forward(self, pred, batch):
-        pred = pred.reshape([-1, pred.shape[2]])
-        max_len = batch[2].max()
-        tgt = batch[1][:, 1:2 + max_len]
-        tgt = tgt.reshape([-1])
-        if self.smoothing:
-            eps = 0.1
-            n_class = pred.shape[1]
-            one_hot = F.one_hot(tgt, pred.shape[1])
-            one_hot = one_hot * (1 - eps) + (1 - one_hot) * eps / (n_class - 1)
-            log_prb = F.log_softmax(pred, axis=1)
-            non_pad_mask = paddle.not_equal(
-                tgt, paddle.zeros(
-                    tgt.shape, dtype=tgt.dtype))
-            loss = -(one_hot * log_prb).sum(axis=1)
-            loss = loss.masked_select(non_pad_mask).mean()
-        else:
-            loss = self.loss_func(pred, tgt)
-        return {'loss': loss}

ppocr/modeling/backbones/__init__.py

@@ -29,35 +29,37 @@ def build_backbone(config, model_type):
         from .rec_mv1_enhance import MobileNetV1Enhance
         from .rec_nrtr_mtb import MTB
         from .rec_resnet_31 import ResNet31
+        from .rec_resnet_45 import ResNet45
         from .rec_resnet_aster import ResNet_ASTER
         from .rec_micronet import MicroNet
         from .rec_efficientb3_pren import EfficientNetb3_PREN
         from .rec_svtrnet import SVTRNet
+        from .rec_vitstr import ViTSTR
         support_dict = [
             'MobileNetV1Enhance', 'MobileNetV3', 'ResNet', 'ResNetFPN', 'MTB',
-            "ResNet31", "ResNet_ASTER", 'MicroNet', 'EfficientNetb3_PREN',
-            'SVTRNet'
+            'ResNet31', 'ResNet45', 'ResNet_ASTER', 'MicroNet',
+            'EfficientNetb3_PREN', 'SVTRNet', 'ViTSTR'
         ]
-    elif model_type == "e2e":
+    elif model_type == 'e2e':
         from .e2e_resnet_vd_pg import ResNet
         support_dict = ['ResNet']
     elif model_type == 'kie':
         from .kie_unet_sdmgr import Kie_backbone
         support_dict = ['Kie_backbone']
-    elif model_type == "table":
+    elif model_type == 'table':
         from .table_resnet_vd import ResNet
         from .table_mobilenet_v3 import MobileNetV3
-        support_dict = ["ResNet", "MobileNetV3"]
+        support_dict = ['ResNet', 'MobileNetV3']
     elif model_type == 'vqa':
         from .vqa_layoutlm import LayoutLMForSer, LayoutLMv2ForSer, LayoutLMv2ForRe, LayoutXLMForSer, LayoutXLMForRe
         support_dict = [
-            "LayoutLMForSer", "LayoutLMv2ForSer", 'LayoutLMv2ForRe',
-            "LayoutXLMForSer", 'LayoutXLMForRe'
+            'LayoutLMForSer', 'LayoutLMv2ForSer', 'LayoutLMv2ForRe',
+            'LayoutXLMForSer', 'LayoutXLMForRe'
         ]
     else:
         raise NotImplementedError
 
-    module_name = config.pop("name")
+    module_name = config.pop('name')
     assert module_name in support_dict, Exception(
         "when model typs is {}, backbone only support {}".format(model_type,
                                                                  support_dict))

ppocr/modeling/backbones/rec_resnet_45.py

+# 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.
+"""
+This code is refer from: 
+https://github.com/FangShancheng/ABINet/tree/main/modules
+"""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import paddle
+from paddle import ParamAttr
+from paddle.nn.initializer import KaimingNormal
+import paddle.nn as nn
+import paddle.nn.functional as F
+import numpy as np
+import math
+
+__all__ = ["ResNet45"]
+
+
+def conv1x1(in_planes, out_planes, stride=1):
+    return nn.Conv2D(
+        in_planes,
+        out_planes,
+        kernel_size=1,
+        stride=1,
+        weight_attr=ParamAttr(initializer=KaimingNormal()),
+        bias_attr=False)
+
+
+def conv3x3(in_channel, out_channel, stride=1):
+    return nn.Conv2D(
+        in_channel,
+        out_channel,
+        kernel_size=3,
+        stride=stride,
+        padding=1,
+        weight_attr=ParamAttr(initializer=KaimingNormal()),
+        bias_attr=False)
+
+
+class BasicBlock(nn.Layer):
+    expansion = 1
+
+    def __init__(self, in_channels, channels, stride=1, downsample=None):
+        super().__init__()
+        self.conv1 = conv1x1(in_channels, channels)
+        self.bn1 = nn.BatchNorm2D(channels)
+        self.relu = nn.ReLU()
+        self.conv2 = conv3x3(channels, channels, stride)
+        self.bn2 = nn.BatchNorm2D(channels)
+        self.downsample = downsample
+        self.stride = stride
+
+    def forward(self, x):
+        residual = x
+
+        out = self.conv1(x)
+        out = self.bn1(out)
+        out = self.relu(out)
+
+        out = self.conv2(out)
+        out = self.bn2(out)
+
+        if self.downsample is not None:
+            residual = self.downsample(x)
+        out += residual
+        out = self.relu(out)
+
+        return out
+
+
+class ResNet45(nn.Layer):
+    def __init__(self, block=BasicBlock, layers=[3, 4, 6, 6, 3], in_channels=3):
+        self.inplanes = 32
+        super(ResNet45, self).__init__()
+        self.conv1 = nn.Conv2D(
+            3,
+            32,
+            kernel_size=3,
+            stride=1,
+            padding=1,
+            weight_attr=ParamAttr(initializer=KaimingNormal()),
+            bias_attr=False)
+        self.bn1 = nn.BatchNorm2D(32)
+        self.relu = nn.ReLU()
+
+        self.layer1 = self._make_layer(block, 32, layers[0], stride=2)
+        self.layer2 = self._make_layer(block, 64, layers[1], stride=1)
+        self.layer3 = self._make_layer(block, 128, layers[2], stride=2)
+        self.layer4 = self._make_layer(block, 256, layers[3], stride=1)
+        self.layer5 = self._make_layer(block, 512, layers[4], stride=1)
+        self.out_channels = 512
+
+        # for m in self.modules():
+        #     if isinstance(m, nn.Conv2D):
+        #         n = m._kernel_size[0] * m._kernel_size[1] * m._out_channels
+        #         m.weight.data.normal_(0, math.sqrt(2. / n))
+
+    def _make_layer(self, block, planes, blocks, stride=1):
+        downsample = None
+        if stride != 1 or self.inplanes != planes * block.expansion:
+            # downsample = True
+            downsample = nn.Sequential(
+                nn.Conv2D(
+                    self.inplanes,
+                    planes * block.expansion,
+                    kernel_size=1,
+                    stride=stride,
+                    weight_attr=ParamAttr(initializer=KaimingNormal()),
+                    bias_attr=False),
+                nn.BatchNorm2D(planes * block.expansion), )
+
+        layers = []
+        layers.append(block(self.inplanes, planes, stride, downsample))
+        self.inplanes = planes * block.expansion
+        for i in range(1, blocks):
+            layers.append(block(self.inplanes, planes))
+
+        return nn.Sequential(*layers)
+
+    def forward(self, x):
+
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = self.relu(x)
+        # print(x)
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+        # print(x)
+        x = self.layer4(x)
+        x = self.layer5(x)
+        return x

ppocr/modeling/backbones/rec_svtrnet.py

@@ -147,7 +147,7 @@ class Attention(nn.Layer):
                  dim,
                  num_heads=8,
                  mixer='Global',
-                 HW=[8, 25],
+                 HW=None,
                  local_k=[7, 11],
                  qkv_bias=False,
                  qk_scale=None,
@@ -210,7 +210,7 @@ class Block(nn.Layer):
                  num_heads,
                  mixer='Global',
                  local_mixer=[7, 11],
-                 HW=[8, 25],
+                 HW=None,
                  mlp_ratio=4.,
                  qkv_bias=False,
                  qk_scale=None,
@@ -274,7 +274,9 @@ class PatchEmbed(nn.Layer):
                  img_size=[32, 100],
                  in_channels=3,
                  embed_dim=768,
-                 sub_num=2):
+                 sub_num=2,
+                 patch_size=[4, 4],
+                 mode='pope'):
         super().__init__()
         num_patches = (img_size[1] // (2 ** sub_num)) * \
                       (img_size[0] // (2 ** sub_num))
@@ -282,50 +284,56 @@ class PatchEmbed(nn.Layer):
         self.num_patches = num_patches
         self.embed_dim = embed_dim
         self.norm = None
-        if sub_num == 2:
-            self.proj = nn.Sequential(
-                ConvBNLayer(
-                    in_channels=in_channels,
-                    out_channels=embed_dim // 2,
-                    kernel_size=3,
-                    stride=2,
-                    padding=1,
-                    act=nn.GELU,
-                    bias_attr=None),
-                ConvBNLayer(
-                    in_channels=embed_dim // 2,
-                    out_channels=embed_dim,
-                    kernel_size=3,
-                    stride=2,
-                    padding=1,
-                    act=nn.GELU,
-                    bias_attr=None))
-        if sub_num == 3:
-            self.proj = nn.Sequential(
-                ConvBNLayer(
-                    in_channels=in_channels,
-                    out_channels=embed_dim // 4,
-                    kernel_size=3,
-                    stride=2,
-                    padding=1,
-                    act=nn.GELU,
-                    bias_attr=None),
-                ConvBNLayer(
-                    in_channels=embed_dim // 4,
-                    out_channels=embed_dim // 2,
-                    kernel_size=3,
-                    stride=2,
-                    padding=1,
-                    act=nn.GELU,
-                    bias_attr=None),
-                ConvBNLayer(
-                    in_channels=embed_dim // 2,
-                    out_channels=embed_dim,
-                    kernel_size=3,
-                    stride=2,
-                    padding=1,
-                    act=nn.GELU,
-                    bias_attr=None))
+        if mode == 'pope':
+            if sub_num == 2:
+                self.proj = nn.Sequential(
+                    ConvBNLayer(
+                        in_channels=in_channels,
+                        out_channels=embed_dim // 2,
+                        kernel_size=3,
+                        stride=2,
+                        padding=1,
+                        act=nn.GELU,
+                        bias_attr=None),
+                    ConvBNLayer(
+                        in_channels=embed_dim // 2,
+                        out_channels=embed_dim,
+                        kernel_size=3,
+                        stride=2,
+                        padding=1,
+                        act=nn.GELU,
+                        bias_attr=None))
+            if sub_num == 3:
+                self.proj = nn.Sequential(
+                    ConvBNLayer(
+                        in_channels=in_channels,
+                        out_channels=embed_dim // 4,
+                        kernel_size=3,
+                        stride=2,
+                        padding=1,
+                        act=nn.GELU,
+                        bias_attr=None),
+                    ConvBNLayer(
+                        in_channels=embed_dim // 4,
+                        out_channels=embed_dim // 2,
+                        kernel_size=3,
+                        stride=2,
+                        padding=1,
+                        act=nn.GELU,
+                        bias_attr=None),
+                    ConvBNLayer(
+                        in_channels=embed_dim // 2,
+                        out_channels=embed_dim,
+                        kernel_size=3,
+                        stride=2,
+                        padding=1,
+                        act=nn.GELU,
+                        bias_attr=None))
+        elif mode == 'linear':
+            self.proj = nn.Conv2D(
+                1, embed_dim, kernel_size=patch_size, stride=patch_size)
+            self.num_patches = img_size[0] // patch_size[0] * img_size[
+                1] // patch_size[1]
 
     def forward(self, x):
         B, C, H, W = x.shape

ppocr/modeling/backbones/rec_vitstr.py

+# 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.
+"""
+This code is refer from: 
+https://github.com/roatienza/deep-text-recognition-benchmark/blob/master/modules/vitstr.py
+"""
+
+import numpy as np
+import paddle
+import paddle.nn as nn
+from ppocr.modeling.backbones.rec_svtrnet import Block, PatchEmbed, zeros_, trunc_normal_, ones_
+
+scale_dim_heads = {'tiny': [192, 3], 'small': [384, 6], 'base': [768, 12]}
+
+
+class ViTSTR(nn.Layer):
+    def __init__(self,
+                 img_size=[224, 224],
+                 in_channels=1,
+                 scale='tiny',
+                 seqlen=27,
+                 patch_size=[16, 16],
+                 embed_dim=None,
+                 depth=12,
+                 num_heads=None,
+                 mlp_ratio=4,
+                 qkv_bias=True,
+                 qk_scale=None,
+                 drop_path_rate=0.,
+                 drop_rate=0.,
+                 attn_drop_rate=0.,
+                 norm_layer='nn.LayerNorm',
+                 act_layer='nn.GELU',
+                 epsilon=1e-6,
+                 out_channels=None,
+                 **kwargs):
+        super().__init__()
+        self.seqlen = seqlen
+        embed_dim = embed_dim if embed_dim is not None else scale_dim_heads[
+            scale][0]
+        num_heads = num_heads if num_heads is not None else scale_dim_heads[
+            scale][1]
+        out_channels = out_channels if out_channels is not None else embed_dim
+        self.patch_embed = PatchEmbed(
+            img_size=img_size,
+            in_channels=in_channels,
+            embed_dim=embed_dim,
+            patch_size=patch_size,
+            mode='linear')
+        num_patches = self.patch_embed.num_patches
+
+        self.pos_embed = self.create_parameter(
+            shape=[1, num_patches + 1, embed_dim], default_initializer=zeros_)
+        self.add_parameter("pos_embed", self.pos_embed)
+        self.cls_token = self.create_parameter(
+            shape=[1, 1, embed_dim], default_initializer=zeros_)
+        self.add_parameter("cls_token", self.cls_token)
+
+        self.pos_drop = nn.Dropout(p=drop_rate)
+
+        dpr = np.linspace(0, drop_path_rate, depth)
+        self.blocks = nn.LayerList([
+            Block(
+                dim=embed_dim,
+                num_heads=num_heads,
+                mlp_ratio=mlp_ratio,
+                qkv_bias=qkv_bias,
+                qk_scale=qk_scale,
+                drop=drop_rate,
+                attn_drop=attn_drop_rate,
+                drop_path=dpr[i],
+                norm_layer=norm_layer,
+                act_layer=eval(act_layer),
+                epsilon=epsilon,
+                prenorm=False) for i in range(depth)
+        ])
+        self.norm = eval(norm_layer)(embed_dim, epsilon=epsilon)
+
+        self.out_channels = out_channels
+
+        trunc_normal_(self.pos_embed)
+        trunc_normal_(self.cls_token)
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                zeros_(m.bias)
+        elif isinstance(m, nn.LayerNorm):
+            zeros_(m.bias)
+            ones_(m.weight)
+
+    def forward_features(self, x):
+        B = x.shape[0]
+        x = self.patch_embed(x)
+        cls_tokens = paddle.tile(self.cls_token, repeat_times=[B, 1, 1])
+        x = paddle.concat((cls_tokens, x), axis=1)
+        x = x + self.pos_embed
+        x = self.pos_drop(x)
+        for blk in self.blocks:
+            x = blk(x)
+        x = self.norm(x)
+        return x
+
+    def forward(self, x):
+        x = self.forward_features(x)
+        x = x[:, :self.seqlen]
+        return x.transpose([0, 2, 1]).unsqueeze(2)

ppocr/modeling/heads/__init__.py

@@ -33,6 +33,7 @@ def build_head(config):
     from .rec_aster_head import AsterHead
     from .rec_pren_head import PRENHead
     from .rec_multi_head import MultiHead
+    from .rec_abinet_head import ABINetHead
 
