🎉 Initial commit.

.gitignore

@@ -0,0 +1,3 @@
+.idea/
+__pycache__/
+runs/

LICENSE

@@ -0,0 +1,165 @@
+                   GNU LESSER GENERAL PUBLIC LICENSE
+                       Version 3, 29 June 2007
+
+ Copyright (C) 2007 Free Software Foundation, Inc. <http://fsf.org/>
+ Everyone is permitted to copy and distribute verbatim copies
+ of this license document, but changing it is not allowed.
+
+
+  This version of the GNU Lesser General Public License incorporates
+the terms and conditions of version 3 of the GNU General Public
+License, supplemented by the additional permissions listed below.
+
+  0. Additional Definitions.
+
+  As used herein, "this License" refers to version 3 of the GNU Lesser
+General Public License, and the "GNU GPL" refers to version 3 of the GNU
+General Public License.
+
+  "The Library" refers to a covered work governed by this License,
+other than an Application or a Combined Work as defined below.
+
+  An "Application" is any work that makes use of an interface provided
+by the Library, but which is not otherwise based on the Library.
+Defining a subclass of a class defined by the Library is deemed a mode
+of using an interface provided by the Library.
+
+  A "Combined Work" is a work produced by combining or linking an
+Application with the Library.  The particular version of the Library
+with which the Combined Work was made is also called the "Linked
+Version".
+
+  The "Minimal Corresponding Source" for a Combined Work means the
+Corresponding Source for the Combined Work, excluding any source code
+for portions of the Combined Work that, considered in isolation, are
+based on the Application, and not on the Linked Version.
+
+  The "Corresponding Application Code" for a Combined Work means the
+object code and/or source code for the Application, including any data
+and utility programs needed for reproducing the Combined Work from the
+Application, but excluding the System Libraries of the Combined Work.
+
+  1. Exception to Section 3 of the GNU GPL.
+
+  You may convey a covered work under sections 3 and 4 of this License
+without being bound by section 3 of the GNU GPL.
+
+  2. Conveying Modified Versions.
+
+  If you modify a copy of the Library, and, in your modifications, a
+facility refers to a function or data to be supplied by an Application
+that uses the facility (other than as an argument passed when the
+facility is invoked), then you may convey a copy of the modified
+version:
+
+   a) under this License, provided that you make a good faith effort to
+   ensure that, in the event an Application does not supply the
+   function or data, the facility still operates, and performs
+   whatever part of its purpose remains meaningful, or
+
+   b) under the GNU GPL, with none of the additional permissions of
+   this License applicable to that copy.
+
+  3. Object Code Incorporating Material from Library Header Files.
+
+  The object code form of an Application may incorporate material from
+a header file that is part of the Library.  You may convey such object
+code under terms of your choice, provided that, if the incorporated
+material is not limited to numerical parameters, data structure
+layouts and accessors, or small macros, inline functions and templates
+(ten or fewer lines in length), you do both of the following:
+
+   a) Give prominent notice with each copy of the object code that the
+   Library is used in it and that the Library and its use are
+   covered by this License.
+
+   b) Accompany the object code with a copy of the GNU GPL and this license
+   document.
+
+  4. Combined Works.
+
+  You may convey a Combined Work under terms of your choice that,
+taken together, effectively do not restrict modification of the
+portions of the Library contained in the Combined Work and reverse
+engineering for debugging such modifications, if you also do each of
+the following:
+
+   a) Give prominent notice with each copy of the Combined Work that
+   the Library is used in it and that the Library and its use are
+   covered by this License.
+
+   b) Accompany the Combined Work with a copy of the GNU GPL and this license
+   document.
+
+   c) For a Combined Work that displays copyright notices during
+   execution, include the copyright notice for the Library among
+   these notices, as well as a reference directing the user to the
+   copies of the GNU GPL and this license document.
+
+   d) Do one of the following:
+
+       0) Convey the Minimal Corresponding Source under the terms of this
+       License, and the Corresponding Application Code in a form
+       suitable for, and under terms that permit, the user to
+       recombine or relink the Application with a modified version of
+       the Linked Version to produce a modified Combined Work, in the
+       manner specified by section 6 of the GNU GPL for conveying
+       Corresponding Source.
+
+       1) Use a suitable shared library mechanism for linking with the
+       Library.  A suitable mechanism is one that (a) uses at run time
+       a copy of the Library already present on the user's computer
+       system, and (b) will operate properly with a modified version
+       of the Library that is interface-compatible with the Linked
+       Version.
+
+   e) Provide Installation Information, but only if you would otherwise
+   be required to provide such information under section 6 of the
+   GNU GPL, and only to the extent that such information is
+   necessary to install and execute a modified version of the
+   Combined Work produced by recombining or relinking the
+   Application with a modified version of the Linked Version. (If
+   you use option 4d0, the Installation Information must accompany
+   the Minimal Corresponding Source and Corresponding Application
+   Code. If you use option 4d1, you must provide the Installation
+   Information in the manner specified by section 6 of the GNU GPL
+   for conveying Corresponding Source.)
+
+  5. Combined Libraries.
+
+  You may place library facilities that are a work based on the
+Library side by side in a single library together with other library
+facilities that are not Applications and are not covered by this
+License, and convey such a combined library under terms of your
+choice, if you do both of the following:
+
+   a) Accompany the combined library with a copy of the same work based
+   on the Library, uncombined with any other library facilities,
+   conveyed under the terms of this License.
+
+   b) Give prominent notice with the combined library that part of it
+   is a work based on the Library, and explaining where to find the
+   accompanying uncombined form of the same work.
+
+  6. Revised Versions of the GNU Lesser General Public License.
+
+  The Free Software Foundation may publish revised and/or new versions
+of the GNU Lesser General Public License from time to time. Such new
+versions will be similar in spirit to the present version, but may
+differ in detail to address new problems or concerns.
+
+  Each version is given a distinguishing version number. If the
+Library as you received it specifies that a certain numbered version
+of the GNU Lesser General Public License "or any later version"
+applies to it, you have the option of following the terms and
+conditions either of that published version or of any later version
+published by the Free Software Foundation. If the Library as you
+received it does not specify a version number of the GNU Lesser
+General Public License, you may choose any version of the GNU Lesser
+General Public License ever published by the Free Software Foundation.
+
+  If the Library as you received it specifies that a proxy can decide
+whether future versions of the GNU Lesser General Public License shall
+apply, that proxy's public statement of acceptance of any version is
+permanent authorization for you to choose that version for the
+Library.