     # cls head
     from .cls_head import ClsHead
@@ -46,7 +47,7 @@ def build_head(config):
         'DBHead', 'PSEHead', 'FCEHead', 'EASTHead', 'SASTHead', 'CTCHead',
         'ClsHead', 'AttentionHead', 'SRNHead', 'PGHead', 'Transformer',
         'TableAttentionHead', 'SARHead', 'AsterHead', 'SDMGRHead', 'PRENHead',
-        'MultiHead'
+        'MultiHead', 'ABINetHead'
     ]
 
     #table head

ppocr/modeling/heads/multiheadAttention.py

-# 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 paddle
-from paddle import nn
-import paddle.nn.functional as F
-from paddle.nn import Linear
-from paddle.nn.initializer import XavierUniform as xavier_uniform_
-from paddle.nn.initializer import Constant as constant_
-from paddle.nn.initializer import XavierNormal as xavier_normal_
-
-zeros_ = constant_(value=0.)
-ones_ = constant_(value=1.)
-
-
-class MultiheadAttention(nn.Layer):
-    """Allows the model to jointly attend to information
-    from different representation subspaces.
-    See reference: Attention Is All You Need
-
-    .. math::
-        \text{MultiHead}(Q, K, V) = \text{Concat}(head_1,\dots,head_h)W^O
-        \text{where} head_i = \text{Attention}(QW_i^Q, KW_i^K, VW_i^V)
-
-    Args:
-        embed_dim: total dimension of the model
-        num_heads: parallel attention layers, or heads
-
-    """
-
-    def __init__(self,
-                 embed_dim,
-                 num_heads,
-                 dropout=0.,
-                 bias=True,
-                 add_bias_kv=False,
-                 add_zero_attn=False):
-        super(MultiheadAttention, self).__init__()
-        self.embed_dim = embed_dim
-        self.num_heads = num_heads
-        self.dropout = dropout
-        self.head_dim = embed_dim // num_heads
-        assert self.head_dim * num_heads == self.embed_dim, "embed_dim must be divisible by num_heads"
-        self.scaling = self.head_dim**-0.5
-        self.out_proj = Linear(embed_dim, embed_dim, bias_attr=bias)
-        self._reset_parameters()
-        self.conv1 = paddle.nn.Conv2D(
-            in_channels=embed_dim, out_channels=embed_dim, kernel_size=(1, 1))
-        self.conv2 = paddle.nn.Conv2D(
-            in_channels=embed_dim, out_channels=embed_dim, kernel_size=(1, 1))
-        self.conv3 = paddle.nn.Conv2D(
-            in_channels=embed_dim, out_channels=embed_dim, kernel_size=(1, 1))
-
-    def _reset_parameters(self):
-        xavier_uniform_(self.out_proj.weight)
-
-    def forward(self,
-                query,
-                key,
-                value,
-                key_padding_mask=None,
-                incremental_state=None,
-                attn_mask=None):
-        """
-        Inputs of forward function
-            query: [target length, batch size, embed dim]
-            key: [sequence length, batch size, embed dim]
-            value: [sequence length, batch size, embed dim]
-            key_padding_mask: if True, mask padding based on batch size
-            incremental_state: if provided, previous time steps are cashed
-            need_weights: output attn_output_weights
-            static_kv: key and value are static
-
-        Outputs of forward function
-            attn_output: [target length, batch size, embed dim]
-            attn_output_weights: [batch size, target length, sequence length]
-        """
-        q_shape = paddle.shape(query)
-        src_shape = paddle.shape(key)
-        q = self._in_proj_q(query)
-        k = self._in_proj_k(key)
-        v = self._in_proj_v(value)
-        q *= self.scaling
-        q = paddle.transpose(
-            paddle.reshape(
-                q, [q_shape[0], q_shape[1], self.num_heads, self.head_dim]),
-            [1, 2, 0, 3])
-        k = paddle.transpose(
-            paddle.reshape(
-                k, [src_shape[0], q_shape[1], self.num_heads, self.head_dim]),
-            [1, 2, 0, 3])
-        v = paddle.transpose(
-            paddle.reshape(
-                v, [src_shape[0], q_shape[1], self.num_heads, self.head_dim]),
-            [1, 2, 0, 3])
-        if key_padding_mask is not None:
-            assert key_padding_mask.shape[0] == q_shape[1]
-            assert key_padding_mask.shape[1] == src_shape[0]
-        attn_output_weights = paddle.matmul(q,
-                                            paddle.transpose(k, [0, 1, 3, 2]))
-        if attn_mask is not None:
-            attn_mask = paddle.unsqueeze(paddle.unsqueeze(attn_mask, 0), 0)
-            attn_output_weights += attn_mask
-        if key_padding_mask is not None:
-            attn_output_weights = paddle.reshape(
-                attn_output_weights,
-                [q_shape[1], self.num_heads, q_shape[0], src_shape[0]])
-            key = paddle.unsqueeze(paddle.unsqueeze(key_padding_mask, 1), 2)
-            key = paddle.cast(key, 'float32')
-            y = paddle.full(
-                shape=paddle.shape(key), dtype='float32', fill_value='-inf')
-            y = paddle.where(key == 0., key, y)
-            attn_output_weights += y
-        attn_output_weights = F.softmax(
-            attn_output_weights.astype('float32'),
-            axis=-1,
-            dtype=paddle.float32 if attn_output_weights.dtype == paddle.float16
-            else attn_output_weights.dtype)
-        attn_output_weights = F.dropout(
-            attn_output_weights, p=self.dropout, training=self.training)
-
-        attn_output = paddle.matmul(attn_output_weights, v)
-        attn_output = paddle.reshape(
-            paddle.transpose(attn_output, [2, 0, 1, 3]),
-            [q_shape[0], q_shape[1], self.embed_dim])
-        attn_output = self.out_proj(attn_output)
-
-        return attn_output
-
-    def _in_proj_q(self, query):
-        query = paddle.transpose(query, [1, 2, 0])
-        query = paddle.unsqueeze(query, axis=2)
-        res = self.conv1(query)
-        res = paddle.squeeze(res, axis=2)
-        res = paddle.transpose(res, [2, 0, 1])
-        return res
-
-    def _in_proj_k(self, key):
-        key = paddle.transpose(key, [1, 2, 0])
-        key = paddle.unsqueeze(key, axis=2)
-        res = self.conv2(key)
-        res = paddle.squeeze(res, axis=2)
-        res = paddle.transpose(res, [2, 0, 1])
-        return res
-
-    def _in_proj_v(self, value):
-        value = paddle.transpose(value, [1, 2, 0])  #(1, 2, 0)
-        value = paddle.unsqueeze(value, axis=2)
-        res = self.conv3(value)
-        res = paddle.squeeze(res, axis=2)
-        res = paddle.transpose(res, [2, 0, 1])
-        return res

ppocr/modeling/heads/rec_abinet_head.py

+# 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.
+"""
+This code is refer from: 
+https://github.com/FangShancheng/ABINet/tree/main/modules
+"""
+
+import math
+import paddle
+from paddle import nn
+import paddle.nn.functional as F
+from paddle.nn import LayerList
+from ppocr.modeling.heads.rec_nrtr_head import TransformerBlock, PositionalEncoding
+
+
+class BCNLanguage(nn.Layer):
+    def __init__(self,
+                 d_model=512,
+                 nhead=8,
+                 num_layers=4,
+                 dim_feedforward=2048,
+                 dropout=0.,
+                 max_length=25,
+                 detach=True,
+                 num_classes=37):
+        super().__init__()
+
+        self.d_model = d_model
+        self.detach = detach
+        self.max_length = max_length + 1  # additional stop token
+        self.proj = nn.Linear(num_classes, d_model, bias_attr=False)
+        self.token_encoder = PositionalEncoding(
+            dropout=0.1, dim=d_model, max_len=self.max_length)
+        self.pos_encoder = PositionalEncoding(
+            dropout=0, dim=d_model, max_len=self.max_length)
+
+        self.decoder = nn.LayerList([
+            TransformerBlock(
+                d_model=d_model,
+                nhead=nhead,
+                dim_feedforward=dim_feedforward,
+                attention_dropout_rate=dropout,
+                residual_dropout_rate=dropout,
+                with_self_attn=False,
+                with_cross_attn=True) for i in range(num_layers)
+        ])
+
+        self.cls = nn.Linear(d_model, num_classes)
+
+    def forward(self, tokens, lengths):
+        """
+        Args:
+            tokens: (B, N, C) where N is length, B is batch size and C is classes number
+            lengths: (B,)
+        """
+        if self.detach: tokens = tokens.detach()
+        embed = self.proj(tokens)  # (B, N, C)
+        embed = self.token_encoder(embed)  # (B, N, C)
+        padding_mask = _get_mask(lengths, self.max_length)
+        zeros = paddle.zeros_like(embed)  # (B, N, C)
+        qeury = self.pos_encoder(zeros)
+        for decoder_layer in self.decoder:
+            qeury = decoder_layer(qeury, embed, cross_mask=padding_mask)
+        output = qeury  # (B, N, C)
+
+        logits = self.cls(output)  # (B, N, C)
+
+        return output, logits
+
+
+def encoder_layer(in_c, out_c, k=3, s=2, p=1):
+    return nn.Sequential(
+        nn.Conv2D(in_c, out_c, k, s, p), nn.BatchNorm2D(out_c), nn.ReLU())
+
+
+def decoder_layer(in_c,
+                  out_c,
+                  k=3,
+                  s=1,
+                  p=1,
+                  mode='nearest',
+                  scale_factor=None,
+                  size=None):
+    align_corners = False if mode == 'nearest' else True
+    return nn.Sequential(
+        nn.Upsample(
+            size=size,
+            scale_factor=scale_factor,
+            mode=mode,
+            align_corners=align_corners),
+        nn.Conv2D(in_c, out_c, k, s, p),
+        nn.BatchNorm2D(out_c),
+        nn.ReLU())
+
+
+class PositionAttention(nn.Layer):
+    def __init__(self,
+                 max_length,
+                 in_channels=512,
+                 num_channels=64,
+                 h=8,
+                 w=32,
+                 mode='nearest',
+                 **kwargs):
+        super().__init__()
+        self.max_length = max_length
+        self.k_encoder = nn.Sequential(
+            encoder_layer(
+                in_channels, num_channels, s=(1, 2)),
+            encoder_layer(
+                num_channels, num_channels, s=(2, 2)),
+            encoder_layer(
+                num_channels, num_channels, s=(2, 2)),
+            encoder_layer(
+                num_channels, num_channels, s=(2, 2)))
+        self.k_decoder = nn.Sequential(
+            decoder_layer(
+                num_channels, num_channels, scale_factor=2, mode=mode),
+            decoder_layer(
+                num_channels, num_channels, scale_factor=2, mode=mode),
+            decoder_layer(
+                num_channels, num_channels, scale_factor=2, mode=mode),
+            decoder_layer(
+                num_channels, in_channels, size=(h, w), mode=mode))
+
+        self.pos_encoder = PositionalEncoding(
+            dropout=0, dim=in_channels, max_len=max_length)
+        self.project = nn.Linear(in_channels, in_channels)
+
+    def forward(self, x):
+        B, C, H, W = x.shape
+        k, v = x, x
+
+        # calculate key vector
+        features = []
+        for i in range(0, len(self.k_encoder)):
+            k = self.k_encoder[i](k)
+            features.append(k)
+        for i in range(0, len(self.k_decoder) - 1):
+            k = self.k_decoder[i](k)
+            # print(k.shape, features[len(self.k_decoder) - 2 - i].shape)
+            k = k + features[len(self.k_decoder) - 2 - i]
+        k = self.k_decoder[-1](k)
+
+        # calculate query vector
+        # TODO q=f(q,k)
+        zeros = paddle.zeros(
+            (B, self.max_length, C), dtype=x.dtype)  # (T, N, C)
+        q = self.pos_encoder(zeros)  # (B, N, C)
+        q = self.project(q)  # (B, N, C)
+
+        # calculate attention
+        attn_scores = q @k.flatten(2)  # (B, N, (H*W))
+        attn_scores = attn_scores / (C**0.5)
+        attn_scores = F.softmax(attn_scores, axis=-1)
+
+        v = v.flatten(2).transpose([0, 2, 1])  # (B, (H*W), C)
+        attn_vecs = attn_scores @v  # (B, N, C)
+
+        return attn_vecs, attn_scores.reshape([0, self.max_length, H, W])
+
+
+class ABINetHead(nn.Layer):
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 d_model=512,
+                 nhead=8,
+                 num_layers=3,
+                 dim_feedforward=2048,
+                 dropout=0.1,
+                 max_length=25,
+                 use_lang=False,
+                 iter_size=1):
+        super().__init__()
+        self.max_length = max_length + 1
+        self.pos_encoder = PositionalEncoding(
+            dropout=0.1, dim=d_model, max_len=8 * 32)
+        self.encoder = nn.LayerList([
+            TransformerBlock(
+                d_model=d_model,
+                nhead=nhead,
+                dim_feedforward=dim_feedforward,
+                attention_dropout_rate=dropout,
+                residual_dropout_rate=dropout,
+                with_self_attn=True,
+                with_cross_attn=False) for i in range(num_layers)
+        ])
+        self.decoder = PositionAttention(
+            max_length=max_length + 1,  # additional stop token
+            mode='nearest', )
+        self.out_channels = out_channels
+        self.cls = nn.Linear(d_model, self.out_channels)
+        self.use_lang = use_lang
+        if use_lang:
+            self.iter_size = iter_size
+            self.language = BCNLanguage(
+                d_model=d_model,
+                nhead=nhead,
+                num_layers=4,
+                dim_feedforward=dim_feedforward,
+                dropout=dropout,
+                max_length=max_length,
+                num_classes=self.out_channels)
+            # alignment
+            self.w_att_align = nn.Linear(2 * d_model, d_model)
+            self.cls_align = nn.Linear(d_model, self.out_channels)
+
+    def forward(self, x, targets=None):
+        x = x.transpose([0, 2, 3, 1])
+        _, H, W, C = x.shape
+        feature = x.flatten(1, 2)
+        feature = self.pos_encoder(feature)
+        for encoder_layer in self.encoder:
+            feature = encoder_layer(feature)
+        feature = feature.reshape([0, H, W, C]).transpose([0, 3, 1, 2])
+        v_feature, attn_scores = self.decoder(
+            feature)  # (B, N, C), (B, C, H, W)
+        vis_logits = self.cls(v_feature)  # (B, N, C)
+        logits = vis_logits
+        vis_lengths = _get_length(vis_logits)
+        if self.use_lang:
+            align_logits = vis_logits
+            align_lengths = vis_lengths
+            all_l_res, all_a_res = [], []
+            for i in range(self.iter_size):
+                tokens = F.softmax(align_logits, axis=-1)
+                lengths = align_lengths
+                lengths = paddle.clip(
+                    lengths, 2, self.max_length)  # TODO:move to langauge model
+                l_feature, l_logits = self.language(tokens, lengths)
+
+                # alignment
+                all_l_res.append(l_logits)
+                fuse = paddle.concat((l_feature, v_feature), -1)
+                f_att = F.sigmoid(self.w_att_align(fuse))
+                output = f_att * v_feature + (1 - f_att) * l_feature
+                align_logits = self.cls_align(output)  # (B, N, C)
+
+                align_lengths = _get_length(align_logits)
+                all_a_res.append(align_logits)
+            if self.training:
+                return {
+                    'align': all_a_res,
+                    'lang': all_l_res,
+                    'vision': vis_logits
+                }
+            else:
+                logits = align_logits
+        if self.training:
+            return logits
+        else:
+            return F.softmax(logits, -1)
+
+
+def _get_length(logit):
+    """ Greed decoder to obtain length from logit"""
+    out = (logit.argmax(-1) == 0)
+    abn = out.any(-1)
+    out_int = out.cast('int32')
+    out = (out_int.cumsum(-1) == 1) & out
+    out = out.cast('int32')
+    out = out.argmax(-1)
+    out = out + 1
+    out = paddle.where(abn, out, paddle.to_tensor(logit.shape[1]))
+    return out
+
+
+def _get_mask(length, max_length):
+    """Generate a square mask for the sequence. The masked positions are filled with float('-inf').
+        Unmasked positions are filled with float(0.0).
+    """
+    length = length.unsqueeze(-1)
+    B = paddle.shape(length)[0]
+    grid = paddle.arange(0, max_length).unsqueeze(0).tile([B, 1])
+    zero_mask = paddle.zeros([B, max_length], dtype='float32')
+    inf_mask = paddle.full([B, max_length], '-inf', dtype='float32')
+    diag_mask = paddle.diag(
+        paddle.full(
+            [max_length], '-inf', dtype=paddle.float32),
+        offset=0,
+        name=None)
+    mask = paddle.where(grid >= length, inf_mask, zero_mask)
+    mask = mask.unsqueeze(1) + diag_mask
+    return mask.unsqueeze(1)

ppocr/modeling/heads/rec_nrtr_head.py

@@ -14,20 +14,15 @@
 
 import math
 import paddle
-import copy
 from paddle import nn
 import paddle.nn.functional as F
 from paddle.nn import LayerList
-from paddle.nn.initializer import XavierNormal as xavier_uniform_
-from paddle.nn import Dropout, Linear, LayerNorm, Conv2D
+# from paddle.nn.initializer import XavierNormal as xavier_uniform_
+from paddle.nn import Dropout, Linear, LayerNorm
 import numpy as np
-from ppocr.modeling.heads.multiheadAttention import MultiheadAttention
-from paddle.nn.initializer import Constant as constant_
+from ppocr.modeling.backbones.rec_svtrnet import Mlp, zeros_, ones_
 from paddle.nn.initializer import XavierNormal as xavier_normal_
 
-zeros_ = constant_(value=0.)
-ones_ = constant_(value=1.)
-
 
 class Transformer(nn.Layer):
     """A transformer model. User is able to modify the attributes as needed. The architechture
@@ -45,7 +40,6 @@ class Transformer(nn.Layer):
         dropout: the dropout value (default=0.1).
         custom_encoder: custom encoder (default=None).
         custom_decoder: custom decoder (default=None).
-
     """
 
     def __init__(self,
@@ -54,45 +48,49 @@ class Transformer(nn.Layer):
                  num_encoder_layers=6,
                  beam_size=0,
                  num_decoder_layers=6,
+                 max_len=25,
                  dim_feedforward=1024,
                  attention_dropout_rate=0.0,
                  residual_dropout_rate=0.1,
-                 custom_encoder=None,
-                 custom_decoder=None,
                  in_channels=0,
                  out_channels=0,
                  scale_embedding=True):
         super(Transformer, self).__init__()
         self.out_channels = out_channels + 1
+        self.max_len = max_len
         self.embedding = Embeddings(
             d_model=d_model,
             vocab=self.out_channels,
             padding_idx=0,
             scale_embedding=scale_embedding)
         self.positional_encoding = PositionalEncoding(
-            dropout=residual_dropout_rate,
-            dim=d_model, )
-        if custom_encoder is not None:
-            self.encoder = custom_encoder
-        else:
-            if num_encoder_layers > 0:
-                encoder_layer = TransformerEncoderLayer(
-                    d_model, nhead, dim_feedforward, attention_dropout_rate,
-                    residual_dropout_rate)
-                self.encoder = TransformerEncoder(encoder_layer,
-                                                  num_encoder_layers)
-            else:
-                self.encoder = None
-
-        if custom_decoder is not None:
-            self.decoder = custom_decoder
+            dropout=residual_dropout_rate, dim=d_model)
+
+        if num_encoder_layers > 0:
+            self.encoder = nn.LayerList([
+                TransformerBlock(
+                    d_model,
+                    nhead,
+                    dim_feedforward,
+                    attention_dropout_rate,
+                    residual_dropout_rate,
+                    with_self_attn=True,
+                    with_cross_attn=False) for i in range(num_encoder_layers)
+            ])
         else:
-            decoder_layer = TransformerDecoderLayer(
-                d_model, nhead, dim_feedforward, attention_dropout_rate,
-                residual_dropout_rate)
-            self.decoder = TransformerDecoder(decoder_layer, num_decoder_layers)
+            self.encoder = None
+
+        self.decoder = nn.LayerList([
+            TransformerBlock(
+                d_model,
+                nhead,
+                dim_feedforward,
+                attention_dropout_rate,
+                residual_dropout_rate,
+                with_self_attn=True,
+                with_cross_attn=True) for i in range(num_decoder_layers)
+        ])
 
-        self._reset_parameters()
         self.beam_size = beam_size
         self.d_model = d_model
         self.nhead = nhead
@@ -105,7 +103,7 @@ class Transformer(nn.Layer):
 
     def _init_weights(self, m):
 
-        if isinstance(m, nn.Conv2D):
+        if isinstance(m, nn.Linear):
             xavier_normal_(m.weight)
             if m.bias is not None:
                 zeros_(m.bias)
@@ -113,24 +111,20 @@ class Transformer(nn.Layer):
     def forward_train(self, src, tgt):
         tgt = tgt[:, :-1]
 
-        tgt_key_padding_mask = self.generate_padding_mask(tgt)
-        tgt = self.embedding(tgt).transpose([1, 0, 2])
+        tgt = self.embedding(tgt)
         tgt = self.positional_encoding(tgt)
-        tgt_mask = self.generate_square_subsequent_mask(tgt.shape[0])
+        tgt_mask = self.generate_square_subsequent_mask(tgt.shape[1])
 
         if self.encoder is not None:
-            src = self.positional_encoding(src.transpose([1, 0, 2]))
-            memory = self.encoder(src)
+            src = self.positional_encoding(src)
+            for encoder_layer in self.encoder:
+                src = encoder_layer(src)
+            memory = src  # B N C
         else:
-            memory = src.squeeze(2).transpose([2, 0, 1])
-        output = self.decoder(
-            tgt,
-            memory,
-            tgt_mask=tgt_mask,
-            memory_mask=None,
-            tgt_key_padding_mask=tgt_key_padding_mask,
-            memory_key_padding_mask=None)
-        output = output.transpose([1, 0, 2])
+            memory = src  # B N C
+        for decoder_layer in self.decoder:
+            tgt = decoder_layer(tgt, memory, self_mask=tgt_mask)
+        output = tgt
         logit = self.tgt_word_prj(output)
         return logit
 
@@ -140,8 +134,8 @@ class Transformer(nn.Layer):
             src: the sequence to the encoder (required).
             tgt: the sequence to the decoder (required).
         Shape:
-            - src: :math:`(S, N, E)`.
-            - tgt: :math:`(T, N, E)`.
+            - src: :math:`(B, sN, C)`.
+            - tgt: :math:`(B, tN, C)`.
         Examples:
             >>> output = transformer_model(src, tgt)
         """
@@ -157,36 +151,35 @@ class Transformer(nn.Layer):
                 return self.forward_test(src)
 
     def forward_test(self, src):
+
         bs = paddle.shape(src)[0]
         if self.encoder is not None:
-            src = self.positional_encoding(paddle.transpose(src, [1, 0, 2]))
-            memory = self.encoder(src)
+            src = self.positional_encoding(src)
+            for encoder_layer in self.encoder:
+                src = encoder_layer(src)
+            memory = src  # B N C
         else:
-            memory = paddle.transpose(paddle.squeeze(src, 2), [2, 0, 1])
+            memory = src
         dec_seq = paddle.full((bs, 1), 2, dtype=paddle.int64)
         dec_prob = paddle.full((bs, 1), 1., dtype=paddle.float32)
-        for len_dec_seq in range(1, 25):
-            dec_seq_embed = paddle.transpose(self.embedding(dec_seq), [1, 0, 2])
+        for len_dec_seq in range(1, self.max_len):
+            dec_seq_embed = self.embedding(dec_seq)
             dec_seq_embed = self.positional_encoding(dec_seq_embed)
             tgt_mask = self.generate_square_subsequent_mask(
-                paddle.shape(dec_seq_embed)[0])
-            output = self.decoder(
-                dec_seq_embed,
-                memory,
-                tgt_mask=tgt_mask,
-                memory_mask=None,
-                tgt_key_padding_mask=None,
-                memory_key_padding_mask=None)
-            dec_output = paddle.transpose(output, [1, 0, 2])
+                paddle.shape(dec_seq_embed)[1])
+            tgt = dec_seq_embed
+            for decoder_layer in self.decoder:
+                tgt = decoder_layer(tgt, memory, self_mask=tgt_mask)
+            dec_output = tgt
             dec_output = dec_output[:, -1, :]
-            word_prob = F.softmax(self.tgt_word_prj(dec_output), axis=1)
-            preds_idx = paddle.argmax(word_prob, axis=1)
+            word_prob = F.softmax(self.tgt_word_prj(dec_output), axis=-1)
+            preds_idx = paddle.argmax(word_prob, axis=-1)
             if paddle.equal_all(
                     preds_idx,
                     paddle.full(
                         paddle.shape(preds_idx), 3, dtype='int64')):
                 break
-            preds_prob = paddle.max(word_prob, axis=1)
+            preds_prob = paddle.max(word_prob, axis=-1)
             dec_seq = paddle.concat(
                 [dec_seq, paddle.reshape(preds_idx, [-1, 1])], axis=1)
             dec_prob = paddle.concat(
@@ -194,10 +187,10 @@ class Transformer(nn.Layer):
         return [dec_seq, dec_prob]
 
     def forward_beam(self, images):
-        ''' Translation work in one batch '''
+        """ Translation work in one batch """
 
         def get_inst_idx_to_tensor_position_map(inst_idx_list):
-            ''' Indicate the position of an instance in a tensor. '''
+            """ Indicate the position of an instance in a tensor. """
             return {
                 inst_idx: tensor_position
                 for tensor_position, inst_idx in enumerate(inst_idx_list)
@@ -205,7 +198,7 @@ class Transformer(nn.Layer):
 
         def collect_active_part(beamed_tensor, curr_active_inst_idx,
                                 n_prev_active_inst, n_bm):
-            ''' Collect tensor parts associated to active instances. '''
+            """ Collect tensor parts associated to active instances. """
 
             beamed_tensor_shape = paddle.shape(beamed_tensor)
             n_curr_active_inst = len(curr_active_inst_idx)
@@ -237,9 +230,8 @@ class Transformer(nn.Layer):
             return active_src_enc, active_inst_idx_to_position_map
 
         def beam_decode_step(inst_dec_beams, len_dec_seq, enc_output,
-                             inst_idx_to_position_map, n_bm,
-                             memory_key_padding_mask):
-            ''' Decode and update beam status, and then return active beam idx '''
+                             inst_idx_to_position_map, n_bm):
+            """ Decode and update beam status, and then return active beam idx """
 
             def prepare_beam_dec_seq(inst_dec_beams, len_dec_seq):
                 dec_partial_seq = [
@@ -249,19 +241,15 @@ class Transformer(nn.Layer):
                 dec_partial_seq = dec_partial_seq.reshape([-1, len_dec_seq])
                 return dec_partial_seq
 