README.md

@@ -0,0 +1,104 @@
+# STNet + LPRNet
+STNet 进行空间校正 + LPRNet 进行车牌识别
+
+目前在 CCPD 数据集下，各类作为一个整体，以 half 模式运行测试，最终 macc 约为 93.5%
+
+## 使用
+1. 下载并安装 [Conda](https://www.anaconda.com/products/individual)
+2. 运行下列命令安装依赖
+
+    ```sh
+    conda create --name py38 python=3.8
+    conda activate py38  # 建议加到 .bashrc 中，不然每次都要先执行
+
+    conda install -c pytorch pytorch
+    conda install -c conda-forge accimage opencv onnx
+    conda install tqdm tensorboard numpy
+    conda install -c anaconda flask
+    pip install thop
+    ```
+3. 下载 [CCPD](https://github.com/detectRecog/CCPD) 数据集，放到任意位置
+4. 运行下列命令进行训练
+
+    ```sh
+    python train.py --source-dir /data/CCPD2019
+    ```
+5. 运行下列命令进行测试
+
+    ```sh
+    python test.py --source-dir /data/CCPD2019 --weights weights/final.pt
+    ```
+
+    * 如果之前训练时中断了，使用 --weights 加载之前的模型继续训练
+    * 如果内存充足，建议启用 --cache-images 参数来加快训练，这大约需要 5.5GiB 的内存
+    * 过小的 batch-size 会引起不稳定，梯度震荡比较严重
+    * 过大的 batch-size 会引起收敛速度慢，需要更多的 epochs 才能得到同样的精度
+    * 不建议手动指定 --workers
+    * 由于现在 ST 和 LPR 是一起训练的，所以多少还有些玄学的东西，如果到 50 epochs，macc 还没上 0.8，就干脆弃了重炼一炉吧
+6. 运行下列命令进行运行下列命令提供 Web API
+
+    ```sh
+    python web.py --weights weights/final.pt
+    ```
+7. 剪枝
+
+    ```sh
+    python prune.py --source-dir /data/CCPD2019 --weights weights/final.pt
+    ```
+
+    * 效果极差，不建议使用
+7. ONNX
+
+    ```sh
+    python export.py --weights weights/final.pt
+    ```
+
+    * 目前的 pytorch 无法正确的导出 STNet，因此实际上它是不可用的
+
+## Web API
+### 识别图片中的车牌
+> 请先将车牌部分切割出来再提交，否则基本无法识别
+
+* POST /api/ml/st_lpr
+* 请求
+    * img 被识别的图片，采用常规 upload 文件的方式上传
+    * rect 可选，用逗号分割的坐标点列表
+        * 设 r1 为区域 1，r2 为区域 2，依此类推
+        * r1x1,r1y1,r1x2,r1y2,r2x1,r2y1,r2x2,r2y2,r3x1,r3y1,r3x2,r3y2...
+        * 会通过这组坐标从图片中切割出车牌部分
+* 响应
+    * 识别到的车牌内容列表
+
+## 目录结构
+```
+/data 数据加载器
+|-- ccpd.py 用于遍历 source-dir 目录中的 CCPD 数据集，提供基本的数据解析
+|-- dataset.py 基于 CCPD 数据集制作的 Pytorch Dataset 实现
+/model 模型
+|-- lprnet.py LPRNet 实现
+|-- st.py STNet 实现
+/runs
+|-- /api WebAPI 的日志文件
+|-- /cache 缓存文件
+|-- /exp* 训练时的数据输出目录，每次启动会自动自增 + 1
+    |-- weights 输出的模型文件
+        |-- last.pt 目前最新一次完成的 epoch 保存的数据
+        |-- best.pt 训练过程中，测试得到最高 acc 的模型
+        |-- final.pt 跑完所有 epoch 后保存的最终模型文件，注意，不应使用此文件继续训练，请使用 last.pt 继续训练
+/weights
+|-- final.pt 已训练好的模型，方便快速开始测试、产品部署，但不应用其继续训练
+```
+
+## Links
+* [原仓库](https://github.com/xuexingyu24/License_Plate_Detection_Pytorch)
+    * 模型和原始代码来自这里
+* [YOLO5](https://github.com/ultralytics/yolov5)
+    * 代码中大部分改动都照抄的这边
+* [Pytorch Tutorials (ST)](https://pytorch.org/tutorials/intermediate/spatial_transformer_tutorial.html)
+    * 注意，此文写于 pytorch 1.3 发布之前
+* [Pytorch Tutorials (Flask)](https://pytorch.org/tutorials/intermediate/flask_rest_api_tutorial.html)
+* [Pytorch Tutorials (Prune)](https://pytorch.org/tutorials/intermediate/pruning_tutorial.html)
+
+## TODO
+* [ ] 分别训练 STNet 和 LPRNet
+* [ ] 双行车牌支持

data/__init__.py

@@ -0,0 +1,2 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-