-            def predict_word(dec_seq, enc_output, n_active_inst, n_bm,
-                             memory_key_padding_mask):
-                dec_seq = paddle.transpose(self.embedding(dec_seq), [1, 0, 2])
+            def predict_word(dec_seq, enc_output, n_active_inst, n_bm):
+                dec_seq = self.embedding(dec_seq)
                 dec_seq = self.positional_encoding(dec_seq)
                 tgt_mask = self.generate_square_subsequent_mask(
-                    paddle.shape(dec_seq)[0])
-                dec_output = self.decoder(
-                    dec_seq,
-                    enc_output,
-                    tgt_mask=tgt_mask,
-                    tgt_key_padding_mask=None,
-                    memory_key_padding_mask=memory_key_padding_mask, )
-                dec_output = paddle.transpose(dec_output, [1, 0, 2])
+                    paddle.shape(dec_seq)[1])
+                tgt = dec_seq
+                for decoder_layer in self.decoder:
+                    tgt = decoder_layer(tgt, enc_output, self_mask=tgt_mask)
+                dec_output = tgt
                 dec_output = dec_output[:,
                                         -1, :]  # Pick the last step: (bh * bm) * d_h
                 word_prob = F.softmax(self.tgt_word_prj(dec_output), axis=1)
@@ -281,8 +269,7 @@ class Transformer(nn.Layer):
 
             n_active_inst = len(inst_idx_to_position_map)
             dec_seq = prepare_beam_dec_seq(inst_dec_beams, len_dec_seq)
-            word_prob = predict_word(dec_seq, enc_output, n_active_inst, n_bm,
-                                     None)
+            word_prob = predict_word(dec_seq, enc_output, n_active_inst, n_bm)
             # Update the beam with predicted word prob information and collect incomplete instances
             active_inst_idx_list = collect_active_inst_idx_list(
                 inst_dec_beams, word_prob, inst_idx_to_position_map)
@@ -303,10 +290,10 @@ class Transformer(nn.Layer):
         with paddle.no_grad():
             #-- Encode
             if self.encoder is not None:
-                src = self.positional_encoding(images.transpose([1, 0, 2]))
+                src = self.positional_encoding(images)
                 src_enc = self.encoder(src)
             else:
-                src_enc = images.squeeze(2).transpose([0, 2, 1])
+                src_enc = images
 
             n_bm = self.beam_size
             src_shape = paddle.shape(src_enc)
@@ -317,11 +304,11 @@ class Transformer(nn.Layer):
             inst_idx_to_position_map = get_inst_idx_to_tensor_position_map(
                 active_inst_idx_list)
             # Decode
-            for len_dec_seq in range(1, 25):
+            for len_dec_seq in range(1, self.max_len):
                 src_enc_copy = src_enc.clone()
                 active_inst_idx_list = beam_decode_step(
                     inst_dec_beams, len_dec_seq, src_enc_copy,
-                    inst_idx_to_position_map, n_bm, None)
+                    inst_idx_to_position_map, n_bm)
                 if not active_inst_idx_list:
                     break  # all instances have finished their path to <EOS>
                 src_enc, inst_idx_to_position_map = collate_active_info(
@@ -354,261 +341,124 @@ class Transformer(nn.Layer):
                 shape=[sz, sz], dtype='float32', fill_value='-inf'),
             diagonal=1)
         mask = mask + mask_inf
-        return mask
-
-    def generate_padding_mask(self, x):
-        padding_mask = paddle.equal(x, paddle.to_tensor(0, dtype=x.dtype))
-        return padding_mask
+        return mask.unsqueeze([0, 1])
 
-    def _reset_parameters(self):
-        """Initiate parameters in the transformer model."""
 
-        for p in self.parameters():
-            if p.dim() > 1:
-                xavier_uniform_(p)
+class MultiheadAttention(nn.Layer):
+    """Allows the model to jointly attend to information
+    from different representation subspaces.
+    See reference: Attention Is All You Need
 
+    .. math::
+        \text{MultiHead}(Q, K, V) = \text{Concat}(head_1,\dots,head_h)W^O
+        \text{where} head_i = \text{Attention}(QW_i^Q, KW_i^K, VW_i^V)
 
-class TransformerEncoder(nn.Layer):
-    """TransformerEncoder is a stack of N encoder layers
     Args:
-        encoder_layer: an instance of the TransformerEncoderLayer() class (required).
-        num_layers: the number of sub-encoder-layers in the encoder (required).
-        norm: the layer normalization component (optional).
-    """
+        embed_dim: total dimension of the model
+        num_heads: parallel attention layers, or heads
 
-    def __init__(self, encoder_layer, num_layers):
-        super(TransformerEncoder, self).__init__()
-        self.layers = _get_clones(encoder_layer, num_layers)
-        self.num_layers = num_layers
+    """
 
-    def forward(self, src):
-        """Pass the input through the endocder layers in turn.
-        Args:
-            src: the sequnce to the encoder (required).
-            mask: the mask for the src sequence (optional).
-            src_key_padding_mask: the mask for the src keys per batch (optional).
-        """
-        output = src
+    def __init__(self, embed_dim, num_heads, dropout=0., self_attn=False):
+        super(MultiheadAttention, self).__init__()
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        # self.dropout = dropout
+        self.head_dim = embed_dim // num_heads
+        assert self.head_dim * num_heads == self.embed_dim, "embed_dim must be divisible by num_heads"
+        self.scale = self.head_dim**-0.5
+        self.self_attn = self_attn
+        if self_attn:
+            self.qkv = nn.Linear(embed_dim, embed_dim * 3)
+        else:
+            self.q = nn.Linear(embed_dim, embed_dim)
+            self.kv = nn.Linear(embed_dim, embed_dim * 2)
+        self.attn_drop = nn.Dropout(dropout)
+        self.out_proj = nn.Linear(embed_dim, embed_dim)
 
-        for i in range(self.num_layers):
-            output = self.layers[i](output,
-                                    src_mask=None,
-                                    src_key_padding_mask=None)
+    def forward(self, query, key=None, attn_mask=None):
 
-        return output
+        qN = query.shape[1]
 
+        if self.self_attn:
+            qkv = self.qkv(query).reshape(
+                (0, qN, 3, self.num_heads, self.head_dim)).transpose(
+                    (2, 0, 3, 1, 4))
+            q, k, v = qkv[0], qkv[1], qkv[2]
+        else:
+            kN = key.shape[1]
+            q = self.q(query).reshape(
+                [0, qN, self.num_heads, self.head_dim]).transpose([0, 2, 1, 3])
+            kv = self.kv(key).reshape(
+                (0, kN, 2, self.num_heads, self.head_dim)).transpose(
+                    (2, 0, 3, 1, 4))
+            k, v = kv[0], kv[1]
 
-class TransformerDecoder(nn.Layer):
-    """TransformerDecoder is a stack of N decoder layers
+        attn = (q.matmul(k.transpose((0, 1, 3, 2)))) * self.scale
 
-    Args:
-        decoder_layer: an instance of the TransformerDecoderLayer() class (required).
-        num_layers: the number of sub-decoder-layers in the decoder (required).
-        norm: the layer normalization component (optional).
+        if attn_mask is not None:
+            attn += attn_mask
 
-    """
+        attn = F.softmax(attn, axis=-1)
+        attn = self.attn_drop(attn)
 
-    def __init__(self, decoder_layer, num_layers):
-        super(TransformerDecoder, self).__init__()
-        self.layers = _get_clones(decoder_layer, num_layers)
-        self.num_layers = num_layers
+        x = (attn.matmul(v)).transpose((0, 2, 1, 3)).reshape(
+            (0, qN, self.embed_dim))
+        x = self.out_proj(x)
 
-    def forward(self,
-                tgt,
-                memory,
-                tgt_mask=None,
-                memory_mask=None,
-                tgt_key_padding_mask=None,
-                memory_key_padding_mask=None):
-        """Pass the inputs (and mask) through the decoder layer in turn.
+        return x
 
-        Args:
-            tgt: the sequence to the decoder (required).
-            memory: the sequnce from the last layer of the encoder (required).
-            tgt_mask: the mask for the tgt sequence (optional).
-            memory_mask: the mask for the memory sequence (optional).
-            tgt_key_padding_mask: the mask for the tgt keys per batch (optional).
-            memory_key_padding_mask: the mask for the memory keys per batch (optional).
-        """
-        output = tgt
-        for i in range(self.num_layers):
-            output = self.layers[i](
-                output,
-                memory,
-                tgt_mask=tgt_mask,
-                memory_mask=memory_mask,
-                tgt_key_padding_mask=tgt_key_padding_mask,
-                memory_key_padding_mask=memory_key_padding_mask)
-
-        return output
-
-
-class TransformerEncoderLayer(nn.Layer):
-    """TransformerEncoderLayer is made up of self-attn and feedforward network.
-    This standard encoder layer is based on the paper "Attention Is All You Need".
-    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez,
-    Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in
-    Neural Information Processing Systems, pages 6000-6010. Users may modify or implement
-    in a different way during application.
-
-    Args:
-        d_model: the number of expected features in the input (required).
-        nhead: the number of heads in the multiheadattention models (required).
-        dim_feedforward: the dimension of the feedforward network model (default=2048).
-        dropout: the dropout value (default=0.1).
-
-    """
 
+class TransformerBlock(nn.Layer):
     def __init__(self,
                  d_model,
                  nhead,
                  dim_feedforward=2048,
                  attention_dropout_rate=0.0,
-                 residual_dropout_rate=0.1):
-        super(TransformerEncoderLayer, self).__init__()
-        self.self_attn = MultiheadAttention(
-            d_model, nhead, dropout=attention_dropout_rate)
-
-        self.conv1 = Conv2D(
-            in_channels=d_model,
-            out_channels=dim_feedforward,
-            kernel_size=(1, 1))
-        self.conv2 = Conv2D(
-            in_channels=dim_feedforward,
-            out_channels=d_model,
-            kernel_size=(1, 1))
-
-        self.norm1 = LayerNorm(d_model)
-        self.norm2 = LayerNorm(d_model)
-        self.dropout1 = Dropout(residual_dropout_rate)
-        self.dropout2 = Dropout(residual_dropout_rate)
-
-    def forward(self, src, src_mask=None, src_key_padding_mask=None):
-        """Pass the input through the endocder layer.
-        Args:
-            src: the sequnce to the encoder layer (required).
-            src_mask: the mask for the src sequence (optional).
-            src_key_padding_mask: the mask for the src keys per batch (optional).
-        """
-        src2 = self.self_attn(
-            src,
-            src,
-            src,
-            attn_mask=src_mask,
-            key_padding_mask=src_key_padding_mask)
-        src = src + self.dropout1(src2)
-        src = self.norm1(src)
-
-        src = paddle.transpose(src, [1, 2, 0])
-        src = paddle.unsqueeze(src, 2)
-        src2 = self.conv2(F.relu(self.conv1(src)))
-        src2 = paddle.squeeze(src2, 2)
-        src2 = paddle.transpose(src2, [2, 0, 1])
-        src = paddle.squeeze(src, 2)
-        src = paddle.transpose(src, [2, 0, 1])
-
-        src = src + self.dropout2(src2)
-        src = self.norm2(src)
-        return src
-
-
-class TransformerDecoderLayer(nn.Layer):
-    """TransformerDecoderLayer is made up of self-attn, multi-head-attn and feedforward network.
-    This standard decoder layer is based on the paper "Attention Is All You Need".
-    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez,
-    Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in
-    Neural Information Processing Systems, pages 6000-6010. Users may modify or implement
-    in a different way during application.
-
-    Args:
-        d_model: the number of expected features in the input (required).
-        nhead: the number of heads in the multiheadattention models (required).
-        dim_feedforward: the dimension of the feedforward network model (default=2048).
-        dropout: the dropout value (default=0.1).
-
-    """
+                 residual_dropout_rate=0.1,
+                 with_self_attn=True,
+                 with_cross_attn=False,
+                 epsilon=1e-5):
+        super(TransformerBlock, self).__init__()
+        self.with_self_attn = with_self_attn
+        if with_self_attn:
+            self.self_attn = MultiheadAttention(
+                d_model,
+                nhead,
+                dropout=attention_dropout_rate,
+                self_attn=with_self_attn)
+            self.norm1 = LayerNorm(d_model, epsilon=epsilon)
+            self.dropout1 = Dropout(residual_dropout_rate)
+        self.with_cross_attn = with_cross_attn
+        if with_cross_attn:
+            self.cross_attn = MultiheadAttention(  #for self_attn of encoder or cross_attn of decoder
+                d_model,
+                nhead,
+                dropout=attention_dropout_rate)
+            self.norm2 = LayerNorm(d_model, epsilon=epsilon)
+            self.dropout2 = Dropout(residual_dropout_rate)
+
+        self.mlp = Mlp(in_features=d_model,
+                       hidden_features=dim_feedforward,
+                       act_layer=nn.ReLU,
+                       drop=residual_dropout_rate)
+
+        self.norm3 = LayerNorm(d_model, epsilon=epsilon)
 
-    def __init__(self,
-                 d_model,
-                 nhead,
-                 dim_feedforward=2048,
-                 attention_dropout_rate=0.0,
-                 residual_dropout_rate=0.1):
-        super(TransformerDecoderLayer, self).__init__()
-        self.self_attn = MultiheadAttention(
-            d_model, nhead, dropout=attention_dropout_rate)
-        self.multihead_attn = MultiheadAttention(
-            d_model, nhead, dropout=attention_dropout_rate)
-
-        self.conv1 = Conv2D(
-            in_channels=d_model,
-            out_channels=dim_feedforward,
-            kernel_size=(1, 1))
-        self.conv2 = Conv2D(
-            in_channels=dim_feedforward,
-            out_channels=d_model,
-            kernel_size=(1, 1))
-
-        self.norm1 = LayerNorm(d_model)
-        self.norm2 = LayerNorm(d_model)
-        self.norm3 = LayerNorm(d_model)
-        self.dropout1 = Dropout(residual_dropout_rate)
-        self.dropout2 = Dropout(residual_dropout_rate)
         self.dropout3 = Dropout(residual_dropout_rate)
 
-    def forward(self,
-                tgt,
-                memory,
-                tgt_mask=None,
-                memory_mask=None,
-                tgt_key_padding_mask=None,
-                memory_key_padding_mask=None):
-        """Pass the inputs (and mask) through the decoder layer.
+    def forward(self, tgt, memory=None, self_mask=None, cross_mask=None):
+        if self.with_self_attn:
+            tgt1 = self.self_attn(tgt, attn_mask=self_mask)
+            tgt = self.norm1(tgt + self.dropout1(tgt1))
 
-        Args:
-            tgt: the sequence to the decoder layer (required).
-            memory: the sequnce from the last layer of the encoder (required).
-            tgt_mask: the mask for the tgt sequence (optional).
-            memory_mask: the mask for the memory sequence (optional).
-            tgt_key_padding_mask: the mask for the tgt keys per batch (optional).
-            memory_key_padding_mask: the mask for the memory keys per batch (optional).
-
-        """
-        tgt2 = self.self_attn(
-            tgt,
-            tgt,
-            tgt,
-            attn_mask=tgt_mask,
-            key_padding_mask=tgt_key_padding_mask)
-        tgt = tgt + self.dropout1(tgt2)
-        tgt = self.norm1(tgt)
-        tgt2 = self.multihead_attn(
-            tgt,
-            memory,
-            memory,
-            attn_mask=memory_mask,
-            key_padding_mask=memory_key_padding_mask)
-        tgt = tgt + self.dropout2(tgt2)
-        tgt = self.norm2(tgt)
-
-        # default
-        tgt = paddle.transpose(tgt, [1, 2, 0])
-        tgt = paddle.unsqueeze(tgt, 2)
-        tgt2 = self.conv2(F.relu(self.conv1(tgt)))
-        tgt2 = paddle.squeeze(tgt2, 2)
-        tgt2 = paddle.transpose(tgt2, [2, 0, 1])
-        tgt = paddle.squeeze(tgt, 2)
-        tgt = paddle.transpose(tgt, [2, 0, 1])
-
-        tgt = tgt + self.dropout3(tgt2)
-        tgt = self.norm3(tgt)
+        if self.with_cross_attn:
+            tgt2 = self.cross_attn(tgt, key=memory, attn_mask=cross_mask)
+            tgt = self.norm2(tgt + self.dropout2(tgt2))
+        tgt = self.norm3(tgt + self.dropout3(self.mlp(tgt)))
         return tgt
 
 
-def _get_clones(module, N):
-    return LayerList([copy.deepcopy(module) for i in range(N)])
-
-
 class PositionalEncoding(nn.Layer):
     """Inject some information about the relative or absolute position of the tokens
         in the sequence. The positional encodings have the same dimension as
@@ -651,8 +501,9 @@ class PositionalEncoding(nn.Layer):
         Examples:
             >>> output = pos_encoder(x)
         """
+        x = x.transpose([1, 0, 2])
         x = x + self.pe[:paddle.shape(x)[0], :]
-        return self.dropout(x)
+        return self.dropout(x).transpose([1, 0, 2])
 
 
 class PositionalEncoding_2d(nn.Layer):
@@ -725,7 +576,7 @@ class PositionalEncoding_2d(nn.Layer):
 
 
 class Embeddings(nn.Layer):
-    def __init__(self, d_model, vocab, padding_idx, scale_embedding):
+    def __init__(self, d_model, vocab, padding_idx=None, scale_embedding=True):
         super(Embeddings, self).__init__()
         self.embedding = nn.Embedding(vocab, d_model, padding_idx=padding_idx)
         w0 = np.random.normal(0.0, d_model**-0.5,
@@ -742,7 +593,7 @@ class Embeddings(nn.Layer):
 
 
 class Beam():
-    ''' Beam search '''
+    """ Beam search """
 
     def __init__(self, size, device=False):
 

ppocr/postprocess/__init__.py

@@ -27,7 +27,7 @@ from .sast_postprocess import SASTPostProcess
 from .fce_postprocess import FCEPostProcess
 from .rec_postprocess import CTCLabelDecode, AttnLabelDecode, SRNLabelDecode, \
     DistillationCTCLabelDecode, TableLabelDecode, NRTRLabelDecode, SARLabelDecode, \
-    SEEDLabelDecode, PRENLabelDecode
+    SEEDLabelDecode, PRENLabelDecode, ViTSTRLabelDecode, ABINetLabelDecode
 from .cls_postprocess import ClsPostProcess
 from .pg_postprocess import PGPostProcess
 from .vqa_token_ser_layoutlm_postprocess import VQASerTokenLayoutLMPostProcess
@@ -42,7 +42,7 @@ def build_post_process(config, global_config=None):
         'DistillationDBPostProcess', 'NRTRLabelDecode', 'SARLabelDecode',
         'SEEDLabelDecode', 'VQASerTokenLayoutLMPostProcess',
         'VQAReTokenLayoutLMPostProcess', 'PRENLabelDecode',
-        'DistillationSARLabelDecode'
+        'DistillationSARLabelDecode', 'ViTSTRLabelDecode', 'ABINetLabelDecode'
     ]
 
     if config['name'] == 'PSEPostProcess':

ppocr/postprocess/rec_postprocess.py

@@ -140,70 +140,6 @@ class DistillationCTCLabelDecode(CTCLabelDecode):
         return output
 
 
-class NRTRLabelDecode(BaseRecLabelDecode):
-    """ Convert between text-label and text-index """
-
-    def __init__(self, character_dict_path=None, use_space_char=True, **kwargs):
-        super(NRTRLabelDecode, self).__init__(character_dict_path,
-                                              use_space_char)
-
-    def __call__(self, preds, label=None, *args, **kwargs):
-
-        if len(preds) == 2:
-            preds_id = preds[0]
-            preds_prob = preds[1]
-            if isinstance(preds_id, paddle.Tensor):
-                preds_id = preds_id.numpy()
-            if isinstance(preds_prob, paddle.Tensor):
-                preds_prob = preds_prob.numpy()
-            if preds_id[0][0] == 2:
-                preds_idx = preds_id[:, 1:]
-                preds_prob = preds_prob[:, 1:]
-            else:
-                preds_idx = preds_id
-            text = self.decode(preds_idx, preds_prob, is_remove_duplicate=False)
-            if label is None:
-                return text
-            label = self.decode(label[:, 1:])
-        else:
-            if isinstance(preds, paddle.Tensor):
-                preds = preds.numpy()
-            preds_idx = preds.argmax(axis=2)
-            preds_prob = preds.max(axis=2)
-            text = self.decode(preds_idx, preds_prob, is_remove_duplicate=False)
-            if label is None:
-                return text
-            label = self.decode(label[:, 1:])
-        return text, label
-
-    def add_special_char(self, dict_character):
-        dict_character = ['blank', '<unk>', '<s>', '</s>'] + dict_character
-        return dict_character
-
-    def decode(self, text_index, text_prob=None, is_remove_duplicate=False):
-        """ convert text-index into text-label. """
-        result_list = []
-        batch_size = len(text_index)
-        for batch_idx in range(batch_size):
-            char_list = []
-            conf_list = []
-            for idx in range(len(text_index[batch_idx])):
-                if text_index[batch_idx][idx] == 3:  # end
-                    break
-                try:
-                    char_list.append(self.character[int(text_index[batch_idx][
-                        idx])])
-                except:
-                    continue
-                if text_prob is not None:
-                    conf_list.append(text_prob[batch_idx][idx])
-                else:
-                    conf_list.append(1)
-            text = ''.join(char_list)
-            result_list.append((text.lower(), np.mean(conf_list).tolist()))
-        return result_list
-
-
 class AttnLabelDecode(BaseRecLabelDecode):
     """ Convert between text-label and text-index """
 
@@ -752,3 +688,122 @@ class PRENLabelDecode(BaseRecLabelDecode):
             return text
         label = self.decode(label)
         return text, label
+
+
+class NRTRLabelDecode(BaseRecLabelDecode):
+    """ Convert between text-label and text-index """
+
+    def __init__(self, character_dict_path=None, use_space_char=True, **kwargs):
+        super(NRTRLabelDecode, self).__init__(character_dict_path,
+                                              use_space_char)
+
+    def __call__(self, preds, label=None, *args, **kwargs):
+
+        if len(preds) == 2:
+            preds_id = preds[0]
+            preds_prob = preds[1]
+            if isinstance(preds_id, paddle.Tensor):
+                preds_id = preds_id.numpy()
+            if isinstance(preds_prob, paddle.Tensor):
+                preds_prob = preds_prob.numpy()
+            if preds_id[0][0] == 2:
+                preds_idx = preds_id[:, 1:]
+                preds_prob = preds_prob[:, 1:]
+            else:
+                preds_idx = preds_id
+            text = self.decode(preds_idx, preds_prob, is_remove_duplicate=False)
+            if label is None:
+                return text
+            label = self.decode(label[:, 1:])
+        else:
+            if isinstance(preds, paddle.Tensor):
+                preds = preds.numpy()
+            preds_idx = preds.argmax(axis=2)
+            preds_prob = preds.max(axis=2)
+            text = self.decode(preds_idx, preds_prob, is_remove_duplicate=False)
+            if label is None:
+                return text
+            label = self.decode(label[:, 1:])
+        return text, label
+
+    def add_special_char(self, dict_character):
+        dict_character = ['blank', '<unk>', '<s>', '</s>'] + dict_character
+        return dict_character
+
+    def decode(self, text_index, text_prob=None, is_remove_duplicate=False):
+        """ convert text-index into text-label. """
+        result_list = []
+        batch_size = len(text_index)
+        for batch_idx in range(batch_size):
+            char_list = []
+            conf_list = []
+            for idx in range(len(text_index[batch_idx])):
+                try:
+                    char_idx = self.character[int(text_index[batch_idx][idx])]
+                except:
+                    continue
+                if char_idx == '</s>':  # end
+                    break
+                char_list.append(char_idx)
+                if text_prob is not None:
+                    conf_list.append(text_prob[batch_idx][idx])
+                else:
+                    conf_list.append(1)
+            text = ''.join(char_list)
+            result_list.append((text.lower(), np.mean(conf_list).tolist()))
+        return result_list
+
+
+class ViTSTRLabelDecode(NRTRLabelDecode):
+    """ Convert between text-label and text-index """
+
+    def __init__(self, character_dict_path=None, use_space_char=False,
+                 **kwargs):
+        super(ViTSTRLabelDecode, self).__init__(character_dict_path,
+                                                use_space_char)
+
+    def __call__(self, preds, label=None, *args, **kwargs):
+        if isinstance(preds, paddle.Tensor):
+            preds = preds[:, 1:].numpy()
+        else:
+            preds = preds[:, 1:]
+        preds_idx = preds.argmax(axis=2)
+        preds_prob = preds.max(axis=2)
+        text = self.decode(preds_idx, preds_prob, is_remove_duplicate=False)
+        if label is None:
+            return text
+        label = self.decode(label[:, 1:])
+        return text, label
+
+    def add_special_char(self, dict_character):
+        dict_character = ['<s>', '</s>'] + dict_character
+        return dict_character
+
+
+class ABINetLabelDecode(NRTRLabelDecode):
+    """ Convert between text-label and text-index """
+
+    def __init__(self, character_dict_path=None, use_space_char=False,
+                 **kwargs):
+        super(ABINetLabelDecode, self).__init__(character_dict_path,
+                                                use_space_char)
+
+    def __call__(self, preds, label=None, *args, **kwargs):
+        if isinstance(preds, dict):
+            preds = preds['align'][-1].numpy()
+        elif isinstance(preds, paddle.Tensor):
+            preds = preds.numpy()
+        else:
+            preds = preds
+
+        preds_idx = preds.argmax(axis=2)
+        preds_prob = preds.max(axis=2)
+        text = self.decode(preds_idx, preds_prob, is_remove_duplicate=False)
+        if label is None:
+            return text
+        label = self.decode(label)
+        return text, label
+
+    def add_special_char(self, dict_character):
+        dict_character = ['</s>'] + dict_character
+        return dict_character

ppocr/utils/utility.py

@@ -91,18 +91,19 @@ def check_and_read_gif(img_path):
 def load_vqa_bio_label_maps(label_map_path):
     with open(label_map_path, "r", encoding='utf-8') as fin:
         lines = fin.readlines()
-    lines = [line.strip() for line in lines]
-    if "O" not in lines:
-        lines.insert(0, "O")
-    labels = []
-    for line in lines:
-        if line == "O":
-            labels.append("O")
-        else:
-            labels.append("B-" + line)
-            labels.append("I-" + line)
-    label2id_map = {label: idx for idx, label in enumerate(labels)}
-    id2label_map = {idx: label for idx, label in enumerate(labels)}
+    old_lines = [line.strip() for line in lines]
+    lines = ["O"]
+    for line in old_lines:
+        # "O" has already been in lines
+        if line.upper() in ["OTHER", "OTHERS", "IGNORE"]:
+            continue
+        lines.append(line)
+    labels = ["O"]
+    for line in lines[1:]:
+        labels.append("B-" + line)
+        labels.append("I-" + line)
+    label2id_map = {label.upper(): idx for idx, label in enumerate(labels)}
+    id2label_map = {idx: label.upper() for idx, label in enumerate(labels)}
     return label2id_map, id2label_map
 
 

ppocr/utils/visual.py

@@ -19,7 +19,7 @@ from PIL import Image, ImageDraw, ImageFont
 def draw_ser_results(image,
                      ocr_results,
                      font_path="doc/fonts/simfang.ttf",
-                     font_size=18):
+                     font_size=14):
     np.random.seed(2021)
     color = (np.random.permutation(range(255)),
              np.random.permutation(range(255)),
@@ -40,9 +40,15 @@ def draw_ser_results(image,
         if ocr_info["pred_id"] not in color_map:
             continue
         color = color_map[ocr_info["pred_id"]]
-        text = "{}: {}".format(ocr_info["pred"], ocr_info["text"])
+        text = "{}: {}".format(ocr_info["pred"], ocr_info["transcription"])
 
-        draw_box_txt(ocr_info["bbox"], text, draw, font, font_size, color)
+        if "bbox" in ocr_info:
+            # draw with ocr engine
+            bbox = ocr_info["bbox"]
+        else:
+            # draw with ocr groundtruth
+            bbox = trans_poly_to_bbox(ocr_info["points"])
+        draw_box_txt(bbox, text, draw, font, font_size, color)
 
     img_new = Image.blend(image, img_new, 0.5)
     return np.array(img_new)
@@ -62,6 +68,14 @@ def draw_box_txt(bbox, text, draw, font, font_size, color):
     draw.text((bbox[0][0] + 1, start_y), text, fill=(255, 255, 255), font=font)
 
 
+def trans_poly_to_bbox(poly):
+    x1 = np.min([p[0] for p in poly])
+    x2 = np.max([p[0] for p in poly])
+    y1 = np.min([p[1] for p in poly])
+    y2 = np.max([p[1] for p in poly])
+    return [x1, y1, x2, y2]
+
+
 def draw_re_results(image,
                     result,
                     font_path="doc/fonts/simfang.ttf",
@@ -80,10 +94,10 @@ def draw_re_results(image,
     color_line = (0, 255, 0)
 
     for ocr_info_head, ocr_info_tail in result:
-        draw_box_txt(ocr_info_head["bbox"], ocr_info_head["text"], draw, font,
-                     font_size, color_head)
-        draw_box_txt(ocr_info_tail["bbox"], ocr_info_tail["text"], draw, font,
-                     font_size, color_tail)
+        draw_box_txt(ocr_info_head["bbox"], ocr_info_head["transcription"],
+                     draw, font, font_size, color_head)
+        draw_box_txt(ocr_info_tail["bbox"], ocr_info_tail["transcription"],
+                     draw, font, font_size, color_tail)
 
         center_head = (
             (ocr_info_head['bbox'][0] + ocr_info_head['bbox'][2]) // 2,

ppstructure/docs/kie.md

@@ -16,7 +16,7 @@ SDMGR是一个关键信息提取算法，将每个检测到的文本区域分类
 训练和测试的数据采用wildreceipt数据集，通过如下指令下载数据集：
 
 ```
-wget https://paddleocr.bj.bcebos.com/dygraph_v2.1/kie/wildreceipt.tar && tar xf wildreceipt.tar
+wget https://paddleocr.bj.bcebos.com/ppstructure/dataset/wildreceipt.tar && tar xf wildreceipt.tar
 ```
 
 执行预测：

ppstructure/docs/kie_en.md

@@ -15,7 +15,7 @@ This section provides a tutorial example on how to quickly use, train, and evalu
 [Wildreceipt dataset](https://paperswithcode.com/dataset/wildreceipt) is used for this tutorial. It contains 1765 photos, with 25 classes, and 50000 text boxes, which can be downloaded by wget:
 
 ```shell
-wget https://paddleocr.bj.bcebos.com/dygraph_v2.1/kie/wildreceipt.tar && tar xf wildreceipt.tar
+wget https://paddleocr.bj.bcebos.com/ppstructure/dataset/wildreceipt.tar && tar xf wildreceipt.tar
 ```
 
 Download the pretrained model and predict the result:

ppstructure/vqa/README.md

@@ -125,13 +125,13 @@ If you want to experience the prediction process directly, you can download the
 
 * Download the processed dataset
 
-The download address of the processed XFUND Chinese dataset: [https://paddleocr.bj.bcebos.com/dataset/XFUND.tar](https://paddleocr.bj.bcebos.com/dataset/XFUND.tar).
+The download address of the processed XFUND Chinese dataset: [link](https://paddleocr.bj.bcebos.com/ppstructure/dataset/XFUND.tar).
 
 
 Download and unzip the dataset, and place the dataset in the current directory after unzipping.
 
 ```shell
-wget https://paddleocr.bj.bcebos.com/dataset/XFUND.tar
+wget https://paddleocr.bj.bcebos.com/ppstructure/dataset/XFUND.tar
 ````
 
 * Convert the dataset

ppstructure/vqa/README_ch.md

@@ -122,13 +122,13 @@ python3 -m pip install -r ppstructure/vqa/requirements.txt
 
 * 下载处理好的数据集
 
-处理好的XFUND中文数据集下载地址：[https://paddleocr.bj.bcebos.com/dataset/XFUND.tar](https://paddleocr.bj.bcebos.com/dataset/XFUND.tar)。
+处理好的XFUND中文数据集下载地址：[链接](https://paddleocr.bj.bcebos.com/ppstructure/dataset/XFUND.tar)。
 
 
 下载并解压该数据集，解压后将数据集放置在当前目录下。
 
 ```shell
-wget https://paddleocr.bj.bcebos.com/dataset/XFUND.tar
+wget https://paddleocr.bj.bcebos.com/ppstructure/dataset/XFUND.tar
 ```
 
 * 转换数据集

ppstructure/vqa/labels/labels_ser.txt

-QUESTION
-ANSWER
-HEADER

ppstructure/vqa/tools/trans_xfun_data.py

@@ -21,26 +21,22 @@ def transfer_xfun_data(json_path=None, output_file=None):
 
     json_info = json.loads(lines[0])
     documents = json_info["documents"]
-    label_info = {}
     with open(output_file, "w", encoding='utf-8') as fout:
         for idx, document in enumerate(documents):
+            label_info = []
             img_info = document["img"]
             document = document["document"]
             image_path = img_info["fname"]
 
-            label_info["height"] = img_info["height"]
-            label_info["width"] = img_info["width"]
-
-            label_info["ocr_info"] = []
-
             for doc in document:
-                label_info["ocr_info"].append({
-                    "text": doc["text"],
+                x1, y1, x2, y2 = doc["box"]
+                points = [[x1, y1], [x2, y1], [x2, y2], [x1, y2]]
+                label_info.append({
+                    "transcription": doc["text"],
                     "label": doc["label"],
-                    "bbox": doc["box"],
+                    "points": points,
                     "id": doc["id"],
-                    "linking": doc["linking"],
-                    "words": doc["words"]
+                    "linking": doc["linking"]
                 })
 
             fout.write(image_path + "\t" + json.dumps(

test_tipc/configs/rec_mtb_nrtr/rec_mtb_nrtr.yml

@@ -49,7 +49,7 @@ Architecture:
     
 
 Loss:
-  name: NRTRLoss
+  name: CELoss
   smoothing: True
 
 PostProcess:
@@ -69,7 +69,7 @@ Train:
           img_mode: BGR
           channel_first: False
       - NRTRLabelEncode: # Class handling label
-      - NRTRRecResizeImg:
+      - GrayRecResizeImg:
           image_shape: [100, 32]
           resize_type: PIL # PIL or OpenCV
       - KeepKeys:
@@ -90,7 +90,7 @@ Eval:
           img_mode: BGR
           channel_first: False
       - NRTRLabelEncode: # Class handling label
-      - NRTRRecResizeImg:
+      - GrayRecResizeImg:
           image_shape: [100, 32]
           resize_type: PIL # PIL or OpenCV
       - KeepKeys:
@@ -99,5 +99,5 @@ Eval:
     shuffle: False
     drop_last: False
     batch_size_per_card: 256
-    num_workers: 1
+    num_workers: 4
     use_shared_memory: False

test_tipc/configs/rec_r45_abinet/rec_r45_abinet.yml

+Global:
+  use_gpu: True
+  epoch_num: 10
+  log_smooth_window: 20
+  print_batch_step: 10
+  save_model_dir: ./output/rec/r45_abinet/
+  save_epoch_step: 1
+  # evaluation is run every 2000 iterations
+  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
+  max_text_length: 25
+  infer_mode: False
+  use_space_char: False
+  save_res_path: ./output/rec/predicts_abinet.txt
+
+Optimizer:
+  name: Adam
+  beta1: 0.9
+  beta2: 0.99
+  clip_norm: 20.0
+  lr:
+    name: Piecewise
+    decay_epochs: [6]
+    values: [0.0001, 0.00001] 
+  regularizer:
+    name: 'L2'
+    factor: 0.
+
+Architecture:
+  model_type: rec
+  algorithm: ABINet
+  in_channels: 3
+  Transform:
+  Backbone:
+    name: ResNet45
+
+  Head:
+    name: ABINetHead
+    use_lang: True
+    iter_size: 3
+    
+
+Loss:
+  name: CELoss
+  ignore_index: &ignore_index 100 # Must be greater than the number of character classes
+
+PostProcess:
+  name: ABINetLabelDecode
+
+Metric:
+  name: RecMetric
+  main_indicator: acc
+
+Train:
+  dataset:
+    name: SimpleDataSet
+    data_dir: ./train_data/ic15_data/
+    label_file_list: ["./train_data/ic15_data/rec_gt_train.txt"]
+    transforms:
+      - DecodeImage: # load image
+          img_mode: RGB
+          channel_first: False
+      - ABINetRecAug:
+      - ABINetLabelEncode: # Class handling label
+          ignore_index: *ignore_index
+      - ABINetRecResizeImg:
+          image_shape: [3, 32, 128]
+          padding: False
+      - KeepKeys:
+          keep_keys: ['image', 'label', 'length'] # dataloader will return list in this order
+  loader:
+    shuffle: True
+    batch_size_per_card: 96
+    drop_last: True
+    num_workers: 4
+
+Eval:
+  dataset:
+    name: SimpleDataSet
+    data_dir: ./train_data/ic15_data
+    label_file_list: ["./train_data/ic15_data/rec_gt_test.txt"]
+    transforms:
+      - DecodeImage: # load image
+          img_mode: RGB
+          channel_first: False
+      - ABINetLabelEncode: # Class handling label
+          ignore_index: *ignore_index
+      - ABINetRecResizeImg:
+          image_shape: [3, 32, 128]
+          padding: False
+      - KeepKeys:
+          keep_keys: ['image', 'label', 'length'] # dataloader will return list in this order
+  loader:
+    shuffle: False
+    drop_last: False
+    batch_size_per_card: 256
+    num_workers: 4
+    use_shared_memory: False

test_tipc/configs/rec_r45_abinet/train_infer_python.txt

+===========================train_params===========================
+model_name:rec_abinet
+python:python3.7
+gpu_list:0|0,1
+Global.use_gpu:True|True
+Global.auto_cast:null
+Global.epoch_num:lite_train_lite_infer=2|whole_train_whole_infer=300
+Global.save_model_dir:./output/
+Train.loader.batch_size_per_card:lite_train_lite_infer=16|whole_train_whole_infer=64
+Global.pretrained_model:null
+train_model_name:latest
+train_infer_img_dir:./inference/rec_inference
+null:null
+##
+trainer:norm_train
+norm_train:tools/train.py -c test_tipc/configs/rec_r45_abinet/rec_r45_abinet.yml -o
+pact_train:null
+fpgm_train:null
+distill_train:null
+null:null
+null:null
+##
+===========================eval_params===========================
+eval:tools/eval.py -c test_tipc/configs/rec_r45_abinet/rec_r45_abinet.yml -o
+null:null
+##
+===========================infer_params===========================
+Global.save_inference_dir:./output/
+Global.checkpoints:
+norm_export:tools/export_model.py -c test_tipc/configs/rec_r45_abinet/rec_r45_abinet.yml -o
+quant_export:null
+fpgm_export:null
+distill_export:null
+export1:null
+export2:null
+##
+train_model:./inference/rec_r45_abinet_train/best_accuracy
+infer_export:tools/export_model.py -c test_tipc/configs/rec_r45_abinet/rec_r45_abinet.yml -o
+infer_quant:False
+inference:tools/infer/predict_rec.py --rec_char_dict_path=./ppocr/utils/ic15_dict.txt --rec_image_shape="3,32,128" --rec_algorithm="ABINet"
+--use_gpu:True|False
+--enable_mkldnn:False
+--cpu_threads:6
+--rec_batch_num:1|6
+--use_tensorrt:False
+--precision:fp32
+--rec_model_dir:
+--image_dir:./inference/rec_inference
+--save_log_path:./test/output/
+--benchmark:True
+null:null
+===========================infer_benchmark_params==========================
+random_infer_input:[{float32,[3,32,128]}]

test_tipc/configs/rec_svtrnet/rec_svtrnet.yml

+Global:
+  use_gpu: True
+  epoch_num: 20
+  log_smooth_window: 20
+  print_batch_step: 10
+  save_model_dir: ./output/rec/svtr/
+  save_epoch_step: 1
+  # evaluation is run every 2000 iterations after the 0th 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
+  max_text_length: 25
+  infer_mode: False
+  use_space_char: False
+  save_res_path: ./output/rec/predicts_svtr_tiny.txt
+
+
+Optimizer:
+  name: AdamW
+  beta1: 0.9
+  beta2: 0.99
+  epsilon: 8.e-8
+  weight_decay: 0.05
+  no_weight_decay_name: norm pos_embed
+  one_dim_param_no_weight_decay: true
+  lr:
+    name: Cosine
+    learning_rate: 0.0005
+    warmup_epoch: 2
+
+Architecture:
+  model_type: rec
+  algorithm: SVTR
+  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: SVTRNet
+    img_size: [32, 100]
+    out_char_num: 25
+    out_channels: 192
+    patch_merging: 'Conv'
+    embed_dim: [64, 128, 256]
+    depth: [3, 6, 3]
+    num_heads: [2, 4, 8]
+    mixer: ['Local','Local','Local','Local','Local','Local','Global','Global','Global','Global','Global','Global']
+    local_mixer: [[7, 11], [7, 11], [7, 11]]
+    last_stage: True
+    prenorm: false
+  Neck:
+    name: SequenceEncoder
+    encoder_type: reshape
+  Head:
+    name: CTCHead
+
+Loss:
+  name: CTCLoss
+
+PostProcess:
+  name: CTCLabelDecode
+
+Metric:
+  name: RecMetric
+  main_indicator: acc
+
+Train:
+  dataset:
+    name: SimpleDataSet
+    data_dir: ./train_data/ic15_data/
+    label_file_list: ["./train_data/ic15_data/rec_gt_train.txt"]
+    transforms:
+      - DecodeImage: # load image
+          img_mode: BGR
+          channel_first: False
+      - CTCLabelEncode: # Class handling label
+      - SVTRRecResizeImg:
+          image_shape: [3, 64, 256]
+          padding: False
+      - KeepKeys:
+          keep_keys: ['image', 'label', 'length'] # dataloader will return list in this order
+  loader:
+    shuffle: True
+    batch_size_per_card: 512
+    drop_last: True
+    num_workers: 4
+
+Eval:
+  dataset:
+    name: SimpleDataSet
+    data_dir: ./train_data/ic15_data
+    label_file_list: ["./train_data/ic15_data/rec_gt_test.txt"]
+    transforms:
+      - DecodeImage: # load image
+          img_mode: BGR
+          channel_first: False
+      - CTCLabelEncode: # Class handling label
+      - SVTRRecResizeImg:
+          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: False
+    batch_size_per_card: 256
+    num_workers: 2

test_tipc/configs/rec_svtrnet/train_infer_python.txt

+===========================train_params===========================
+model_name:rec_svtrnet
+python:python3.7
+gpu_list:0|0,1
+Global.use_gpu:True|True
+Global.auto_cast:null
+Global.epoch_num:lite_train_lite_infer=2|whole_train_whole_infer=300
+Global.save_model_dir:./output/
+Train.loader.batch_size_per_card:lite_train_lite_infer=16|whole_train_whole_infer=64
+Global.pretrained_model:null
+train_model_name:latest
+train_infer_img_dir:./inference/rec_inference
+null:null
+##
+trainer:norm_train
+norm_train:tools/train.py -c test_tipc/configs/rec_svtrnet/rec_svtrnet.yml -o
+pact_train:null
+fpgm_train:null
+distill_train:null
+null:null
+null:null
+##
+===========================eval_params===========================
+eval:tools/eval.py -c test_tipc/configs/rec_svtrnet/rec_svtrnet.yml -o
+null:null
+##
+===========================infer_params===========================
+Global.save_inference_dir:./output/
+Global.checkpoints:
+norm_export:tools/export_model.py -c test_tipc/configs/rec_svtrnet/rec_svtrnet.yml -o
+quant_export:null
+fpgm_export:null
+distill_export:null
+export1:null
+export2:null
+##
+train_model:./inference/rec_svtrnet_train/best_accuracy
+infer_export:tools/export_model.py -c test_tipc/configs/rec_svtrnet/rec_svtrnet.yml -o
+infer_quant:False
+inference:tools/infer/predict_rec.py --rec_char_dict_path=./ppocr/utils/ic15_dict.txt --rec_image_shape="3,64,256" --rec_algorithm="SVTR"
+--use_gpu:True|False
+--enable_mkldnn:False
+--cpu_threads:6
+--rec_batch_num:1|6
+--use_tensorrt:False
+--precision:fp32
+--rec_model_dir:
+--image_dir:./inference/rec_inference
+--save_log_path:./test/output/
+--benchmark:True
+null:null
+===========================infer_benchmark_params==========================
+random_infer_input:[{float32,[3,64,256]}]

test_tipc/configs/rec_vitstr_none_ce/rec_vitstr_none_ce.yml

+Global:
+  use_gpu: True
+  epoch_num: 20
+  log_smooth_window: 20
+  print_batch_step: 10
+  save_model_dir: ./output/rec/vitstr_none_ce/
+  save_epoch_step: 1
+  # evaluation is run every 2000 iterations after the 0th 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: ppocr/utils/EN_symbol_dict.txt
+  max_text_length: 25
+  infer_mode: False
+  use_space_char: False
+  save_res_path: ./output/rec/predicts_vitstr.txt
+
+
+Optimizer:
+  name: Adadelta
+  epsilon: 1.e-8
+  rho: 0.95
+  clip_norm: 5.0
+  lr:
+    learning_rate: 1.0
+
+Architecture:
+  model_type: rec
+  algorithm: ViTSTR
+  in_channels: 1
+  Transform:
+  Backbone:
+    name: ViTSTR
+  Neck:
+    name: SequenceEncoder
+    encoder_type: reshape
+  Head:
+    name: CTCHead
+
+Loss:
+  name: CELoss
+  smoothing: False
+  with_all: True
+  ignore_index: &ignore_index 0 # Must be zero or greater than the number of character classes
+
+PostProcess:
+  name: ViTSTRLabelDecode
+
+Metric:
+  name: RecMetric
+  main_indicator: acc
+
+Train:
+  dataset:
+    name: SimpleDataSet
+    data_dir: ./train_data/ic15_data/
+    label_file_list: ["./train_data/ic15_data/rec_gt_train.txt"]
+    transforms:
+      - DecodeImage: # load image
+          img_mode: BGR
+          channel_first: False
+      - ViTSTRLabelEncode: # Class handling label
+          ignore_index: *ignore_index
+      - GrayRecResizeImg:
+          image_shape: [224, 224] # W H
+          resize_type: PIL # PIL or OpenCV
+          inter_type: 'Image.BICUBIC'
+          scale: false
+      - KeepKeys:
+          keep_keys: ['image', 'label', 'length'] # dataloader will return list in this order
+  loader:
+    shuffle: True
+    batch_size_per_card: 48
+    drop_last: True
+    num_workers: 8
+
+Eval:
+  dataset:
+    name: SimpleDataSet
+    data_dir: ./train_data/ic15_data
+    label_file_list: ["./train_data/ic15_data/rec_gt_test.txt"]
+    transforms:
+      - DecodeImage: # load image
+          img_mode: BGR
+          channel_first: False
+      - ViTSTRLabelEncode: # Class handling label
+          ignore_index: *ignore_index
+      - GrayRecResizeImg:
+          image_shape: [224, 224] # W H
+          resize_type: PIL # PIL or OpenCV
+          inter_type: 'Image.BICUBIC'
+          scale: false
+      - KeepKeys:
+          keep_keys: ['image', 'label', 'length'] # dataloader will return list in this order
+  loader:
+    shuffle: False
+    drop_last: False
+    batch_size_per_card: 256
+    num_workers: 2

test_tipc/configs/rec_vitstr_none_ce/train_infer_python.txt

+===========================train_params===========================
+model_name:rec_vitstr
+python:python3.7
+gpu_list:0|0,1
+Global.use_gpu:True|True
+Global.auto_cast:null
+Global.epoch_num:lite_train_lite_infer=2|whole_train_whole_infer=300
+Global.save_model_dir:./output/
+Train.loader.batch_size_per_card:lite_train_lite_infer=16|whole_train_whole_infer=64
+Global.pretrained_model:null
+train_model_name:latest
+train_infer_img_dir:./inference/rec_inference
+null:null
+##
+trainer:norm_train
+norm_train:tools/train.py -c test_tipc/configs/rec_vitstr_none_ce/rec_vitstr_none_ce.yml -o
+pact_train:null
+fpgm_train:null
+distill_train:null
+null:null
+null:null
+##
+===========================eval_params===========================
+eval:tools/eval.py -c test_tipc/configs/rec_vitstr_none_ce/rec_vitstr_none_ce.yml -o
+null:null
+##
+===========================infer_params===========================
+Global.save_inference_dir:./output/
+Global.checkpoints:
+norm_export:tools/export_model.py -c test_tipc/configs/rec_vitstr_none_ce/rec_vitstr_none_ce.yml -o
+quant_export:null
+fpgm_export:null
+distill_export:null
+export1:null
+export2:null
+##
+train_model:./inference/rec_vitstr_none_ce_train/best_accuracy
+infer_export:tools/export_model.py -c test_tipc/configs/rec_vitstr_none_ce/rec_vitstr_none_ce.yml -o
+infer_quant:False
+inference:tools/infer/predict_rec.py --rec_char_dict_path=./ppocr/utils/EN_symbol_dict.txt --rec_image_shape="1,224,224" --rec_algorithm="ViTSTR"
+--use_gpu:True|False
+--enable_mkldnn:False
+--cpu_threads:6
+--rec_batch_num:1|6
+--use_tensorrt:False
+--precision:fp32
+--rec_model_dir:
+--image_dir:./inference/rec_inference
+--save_log_path:./test/output/
+--benchmark:True
+null:null
+===========================infer_benchmark_params==========================
+random_infer_input:[{float32,[1,224,224]}]

test_tipc/test_serving_infer_cpp.sh

@@ -87,11 +87,12 @@ function func_serving(){
     set_image_dir=$(func_set_params "${image_dir_key}" "${image_dir_value}")
     python_list=(${python_list})
     cd ${serving_dir_value}
+    
     # cpp serving
     for gpu_id in ${gpu_value[*]}; do
         if [ ${gpu_id} = "null" ]; then
             server_log_path="${LOG_PATH}/cpp_server_cpu.log"
-            web_service_cpp_cmd="${python_list[0]} ${web_service_py} --model ${det_server_value} ${rec_server_value} ${op_key} ${op_value} ${port_key} ${port_value} > ${server_log_path} 2>&1 "
+            web_service_cpp_cmd="nohup ${python_list[0]} ${web_service_py} --model ${det_server_value} ${rec_server_value} ${op_key} ${op_value} ${port_key} ${port_value} > ${server_log_path} 2>&1 &"
             eval $web_service_cpp_cmd
             last_status=${PIPESTATUS[0]}
             status_check $last_status "${web_service_cpp_cmd}" "${status_log}" "${model_name}"
@@ -105,7 +106,7 @@ function func_serving(){
             ps ux | grep -i ${port_value} | awk '{print $2}' | xargs kill -s 9
         else
             server_log_path="${LOG_PATH}/cpp_server_gpu.log"
-            web_service_cpp_cmd="${python_list[0]} ${web_service_py} --model ${det_server_value} ${rec_server_value} ${op_key} ${op_value} ${port_key} ${port_value} ${gpu_key} ${gpu_id} > ${server_log_path} 2>&1 "
+            web_service_cpp_cmd="nohup ${python_list[0]} ${web_service_py} --model ${det_server_value} ${rec_server_value} ${op_key} ${op_value} ${port_key} ${port_value} ${gpu_key} ${gpu_id} > ${server_log_path} 2>&1 &"
             eval $web_service_cpp_cmd
             sleep 5s
             _save_log_path="${LOG_PATH}/cpp_client_gpu.log"

test_tipc/test_serving_infer_python.sh

@@ -112,7 +112,7 @@ function func_serving(){
     
     cd ${serving_dir_value}
     python=${python_list[0]}
-        
+    
     # python serving
     for use_gpu in ${web_use_gpu_list[*]}; do
         if [ ${use_gpu} = "null" ]; then
@@ -123,19 +123,19 @@ function func_serving(){
                     if [ ${model_name} = "ch_PP-OCRv2" ] || [ ${model_name} = "ch_PP-OCRv3" ] || [ ${model_name} = "ch_ppocr_mobile_v2.0" ] || [ ${model_name} = "ch_ppocr_server_v2.0" ]; then
                         set_det_model_config=$(func_set_params "${det_server_key}" "${det_server_value}")
                         set_rec_model_config=$(func_set_params "${rec_server_key}" "${rec_server_value}")
-                        web_service_cmd="${python} ${web_service_py} ${web_use_gpu_key}="" ${web_use_mkldnn_key}=${use_mkldnn} ${set_cpu_threads} ${set_det_model_config} ${set_rec_model_config} > ${server_log_path} 2>&1 "
+                        web_service_cmd="nohup ${python} ${web_service_py} ${web_use_gpu_key}="" ${web_use_mkldnn_key}=${use_mkldnn} ${set_cpu_threads} ${set_det_model_config} ${set_rec_model_config} > ${server_log_path} 2>&1 &"
                         eval $web_service_cmd
                         last_status=${PIPESTATUS[0]}
                         status_check $last_status "${web_service_cmd}" "${status_log}" "${model_name}"
                     elif [[ ${model_name} =~ "det" ]]; then
                         set_det_model_config=$(func_set_params "${det_server_key}" "${det_server_value}")
-                        web_service_cmd="${python} ${web_service_py} ${web_use_gpu_key}="" ${web_use_mkldnn_key}=${use_mkldnn} ${set_cpu_threads} ${set_det_model_config} > ${server_log_path} 2>&1 "
+                        web_service_cmd="nohup ${python} ${web_service_py} ${web_use_gpu_key}="" ${web_use_mkldnn_key}=${use_mkldnn} ${set_cpu_threads} ${set_det_model_config} > ${server_log_path} 2>&1 &"
                         eval $web_service_cmd
                         last_status=${PIPESTATUS[0]}
                         status_check $last_status "${web_service_cmd}" "${status_log}" "${model_name}"
                     elif [[ ${model_name} =~ "rec" ]]; then
                         set_rec_model_config=$(func_set_params "${rec_server_key}" "${rec_server_value}")
-                        web_service_cmd="${python} ${web_service_py} ${web_use_gpu_key}="" ${web_use_mkldnn_key}=${use_mkldnn} ${set_cpu_threads} ${set_rec_model_config} > ${server_log_path} 2>&1 "
+                        web_service_cmd="nohup ${python} ${web_service_py} ${web_use_gpu_key}="" ${web_use_mkldnn_key}=${use_mkldnn} ${set_cpu_threads} ${set_rec_model_config} > ${server_log_path} 2>&1 &"
                         eval $web_service_cmd
                         last_status=${PIPESTATUS[0]}
                         status_check $last_status "${web_service_cmd}" "${status_log}" "${model_name}"
@@ -174,19 +174,19 @@ function func_serving(){
                     if [ ${model_name} = "ch_PP-OCRv2" ] || [ ${model_name} = "ch_PP-OCRv3" ] || [ ${model_name} = "ch_ppocr_mobile_v2.0" ] || [ ${model_name} = "ch_ppocr_server_v2.0" ]; then
                         set_det_model_config=$(func_set_params "${det_server_key}" "${det_server_value}")
                         set_rec_model_config=$(func_set_params "${rec_server_key}" "${rec_server_value}")
-                        web_service_cmd="${python} ${web_service_py} ${set_tensorrt} ${set_precision} ${set_det_model_config} ${set_rec_model_config} > ${server_log_path} 2>&1 "
+                        web_service_cmd="nohup ${python} ${web_service_py} ${set_tensorrt} ${set_precision} ${set_det_model_config} ${set_rec_model_config} > ${server_log_path} 2>&1 &"
                         eval $web_service_cmd
                         last_status=${PIPESTATUS[0]}
                         status_check $last_status "${web_service_cmd}" "${status_log}" "${model_name}"
                     elif [[ ${model_name} =~ "det" ]]; then
                         set_det_model_config=$(func_set_params "${det_server_key}" "${det_server_value}")
-                        web_service_cmd="${python} ${web_service_py} ${set_tensorrt} ${set_precision} ${set_det_model_config} > ${server_log_path} 2>&1 "
+                        web_service_cmd="nohup ${python} ${web_service_py} ${set_tensorrt} ${set_precision} ${set_det_model_config} > ${server_log_path} 2>&1 &"
                         eval $web_service_cmd
                         last_status=${PIPESTATUS[0]}
                         status_check $last_status "${web_service_cmd}" "${status_log}" "${model_name}"
                     elif [[ ${model_name} =~ "rec" ]]; then
                         set_rec_model_config=$(func_set_params "${rec_server_key}" "${rec_server_value}")
-                        web_service_cmd="${python} ${web_service_py} ${set_tensorrt} ${set_precision} ${set_rec_model_config} > ${server_log_path} 2>&1 "
+                        web_service_cmd="nohup ${python} ${web_service_py} ${set_tensorrt} ${set_precision} ${set_rec_model_config} > ${server_log_path} 2>&1 &"
                         eval $web_service_cmd
                         last_status=${PIPESTATUS[0]}
                         status_check $last_status "${web_service_cmd}" "${status_log}" "${model_name}"

tools/export_model.py

@@ -31,7 +31,12 @@ from ppocr.utils.logging import get_logger
 from tools.program import load_config, merge_config, ArgsParser
 
 
-def export_single_model(model, arch_config, save_path, logger, quanter=None):
+def export_single_model(model,
+                        arch_config,
+                        save_path,
+                        logger,
+                        input_shape=None,
+                        quanter=None):
     if arch_config["algorithm"] == "SRN":
         max_text_length = arch_config["Head"]["max_text_length"]
         other_shape = [
@@ -64,7 +69,7 @@ def export_single_model(model, arch_config, save_path, logger, quanter=None):
         else:
             other_shape = [
                 paddle.static.InputSpec(
-                    shape=[None, 3, 64, 256], dtype="float32"),
+                    shape=[None] + input_shape, dtype="float32"),
             ]
         model = to_static(model, input_spec=other_shape)
     elif arch_config["algorithm"] == "PREN":
@@ -73,6 +78,25 @@ def export_single_model(model, arch_config, save_path, logger, quanter=None):
                 shape=[None, 3, 64, 512], dtype="float32"),
         ]
         model = to_static(model, input_spec=other_shape)
+    elif arch_config["algorithm"] == "ViTSTR":
+        other_shape = [
+            paddle.static.InputSpec(
+                shape=[None, 1, 224, 224], dtype="float32"),
+        ]
+        model = to_static(model, input_spec=other_shape)
+    elif arch_config["algorithm"] == "ABINet":
+        other_shape = [
+            paddle.static.InputSpec(
+                shape=[None, 3, 32, 128], dtype="float32"),
+        ]
+        # print([None, 3, 32, 128])
+        model = to_static(model, input_spec=other_shape)
+    elif arch_config["algorithm"] == "NRTR":
+        other_shape = [
+            paddle.static.InputSpec(
+                shape=[None, 1, 32, 100], dtype="float32"),
+        ]
+        model = to_static(model, input_spec=other_shape)
     else:
         infer_shape = [3, -1, -1]
         if arch_config["model_type"] == "rec":
@@ -84,8 +108,6 @@ def export_single_model(model, arch_config, save_path, logger, quanter=None):
                     "When there is tps in the network, variable length input is not supported, and the input size needs to be the same as during training"
                 )
                 infer_shape[-1] = 100
-            if arch_config["algorithm"] == "NRTR":
-                infer_shape = [1, 32, 100]
         elif arch_config["model_type"] == "table":
             infer_shape = [3, 488, 488]
         model = to_static(
@@ -157,6 +179,13 @@ def main():
 
     arch_config = config["Architecture"]
 
+    if arch_config["algorithm"] == "SVTR" and arch_config["Head"][
+            "name"] != 'MultiHead':
+        input_shape = config["Eval"]["dataset"]["transforms"][-2][
+            'SVTRRecResizeImg']['image_shape']
+    else:
+        input_shape = None
+
     if arch_config["algorithm"] in ["Distillation", ]:  # distillation model
         archs = list(arch_config["Models"].values())
         for idx, name in enumerate(model.model_name_list):
@@ -165,7 +194,8 @@ def main():
                                 sub_model_save_path, logger)
     else:
         save_path = os.path.join(save_path, "inference")
-        export_single_model(model, arch_config, save_path, logger)
+        export_single_model(
+            model, arch_config, save_path, logger, input_shape=input_shape)
 
 
 if __name__ == "__main__":

tools/infer/predict_rec.py

@@ -69,6 +69,18 @@ class TextRecognizer(object):
                 "character_dict_path": args.rec_char_dict_path,
                 "use_space_char": args.use_space_char
             }
+        elif self.rec_algorithm == 'ViTSTR':
+            postprocess_params = {
+                'name': 'ViTSTRLabelDecode',
+                "character_dict_path": args.rec_char_dict_path,
+                "use_space_char": args.use_space_char
+            }
+        elif self.rec_algorithm == 'ABINet':
+            postprocess_params = {
+                'name': 'ABINetLabelDecode',
+                "character_dict_path": args.rec_char_dict_path,
+                "use_space_char": args.use_space_char
+            }
         self.postprocess_op = build_post_process(postprocess_params)
         self.predictor, self.input_tensor, self.output_tensors, self.config = \
             utility.create_predictor(args, 'rec', logger)
@@ -96,15 +108,22 @@ class TextRecognizer(object):
 
     def resize_norm_img(self, img, max_wh_ratio):
         imgC, imgH, imgW = self.rec_image_shape
-        if self.rec_algorithm == 'NRTR':
+        if self.rec_algorithm == 'NRTR' or self.rec_algorithm == 'ViTSTR':
             img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
             # return padding_im
             image_pil = Image.fromarray(np.uint8(img))
-            img = image_pil.resize([100, 32], Image.ANTIALIAS)
+            if self.rec_algorithm == 'ViTSTR':
+                img = image_pil.resize([imgW, imgH], Image.BICUBIC)
+            else:
+                img = image_pil.resize([imgW, imgH], Image.ANTIALIAS)
             img = np.array(img)
             norm_img = np.expand_dims(img, -1)
             norm_img = norm_img.transpose((2, 0, 1))
-            return norm_img.astype(np.float32) / 128. - 1.
+            if self.rec_algorithm == 'ViTSTR':
+                norm_img = norm_img.astype(np.float32) / 255.
+            else:
+                norm_img = norm_img.astype(np.float32) / 128. - 1.
+            return norm_img
 
         assert imgC == img.shape[2]
         imgW = int((imgH * max_wh_ratio))
@@ -132,17 +151,6 @@ class TextRecognizer(object):
         padding_im[:, :, 0:resized_w] = resized_image
         return padding_im
 
-    def resize_norm_img_svtr(self, img, image_shape):
-
-        imgC, imgH, imgW = image_shape
-        resized_image = cv2.resize(
-            img, (imgW, imgH), interpolation=cv2.INTER_LINEAR)
-        resized_image = resized_image.astype('float32')
-        resized_image = resized_image.transpose((2, 0, 1)) / 255
-        resized_image -= 0.5
-        resized_image /= 0.5
-        return resized_image
-
     def resize_norm_img_srn(self, img, image_shape):
         imgC, imgH, imgW = image_shape
 
@@ -250,6 +258,35 @@ class TextRecognizer(object):
 
         return padding_im, resize_shape, pad_shape, valid_ratio
 
+    def resize_norm_img_svtr(self, img, image_shape):
+
+        imgC, imgH, imgW = image_shape
+        resized_image = cv2.resize(
+            img, (imgW, imgH), interpolation=cv2.INTER_LINEAR)
+        resized_image = resized_image.astype('float32')
+        resized_image = resized_image.transpose((2, 0, 1)) / 255
+        resized_image -= 0.5
+        resized_image /= 0.5
+        return resized_image
+
+    def resize_norm_img_abinet(self, img, image_shape):
+
+        imgC, imgH, imgW = image_shape
+
+        resized_image = cv2.resize(
+            img, (imgW, imgH), interpolation=cv2.INTER_LINEAR)
+        resized_image = resized_image.astype('float32')
+        resized_image = resized_image / 255.
+
+        mean = np.array([0.485, 0.456, 0.406])
+        std = np.array([0.229, 0.224, 0.225])
+        resized_image = (
+            resized_image - mean[None, None, ...]) / std[None, None, ...]
+        resized_image = resized_image.transpose((2, 0, 1))
+        resized_image = resized_image.astype('float32')
+
+        return resized_image
+
     def __call__(self, img_list):
         img_num = len(img_list)
         # Calculate the aspect ratio of all text bars
@@ -300,6 +337,11 @@ class TextRecognizer(object):
                                                          self.rec_image_shape)
                     norm_img = norm_img[np.newaxis, :]
                     norm_img_batch.append(norm_img)
+                elif self.rec_algorithm == "ABINet":
+                    norm_img = self.resize_norm_img_abinet(
+                        img_list[indices[ino]], self.rec_image_shape)
+                    norm_img = norm_img[np.newaxis, :]
+                    norm_img_batch.append(norm_img)
                 else:
                     norm_img = self.resize_norm_img(img_list[indices[ino]],
                                                     max_wh_ratio)

tools/infer_kie.py

@@ -39,13 +39,12 @@ import time
 
 
 def read_class_list(filepath):
-    dict = {}
+    ret = {}
     with open(filepath, "r") as f:
         lines = f.readlines()
-        for line in lines:
-            key, value = line.split(" ")
-            dict[key] = value.rstrip()
-    return dict
+        for idx, line in enumerate(lines):
+            ret[idx] = line.strip("\n")
+    return ret
 
 
 def draw_kie_result(batch, node, idx_to_cls, count):
@@ -71,7 +70,7 @@ def draw_kie_result(batch, node, idx_to_cls, count):
         x_min = int(min([point[0] for point in new_box]))
         y_min = int(min([point[1] for point in new_box]))
 
-        pred_label = str(node_pred_label[i])
+        pred_label = node_pred_label[i]
         if pred_label in idx_to_cls:
             pred_label = idx_to_cls[pred_label]
         pred_score = '{:.2f}'.format(node_pred_score[i])
@@ -109,8 +108,7 @@ def main():
     save_res_path = config['Global']['save_res_path']
     class_path = config['Global']['class_path']
     idx_to_cls = read_class_list(class_path)
-    if not os.path.exists(os.path.dirname(save_res_path)):
-        os.makedirs(os.path.dirname(save_res_path))
+    os.makedirs(os.path.dirname(save_res_path), exist_ok=True)
 
     model.eval()
 

tools/infer_vqa_token_ser.py

@@ -86,15 +86,16 @@ class SerPredictor(object):
                 ]
 
             transforms.append(op)
-        global_config['infer_mode'] = True
+        if config["Global"].get("infer_mode", None) is None:
+            global_config['infer_mode'] = True
         self.ops = create_operators(config['Eval']['dataset']['transforms'],
                                     global_config)
         self.model.eval()
 
-    def __call__(self, img_path):
-        with open(img_path, 'rb') as f:
+    def __call__(self, data):
+        with open(data["img_path"], 'rb') as f:
             img = f.read()
-            data = {'image': img}
+        data["image"] = img
         batch = transform(data, self.ops)
         batch = to_tensor(batch)
         preds = self.model(batch)
@@ -112,20 +113,35 @@ if __name__ == '__main__':
 
     ser_engine = SerPredictor(config)
 
-    infer_imgs = get_image_file_list(config['Global']['infer_img'])
+    if config["Global"].get("infer_mode", None) is False:
+        data_dir = config['Eval']['dataset']['data_dir']
+        with open(config['Global']['infer_img'], "rb") as f:
+            infer_imgs = f.readlines()
+    else:
+        infer_imgs = get_image_file_list(config['Global']['infer_img'])
+
     with open(
             os.path.join(config['Global']['save_res_path'],
                          "infer_results.txt"),
             "w",
             encoding='utf-8') as fout:
-        for idx, img_path in enumerate(infer_imgs):
+        for idx, info in enumerate(infer_imgs):
+            if config["Global"].get("infer_mode", None) is False:
+                data_line = info.decode('utf-8')
+                substr = data_line.strip("\n").split("\t")
+                img_path = os.path.join(data_dir, substr[0])
+                data = {'img_path': img_path, 'label': substr[1]}
+            else:
+                img_path = info
+                data = {'img_path': img_path}
+
             save_img_path = os.path.join(
                 config['Global']['save_res_path'],
                 os.path.splitext(os.path.basename(img_path))[0] + "_ser.jpg")
             logger.info("process: [{}/{}], save result to {}".format(
                 idx, len(infer_imgs), save_img_path))
 
-            result, _ = ser_engine(img_path)
+            result, _ = ser_engine(data)
             result = result[0]
             fout.write(img_path + "\t" + json.dumps(
                 {

tools/program.py

@@ -255,6 +255,8 @@ def train(config,
                 with paddle.amp.auto_cast():
                     if model_type == 'table' or extra_input:
                         preds = model(images, data=batch[1:])
+                    elif model_type in ["kie", 'vqa']:
+                        preds = model(batch)
                     else:
                         preds = model(images)
             else: