[deps] Grouped deps & setup vcs

- Add latex_formatter.py based on tex-fmt (https://github.com/WGUNDERWOOD/tex-fmt)
- Update to_katex.py to use the new formatter
- Enhance LaTeX formula output with better formatting and readability

This integration helps make generated LaTeX formulas more readable and
maintainable by applying consistent formatting rules.

.github/workflows/deploy-doc.yml

@@ -0,0 +1,36 @@
+name: "Sphinx: Render docs"
+
+on: push
+
+jobs:
+  build:
+    runs-on: ubuntu-latest
+    permissions:
+      contents: write
+    steps:
+    - uses: actions/checkout@v4
+      with:
+        persist-credentials: false
+    - name: Set up Python
+      uses: actions/setup-python@v4
+      with:
+        python-version: '3.10'
+    - name: Install uv
+      run: pip install uv
+    - name: Install docs dependencies
+      run: uv pip install --system -e ".[docs]"
+    - name: Build HTML
+      run: |
+        cd docs
+        make html
+    - name: Upload artifacts
+      uses: actions/upload-artifact@v4
+      with:
+        name: html-docs
+        path: docs/build/html/
+    - name: Deploy
+      uses: peaceiris/actions-gh-pages@v3
+      if: github.ref == 'refs/heads/main'
+      with:
+        github_token: ${{ secrets.GITHUB_TOKEN }}
+        publish_dir: docs/build/html/

.github/workflows/pr-welcome.yml

@@ -0,0 +1,37 @@
+name: PR Welcome Bot
+
+on:
+  pull_request:
+    types: [opened]
+
+jobs:
+  welcome:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Post Welcome Comment
+        uses: actions/github-script@v6
+        with:
+          github-token: ${{ secrets.GITHUB_TOKEN }}
+          script: |
+            const prNumber = context.issue.number;
+            const prAuthor = context.payload.pull_request.user.login;
+
+            const welcomeMessage = `
+            👋 Hello @${prAuthor}, thank you for contributing to this project! 🎉
+
+            We've received your Pull Request and the team will review it as soon as possible.
+
+            In the meantime, please ensure:
+            - [ ] Your code follows the project's coding style
+            - [ ] Relevant tests have been added and are passing
+            - [ ] Documentation has been updated if needed
+
+            If you have any questions, feel free to ask here. Happy coding! 😊
+            `;
+
+            await github.rest.issues.createComment({
+              owner: context.repo.owner,
+              repo: context.repo.repo,
+              issue_number: prNumber,
+              body: welcomeMessage
+            });

.github/workflows/publish.yml

@@ -0,0 +1,36 @@
+name: Publish to PyPI
+
+on:
+  push:
+    branches:
+      - 'main'
+    tags:
+      - 'v*'
+
+jobs:
+  publish:
+    runs-on: ubuntu-latest
+
+    steps:
+    - name: Checkout code
+      uses: actions/checkout@v4
+
+    - name: Set up Python
+      uses: actions/setup-python@v5
+      with:
+        python-version: '3.x'
+
+    - name: Install uv
+      run: |
+        curl -LsSf https://astral.sh/uv/install.sh | sh
+        echo "$HOME/.cargo/bin" >> $GITHUB_PATH
+
+    - name: Build package with uv
+      run: |
+        uv build
+
+    - name: Publish to PyPI
+      env:
+        UV_PUBLISH_TOKEN: ${{ secrets.PYPI_TOKEN }}
+      run: |
+        uv publish --token $UV_PUBLISH_TOKEN

.github/workflows/python-lint.yml

@@ -0,0 +1,27 @@
+name: Python Linting
+
+on:
+  push:
+    branches: [main]
+  pull_request:
+    branches: [main]
+
+jobs:
+  lint:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: '3.10'
+          cache: 'pip'
+
+      - name: Install dependencies
+        run: |
+          python -m pip install --upgrade pip
+          pip install pre-commit
+
+      - name: Run pre-commit
+        run: pre-commit run --all-files

.github/workflows/test.yaml

@@ -0,0 +1,35 @@
+name: Run Tests with Pytest
+
+on:
+  push:
+    branches:
+      - main
+  pull_request:
+    branches:
+      - main
+
+jobs:
+  test:
+    runs-on: ubuntu-latest
+
+    steps:
+    - name: Checkout code
+      uses: actions/checkout@v4
+
+    - name: Set up Python
+      uses: actions/setup-python@v5
+      with:
+        python-version: '3.x'
+
+    - name: Install uv
+      run: |
+        curl -LsSf https://astral.sh/uv/install.sh | sh
+        echo "$HOME/.cargo/bin" >> $GITHUB_PATH
+
+    - name: Install dependencies
+      run: |
+        uv sync --group test
+
+    - name: Run tests with pytest
+      run: |
+        uv run pytest tests/

.gitignore

@@ -1,10 +1,349 @@
-**/__pycache__
-**/.vscode
-**/train_result
+# Created by https://www.toptal.com/developers/gitignore/api/macos,visualstudiocode,pycharm,python
+# Edit at https://www.toptal.com/developers/gitignore?templates=macos,visualstudiocode,pycharm,python
 
-**/logs
-**/.cache
-**/tmp*
-**/data
-**/*cache
-**/ckpt
+### macOS ###
+# General
+.DS_Store
+.AppleDouble
+.LSOverride
+
+# Icon must end with two \r
+Icon
+
+
+# Thumbnails
+._*
+
+# Files that might appear in the root of a volume
+.DocumentRevisions-V100
+.fseventsd
+.Spotlight-V100
+.TemporaryItems
+.Trashes
+.VolumeIcon.icns
+.com.apple.timemachine.donotpresent
+
+# Directories potentially created on remote AFP share
+.AppleDB
+.AppleDesktop
+Network Trash Folder
+Temporary Items
+.apdisk
+
+### macOS Patch ###
+# iCloud generated files
+*.icloud
+
+### PyCharm ###
+# Covers JetBrains IDEs: IntelliJ, RubyMine, PhpStorm, AppCode, PyCharm, CLion, Android Studio, WebStorm and Rider
+# Reference: https://intellij-support.jetbrains.com/hc/en-us/articles/206544839
+
+# User-specific stuff
+.idea/**/workspace.xml
+.idea/**/tasks.xml
+.idea/**/usage.statistics.xml
+.idea/**/dictionaries
+.idea/**/shelf
+
+# AWS User-specific
+.idea/**/aws.xml
+
+# Generated files
+.idea/**/contentModel.xml
+
+# Sensitive or high-churn files
+.idea/**/dataSources/
+.idea/**/dataSources.ids
+.idea/**/dataSources.local.xml
+.idea/**/sqlDataSources.xml
+.idea/**/dynamic.xml
+.idea/**/uiDesigner.xml
+.idea/**/dbnavigator.xml
+
+# Gradle
+.idea/**/gradle.xml
+.idea/**/libraries
+
+# Gradle and Maven with auto-import
+# When using Gradle or Maven with auto-import, you should exclude module files,
+# since they will be recreated, and may cause churn.  Uncomment if using
+# auto-import.
+# .idea/artifacts
+# .idea/compiler.xml
+# .idea/jarRepositories.xml
+# .idea/modules.xml
+# .idea/*.iml
+# .idea/modules
+# *.iml
+# *.ipr
+
+# CMake
+cmake-build-*/
+
+# Mongo Explorer plugin
+.idea/**/mongoSettings.xml
+
+# File-based project format
+*.iws
+
+# IntelliJ
+out/
+
+# mpeltonen/sbt-idea plugin
+.idea_modules/
+
+# JIRA plugin
+atlassian-ide-plugin.xml
+
+# Cursive Clojure plugin
+.idea/replstate.xml
+
+# SonarLint plugin
+.idea/sonarlint/
+
+# Crashlytics plugin (for Android Studio and IntelliJ)
+com_crashlytics_export_strings.xml
+crashlytics.properties
+crashlytics-build.properties
+fabric.properties
+
+# Editor-based Rest Client
+.idea/httpRequests
+
+# Android studio 3.1+ serialized cache file
+.idea/caches/build_file_checksums.ser
+
+### PyCharm Patch ###
+# Comment Reason: https://github.com/joeblau/gitignore.io/issues/186#issuecomment-215987721
+
+# *.iml
+# modules.xml
+# .idea/misc.xml
+# *.ipr
+
+# Sonarlint plugin
+# https://plugins.jetbrains.com/plugin/7973-sonarlint
+.idea/**/sonarlint/
+
+# SonarQube Plugin
+# https://plugins.jetbrains.com/plugin/7238-sonarqube-community-plugin
+.idea/**/sonarIssues.xml
+
+# Markdown Navigator plugin
+# https://plugins.jetbrains.com/plugin/7896-markdown-navigator-enhanced
+.idea/**/markdown-navigator.xml
+.idea/**/markdown-navigator-enh.xml
+.idea/**/markdown-navigator/
+
+# Cache file creation bug
+# See https://youtrack.jetbrains.com/issue/JBR-2257
+.idea/$CACHE_FILE$
+
+# CodeStream plugin
+# https://plugins.jetbrains.com/plugin/12206-codestream
+.idea/codestream.xml
+
+# Azure Toolkit for IntelliJ plugin
+# https://plugins.jetbrains.com/plugin/8053-azure-toolkit-for-intellij
+.idea/**/azureSettings.xml
+
+### Python ###
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+cover/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+.pybuilder/
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+#   For a library or package, you might want to ignore these files since the code is
+#   intended to run in multiple environments; otherwise, check them in:
+# .python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# poetry
+#   Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
+#   This is especially recommended for binary packages to ensure reproducibility, and is more
+#   commonly ignored for libraries.
+#   https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
+#poetry.lock
+
+# pdm
+#   Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
+#pdm.lock
+#   pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it
+#   in version control.
+#   https://pdm.fming.dev/#use-with-ide
+.pdm.toml
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+# pytype static type analyzer
+.pytype/
+
+# Cython debug symbols
+cython_debug/
+
+# PyCharm
+#  JetBrains specific template is maintained in a separate JetBrains.gitignore that can
+#  be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
+#  and can be added to the global gitignore or merged into this file.  For a more nuclear
+#  option (not recommended) you can uncomment the following to ignore the entire idea folder.
+#.idea/
+
+### Python Patch ###
+# Poetry local configuration file - https://python-poetry.org/docs/configuration/#local-configuration
+poetry.toml
+
+# ruff
+.ruff_cache/
+
+# LSP config files
+pyrightconfig.json
+
+### VisualStudioCode ###
+**/.vscode
+.vscode/*
+!.vscode/settings.json
+!.vscode/tasks.json
+!.vscode/launch.json
+!.vscode/extensions.json
+!.vscode/*.code-snippets
+
+# Local History for Visual Studio Code
+.history/
+
+# Built Visual Studio Code Extensions
+*.vsix
+
+### VisualStudioCode Patch ###
+# Ignore all local history of files
+.history
+.ionide
+
+# End of https://www.toptal.com/developers/gitignore/api/macos,visualstudiocode,pycharm,python
+
+uv.lock
+**/train_result
+**/*.onnx
+**/*.png
+**/*.jpg
+**/augraphy_cache

.pre-commit-config.yaml

@@ -0,0 +1,22 @@
+repos:
+  - repo: https://github.com/astral-sh/ruff-pre-commit
+    rev: v0.11.6
+    hooks:
+      - id: ruff
+        args: [--fix, --respect-gitignore, --config=pyproject.toml]
+        exclude: ^texteller/models/thrid_party/paddleocr/
+      - id: ruff-format
+        args: [--config=pyproject.toml]
+        exclude: ^texteller/models/thrid_party/paddleocr/
+
+  - repo: https://github.com/pre-commit/pre-commit-hooks
+    rev: v4.5.0
+    hooks:
+      - id: trailing-whitespace
+      - id: end-of-file-fixer
+      - id: check-yaml
+      - id: check-toml
+      - id: check-added-large-files
+      - id: check-case-conflict
+      - id: check-merge-conflict
+      - id: debug-statements

.python-version

@@ -0,0 +1 @@
+3.10

LICENSE

@@ -0,0 +1,202 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright OleehyO
+
+   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.

README.md

@@ -0,0 +1,209 @@
+📄 English | <a href="./assets/README_zh.md">中文</a>
+
+<div align="center">
+    <h1>
+        <img src="./assets/fire.svg" width=30, height=30>
+        𝚃𝚎𝚡𝚃𝚎𝚕𝚕𝚎𝚛
+        <img src="./assets/fire.svg" width=30, height=30>
+    </h1>
+
+  [![](https://img.shields.io/badge/API-Docs-orange.svg?logo=read-the-docs)](https://oleehyo.github.io/TexTeller/)
+  [![](https://img.shields.io/badge/docker-pull-green.svg?logo=docker)](https://hub.docker.com/r/oleehyo/texteller)
+  [![](https://img.shields.io/badge/Data-Texteller1.0-brightgreen.svg?logo=huggingface)](https://huggingface.co/datasets/OleehyO/latex-formulas)
+  [![](https://img.shields.io/badge/Weights-Texteller3.0-yellow.svg?logo=huggingface)](https://huggingface.co/OleehyO/TexTeller)
+  [![](https://img.shields.io/badge/License-Apache_2.0-blue.svg?logo=github)](https://opensource.org/licenses/Apache-2.0)
+
+</div>
+
+https://github.com/OleehyO/TexTeller/assets/56267907/532d1471-a72e-4960-9677-ec6c19db289f
+
+TexTeller is an end-to-end formula recognition model, capable of converting images into corresponding LaTeX formulas.
+
+TexTeller was trained with **80M image-formula pairs** (previous dataset can be obtained [here](https://huggingface.co/datasets/OleehyO/latex-formulas)), compared to [LaTeX-OCR](https://github.com/lukas-blecher/LaTeX-OCR) which used a 100K dataset, TexTeller has **stronger generalization abilities** and **higher accuracy**, covering most use cases.
+
+>[!NOTE]
+> If you would like to provide feedback or suggestions for this project, feel free to start a discussion in the [Discussions section](https://github.com/OleehyO/TexTeller/discussions).
+
+---
+
+<table>
+<tr>
+<td>
+
+## 🔖 Table of Contents
+- [Getting Started](#-getting-started)
+- [Web Demo](#-web-demo)
+- [Server](#-server)
+- [Python API](#-python-api)
+- [Formula Detection](#-formula-detection)
+- [Training](#️️-training)
+
+</td>
+<td>
+
+<div align="center">
+  <figure>
+    <img src="assets/cover.png" width="800">
+    <figcaption>
+      <p>Images that can be recognized by TexTeller</p>
+    </figcaption>
+  </figure>
+  <div>
+  </div>
+</div>
+
+</td>
+</tr>
+</table>
+
+## 🔄 Change Log
+
+- 📮[2024-06-06] **TexTeller3.0 released!** The training data has been increased to **80M** (**10x more than** TexTeller2.0 and also improved in data diversity). TexTeller3.0's new features:
+
+  - Support scanned image, handwritten formulas, English(Chinese) mixed formulas.
+
+  - OCR abilities in both Chinese and English for printed images.
+
+- 📮[2024-05-02] Support **paragraph recognition**.
+
+- 📮[2024-04-12] **Formula detection model** released!
+
+- 📮[2024-03-25] TexTeller2.0 released! The training data for TexTeller2.0 has been increased to 7.5M (15x more than TexTeller1.0 and also improved in data quality). The trained TexTeller2.0 demonstrated **superior performance** in the test set, especially in recognizing rare symbols, complex multi-line formulas, and matrices.
+
+  > [Here](./assets/test.pdf) are more test images and a horizontal comparison of various recognition models.
+
+## 🚀 Getting Started
+
+1. Install the project's dependencies:
+
+   ```bash
+   pip install texteller
+   ```
+
+2. If your are using CUDA backend, you may need to install `onnxruntime-gpu`:
+
+   ```bash
+   pip install texteller[onnxruntime-gpu]
+   ```
+
+3. Run the following command to start inference:
+
+   ```bash
+   texteller inference "/path/to/image.{jpg,png}"
+   ```
+
+   > See `texteller inference --help` for more details
+
+## 🌐 Web Demo
+
+Run the following command:
+
+```bash
+texteller web
+```
+
+Enter `http://localhost:8501` in a browser to view the web demo.
+
+> [!NOTE]
+> Paragraph recognition cannot restore the structure of a document, it can only recognize its content.
+
+## 🖥️ Server
+
+We use [ray serve](https://github.com/ray-project/ray) to provide an API server for TexTeller. To start the server, run the following command:
+
+```bash
+texteller launch
+```
+
+| Parameter | Description |
+| --------- | -------- |
+| `-ckpt` | The path to the weights file,*default is TexTeller's pretrained weights*. |
+| `-tknz` | The path to the tokenizer,*default is TexTeller's tokenizer*. |
+| `-p` | The server's service port,*default is 8000*. |
+| `--num-replicas` | The number of service replicas to run on the server,*default is 1 replica*. You can use more replicas to achieve greater throughput.|
+| `--ncpu-per-replica` | The number of CPU cores used per service replica,*default is 1*.|
+| `--ngpu-per-replica` | The number of GPUs used per service replica,*default is 1*. You can set this value between 0 and 1 to run multiple service replicas on one GPU to share the GPU, thereby improving GPU utilization. (Note, if --num_replicas is 2, --ngpu_per_replica is 0.7, then 2 GPUs must be available) |
+| `--num-beams` | The number of beams for beam search,*default is 1*. |
+| `--use-onnx` | Perform inference using Onnx Runtime, *disabled by default* |
+
+To send requests to the server:
+
+```python
+# client_demo.py
+
+import requests
+
+server_url = "http://127.0.0.1:8000/predict"
+
+img_path = "/path/to/your/image"
+with open(img_path, 'rb') as img:
+    files = {'img': img}
+    response = requests.post(server_url, files=files)
+
+print(response.text)
+```
+
+## 🐍 Python API
+
+We provide several easy-to-use Python APIs for formula OCR scenarios. Please refer to our [documentation](https://oleehyo.github.io/TexTeller/) to learn about the corresponding API interfaces and usage.
+
+## 🔍 Formula Detection
+
+TexTeller's formula detection model is trained on 3,415 images of Chinese materials and 8,272 images from the [IBEM dataset](https://zenodo.org/records/4757865).
+
+<div align="center">
+    <img src="./assets/det_rec.png" width=250>
+</div>
+
+We provide a formula detection interface in the Python API. Please refer to our [API documentation](https://oleehyo.github.io/TexTeller/) for more details.
+
+## 🏋️‍♂️ Training
+
+Please setup your environment before training:
+
+1. Install the dependencies for training:
+
+   ```bash
+   pip install texteller[train]
+   ```
+
+2. Clone the repository:
+
+   ```bash
+   git clone https://github.com/OleehyO/TexTeller.git
+   ```
+
+### Dataset
+
+We provide an example dataset in the `examples/train_texteller/dataset/train` directory, you can place your own training data according to the format of the example dataset.
+
+### Training the Model
+
+In the `examples/train_texteller/` directory, run the following command:
+
+   ```bash
+   accelerate launch train.py
+   ```
+
+Training arguments can be adjusted in [`train_config.yaml`](./examples/train_texteller/train_config.yaml).
+
+## 📅 Plans
+
+- [X] ~~Train the model with a larger dataset~~
+- [X] ~~Recognition of scanned images~~
+- [X] ~~Support for English and Chinese scenarios~~
+- [X] ~~Handwritten formulas support~~
+- [ ] PDF document recognition
+- [ ] Inference acceleration
+
+## ⭐️ Stargazers over time
+
+[![Stargazers over time](https://starchart.cc/OleehyO/TexTeller.svg?variant=adaptive)](https://starchart.cc/OleehyO/TexTeller)
+
+## 👥 Contributors
+
+<a href="https://github.com/OleehyO/TexTeller/graphs/contributors">
+   <a href="https://github.com/OleehyO/TexTeller/graphs/contributors">
+      <img src="https://contrib.rocks/image?repo=OleehyO/TexTeller" />
+   </a>
+</a>

assets/README_zh.md

@@ -0,0 +1,209 @@
+📄 中文 | [English](./README.md)
+
+<div align="center">
+    <h1>
+        <img src="./fire.svg" width=30, height=30>
+        𝚃𝚎𝚡𝚃𝚎𝚕𝚕𝚎𝚛
+        <img src="./fire.svg" width=30, height=30>
+    </h1>
+
+  [![](https://img.shields.io/badge/API-文档-orange.svg?logo=read-the-docs)](https://oleehyo.github.io/TexTeller/)
+  [![](https://img.shields.io/badge/docker-镜像-green.svg?logo=docker)](https://hub.docker.com/r/oleehyo/texteller)
+  [![](https://img.shields.io/badge/数据-Texteller1.0-brightgreen.svg?logo=huggingface)](https://huggingface.co/datasets/OleehyO/latex-formulas)
+  [![](https://img.shields.io/badge/权重-Texteller3.0-yellow.svg?logo=huggingface)](https://huggingface.co/OleehyO/TexTeller)
+  [![](https://img.shields.io/badge/协议-Apache_2.0-blue.svg?logo=github)](https://opensource.org/licenses/Apache-2.0)
+
+</div>
+
+https://github.com/OleehyO/TexTeller/assets/56267907/532d1471-a72e-4960-9677-ec6c19db289f
+
+TexTeller 是一个端到端的公式识别模型，能够将图像转换为对应的 LaTeX 公式。
+
+TexTeller 使用 **8千万图像-公式对** 进行训练（前代数据集可在此[获取](https://huggingface.co/datasets/OleehyO/latex-formulas)），相较 [LaTeX-OCR](https://github.com/lukas-blecher/LaTeX-OCR) 使用的 10 万量级数据集，TexTeller 具有**更强的泛化能力**和**更高的准确率**，覆盖绝大多数使用场景。
+
+>[!NOTE]
+> 如果您想对本项目提出反馈或建议，欢迎前往 [讨论区](https://github.com/OleehyO/TexTeller/discussions) 发起讨论。
+
+---
+
+<table>
+<tr>
+<td>
+
+## 🔖 目录
+- [快速开始](#-快速开始)
+- [网页演示](#-网页演示)
+- [服务部署](#-服务部署)
+- [Python接口](#-python接口)
+- [公式检测](#-公式检测)
+- [模型训练](#️️-模型训练)
+
+</td>
+<td>
+
+<div align="center">
+  <figure>
+    <img src="cover.png" width="800">
+    <figcaption>
+      <p>TexTeller 可识别的图像示例</p>
+    </figcaption>
+  </figure>
+  <div>
+  </div>
+</div>
+
+</td>
+</tr>
+</table>
+
+## 📮 更新日志
+
+- [2024-06-06] **TexTeller3.0 发布！** 训练数据增至 **8千万**（是 TexTeller2.0 的 **10倍** 并提升了数据多样性）。TexTeller3.0 新特性：
+
+  - 支持扫描件、手写公式、中英文混合公式识别
+
+  - 支持印刷体中英文混排公式的OCR识别
+
+- [2024-05-02] 支持**段落识别**功能
+
+- [2024-04-12] **公式检测模型**发布！
+
+- [2024-03-25] TexTeller2.0 发布！TexTeller2.0 的训练数据增至750万（是前代的15倍并提升了数据质量）。训练后的 TexTeller2.0 在测试集中展现了**更优性能**，特别是在识别罕见符号、复杂多行公式和矩阵方面表现突出。
+
+  > [此处](./assets/test.pdf) 展示了更多测试图像及各类识别模型的横向对比。
+
+## 🚀 快速开始
+
+1. 安装项目依赖：
+
+   ```bash
+   pip install texteller
+   ```
+
+2. 若使用 CUDA 后端，可能需要安装 `onnxruntime-gpu`：
+
+   ```bash
+   pip install texteller[onnxruntime-gpu]
+   ```
+
+3. 运行以下命令开始推理：
+
+   ```bash
+   texteller inference "/path/to/image.{jpg,png}"
+   ```
+
+   > 更多参数请查看 `texteller inference --help`
+
+## 🌐 网页演示
+
+运行命令：
+
+```bash
+texteller web
+```
+
+在浏览器中输入 `http://localhost:8501` 查看网页演示。
+
+> [!NOTE]
+> 段落识别无法还原文档结构，仅能识别其内容。
+
+## 🖥️ 服务部署
+
+我们使用 [ray serve](https://github.com/ray-project/ray) 为 TexTeller 提供 API 服务。启动服务：
+
+```bash
+texteller launch
+```
+
+| 参数 | 说明 |
+| --------- | -------- |
+| `-ckpt` | 权重文件路径，*默认为 TexTeller 预训练权重* |
+| `-tknz` | 分词器路径，*默认为 TexTeller 分词器* |
+| `-p` | 服务端口，*默认 8000* |
+| `--num-replicas` | 服务副本数，*默认 1*。可使用更多副本来提升吞吐量 |
+| `--ncpu-per-replica` | 单个副本使用的CPU核数，*默认 1* |
+| `--ngpu-per-replica` | 单个副本使用的GPU数，*默认 1*。可设置为0~1之间的值来在单卡上运行多个服务副本共享GPU，提升GPU利用率（注意，若--num_replicas为2，--ngpu_per_replica为0.7，则需有2块可用GPU） |
+| `--num-beams` | beam search的束宽，*默认 1* |
+| `--use-onnx` | 使用Onnx Runtime进行推理，*默认关闭* |
+
+向服务发送请求：
+
+```python
+# client_demo.py
+
+import requests
+
+server_url = "http://127.0.0.1:8000/predict"
+
+img_path = "/path/to/your/image"
+with open(img_path, 'rb') as img:
+    files = {'img': img}
+    response = requests.post(server_url, files=files)
+
+print(response.text)
+```
+
+## 🐍 Python接口
+
+我们为公式OCR场景提供了多个易用的Python API接口，请参考[接口文档](https://oleehyo.github.io/TexTeller/)了解对应的API接口及使用方法。
+
+## 🔍 公式检测
+
+TexTeller的公式检测模型在3415张中文资料图像和8272张[IBEM数据集](https://zenodo.org/records/4757865)图像上训练。
+
+<div align="center">
+    <img src="./assets/det_rec.png" width=250>
+</div>
+
+我们在Python接口中提供了公式检测接口，详见[接口文档](https://oleehyo.github.io/TexTeller/)。
+
+## 🏋️‍♂️ 模型训练
+
+请按以下步骤配置训练环境：
+
+1. 安装训练依赖：
+
+   ```bash
+   pip install texteller[train]
+   ```
+
+2. 克隆仓库：
+
+   ```bash
+   git clone https://github.com/OleehyO/TexTeller.git
+   ```
+
+### 数据集准备
+
+我们在`examples/train_texteller/dataset/train`目录中提供了示例数据集，您可按照示例数据集的格式放置自己的训练数据。
+
+### 开始训练
+
+在`examples/train_texteller/`目录下运行：
+
+   ```bash
+   accelerate launch train.py
+   ```
+
+训练参数可通过[`train_config.yaml`](./examples/train_texteller/train_config.yaml)调整。
+
+## 📅 计划列表
+
+- [X] ~~使用更大规模数据集训练模型~~
+- [X] ~~扫描件识别支持~~
+- [X] ~~中英文场景支持~~
+- [X] ~~手写公式支持~~
+- [ ] PDF文档识别
+- [ ] 推理加速
+
+## ⭐️ 项目星标
+
+[![Star增长曲线](https://starchart.cc/OleehyO/TexTeller.svg?variant=adaptive)](https://starchart.cc/OleehyO/TexTeller)
+
+## 👥 贡献者
+
+<a href="https://github.com/OleehyO/TexTeller/graphs/contributors">
+   <a href="https://github.com/OleehyO/TexTeller/graphs/contributors">
+      <img src="https://contrib.rocks/image?repo=OleehyO/TexTeller" />
+   </a>
+</a>

assets/fire.svg

@@ -0,0 +1,460 @@
+<svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" style="" width="200px" height="100px" viewBox="0 0 100 100" preserveAspectRatio="xMidYMid">
+<defs>
+  <filter id="ldio-ekpf7uvh2aq-filter" filterUnits="userSpaceOnUse" x="0" y="0" width="100" height="100">
+    <feGaussianBlur in="SourceGraphic" stdDeviation="3"></feGaussianBlur>
+    <feComponentTransfer result="cutoff">
+      <feFuncA type="linear" slope="10" intercept="-5"></feFuncA>
+    </feComponentTransfer>
+  </filter>
+</defs><g filter="url(#ldio-ekpf7uvh2aq-filter)"><circle cx="45" cy="154.67770829199992" r="42" fill="#e15b64">
+  <animate attributeName="cy" values="154.67770829199992;-27.568110790210763" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.7914508173328552s"></animate>
+  <animate attributeName="r" values="42;0;0" keyTimes="0;0.6593879177915443;1" dur="1s" repeatCount="indefinite" begin="-0.7914508173328552s"></animate>
+</circle><circle cx="53" cy="156.51873756667007" r="43" fill="#e15b64">
+  <animate attributeName="cy" values="156.51873756667007;-28.593472199379597" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.8990601299952956s"></animate>
+  <animate attributeName="r" values="43;0;0" keyTimes="0;0.9199190750649376;1" dur="1s" repeatCount="indefinite" begin="-0.8990601299952956s"></animate>
+</circle><circle cx="22" cy="118.4676277511406" r="6" fill="#e15b64">
+  <animate attributeName="cy" values="118.4676277511406;-1.812134766063739" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.2574158626531723s"></animate>
+  <animate attributeName="r" values="6;0;0" keyTimes="0;0.7424894336620584;1" dur="1s" repeatCount="indefinite" begin="-0.2574158626531723s"></animate>
+</circle><circle cx="56" cy="143.3980016480395" r="34" fill="#e15b64">
+  <animate attributeName="cy" values="143.3980016480395;-23.264651741765398" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.5292591072219247s"></animate>
+  <animate attributeName="r" values="34;0;0" keyTimes="0;0.8257208789488842;1" dur="1s" repeatCount="indefinite" begin="-0.5292591072219247s"></animate>
+</circle><circle cx="43" cy="154.61226210156264" r="43" fill="#e15b64">
+  <animate attributeName="cy" values="154.61226210156264;-39.72257238426019" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.9349241678635103s"></animate>
+  <animate attributeName="r" values="43;0;0" keyTimes="0;0.6655411648349204;1" dur="1s" repeatCount="indefinite" begin="-0.9349241678635103s"></animate>
+</circle><circle cx="36" cy="141.18233539125538" r="23" fill="#e15b64">
+  <animate attributeName="cy" values="141.18233539125538;-11.919782601799477" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.9661184430026497s"></animate>
+  <animate attributeName="r" values="23;0;0" keyTimes="0;0.7340510315067473;1" dur="1s" repeatCount="indefinite" begin="-0.9661184430026497s"></animate>
+</circle><circle cx="55" cy="137.61381349909033" r="35" fill="#e15b64">
+  <animate attributeName="cy" values="137.61381349909033;-27.023105799592948" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.7882390392923937s"></animate>
+  <animate attributeName="r" values="35;0;0" keyTimes="0;0.5596286394923506;1" dur="1s" repeatCount="indefinite" begin="-0.7882390392923937s"></animate>
+</circle><circle cx="81" cy="116.42482869722863" r="6" fill="#e15b64">
+  <animate attributeName="cy" values="116.42482869722863;2.642571962973477" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.6838551001109257s"></animate>
+  <animate attributeName="r" values="6;0;0" keyTimes="0;0.8530428185299654;1" dur="1s" repeatCount="indefinite" begin="-0.6838551001109257s"></animate>
+</circle><circle cx="51" cy="144.1337397120671" r="41" fill="#e15b64">
+  <animate attributeName="cy" values="144.1337397120671;-35.62888188299487" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.8931867510460544s"></animate>
+  <animate attributeName="r" values="41;0;0" keyTimes="0;0.9351064787950636;1" dur="1s" repeatCount="indefinite" begin="-0.8931867510460544s"></animate>
+</circle><circle cx="22" cy="127.94124738258117" r="20" fill="#e15b64">
+  <animate attributeName="cy" values="127.94124738258117;-4.588101238414598" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.9129507531699166s"></animate>
+  <animate attributeName="r" values="20;0;0" keyTimes="0;0.9626971761152365;1" dur="1s" repeatCount="indefinite" begin="-0.9129507531699166s"></animate>
+</circle><circle cx="51" cy="130.13871763314205" r="21" fill="#e15b64">
+  <animate attributeName="cy" values="130.13871763314205;-2.771870373434613" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.16276671760313832s"></animate>
+  <animate attributeName="r" values="21;0;0" keyTimes="0;0.6367210977937845;1" dur="1s" repeatCount="indefinite" begin="-0.16276671760313832s"></animate>
+</circle><circle cx="28" cy="130.94671647108635" r="26" fill="#e15b64">
+  <animate attributeName="cy" values="130.94671647108635;-20.54470862263146" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.010777607623041363s"></animate>
+  <animate attributeName="r" values="26;0;0" keyTimes="0;0.5986827903483527;1" dur="1s" repeatCount="indefinite" begin="-0.010777607623041363s"></animate>
+</circle><circle cx="32" cy="133.57559887485095" r="18" fill="#e15b64">
+  <animate attributeName="cy" values="133.57559887485095;-13.998747273650661" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.6849903294560423s"></animate>
+  <animate attributeName="r" values="18;0;0" keyTimes="0;0.9272684317035897;1" dur="1s" repeatCount="indefinite" begin="-0.6849903294560423s"></animate>
+</circle><circle cx="50" cy="129.2368025879272" r="29" fill="#e15b64">
+  <animate attributeName="cy" values="129.2368025879272;-21.38222818211007" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.2570532837614655s"></animate>
+  <animate attributeName="r" values="29;0;0" keyTimes="0;0.5349692982819836;1" dur="1s" repeatCount="indefinite" begin="-0.2570532837614655s"></animate>
+</circle><circle cx="54" cy="147.67203918209864" r="32" fill="#e15b64">
+  <animate attributeName="cy" values="147.67203918209864;-23.292000640460095" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.8840781999829185s"></animate>
+  <animate attributeName="r" values="32;0;0" keyTimes="0;0.9905440228534627;1" dur="1s" repeatCount="indefinite" begin="-0.8840781999829185s"></animate>
+</circle><circle cx="49" cy="156.33097983975816" r="43" fill="#e15b64">
+  <animate attributeName="cy" values="156.33097983975816;-30.688836209655307" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.6363282840605137s"></animate>
+  <animate attributeName="r" values="43;0;0" keyTimes="0;0.578321371334853;1" dur="1s" repeatCount="indefinite" begin="-0.6363282840605137s"></animate>
+</circle><circle cx="53" cy="150.73132612778645" r="38" fill="#e15b64">
+  <animate attributeName="cy" values="150.73132612778645;-24.243875812169208" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.6889884148164682s"></animate>
+  <animate attributeName="r" values="38;0;0" keyTimes="0;0.9820908894527897;1" dur="1s" repeatCount="indefinite" begin="-0.6889884148164682s"></animate>
+</circle><circle cx="58" cy="136.92364235316566" r="30" fill="#e15b64">
+  <animate attributeName="cy" values="136.92364235316566;-14.514104757207221" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.3274028295945308s"></animate>
+  <animate attributeName="r" values="30;0;0" keyTimes="0;0.9109990458833535;1" dur="1s" repeatCount="indefinite" begin="-0.3274028295945308s"></animate>
+</circle><circle cx="21" cy="125.47085228007643" r="18" fill="#e15b64">
+  <animate attributeName="cy" values="125.47085228007643;-8.232426956653288" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.11103461733078768s"></animate>
+  <animate attributeName="r" values="18;0;0" keyTimes="0;0.7718042613876622;1" dur="1s" repeatCount="indefinite" begin="-0.11103461733078768s"></animate>
+</circle><circle cx="57" cy="154.13251799723747" r="37" fill="#e15b64">
+  <animate attributeName="cy" values="154.13251799723747;-18.665203993986026" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.8263441768461145s"></animate>
+  <animate attributeName="r" values="37;0;0" keyTimes="0;0.7148325280461965;1" dur="1s" repeatCount="indefinite" begin="-0.8263441768461145s"></animate>
+</circle><circle cx="52" cy="163.55969451733722" r="47" fill="#e15b64">
+  <animate attributeName="cy" values="163.55969451733722;-45.32343944696123" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.08605155305311041s"></animate>
+  <animate attributeName="r" values="47;0;0" keyTimes="0;0.8554524873372089;1" dur="1s" repeatCount="indefinite" begin="-0.08605155305311041s"></animate>
+</circle><circle cx="43" cy="150.72861891310126" r="42" fill="#e15b64">
+  <animate attributeName="cy" values="150.72861891310126;-23.942286768617272" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.8013052401764136s"></animate>
+  <animate attributeName="r" values="42;0;0" keyTimes="0;0.6681090498432822;1" dur="1s" repeatCount="indefinite" begin="-0.8013052401764136s"></animate>
+</circle><circle cx="62" cy="109.2607457626771" r="2" fill="#e15b64">
+  <animate attributeName="cy" values="109.2607457626771;3.194634855160243" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.7901767326521292s"></animate>
+  <animate attributeName="r" values="2;0;0" keyTimes="0;0.7018579919397697;1" dur="1s" repeatCount="indefinite" begin="-0.7901767326521292s"></animate>
+</circle><circle cx="29" cy="132.04950518708117" r="26" fill="#e15b64">
+  <animate attributeName="cy" values="132.04950518708117;-24.268419710129816" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.9729317633977274s"></animate>
+  <animate attributeName="r" values="26;0;0" keyTimes="0;0.8277305604086497;1" dur="1s" repeatCount="indefinite" begin="-0.9729317633977274s"></animate>
+</circle><circle cx="54" cy="150.69697127653222" r="41" fill="#e15b64">
+  <animate attributeName="cy" values="150.69697127653222;-27.168516505190766" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.5902016146688314s"></animate>
+  <animate attributeName="r" values="41;0;0" keyTimes="0;0.8175867220161461;1" dur="1s" repeatCount="indefinite" begin="-0.5902016146688314s"></animate>
+</circle><circle cx="50" cy="115.01352405454155" r="7" fill="#e15b64">
+  <animate attributeName="cy" values="115.01352405454155;-4.5076288690789195" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.5091907734741129s"></animate>
+  <animate attributeName="r" values="7;0;0" keyTimes="0;0.6751846924914742;1" dur="1s" repeatCount="indefinite" begin="-0.5091907734741129s"></animate>
+</circle><circle cx="65" cy="137.6419430633514" r="34" fill="#e15b64">
+  <animate attributeName="cy" values="137.6419430633514;-17.00344965868893" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.34747192063247945s"></animate>
+  <animate attributeName="r" values="34;0;0" keyTimes="0;0.5212737600536792;1" dur="1s" repeatCount="indefinite" begin="-0.34747192063247945s"></animate>
+</circle><circle cx="34" cy="127.0455079544209" r="14" fill="#e15b64">
+  <animate attributeName="cy" values="127.0455079544209;-3.6990759299641454" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.4890615261218786s"></animate>
+  <animate attributeName="r" values="14;0;0" keyTimes="0;0.6183470012170013;1" dur="1s" repeatCount="indefinite" begin="-0.4890615261218786s"></animate>
+</circle><circle cx="12" cy="120.43345098845494" r="3" fill="#e15b64">
+  <animate attributeName="cy" values="120.43345098845494;9.74374931913883" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.3026505339978601s"></animate>
+  <animate attributeName="r" values="3;0;0" keyTimes="0;0.5414300978949788;1" dur="1s" repeatCount="indefinite" begin="-0.3026505339978601s"></animate>
+</circle><circle cx="49" cy="161.35205628493102" r="43" fill="#e15b64">
+  <animate attributeName="cy" values="161.35205628493102;-37.872089939512506" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.38741962448531564s"></animate>
+  <animate attributeName="r" values="43;0;0" keyTimes="0;0.5096615889177538;1" dur="1s" repeatCount="indefinite" begin="-0.38741962448531564s"></animate>
+</circle><circle cx="54" cy="146.5769009919314" r="44" fill="#e15b64">
+  <animate attributeName="cy" values="146.5769009919314;-38.33530354334875" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.34335748774106034s"></animate>
+  <animate attributeName="r" values="44;0;0" keyTimes="0;0.743420827137904;1" dur="1s" repeatCount="indefinite" begin="-0.34335748774106034s"></animate>
+</circle><circle cx="20" cy="111.24659457696168" r="7" fill="#e15b64">
+  <animate attributeName="cy" values="111.24659457696168;10.851798254886354" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.6282307990647713s"></animate>
+  <animate attributeName="r" values="7;0;0" keyTimes="0;0.8297799829349941;1" dur="1s" repeatCount="indefinite" begin="-0.6282307990647713s"></animate>
+</circle><circle cx="50" cy="164.0676485495781" r="45" fill="#e15b64">
+  <animate attributeName="cy" values="164.0676485495781;-31.499414285176986" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.7760446285439819s"></animate>
+  <animate attributeName="r" values="45;0;0" keyTimes="0;0.5740694195049653;1" dur="1s" repeatCount="indefinite" begin="-0.7760446285439819s"></animate>
+</circle><circle cx="63" cy="121.15583070803987" r="16" fill="#e15b64">
+  <animate attributeName="cy" values="121.15583070803987;-2.1042758907266066" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.2305276534763374s"></animate>
+  <animate attributeName="r" values="16;0;0" keyTimes="0;0.5205278426126575;1" dur="1s" repeatCount="indefinite" begin="-0.2305276534763374s"></animate>
+</circle><circle cx="70" cy="143.94247592516618" r="29" fill="#e15b64">
+  <animate attributeName="cy" values="143.94247592516618;-23.62297573618442" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.5284797120514513s"></animate>
+  <animate attributeName="r" values="29;0;0" keyTimes="0;0.9336811516026573;1" dur="1s" repeatCount="indefinite" begin="-0.5284797120514513s"></animate>
+</circle><circle cx="21" cy="122.79868387744153" r="20" fill="#e15b64">
+  <animate attributeName="cy" values="122.79868387744153;-13.104461771681535" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.8845782118773111s"></animate>
+  <animate attributeName="r" values="20;0;0" keyTimes="0;0.904216846935756;1" dur="1s" repeatCount="indefinite" begin="-0.8845782118773111s"></animate>
+</circle><circle cx="46" cy="143.70707265719267" r="24" fill="#e15b64">
+  <animate attributeName="cy" values="143.70707265719267;-20.28891701845349" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.23245576862802375s"></animate>
+  <animate attributeName="r" values="24;0;0" keyTimes="0;0.6586288079548765;1" dur="1s" repeatCount="indefinite" begin="-0.23245576862802375s"></animate>
+</circle><circle cx="65" cy="140.13731645312657" r="22" fill="#e15b64">
+  <animate attributeName="cy" values="140.13731645312657;-5.338876455584764" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.7182419259629308s"></animate>
+  <animate attributeName="r" values="22;0;0" keyTimes="0;0.8813907372203135;1" dur="1s" repeatCount="indefinite" begin="-0.7182419259629308s"></animate>
+</circle><circle cx="37" cy="139.00958710472267" r="35" fill="#e15b64">
+  <animate attributeName="cy" values="139.00958710472267;-25.68265144780311" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.7030100698848409s"></animate>
+  <animate attributeName="r" values="35;0;0" keyTimes="0;0.7320613459176248;1" dur="1s" repeatCount="indefinite" begin="-0.7030100698848409s"></animate>
+</circle><circle cx="45" cy="146.6744507961619" r="44" fill="#e15b64">
+  <animate attributeName="cy" values="146.6744507961619;-38.087338695486295" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.8319540053556033s"></animate>
+  <animate attributeName="r" values="44;0;0" keyTimes="0;0.5904241586083279;1" dur="1s" repeatCount="indefinite" begin="-0.8319540053556033s"></animate>
+</circle><circle cx="53" cy="116.16529146873187" r="15" fill="#e15b64">
+  <animate attributeName="cy" values="116.16529146873187;-3.17669223153381" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.7864341362651808s"></animate>
+  <animate attributeName="r" values="15;0;0" keyTimes="0;0.589186107816807;1" dur="1s" repeatCount="indefinite" begin="-0.7864341362651808s"></animate>
+</circle><circle cx="29" cy="141.6902909599232" r="23" fill="#e15b64">
+  <animate attributeName="cy" values="141.6902909599232;-16.250272669063218" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.18084365714200346s"></animate>
+  <animate attributeName="r" values="23;0;0" keyTimes="0;0.8116571311237253;1" dur="1s" repeatCount="indefinite" begin="-0.18084365714200346s"></animate>
+</circle><circle cx="65" cy="143.73302386926983" r="32" fill="#e15b64">
+  <animate attributeName="cy" values="143.73302386926983;-24.229369251904558" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.5786484558188305s"></animate>
+  <animate attributeName="r" values="32;0;0" keyTimes="0;0.8515606125902615;1" dur="1s" repeatCount="indefinite" begin="-0.5786484558188305s"></animate>
+</circle><circle cx="39" cy="143.3951504366216" r="33" fill="#e15b64">
+  <animate attributeName="cy" values="143.3951504366216;-27.75171362166084" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.1481578769905092s"></animate>
+  <animate attributeName="r" values="33;0;0" keyTimes="0;0.797255218191478;1" dur="1s" repeatCount="indefinite" begin="-0.1481578769905092s"></animate>
+</circle><circle cx="59" cy="129.28605384114482" r="27" fill="#e15b64">
+  <animate attributeName="cy" values="129.28605384114482;-12.095864862844131" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.23581997562886903s"></animate>
+  <animate attributeName="r" values="27;0;0" keyTimes="0;0.8271538616610963;1" dur="1s" repeatCount="indefinite" begin="-0.23581997562886903s"></animate>
+</circle><circle cx="70" cy="144.09835508207823" r="28" fill="#e15b64">
+  <animate attributeName="cy" values="144.09835508207823;-13.162793363728145" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.23606519556482253s"></animate>
+  <animate attributeName="r" values="28;0;0" keyTimes="0;0.73085815703799;1" dur="1s" repeatCount="indefinite" begin="-0.23606519556482253s"></animate>
+</circle><circle cx="48" cy="145.01565757702042" r="44" fill="#e15b64">
+  <animate attributeName="cy" values="145.01565757702042;-32.30510020024561" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.8615348704203486s"></animate>
+  <animate attributeName="r" values="44;0;0" keyTimes="0;0.9694373671371078;1" dur="1s" repeatCount="indefinite" begin="-0.8615348704203486s"></animate>
+</circle><circle cx="95" cy="113.78554320990165" r="4" fill="#e15b64">
+  <animate attributeName="cy" values="113.78554320990165;-1.2652564238335904" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.21370544900580335s"></animate>
+  <animate attributeName="r" values="4;0;0" keyTimes="0;0.5334621383741172;1" dur="1s" repeatCount="indefinite" begin="-0.21370544900580335s"></animate>
+</circle><circle cx="57" cy="136.06708935936715" r="34" fill="#e15b64">
+  <animate attributeName="cy" values="136.06708935936715;-19.758990054858902" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.7755376997281404s"></animate>
+  <animate attributeName="r" values="34;0;0" keyTimes="0;0.9943252777203475;1" dur="1s" repeatCount="indefinite" begin="-0.7755376997281404s"></animate>
+</circle><circle cx="72" cy="123.8422572942333" r="19" fill="#e15b64">
+  <animate attributeName="cy" values="123.8422572942333;-1.0000700639794928" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.9670461872772004s"></animate>
+  <animate attributeName="r" values="19;0;0" keyTimes="0;0.7801926792335607;1" dur="1s" repeatCount="indefinite" begin="-0.9670461872772004s"></animate>
+</circle></g><g filter="url(#ldio-ekpf7uvh2aq-filter)"><circle cx="27" cy="136.75172282051147" r="17" fill="#f47e60">
+  <animate attributeName="cy" values="136.75172282051147;-5.48853662281188" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.4403846891955857s"></animate>
+  <animate attributeName="r" values="17;0;0" keyTimes="0;0.7894732341719188;1" dur="1s" repeatCount="indefinite" begin="-0.4403846891955857s"></animate>
+</circle><circle cx="34" cy="132.08290473906044" r="28" fill="#f47e60">
+  <animate attributeName="cy" values="132.08290473906044;-16.339029232048958" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.7882134883361418s"></animate>
+  <animate attributeName="r" values="28;0;0" keyTimes="0;0.5035175026787356;1" dur="1s" repeatCount="indefinite" begin="-0.7882134883361418s"></animate>
+</circle><circle cx="66" cy="127.45606892584162" r="23" fill="#f47e60">
+  <animate attributeName="cy" values="127.45606892584162;-11.56763185745981" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.23537267190332678s"></animate>
+  <animate attributeName="r" values="23;0;0" keyTimes="0;0.7818578332234903;1" dur="1s" repeatCount="indefinite" begin="-0.23537267190332678s"></animate>
+</circle><circle cx="29" cy="124.28337961013858" r="15" fill="#f47e60">
+  <animate attributeName="cy" values="124.28337961013858;0.8461921465181206" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.30918442080681285s"></animate>
+  <animate attributeName="r" values="15;0;0" keyTimes="0;0.9741475377259025;1" dur="1s" repeatCount="indefinite" begin="-0.30918442080681285s"></animate>
+</circle><circle cx="61" cy="147.91603256008383" r="31" fill="#f47e60">
+  <animate attributeName="cy" values="147.91603256008383;-14.754981670358578" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.0033816756583812113s"></animate>
+  <animate attributeName="r" values="31;0;0" keyTimes="0;0.6463193577485268;1" dur="1s" repeatCount="indefinite" begin="-0.0033816756583812113s"></animate>
+</circle><circle cx="25" cy="120.64483537229628" r="9" fill="#f47e60">
+  <animate attributeName="cy" values="120.64483537229628;-7.193123212298179" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.6891092543031828s"></animate>
+  <animate attributeName="r" values="9;0;0" keyTimes="0;0.8637808572418493;1" dur="1s" repeatCount="indefinite" begin="-0.6891092543031828s"></animate>
+</circle><circle cx="12" cy="121.18727231753691" r="4" fill="#f47e60">
+  <animate attributeName="cy" values="121.18727231753691;15.883181236637633" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.24454851002004097s"></animate>
+  <animate attributeName="r" values="4;0;0" keyTimes="0;0.8215012014926046;1" dur="1s" repeatCount="indefinite" begin="-0.24454851002004097s"></animate>
+</circle><circle cx="58" cy="136.64954415018815" r="19" fill="#f47e60">
+  <animate attributeName="cy" values="136.64954415018815;-13.637628862199563" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.7672442553828805s"></animate>
+  <animate attributeName="r" values="19;0;0" keyTimes="0;0.7534841891330046;1" dur="1s" repeatCount="indefinite" begin="-0.7672442553828805s"></animate>
+</circle><circle cx="69" cy="120.72538023727738" r="10" fill="#f47e60">
+  <animate attributeName="cy" values="120.72538023727738;-5.651458016294906" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.6587915764098667s"></animate>
+  <animate attributeName="r" values="10;0;0" keyTimes="0;0.5977129956186352;1" dur="1s" repeatCount="indefinite" begin="-0.6587915764098667s"></animate>
+</circle><circle cx="46" cy="122.63158963579554" r="20" fill="#f47e60">
+  <animate attributeName="cy" values="122.63158963579554;-8.99196405151625" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.3698350873089088s"></animate>
+  <animate attributeName="r" values="20;0;0" keyTimes="0;0.5563937567659611;1" dur="1s" repeatCount="indefinite" begin="-0.3698350873089088s"></animate>
+</circle><circle cx="7" cy="121.15700947168602" r="2" fill="#f47e60">
+  <animate attributeName="cy" values="121.15700947168602;0.605011189845321" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.514133243834255s"></animate>
+  <animate attributeName="r" values="2;0;0" keyTimes="0;0.7510335363256938;1" dur="1s" repeatCount="indefinite" begin="-0.514133243834255s"></animate>
+</circle><circle cx="19" cy="117.69071117783832" r="7" fill="#f47e60">
+  <animate attributeName="cy" values="117.69071117783832;-2.4512162536532234" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.4163222368875168s"></animate>
+  <animate attributeName="r" values="7;0;0" keyTimes="0;0.9697983093212361;1" dur="1s" repeatCount="indefinite" begin="-0.4163222368875168s"></animate>
+</circle><circle cx="34" cy="122.22172344680293" r="22" fill="#f47e60">
+  <animate attributeName="cy" values="122.22172344680293;-14.875000336072436" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.8346904488502503s"></animate>
+  <animate attributeName="r" values="22;0;0" keyTimes="0;0.9284864899458874;1" dur="1s" repeatCount="indefinite" begin="-0.8346904488502503s"></animate>
+</circle><circle cx="48" cy="118.34245443793573" r="12" fill="#f47e60">
+  <animate attributeName="cy" values="118.34245443793573;6.1569446890589035" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.7372012265846987s"></animate>
+  <animate attributeName="r" values="12;0;0" keyTimes="0;0.9146509122657862;1" dur="1s" repeatCount="indefinite" begin="-0.7372012265846987s"></animate>
+</circle><circle cx="38" cy="108.37260349538107" r="4" fill="#f47e60">
+  <animate attributeName="cy" values="108.37260349538107;-3.9166184571860483" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.6955752887050161s"></animate>
+  <animate attributeName="r" values="4;0;0" keyTimes="0;0.9793871272170744;1" dur="1s" repeatCount="indefinite" begin="-0.6955752887050161s"></animate>
+</circle><circle cx="50" cy="120.05611377372627" r="20" fill="#f47e60">
+  <animate attributeName="cy" values="120.05611377372627;-19.59128463520709" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.8198691615147322s"></animate>
+  <animate attributeName="r" values="20;0;0" keyTimes="0;0.6017320767396992;1" dur="1s" repeatCount="indefinite" begin="-0.8198691615147322s"></animate>
+</circle><circle cx="69" cy="133.11553485199934" r="21" fill="#f47e60">
+  <animate attributeName="cy" values="133.11553485199934;-7.230262198733577" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.6502042470386947s"></animate>
+  <animate attributeName="r" values="21;0;0" keyTimes="0;0.9802383350633911;1" dur="1s" repeatCount="indefinite" begin="-0.6502042470386947s"></animate>
+</circle><circle cx="60" cy="138.10205797824347" r="31" fill="#f47e60">
+  <animate attributeName="cy" values="138.10205797824347;-21.149182634283513" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.8527464543018912s"></animate>
+  <animate attributeName="r" values="31;0;0" keyTimes="0;0.5593223005306734;1" dur="1s" repeatCount="indefinite" begin="-0.8527464543018912s"></animate>
+</circle><circle cx="72" cy="121.45841247692351" r="16" fill="#f47e60">
+  <animate attributeName="cy" values="121.45841247692351;-5.0851516529984195" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.4077549975882817s"></animate>
+  <animate attributeName="r" values="16;0;0" keyTimes="0;0.5763111141098053;1" dur="1s" repeatCount="indefinite" begin="-0.4077549975882817s"></animate>
+</circle><circle cx="56" cy="118.12349945951125" r="10" fill="#f47e60">
+  <animate attributeName="cy" values="118.12349945951125;-7.082779421666896" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.21747152423150562s"></animate>
+  <animate attributeName="r" values="10;0;0" keyTimes="0;0.6868094744383062;1" dur="1s" repeatCount="indefinite" begin="-0.21747152423150562s"></animate>
+</circle><circle cx="77" cy="119.41951761904794" r="17" fill="#f47e60">
+  <animate attributeName="cy" values="119.41951761904794;-9.114276721599797" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.48345793287516814s"></animate>
+  <animate attributeName="r" values="17;0;0" keyTimes="0;0.5135663211192452;1" dur="1s" repeatCount="indefinite" begin="-0.48345793287516814s"></animate>
+</circle><circle cx="78" cy="125.60192795392818" r="11" fill="#f47e60">
+  <animate attributeName="cy" values="125.60192795392818;-6.73068982191926" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.23667812050200931s"></animate>
+  <animate attributeName="r" values="11;0;0" keyTimes="0;0.9898092475181265;1" dur="1s" repeatCount="indefinite" begin="-0.23667812050200931s"></animate>
+</circle><circle cx="51" cy="138.224179154187" r="24" fill="#f47e60">
+  <animate attributeName="cy" values="138.224179154187;-8.55653503677315" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.5735700676741093s"></animate>
+  <animate attributeName="r" values="24;0;0" keyTimes="0;0.9566960986989479;1" dur="1s" repeatCount="indefinite" begin="-0.5735700676741093s"></animate>
+</circle><circle cx="41" cy="131.14944604607328" r="21" fill="#f47e60">
+  <animate attributeName="cy" values="131.14944604607328;-17.847508222350655" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.07696580759865079s"></animate>
+  <animate attributeName="r" values="21;0;0" keyTimes="0;0.6865631531399743;1" dur="1s" repeatCount="indefinite" begin="-0.07696580759865079s"></animate>
+</circle><circle cx="49" cy="128.787268826053" r="17" fill="#f47e60">
+  <animate attributeName="cy" values="128.787268826053;1.143259231969072" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.7890428937034474s"></animate>
+  <animate attributeName="r" values="17;0;0" keyTimes="0;0.5926722445396657;1" dur="1s" repeatCount="indefinite" begin="-0.7890428937034474s"></animate>
+</circle><circle cx="17" cy="120.22416295842616" r="13" fill="#f47e60">
+  <animate attributeName="cy" values="120.22416295842616;5.932998615440596" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.25642472915187764s"></animate>
+  <animate attributeName="r" values="13;0;0" keyTimes="0;0.5738477034101163;1" dur="1s" repeatCount="indefinite" begin="-0.25642472915187764s"></animate>
+</circle><circle cx="73" cy="127.02191586426626" r="24" fill="#f47e60">
+  <animate attributeName="cy" values="127.02191586426626;-19.34982189589097" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.9257599774553938s"></animate>
+  <animate attributeName="r" values="24;0;0" keyTimes="0;0.6060248140675957;1" dur="1s" repeatCount="indefinite" begin="-0.9257599774553938s"></animate>
+</circle><circle cx="29" cy="122.37303701766326" r="22" fill="#f47e60">
+  <animate attributeName="cy" values="122.37303701766326;-17.181874655618834" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.11979523584713825s"></animate>
+  <animate attributeName="r" values="22;0;0" keyTimes="0;0.5778892301319281;1" dur="1s" repeatCount="indefinite" begin="-0.11979523584713825s"></animate>
+</circle><circle cx="30" cy="132.91741320840808" r="18" fill="#f47e60">
+  <animate attributeName="cy" values="132.91741320840808;0.24294121648419775" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.6890213202603488s"></animate>
+  <animate attributeName="r" values="18;0;0" keyTimes="0;0.8587373770805918;1" dur="1s" repeatCount="indefinite" begin="-0.6890213202603488s"></animate>
+</circle><circle cx="80" cy="116.72839679840811" r="14" fill="#f47e60">
+  <animate attributeName="cy" values="116.72839679840811;4.82183707831593" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.08182847032405782s"></animate>
+  <animate attributeName="r" values="14;0;0" keyTimes="0;0.6809633164153448;1" dur="1s" repeatCount="indefinite" begin="-0.08182847032405782s"></animate>
+</circle><circle cx="31" cy="125.20247260666616" r="13" fill="#f47e60">
+  <animate attributeName="cy" values="125.20247260666616;2.008326413572634" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.8369662812852767s"></animate>
+  <animate attributeName="r" values="13;0;0" keyTimes="0;0.5845779670186058;1" dur="1s" repeatCount="indefinite" begin="-0.8369662812852767s"></animate>
+</circle><circle cx="60" cy="125.0794549947879" r="16" fill="#f47e60">
+  <animate attributeName="cy" values="125.0794549947879;0.7338248372355807" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.8948237868324189s"></animate>
+  <animate attributeName="r" values="16;0;0" keyTimes="0;0.9120596722058173;1" dur="1s" repeatCount="indefinite" begin="-0.8948237868324189s"></animate>
+</circle><circle cx="25" cy="126.90612837175388" r="8" fill="#f47e60">
+  <animate attributeName="cy" values="126.90612837175388;4.0472618983783715" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.39581604043317986s"></animate>
+  <animate attributeName="r" values="8;0;0" keyTimes="0;0.8074064845720312;1" dur="1s" repeatCount="indefinite" begin="-0.39581604043317986s"></animate>
+</circle><circle cx="37" cy="131.42028038990128" r="25" fill="#f47e60">
+  <animate attributeName="cy" values="131.42028038990128;-22.403977227715075" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.04301794169924622s"></animate>
+  <animate attributeName="r" values="25;0;0" keyTimes="0;0.524891315929541;1" dur="1s" repeatCount="indefinite" begin="-0.04301794169924622s"></animate>
+</circle><circle cx="41" cy="149.05000141391616" r="31" fill="#f47e60">
+  <animate attributeName="cy" values="149.05000141391616;-19.10046896539864" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.7213401886638007s"></animate>
+  <animate attributeName="r" values="31;0;0" keyTimes="0;0.6890520162965066;1" dur="1s" repeatCount="indefinite" begin="-0.7213401886638007s"></animate>
+</circle><circle cx="36" cy="138.58798523568342" r="27" fill="#f47e60">
+  <animate attributeName="cy" values="138.58798523568342;-15.572058043829461" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.40556498158772736s"></animate>
+  <animate attributeName="r" values="27;0;0" keyTimes="0;0.8506348676044777;1" dur="1s" repeatCount="indefinite" begin="-0.40556498158772736s"></animate>
+</circle><circle cx="78" cy="137.9707233461312" r="20" fill="#f47e60">
+  <animate attributeName="cy" values="137.9707233461312;-3.6945948738885512" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.8880631706610672s"></animate>
+  <animate attributeName="r" values="20;0;0" keyTimes="0;0.9304971995517395;1" dur="1s" repeatCount="indefinite" begin="-0.8880631706610672s"></animate>
+</circle><circle cx="79" cy="134.71673525431498" r="18" fill="#f47e60">
+  <animate attributeName="cy" values="134.71673525431498;-10.261412982322742" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.2848983056723242s"></animate>
+  <animate attributeName="r" values="18;0;0" keyTimes="0;0.7526875949615255;1" dur="1s" repeatCount="indefinite" begin="-0.2848983056723242s"></animate>
+</circle><circle cx="82" cy="111.49802891873294" r="5" fill="#f47e60">
+  <animate attributeName="cy" values="111.49802891873294;12.140748225430922" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.40945179236345397s"></animate>
+  <animate attributeName="r" values="5;0;0" keyTimes="0;0.703997116139137;1" dur="1s" repeatCount="indefinite" begin="-0.40945179236345397s"></animate>
+</circle><circle cx="68" cy="140.96466884045572" r="22" fill="#f47e60">
+  <animate attributeName="cy" values="140.96466884045572;-4.079142984351218" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.40439383112303107s"></animate>
+  <animate attributeName="r" values="22;0;0" keyTimes="0;0.5493704483007363;1" dur="1s" repeatCount="indefinite" begin="-0.40439383112303107s"></animate>
+</circle><circle cx="41" cy="116.24169615516264" r="16" fill="#f47e60">
+  <animate attributeName="cy" values="116.24169615516264;-13.644720096932094" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.22449184929827926s"></animate>
+  <animate attributeName="r" values="16;0;0" keyTimes="0;0.6587866247823291;1" dur="1s" repeatCount="indefinite" begin="-0.22449184929827926s"></animate>
+</circle><circle cx="20" cy="124.66929057881916" r="15" fill="#f47e60">
+  <animate attributeName="cy" values="124.66929057881916;2.5505611618972814" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.017560126563357925s"></animate>
+  <animate attributeName="r" values="15;0;0" keyTimes="0;0.6128429739262174;1" dur="1s" repeatCount="indefinite" begin="-0.017560126563357925s"></animate>
+</circle><circle cx="63" cy="126.5115900704738" r="26" fill="#f47e60">
+  <animate attributeName="cy" values="126.5115900704738;-20.921901271813873" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.5285257319858678s"></animate>
+  <animate attributeName="r" values="26;0;0" keyTimes="0;0.9007468611639214;1" dur="1s" repeatCount="indefinite" begin="-0.5285257319858678s"></animate>
+</circle><circle cx="90" cy="111.61440083571019" r="6" fill="#f47e60">
+  <animate attributeName="cy" values="111.61440083571019;11.61930520437923" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.8167452043810126s"></animate>
+  <animate attributeName="r" values="6;0;0" keyTimes="0;0.9810779841180124;1" dur="1s" repeatCount="indefinite" begin="-0.8167452043810126s"></animate>
+</circle><circle cx="78" cy="122.50775060552778" r="20" fill="#f47e60">
+  <animate attributeName="cy" values="122.50775060552778;-4.59807973956865" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.11755589684814727s"></animate>
+  <animate attributeName="r" values="20;0;0" keyTimes="0;0.6705237343698631;1" dur="1s" repeatCount="indefinite" begin="-0.11755589684814727s"></animate>
+</circle><circle cx="31" cy="127.90703241028092" r="9" fill="#f47e60">
+  <animate attributeName="cy" values="127.90703241028092;0.829718008041219" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.5851309189776632s"></animate>
+  <animate attributeName="r" values="9;0;0" keyTimes="0;0.6889560303799027;1" dur="1s" repeatCount="indefinite" begin="-0.5851309189776632s"></animate>
+</circle><circle cx="65" cy="117.43435709704966" r="4" fill="#f47e60">
+  <animate attributeName="cy" values="117.43435709704966;15.28596080488979" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.8492165554334472s"></animate>
+  <animate attributeName="r" values="4;0;0" keyTimes="0;0.5287459347086204;1" dur="1s" repeatCount="indefinite" begin="-0.8492165554334472s"></animate>
+</circle><circle cx="89" cy="122.93132420091489" r="3" fill="#f47e60">
+  <animate attributeName="cy" values="122.93132420091489;5.980513428860888" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.06884209677796871s"></animate>
+  <animate attributeName="r" values="3;0;0" keyTimes="0;0.5868616814040618;1" dur="1s" repeatCount="indefinite" begin="-0.06884209677796871s"></animate>
+</circle><circle cx="68" cy="129.1441504106191" r="26" fill="#f47e60">
+  <animate attributeName="cy" values="129.1441504106191;-22.781245889673905" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.26191875209122073s"></animate>
+  <animate attributeName="r" values="26;0;0" keyTimes="0;0.6200648439404779;1" dur="1s" repeatCount="indefinite" begin="-0.26191875209122073s"></animate>
+</circle><circle cx="22" cy="130.63745849588264" r="20" fill="#f47e60">
+  <animate attributeName="cy" values="130.63745849588264;-10.695329441338862" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.6192951915425052s"></animate>
+  <animate attributeName="r" values="20;0;0" keyTimes="0;0.6969346125529845;1" dur="1s" repeatCount="indefinite" begin="-0.6192951915425052s"></animate>
+</circle></g><g filter="url(#ldio-ekpf7uvh2aq-filter)"><circle cx="57" cy="123.68953191890479" r="12" fill="#f8b26a">
+  <animate attributeName="cy" values="123.68953191890479;4.854991577389438" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.9097135632734302s"></animate>
+  <animate attributeName="r" values="12;0;0" keyTimes="0;0.9463910575266388;1" dur="1s" repeatCount="indefinite" begin="-0.9097135632734302s"></animate>
+</circle><circle cx="24" cy="124.54645838615471" r="12" fill="#f8b26a">
+  <animate attributeName="cy" values="124.54645838615471;-11.813810322332547" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.007050694143823311s"></animate>
+  <animate attributeName="r" values="12;0;0" keyTimes="0;0.7078891674964196;1" dur="1s" repeatCount="indefinite" begin="-0.007050694143823311s"></animate>
+</circle><circle cx="54" cy="110.08044357995595" r="3" fill="#f8b26a">
+  <animate attributeName="cy" values="110.08044357995595;13.402947007936334" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.994432759852213s"></animate>
+  <animate attributeName="r" values="3;0;0" keyTimes="0;0.8430605754104277;1" dur="1s" repeatCount="indefinite" begin="-0.994432759852213s"></animate>
+</circle><circle cx="49" cy="127.80477114160061" r="16" fill="#f8b26a">
+  <animate attributeName="cy" values="127.80477114160061;2.7658256519770603" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.07188593356616135s"></animate>
+  <animate attributeName="r" values="16;0;0" keyTimes="0;0.6049768163612267;1" dur="1s" repeatCount="indefinite" begin="-0.07188593356616135s"></animate>
+</circle><circle cx="52" cy="112.09746694041411" r="10" fill="#f8b26a">
+  <animate attributeName="cy" values="112.09746694041411;-2.8104821907767574" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.4132445270517203s"></animate>
+  <animate attributeName="r" values="10;0;0" keyTimes="0;0.7843188648425736;1" dur="1s" repeatCount="indefinite" begin="-0.4132445270517203s"></animate>
+</circle><circle cx="68" cy="119.76797510227266" r="15" fill="#f8b26a">
+  <animate attributeName="cy" values="119.76797510227266;-2.3187957684067317" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.6317748306797277s"></animate>
+  <animate attributeName="r" values="15;0;0" keyTimes="0;0.8464277838946668;1" dur="1s" repeatCount="indefinite" begin="-0.6317748306797277s"></animate>
+</circle><circle cx="17" cy="121.7997527406382" r="5" fill="#f8b26a">
+  <animate attributeName="cy" values="121.7997527406382;13.556957891026624" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.9136732084136533s"></animate>
+  <animate attributeName="r" values="5;0;0" keyTimes="0;0.5349721785314134;1" dur="1s" repeatCount="indefinite" begin="-0.9136732084136533s"></animate>
+</circle><circle cx="59" cy="116.30296558149124" r="4" fill="#f8b26a">
+  <animate attributeName="cy" values="116.30296558149124;-1.0433564145924477" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.08891813207741484s"></animate>
+  <animate attributeName="r" values="4;0;0" keyTimes="0;0.6574981312374213;1" dur="1s" repeatCount="indefinite" begin="-0.08891813207741484s"></animate>
+</circle><circle cx="88" cy="113.1583378513422" r="12" fill="#f8b26a">
+  <animate attributeName="cy" values="113.1583378513422;1.456869512308952" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.14992898603700067s"></animate>
+  <animate attributeName="r" values="12;0;0" keyTimes="0;0.9565108058771807;1" dur="1s" repeatCount="indefinite" begin="-0.14992898603700067s"></animate>
+</circle><circle cx="84" cy="112.41279273844411" r="10" fill="#f8b26a">
+  <animate attributeName="cy" values="112.41279273844411;1.6491176590177243" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.5833010262862421s"></animate>
+  <animate attributeName="r" values="10;0;0" keyTimes="0;0.5438806242531744;1" dur="1s" repeatCount="indefinite" begin="-0.5833010262862421s"></animate>
+</circle><circle cx="87" cy="120.26530337145327" r="5" fill="#f8b26a">
+  <animate attributeName="cy" values="120.26530337145327;9.388664939149207" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.05018189342538548s"></animate>
+  <animate attributeName="r" values="5;0;0" keyTimes="0;0.637897648645736;1" dur="1s" repeatCount="indefinite" begin="-0.05018189342538548s"></animate>
+</circle><circle cx="24" cy="123.99448894779877" r="9" fill="#f8b26a">
+  <animate attributeName="cy" values="123.99448894779877;2.3750067806866078" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.8890495329191316s"></animate>
+  <animate attributeName="r" values="9;0;0" keyTimes="0;0.663064102718458;1" dur="1s" repeatCount="indefinite" begin="-0.8890495329191316s"></animate>
+</circle><circle cx="73" cy="120.00019528994846" r="12" fill="#f8b26a">
+  <animate attributeName="cy" values="120.00019528994846;-9.503507375076166" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.6351313241419324s"></animate>
+  <animate attributeName="r" values="12;0;0" keyTimes="0;0.9354194941922095;1" dur="1s" repeatCount="indefinite" begin="-0.6351313241419324s"></animate>
+</circle><circle cx="74" cy="113.88820186698781" r="4" fill="#f8b26a">
+  <animate attributeName="cy" values="113.88820186698781;10.570535200732685" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.7132998998028989s"></animate>
+  <animate attributeName="r" values="4;0;0" keyTimes="0;0.91895021859856;1" dur="1s" repeatCount="indefinite" begin="-0.7132998998028989s"></animate>
+</circle><circle cx="68" cy="129.5841522641359" r="12" fill="#f8b26a">
+  <animate attributeName="cy" values="129.5841522641359;3.894919008898638" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.29330391921510546s"></animate>
+  <animate attributeName="r" values="12;0;0" keyTimes="0;0.9096568793749455;1" dur="1s" repeatCount="indefinite" begin="-0.29330391921510546s"></animate>
+</circle><circle cx="53" cy="119.31720358172306" r="9" fill="#f8b26a">
+  <animate attributeName="cy" values="119.31720358172306;9.73624644875764" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.9958245939061628s"></animate>
+  <animate attributeName="r" values="9;0;0" keyTimes="0;0.8571965277158554;1" dur="1s" repeatCount="indefinite" begin="-0.9958245939061628s"></animate>
+</circle><circle cx="76" cy="134.80739606982607" r="17" fill="#f8b26a">
+  <animate attributeName="cy" values="134.80739606982607;0.3932385595869441" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.8607153243461125s"></animate>
+  <animate attributeName="r" values="17;0;0" keyTimes="0;0.8654455107706405;1" dur="1s" repeatCount="indefinite" begin="-0.8607153243461125s"></animate>
+</circle><circle cx="75" cy="122.61568996754474" r="7" fill="#f8b26a">
+  <animate attributeName="cy" values="122.61568996754474;10.652526875734779" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.959721298983397s"></animate>
+  <animate attributeName="r" values="7;0;0" keyTimes="0;0.6271803990132601;1" dur="1s" repeatCount="indefinite" begin="-0.959721298983397s"></animate>
+</circle><circle cx="87" cy="115.0788054109218" r="12" fill="#f8b26a">
+  <animate attributeName="cy" values="115.0788054109218;-8.15567938666852" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.0690058777440068s"></animate>
+  <animate attributeName="r" values="12;0;0" keyTimes="0;0.6627211388649489;1" dur="1s" repeatCount="indefinite" begin="-0.0690058777440068s"></animate>
+</circle><circle cx="21" cy="118.08738171978098" r="9" fill="#f8b26a">
+  <animate attributeName="cy" values="118.08738171978098;-4.9475469075625504" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.7078831683260647s"></animate>
+  <animate attributeName="r" values="9;0;0" keyTimes="0;0.9501044367725069;1" dur="1s" repeatCount="indefinite" begin="-0.7078831683260647s"></animate>
+</circle><circle cx="24" cy="128.09150085659442" r="9" fill="#f8b26a">
+  <animate attributeName="cy" values="128.09150085659442;2.7320353690265122" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.521121701341132s"></animate>
+  <animate attributeName="r" values="9;0;0" keyTimes="0;0.7357531229285373;1" dur="1s" repeatCount="indefinite" begin="-0.521121701341132s"></animate>
+</circle><circle cx="26" cy="127.49368345428452" r="15" fill="#f8b26a">
+  <animate attributeName="cy" values="127.49368345428452;-10.361246269666196" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.9420307783603239s"></animate>
+  <animate attributeName="r" values="15;0;0" keyTimes="0;0.7467409545014994;1" dur="1s" repeatCount="indefinite" begin="-0.9420307783603239s"></animate>
+</circle><circle cx="39" cy="114.20744515306558" r="6" fill="#f8b26a">
+  <animate attributeName="cy" values="114.20744515306558;5.606516894440285" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.49268347147689695s"></animate>
+  <animate attributeName="r" values="6;0;0" keyTimes="0;0.5874854761603912;1" dur="1s" repeatCount="indefinite" begin="-0.49268347147689695s"></animate>
+</circle><circle cx="61" cy="123.10463246179438" r="11" fill="#f8b26a">
+  <animate attributeName="cy" values="123.10463246179438;-5.189366828773049" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.21359109324800063s"></animate>
+  <animate attributeName="r" values="11;0;0" keyTimes="0;0.6970744691674484;1" dur="1s" repeatCount="indefinite" begin="-0.21359109324800063s"></animate>
+</circle><circle cx="37" cy="115.40335155247101" r="10" fill="#f8b26a">
+  <animate attributeName="cy" values="115.40335155247101;3.4285850566842946" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.5344545499798534s"></animate>
+  <animate attributeName="r" values="10;0;0" keyTimes="0;0.9983685792824288;1" dur="1s" repeatCount="indefinite" begin="-0.5344545499798534s"></animate>
+</circle><circle cx="22" cy="124.59228223795324" r="7" fill="#f8b26a">
+  <animate attributeName="cy" values="124.59228223795324;-3.5076355130396912" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.8102510016775601s"></animate>
+  <animate attributeName="r" values="7;0;0" keyTimes="0;0.6369981578428732;1" dur="1s" repeatCount="indefinite" begin="-0.8102510016775601s"></animate>
+</circle><circle cx="34" cy="111.69621652751701" r="5" fill="#f8b26a">
+  <animate attributeName="cy" values="111.69621652751701;13.965538669421832" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.3819120829819431s"></animate>
+  <animate attributeName="r" values="5;0;0" keyTimes="0;0.9240036927970401;1" dur="1s" repeatCount="indefinite" begin="-0.3819120829819431s"></animate>
+</circle><circle cx="61" cy="121.99207528226256" r="6" fill="#f8b26a">
+  <animate attributeName="cy" values="121.99207528226256;-1.1884130816048284" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.351012424136126s"></animate>
+  <animate attributeName="r" values="6;0;0" keyTimes="0;0.9527855705617168;1" dur="1s" repeatCount="indefinite" begin="-0.351012424136126s"></animate>
+</circle><circle cx="32" cy="115.36386365084275" r="13" fill="#f8b26a">
+  <animate attributeName="cy" values="115.36386365084275;-7.635796261623495" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.22026693987990997s"></animate>
+  <animate attributeName="r" values="13;0;0" keyTimes="0;0.6822821982216503;1" dur="1s" repeatCount="indefinite" begin="-0.22026693987990997s"></animate>
+</circle><circle cx="38" cy="123.93260454500944" r="10" fill="#f8b26a">
+  <animate attributeName="cy" values="123.93260454500944;-9.019646946232784" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.5897767052001425s"></animate>
+  <animate attributeName="r" values="10;0;0" keyTimes="0;0.747643174639248;1" dur="1s" repeatCount="indefinite" begin="-0.5897767052001425s"></animate>
+</circle><circle cx="91" cy="111.20360670124936" r="4" fill="#f8b26a">
+  <animate attributeName="cy" values="111.20360670124936;-2.7511383786778185" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.5936715943771124s"></animate>
+  <animate attributeName="r" values="4;0;0" keyTimes="0;0.5292863982274825;1" dur="1s" repeatCount="indefinite" begin="-0.5936715943771124s"></animate>
+</circle><circle cx="93" cy="109.08688866758263" r="6" fill="#f8b26a">
+  <animate attributeName="cy" values="109.08688866758263;13.986514639855155" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.20182465253134418s"></animate>
+  <animate attributeName="r" values="6;0;0" keyTimes="0;0.9578727930035874;1" dur="1s" repeatCount="indefinite" begin="-0.20182465253134418s"></animate>
+</circle><circle cx="90" cy="115.44258946143852" r="3" fill="#f8b26a">
+  <animate attributeName="cy" values="115.44258946143852;7.971557449807172" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.8138344996352406s"></animate>
+  <animate attributeName="r" values="3;0;0" keyTimes="0;0.822677504532275;1" dur="1s" repeatCount="indefinite" begin="-0.8138344996352406s"></animate>
+</circle><circle cx="24" cy="130.98782632438636" r="15" fill="#f8b26a">
+  <animate attributeName="cy" values="130.98782632438636;-11.868426017755008" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.8574009914089539s"></animate>
+  <animate attributeName="r" values="15;0;0" keyTimes="0;0.8610318085552064;1" dur="1s" repeatCount="indefinite" begin="-0.8574009914089539s"></animate>
+</circle><circle cx="49" cy="122.24309971563434" r="14" fill="#f8b26a">
+  <animate attributeName="cy" values="122.24309971563434;3.5685994935617273" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.4267384904796552s"></animate>
+  <animate attributeName="r" values="14;0;0" keyTimes="0;0.5503829186981541;1" dur="1s" repeatCount="indefinite" begin="-0.4267384904796552s"></animate>
+</circle><circle cx="18" cy="117.38217971971676" r="9" fill="#f8b26a">
+  <animate attributeName="cy" values="117.38217971971676;6.631006164776416" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.6828218424869835s"></animate>
+  <animate attributeName="r" values="9;0;0" keyTimes="0;0.6808177575913787;1" dur="1s" repeatCount="indefinite" begin="-0.6828218424869835s"></animate>
+</circle><circle cx="78" cy="124.28678852303256" r="15" fill="#f8b26a">
+  <animate attributeName="cy" values="124.28678852303256;1.3740946843405304" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.4161035078940827s"></animate>
+  <animate attributeName="r" values="15;0;0" keyTimes="0;0.6388001474427218;1" dur="1s" repeatCount="indefinite" begin="-0.4161035078940827s"></animate>
+</circle><circle cx="44" cy="106.6189204965897" r="3" fill="#f8b26a">
+  <animate attributeName="cy" values="106.6189204965897;16.750815514807034" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.0510803765953457s"></animate>
+  <animate attributeName="r" values="3;0;0" keyTimes="0;0.7907276882734477;1" dur="1s" repeatCount="indefinite" begin="-0.0510803765953457s"></animate>
+</circle><circle cx="41" cy="119.64799537397232" r="5" fill="#f8b26a">
+  <animate attributeName="cy" values="119.64799537397232;6.398667601394809" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.4280945050279754s"></animate>
+  <animate attributeName="r" values="5;0;0" keyTimes="0;0.5751942250658201;1" dur="1s" repeatCount="indefinite" begin="-0.4280945050279754s"></animate>
+</circle><circle cx="19" cy="120.0916729802829" r="10" fill="#f8b26a">
+  <animate attributeName="cy" values="120.0916729802829;-9.513704965243033" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.043405970368113445s"></animate>
+  <animate attributeName="r" values="10;0;0" keyTimes="0;0.5435267537060107;1" dur="1s" repeatCount="indefinite" begin="-0.043405970368113445s"></animate>
+</circle><circle cx="61" cy="123.62714133794762" r="5" fill="#f8b26a">
+  <animate attributeName="cy" values="123.62714133794762;2.362315551662477" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.5256540407430482s"></animate>
+  <animate attributeName="r" values="5;0;0" keyTimes="0;0.9222037100732456;1" dur="1s" repeatCount="indefinite" begin="-0.5256540407430482s"></animate>
+</circle><circle cx="64" cy="115.25525614926073" r="13" fill="#f8b26a">
+  <animate attributeName="cy" values="115.25525614926073;-10.304511881341815" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.6633519944592159s"></animate>
+  <animate attributeName="r" values="13;0;0" keyTimes="0;0.5401283508859178;1" dur="1s" repeatCount="indefinite" begin="-0.6633519944592159s"></animate>
+</circle><circle cx="12" cy="129.13660549492693" r="11" fill="#f8b26a">
+  <animate attributeName="cy" values="129.13660549492693;-7.965594883525825" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.9929282227674491s"></animate>
+  <animate attributeName="r" values="11;0;0" keyTimes="0;0.9536114994321867;1" dur="1s" repeatCount="indefinite" begin="-0.9929282227674491s"></animate>
+</circle><circle cx="39" cy="106.95504126040025" r="2" fill="#f8b26a">
+  <animate attributeName="cy" values="106.95504126040025;5.834416891524681" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.22005892301327157s"></animate>
+  <animate attributeName="r" values="2;0;0" keyTimes="0;0.6089960643653531;1" dur="1s" repeatCount="indefinite" begin="-0.22005892301327157s"></animate>
+</circle><circle cx="30" cy="112.12744151244388" r="8" fill="#f8b26a">
+  <animate attributeName="cy" values="112.12744151244388;-4.465606537168944" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.24710322548242414s"></animate>
+  <animate attributeName="r" values="8;0;0" keyTimes="0;0.7479705418636007;1" dur="1s" repeatCount="indefinite" begin="-0.24710322548242414s"></animate>
+</circle><circle cx="67" cy="124.83294711941956" r="16" fill="#f8b26a">
+  <animate attributeName="cy" values="124.83294711941956;-7.6291463245052284" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.614066023590482s"></animate>
+  <animate attributeName="r" values="16;0;0" keyTimes="0;0.7584434636145084;1" dur="1s" repeatCount="indefinite" begin="-0.614066023590482s"></animate>
+</circle><circle cx="22" cy="119.36463088979876" r="4" fill="#f8b26a">
+  <animate attributeName="cy" values="119.36463088979876;12.12664234343379" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.527385385953813s"></animate>
+  <animate attributeName="r" values="4;0;0" keyTimes="0;0.5661680148267347;1" dur="1s" repeatCount="indefinite" begin="-0.527385385953813s"></animate>
+</circle><circle cx="12" cy="122.52124979151506" r="7" fill="#f8b26a">
+  <animate attributeName="cy" values="122.52124979151506;3.7506712743784085" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.37225883133903837s"></animate>
+  <animate attributeName="r" values="7;0;0" keyTimes="0;0.9003327357718601;1" dur="1s" repeatCount="indefinite" begin="-0.37225883133903837s"></animate>
+</circle><circle cx="69" cy="130.5210986475815" r="14" fill="#f8b26a">
+  <animate attributeName="cy" values="130.5210986475815;-0.30973651460238827" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.6062299863585278s"></animate>
+  <animate attributeName="r" values="14;0;0" keyTimes="0;0.9220180768904789;1" dur="1s" repeatCount="indefinite" begin="-0.6062299863585278s"></animate>
+</circle><circle cx="20" cy="114.80243604193255" r="9" fill="#f8b26a">
+  <animate attributeName="cy" values="114.80243604193255;7.19374553530416" keyTimes="0;1" dur="1s" repeatCount="indefinite" begin="-0.6866227460985781s"></animate>
+  <animate attributeName="r" values="9;0;0" keyTimes="0;0.6690048284116141;1" dur="1s" repeatCount="indefinite" begin="-0.6866227460985781s"></animate>
+</circle></g>
+</svg>

assets/logo.svg

@@ -0,0 +1,13 @@
+<svg xmlns="http://www.w3.org/2000/svg" width="354" height="100" viewBox="0 0 354 100">
+
+  <text
+    x="50%"
+    y="50%"
+    font-family="Arial, sans-serif"
+    font-size="55"
+    text-anchor="middle"
+    dominant-baseline="middle">
+      <tspan fill="#F37726">{</tspan><tspan fill="#616161">Tex</tspan><tspan fill="#F37726">}</tspan><tspan fill="#616161">Teller</tspan>
+  </text>
+
+</svg>

docs/Makefile

@@ -0,0 +1,20 @@
+# Minimal makefile for Sphinx documentation
+#
+
+# You can set these variables from the command line, and also
+# from the environment for the first two.
+SPHINXOPTS    ?=
+SPHINXBUILD   ?= sphinx-build
+SOURCEDIR     = source
+BUILDDIR      = build
+
+# Put it first so that "make" without argument is like "make help".
+help:
+	@$(SPHINXBUILD) -M help "$(SOURCEDIR)" "$(BUILDDIR)" $(SPHINXOPTS) $(O)
+
+.PHONY: help Makefile
+
+# Catch-all target: route all unknown targets to Sphinx using the new
+# "make mode" option.  $(O) is meant as a shortcut for $(SPHINXOPTS).
+%: Makefile
+	@$(SPHINXBUILD) -M $@ "$(SOURCEDIR)" "$(BUILDDIR)" $(SPHINXOPTS) $(O)

docs/make.bat

@@ -0,0 +1,35 @@
+@ECHO OFF
+
+pushd %~dp0
+
+REM Command file for Sphinx documentation
+
+if "%SPHINXBUILD%" == "" (
+	set SPHINXBUILD=sphinx-build
+)
+set SOURCEDIR=source
+set BUILDDIR=build
+
+%SPHINXBUILD% >NUL 2>NUL
+if errorlevel 9009 (
+	echo.
+	echo.The 'sphinx-build' command was not found. Make sure you have Sphinx
+	echo.installed, then set the SPHINXBUILD environment variable to point
+	echo.to the full path of the 'sphinx-build' executable. Alternatively you
+	echo.may add the Sphinx directory to PATH.
+	echo.
+	echo.If you don't have Sphinx installed, grab it from
+	echo.https://www.sphinx-doc.org/
+	exit /b 1
+)
+
+if "%1" == "" goto help
+
+%SPHINXBUILD% -M %1 %SOURCEDIR% %BUILDDIR% %SPHINXOPTS% %O%
+goto end
+
+:help
+%SPHINXBUILD% -M help %SOURCEDIR% %BUILDDIR% %SPHINXOPTS% %O%
+
+:end
+popd

docs/source/api.rst

@@ -0,0 +1,39 @@
+API Reference
+=============
+
+This section provides detailed API documentation for the TexTeller package. TexTeller is a tool for detecting and recognizing LaTeX formulas in images and converting mixed text and formula images to markdown.
+
+.. contents:: Table of Contents
+   :local:
+   :depth: 2
+
+
+Image to LaTeX Conversion
+-------------------------
+
+.. autofunction:: texteller.api.img2latex
+
+Paragraph to Markdown Conversion
+------------------------------
+
+.. autofunction:: texteller.api.paragraph2md
+
+LaTeX Detection
+---------------
+
+.. autofunction:: texteller.api.detection.latex_detect
+
+Model Loading
+-------------
+
+.. autofunction:: texteller.api.load_model
+.. autofunction:: texteller.api.load_tokenizer
+.. autofunction:: texteller.api.load_latexdet_model
+.. autofunction:: texteller.api.load_textdet_model
+.. autofunction:: texteller.api.load_textrec_model
+
+
+KaTeX Conversion
+----------------
+
+.. autofunction:: texteller.api.to_katex

docs/source/conf.py

@@ -0,0 +1,75 @@
+# Configuration file for the Sphinx documentation builder.
+#
+# This file only contains a selection of the most common options. For a full
+# list see the documentation:
+# https://www.sphinx-doc.org/en/master/usage/configuration.html
+
+# -- Path setup --------------------------------------------------------------
+
+# If extensions (or modules to document with autodoc) are in another directory,
+# add these directories to sys.path here. If the directory is relative to the
+# documentation root, use os.path.abspath to make it absolute.
+import os
+import sys
+
+sys.path.insert(0, os.path.abspath('../..'))
+
+# -- Project information -----------------------------------------------------
+
+project = 'TexTeller'
+copyright = '2025, TexTeller Team'
+author = 'TexTeller Team'
+
+# -- General configuration ---------------------------------------------------
+# https://www.sphinx-doc.org/en/master/usage/configuration.html#general-configuration
+
+extensions = [
+    'myst_parser',
+    'sphinx.ext.duration',
+    'sphinx.ext.intersphinx',
+    'sphinx.ext.autosectionlabel',
+    'sphinx.ext.autodoc',
+    'sphinx.ext.viewcode',
+    'sphinx.ext.napoleon',
+    'sphinx.ext.autosummary',
+    'sphinx_copybutton',
+    # 'sphinx.ext.linkcode',
+    # 'sphinxarg.ext',
+    'sphinx_design',
+    'nbsphinx',
+]
+
+templates_path = ['_templates']
+exclude_patterns = []
+
+# Autodoc settings
+autodoc_member_order = 'bysource'
+add_module_names = False
+autoclass_content = 'both'
+autodoc_default_options = {
+    'members': True,
+    'member-order': 'bysource',
+    'undoc-members': True,
+    'show-inheritance': True,
+    'imported-members': True,
+}
+
+# Intersphinx settings
+intersphinx_mapping = {
+    'python': ('https://docs.python.org/3', None),
+    'numpy': ('https://numpy.org/doc/stable', None),
+    'torch': ('https://pytorch.org/docs/stable', None),
+    'transformers': ('https://huggingface.co/docs/transformers/main/en', None),
+}
+
+html_theme = 'sphinx_book_theme'
+
+html_theme_options = {
+    'repository_url': 'https://github.com/OleehyO/TexTeller',
+    'use_repository_button': True,
+    'use_issues_button': True,
+    'use_edit_page_button': True,
+    'use_download_button': True,
+}
+
+html_logo = "../../assets/logo.svg"

docs/source/index.rst

@@ -0,0 +1,76 @@
+.. TexTeller documentation master file, created by
+   sphinx-quickstart on Sun Apr 20 13:05:53 2025.
+   You can adapt this file completely to your liking, but it should at least
+   contain the root `toctree` directive.
+
+TexTeller Documentation
+===========================================
+
+Features
+--------
+
+- **Image to LaTeX Conversion**: Convert images containing LaTeX formulas to LaTeX code
+- **LaTeX Detection**: Detect and locate LaTeX formulas in mixed text/formula images
+- **Paragraph to Markdown**: Convert mixed text and formula images to Markdown format
+
+Installation
+-----------
+
+You can install TexTeller using pip:
+
+.. code-block:: bash
+
+   pip install texteller
+
+Quick Start
+----------
+
+Converting an image to LaTeX:
+
+.. code-block:: python
+
+   from texteller import load_model, load_tokenizer, img2latex
+
+   # Load models
+   model = load_model(use_onnx=False)
+   tokenizer = load_tokenizer()
+
+   # Convert image to LaTeX
+   latex = img2latex(model, tokenizer, ["path/to/image.png"])[0]
+
+Processing a mixed text/formula image:
+
+.. code-block::python
+
+   from texteller import (
+       load_model, load_tokenizer, load_latexdet_model,
+       load_textdet_model, load_textrec_model, paragraph2md
+   )
+
+   # Load all required models
+   latex_model = load_model()
+   tokenizer = load_tokenizer()
+   latex_detector = load_latexdet_model()
+   text_detector = load_textdet_model()
+   text_recognizer = load_textrec_model()
+
+   # Convert to markdown
+   markdown = paragraph2md(
+       "path/to/mixed_image.png",
+       latex_detector,
+       text_detector,
+       text_recognizer,
+       latex_model,
+       tokenizer
+   )
+
+API Documentation
+----------------
+
+For detailed API documentation, please see :doc:`./api`.
+
+.. toctree::
+   :maxdepth: 2
+   :hidden:
+
+   api

examples/client_demo.py

@@ -0,0 +1,10 @@
+import requests
+
+server_url = "http://127.0.0.1:8000/predict"
+
+img_path = "/path/to/your/image"
+with open(img_path, 'rb') as img:
+    files = {'img': img}
+    response = requests.post(server_url, files=files)
+
+print(response.text)

examples/train_texteller/dataset/train/metadata.jsonl

@@ -0,0 +1,35 @@
+{"file_name": "0.png", "latex_formula": "\\[\\mathbb{C}^{4}\\stackrel{{\\pi_{1}}}{{\\longleftarrow}}\\mathcal{ F}\\stackrel{{\\pi_{2}}}{{\\rightarrow}}\\mathcal{PT},\\]"}
+{"file_name": "1.png", "latex_formula": "\\[W^{*}_{Z}(x_{1},x_{2})=W_{f\\lrcorner Z}(y_{1},y_{2})=\\mathcal{P}\\exp\\left( \\int_{\\gamma}A_{\\mu}dx^{\\mu}\\right).\\]"}
+{"file_name": "2.png", "latex_formula": "\\[G=W^{*}_{Z}(q,p)=\\tilde{H}H^{-1}\\]"}
+{"file_name": "3.png", "latex_formula": "\\[H=W^{*}_{Z}(p,x),\\ \\ \\tilde{H}=W^{*}_{Z}(q,x).\\]"}
+{"file_name": "4.png", "latex_formula": "\\[v\\cdot f^{*}A|_{x}=(f\\lrcorner Z)_{*}v\\cdot A|_{f\\lrcorner Z(x)},\\quad x\\in Z, \\ v\\in T_{x}Z.\\]"}
+{"file_name": "5.png", "latex_formula": "\\[(f\\lrcorner Z)_{*}v\\cdot A|_{f\\lrcorner Z(x)}=v^{\\alpha\\dot{\\alpha}}\\Big{(} \\frac{\\partial y^{\\beta\\dot{\\beta}}}{\\partial x^{\\alpha\\dot{\\alpha}}}A_{\\beta \\dot{\\beta}}\\Big{)}\\Big{|}_{f\\lrcorner Z(x)},\\ x\\in Z,\\ v\\in T_{x}Z,\\]"}
+{"file_name": "6.png", "latex_formula": "\\[\\{T_{i},T_{j}\\}=\\{\\tilde{T}^{i},\\tilde{T}^{j}\\}=0,\\ \\ \\{T_{i},\\tilde{T}^{j}\\}=2i \\delta^{j}_{i}D,\\]"}
+{"file_name": "7.png", "latex_formula": "\\[(\\partial_{s},q_{i},\\tilde{q}^{k})\\rightarrow(D,M^{j}_{i}T_{j},\\tilde{M}^{k}_ {l}\\tilde{T}^{l}),\\]"}
+{"file_name": "8.png", "latex_formula": "\\[M^{i}_{j}\\tilde{M}^{j}_{k}=\\delta^{i}_{k}.\\]"}
+{"file_name": "9.png", "latex_formula": "\\[Q_{i\\alpha}=q_{i\\alpha}+\\omega_{i\\alpha},\\ \\tilde{Q}^{i}_{\\dot{\\alpha}}=q^{i}_{ \\dot{\\alpha}}+\\tilde{\\omega}^{i}_{\\dot{\\alpha}},\\ D_{\\alpha\\dot{\\alpha}}= \\partial_{\\alpha\\dot{\\alpha}}+A_{\\alpha\\dot{\\alpha}}.\\]"}
+{"file_name": "10.png", "latex_formula": "\\[\\hat{f}(g,\\theta^{i\\alpha},\\tilde{\\theta}^{\\dot{\\alpha}}_{j})=(f(g),[V^{-1}]^ {\\alpha}_{\\beta}\\theta^{i\\beta},[\\tilde{V}^{-1}]^{\\dot{\\alpha}}_{\\dot{\\beta}} \\tilde{\\theta}^{\\dot{\\beta}}_{j}),\\ g\\in{\\cal G},\\]"}
+{"file_name": "11.png", "latex_formula": "\\[v^{\\beta\\dot{\\beta}}V^{\\alpha}_{\\beta}\\tilde{V}^{\\dot{\\alpha}}_{\\dot{\\beta}} =((f\\lrcorner L_{0})_{*}v)^{\\alpha\\dot{\\alpha}},\\]"}
+{"file_name": "12.png", "latex_formula": "\\[\\omega_{i\\alpha}=\\tilde{\\theta}^{\\dot{\\alpha}}_{i}h_{\\alpha\\dot{\\alpha}}(x^{ \\beta\\dot{\\beta}},\\tau^{\\beta\\dot{\\beta}}),\\ \\ \\tilde{\\omega}^{i}_{\\alpha}=\\theta^{i\\alpha}\\tilde{h}_{\\alpha\\dot{\\alpha}}(x^{ \\beta\\dot{\\beta}},\\tau^{\\beta\\dot{\\beta}}),\\]"}
+{"file_name": "13.png", "latex_formula": "\\[\\begin{split}&\\lambda^{\\alpha}\\hat{f}^{*}\\omega_{i\\alpha}(z)= \\tilde{\\theta}^{\\dot{\\beta}}_{i}\\lambda^{\\alpha}\\left(V^{\\beta}_{\\alpha}h_{ \\beta\\dot{\\beta}}(x^{\\prime},\\tau^{\\prime})\\right),\\\\ &\\tilde{\\lambda}^{\\dot{\\alpha}}\\hat{f}^{*}\\tilde{\\omega}^{i}_{ \\dot{\\alpha}}(z)=\\theta^{i\\beta}\\tilde{\\lambda}^{\\dot{\\alpha}}\\left(\\tilde{V}^ {\\dot{\\beta}}_{\\dot{\\alpha}}\\tilde{h}_{\\beta\\dot{\\beta}}(x^{\\prime},\\tau^{ \\prime})\\right),\\end{split}\\]"}
+{"file_name": "14.png", "latex_formula": "\\[A_{\\alpha\\dot{\\alpha}}=A_{\\alpha\\dot{\\alpha}}(x^{\\beta\\dot{\\beta}},\\tau^{ \\beta\\dot{\\beta}})\\]"}
+{"file_name": "15.png", "latex_formula": "\\[D=\\lambda^{\\alpha}\\tilde{\\lambda}^{\\dot{\\alpha}}D_{\\alpha\\dot{\\alpha}}\\]"}
+{"file_name": "16.png", "latex_formula": "\\[D=\\lambda^{\\alpha}\\tilde{\\lambda}^{\\dot{\\alpha}}\\partial_{\\alpha\\dot{\\alpha}}\\]"}
+{"file_name": "17.png", "latex_formula": "\\[[v_{1}\\cdot D^{*},v_{2}\\cdot D^{*}]=0\\]"}
+{"file_name": "18.png", "latex_formula": "\\[\\Phi_{A}=(\\omega_{i\\alpha},\\tilde{\\omega}^{i}_{\\dot{\\alpha}},A_{\\alpha\\dot{ \\alpha}})\\]"}
+{"file_name": "19.png", "latex_formula": "\\[\\hat{f}:{\\cal F}^{6|4N}\\rightarrow{\\cal F}^{6|4N}\\]"}
+{"file_name": "20.png", "latex_formula": "\\[\\sigma=(s,\\xi^{i},\\tilde{\\xi}_{j})\\in\\mathbb{C}^{1|2N}\\]"}
+{"file_name": "21.png", "latex_formula": "\\[\\tau^{\\alpha\\dot{\\alpha}}(h_{\\alpha\\dot{\\alpha}}+\\tilde{h}_{\\alpha\\dot{\\alpha} })=0\\]"}
+{"file_name": "22.png", "latex_formula": "\\[\\tau^{\\alpha\\dot{\\alpha}}\\rightarrow[V^{-1}]^{\\alpha}_{\\beta}[\\tilde{V}^{-1}]^{ \\dot{\\alpha}}_{\\dot{\\beta}}\\tau^{\\beta\\dot{\\beta}}\\]"}
+{"file_name": "23.png", "latex_formula": "\\[\\tau^{\\beta\\dot{\\beta}}=\\sum_{i}\\theta^{i\\beta}\\tilde{\\theta}^{\\dot{\\beta}}_{i}\\]"}
+{"file_name": "24.png", "latex_formula": "\\[\\theta^{i\\alpha}\\omega_{i\\alpha}+\\tilde{\\theta}^{i}_{\\dot{\\alpha}}\\tilde{ \\omega}^{\\dot{\\alpha}}_{i}=0\\]"}
+{"file_name": "25.png", "latex_formula": "\\[\\tilde{T}^{i}=\\tilde{\\lambda}^{\\dot{\\alpha}}\\tilde{Q}^{i}_{\\dot{\\alpha}}\\]"}
+{"file_name": "26.png", "latex_formula": "\\[\\tilde{T}^{i}=\\tilde{\\lambda}^{\\dot{\\alpha}}\\tilde{q}^{i}_{\\dot{\\alpha}}\\]"}
+{"file_name": "27.png", "latex_formula": "\\[\\tilde{\\lambda}^{\\dot{\\alpha}}f^{*}A_{\\alpha\\dot{\\alpha}}=H^{-1}\\tilde{ \\lambda}^{\\dot{\\alpha}}\\partial_{\\alpha\\dot{\\alpha}}H\\]"}
+{"file_name": "28.png", "latex_formula": "\\[\\tilde{q}^{i}=\\partial_{\\tilde{\\xi}_{i}}+i\\xi^{i}\\partial_{s}\\]"}
+{"file_name": "29.png", "latex_formula": "\\[\\tilde{q}^{i}_{\\dot{\\alpha}}=\\frac{\\partial}{\\partial\\tilde{\\theta}^{\\dot{ \\alpha}}_{i}}+i\\theta^{i\\alpha}\\frac{\\partial}{\\partial x^{\\alpha\\dot{\\alpha}}}\\]"}
+{"file_name": "30.png", "latex_formula": "\\[f\\lrcorner L(z)=\\pi_{1}\\circ f(z,\\lambda,\\tilde{\\lambda})\\ \\forall z\\in L\\]"}
+{"file_name": "31.png", "latex_formula": "\\[q_{i\\alpha}=\\frac{\\partial}{\\partial\\theta^{i\\alpha}}+i\\tilde{\\theta}^{\\dot{ \\alpha}}_{i}\\frac{\\partial}{\\partial x^{\\alpha\\dot{\\alpha}}}\\]"}
+{"file_name": "32.png", "latex_formula": "\\[q_{i}=\\partial_{\\xi^{i}}+i\\tilde{\\xi}_{i}\\partial_{s}\\]"}
+{"file_name": "33.png", "latex_formula": "\\[v^{\\alpha\\dot{\\alpha}}=\\lambda^{\\alpha}\\tilde{\\lambda}^{\\dot{\\alpha}}\\]"}
+{"file_name": "34.png", "latex_formula": "\\[z^{A}=(x^{\\alpha\\dot{\\alpha}},\\theta^{i\\alpha},\\tilde{\\theta}^{\\dot{\\alpha}}_{ j})\\]"}

examples/train_texteller/train.py

@@ -0,0 +1,71 @@
+from functools import partial
+
+import yaml
+from datasets import load_dataset
+from transformers import (
+    Trainer,
+    TrainingArguments,
+)
+
+from texteller import load_model, load_tokenizer
+from texteller.constants import MIN_HEIGHT, MIN_WIDTH
+
+from examples.train_texteller.utils import (
+    collate_fn,
+    filter_fn,
+    img_inf_transform,
+    img_train_transform,
+    tokenize_fn,
+)
+
+
+def train(model, tokenizer, train_dataset, eval_dataset, collate_fn_with_tokenizer):
+    training_args = TrainingArguments(**training_config)
+    trainer = Trainer(
+        model,
+        training_args,
+        train_dataset=train_dataset,
+        eval_dataset=eval_dataset,
+        tokenizer=tokenizer,
+        data_collator=collate_fn_with_tokenizer,
+    )
+
+    trainer.train(resume_from_checkpoint=None)
+
+
+if __name__ == "__main__":
+    dataset = load_dataset("imagefolder", data_dir="dataset")["train"]
+    dataset = dataset.filter(
+        lambda x: x["image"].height > MIN_HEIGHT and x["image"].width > MIN_WIDTH
+    )
+    dataset = dataset.shuffle(seed=42)
+    dataset = dataset.flatten_indices()
+
+    tokenizer = load_tokenizer()
+    # If you want use your own tokenizer, please modify the path to your tokenizer
+    # tokenizer = load_tokenizer("/path/to/your/tokenizer")
+    filter_fn_with_tokenizer = partial(filter_fn, tokenizer=tokenizer)
+    dataset = dataset.filter(filter_fn_with_tokenizer, num_proc=8)
+
+    map_fn = partial(tokenize_fn, tokenizer=tokenizer)
+    tokenized_dataset = dataset.map(
+        map_fn, batched=True, remove_columns=dataset.column_names, num_proc=8
+    )
+
+    # Split dataset into train and eval, ratio 9:1
+    split_dataset = tokenized_dataset.train_test_split(test_size=0.1, seed=42)
+    train_dataset, eval_dataset = split_dataset["train"], split_dataset["test"]
+    train_dataset = train_dataset.with_transform(img_train_transform)
+    eval_dataset = eval_dataset.with_transform(img_inf_transform)
+    collate_fn_with_tokenizer = partial(collate_fn, tokenizer=tokenizer)
+
+    # Train from scratch
+    model = load_model()
+
+    # If you want to train from pre-trained model, please modify the path to your pre-trained checkpoint
+    # model = load_model("/path/to/your/model_checkpoint")
+
+    enable_train = True
+    training_config = yaml.safe_load(open("train_config.yaml"))
+    if enable_train:
+        train(model, tokenizer, train_dataset, eval_dataset, collate_fn_with_tokenizer)

examples/train_texteller/train_config.yaml

@@ -0,0 +1,32 @@
+# For more information, please refer to the official documentation: https://huggingface.co/docs/transformers/main/en/main_classes/trainer#transformers.TrainingArguments
+
+seed: 42  # Random seed for reproducibility
+use_cpu: false  # Whether to use CPU (it's easier to debug with CPU when starting to test the code)
+learning_rate: 5.0e-5  # Learning rate
+num_train_epochs: 10  # Total number of training epochs
+per_device_train_batch_size: 4  # Batch size per GPU for training
+per_device_eval_batch_size: 8  # Batch size per GPU for evaluation
+output_dir: "train_result"  # Output directory
+overwrite_output_dir: false  # If the output directory exists, do not delete its content
+report_to:
+  - tensorboard  # Report logs to TensorBoard
+save_strategy: "steps"  # Strategy to save checkpoints
+save_steps: 500  # Interval of steps to save checkpoints, can be int or a float (0~1), when float it represents the ratio of total training steps (e.g., can set to 1.0 / 2000)
+save_total_limit: 5  # Maximum number of models to save. The oldest models will be deleted if this number is exceeded
+logging_strategy: "steps"  # Log every certain number of steps
+logging_steps: 500  # Number of steps between each log
+logging_nan_inf_filter: false  # Record logs for loss=nan or inf
+optim: "adamw_torch"  # Optimizer
+lr_scheduler_type: "cosine"  # Learning rate scheduler
+warmup_ratio: 0.1  # Ratio of warmup steps in total training steps (e.g., for 1000 steps, the first 100 steps gradually increase lr from 0 to the set lr)
+max_grad_norm: 1.0  # For gradient clipping, ensure the norm of the gradients does not exceed 1.0 (default 1.0)
+fp16: false  # Whether to use 16-bit floating point for training (generally not recommended, as loss can easily explode)
+bf16: false  # Whether to use Brain Floating Point (bfloat16) for training (recommended if architecture supports it)
+gradient_accumulation_steps: 1  # Gradient accumulation steps, consider this parameter to achieve large batch size effects when batch size cannot be large
+jit_mode_eval: false  # Whether to use PyTorch jit trace during eval (can speed up the model, but the model must be static, otherwise will throw errors)
+torch_compile: false  # Whether to use torch.compile to compile the model (for better training and inference performance)
+dataloader_pin_memory: true  # Can speed up data transfer between CPU and GPU
+dataloader_num_workers: 1  # Default is not to use multiprocessing for data loading, usually set to 4*number of GPUs used
+evaluation_strategy: "steps"  # Evaluation strategy, can be "steps" or "epoch"
+eval_steps: 500  # If evaluation_strategy="step"
+remove_unused_columns: false  # Don't change this unless you really know what you are doing.

examples/train_texteller/utils/__init__.py

@@ -0,0 +1,17 @@
+from .functional import (
+    collate_fn,
+    filter_fn,
+    tokenize_fn,
+)
+from .transforms import (
+    img_train_transform,
+    img_inf_transform,
+)
+
+__all__ = [
+    "collate_fn",
+    "filter_fn",
+    "tokenize_fn",
+    "img_train_transform",
+    "img_inf_transform",
+]

examples/train_texteller/utils/augraphy_pipe.py

@@ -1,14 +1,40 @@
-from augraphy import *
+"""
+Custom augraphy pipeline for training
+
+This file implements a custom augraphy data augmentation pipeline. We found that using augraphy's
+default pipeline can cause significant degradation to formula images, potentially losing semantic
+information. Therefore, we carefully selected several common augmentation effects,
+adjusting their parameters and combination methods to preserve the original semantic information
+of the images as much as possible.
+"""
+
+from augraphy import (
+    InkColorSwap,
+    LinesDegradation,
+    OneOf,
+    Dithering,
+    InkBleed,
+    InkShifter,
+    NoiseTexturize,
+    BrightnessTexturize,
+    ColorShift,
+    DirtyDrum,
+    LightingGradient,
+    Brightness,
+    Gamma,
+    SubtleNoise,
+    Jpeg,
+    AugraphyPipeline,
+)
 import random
 
-def ocr_augmentation_pipeline():
-    pre_phase = [
-        # Rescale(scale="optimal", target_dpi = 300,  p = 1.0),
-    ]
+
+def get_custom_augraphy():
+    pre_phase = []
 
     ink_phase = [
         InkColorSwap(
-            ink_swap_color="lhy_custom",
+            ink_swap_color="random",
             ink_swap_sequence_number_range=(5, 10),
             ink_swap_min_width_range=(2, 3),
             ink_swap_max_width_range=(100, 120),
@@ -16,7 +42,7 @@ def ocr_augmentation_pipeline():
             ink_swap_max_height_range=(100, 120),
             ink_swap_min_area_range=(10, 20),
             ink_swap_max_area_range=(400, 500),
-            p=0.2
+            p=0.2,
         ),
         LinesDegradation(
             line_roi=(0.0, 0.0, 1.0, 1.0),
@@ -28,9 +54,8 @@ def ocr_augmentation_pipeline():
             line_long_to_short_ratio=(5, 7),
             line_replacement_probability=(0.4, 0.5),
             line_replacement_thickness=(1, 3),
-            p=0.2
+            p=0.2,
         ),
-
         #  ============================
         OneOf(
             [
@@ -44,10 +69,9 @@ def ocr_augmentation_pipeline():
                     severity=(0.4, 0.6),
                 ),
             ],
-            p=0.2
+            p=0.2,
         ),
         #  ============================
-
         #  ============================
         InkShifter(
             text_shift_scale_range=(18, 27),
@@ -56,38 +80,32 @@ def ocr_augmentation_pipeline():
             blur_kernel_size=(5, 5),
             blur_sigma=0,
             noise_type="perlin",
-            p=0.2
+            p=0.2,
         ),
         #  ============================
-
     ]
 
     paper_phase = [
-        NoiseTexturize(  # tested
+        NoiseTexturize(
             sigma_range=(3, 10),
             turbulence_range=(2, 5),
             texture_width_range=(300, 500),
             texture_height_range=(300, 500),
-            p=0.2
+            p=0.2,
         ),
-        BrightnessTexturize(  # tested
-            texturize_range=(0.9, 0.99),
-            deviation=0.03,
-            p=0.2
-        )
+        BrightnessTexturize(texturize_range=(0.9, 0.99), deviation=0.03, p=0.2),
     ]
 
     post_phase = [
-        ColorShift(  # tested
+        ColorShift(
             color_shift_offset_x_range=(3, 5),
             color_shift_offset_y_range=(3, 5),
             color_shift_iterations=(2, 3),
             color_shift_brightness_range=(0.9, 1.1),
             color_shift_gaussian_kernel_range=(3, 3),
-            p=0.2
+            p=0.2,
         ),
-
-        DirtyDrum(  # tested
+        DirtyDrum(
             line_width_range=(1, 6),
             line_concentration=random.uniform(0.05, 0.15),
             direction=random.randint(0, 2),
@@ -95,9 +113,8 @@ def ocr_augmentation_pipeline():
             noise_value=(64, 224),
             ksize=random.choice([(3, 3), (5, 5), (7, 7)]),
             sigmaX=0,
-            p=0.2
+            p=0.2,
         ),
-
         # =====================================
         OneOf(
             [
@@ -119,10 +136,9 @@ def ocr_augmentation_pipeline():
                     gamma_range=(0.9, 1.1),
                 ),
             ],
-            p=0.2
+            p=0.2,
         ),
         # =====================================
-
         # =====================================
         OneOf(
             [
@@ -130,10 +146,10 @@ def ocr_augmentation_pipeline():
                     subtle_range=random.randint(5, 10),
                 ),
                 Jpeg(
-                    quality_range=(85, 95),
+                    quality_range=(70, 95),
                 ),
             ],
-            p=0.2
+            p=0.2,
         ),
         # =====================================
     ]
@@ -143,7 +159,7 @@ def ocr_augmentation_pipeline():
         paper_phase=paper_phase,
         post_phase=post_phase,
         pre_phase=pre_phase,
-        log=False
+        log=False,
     )
 
     return pipeline

examples/train_texteller/utils/functional.py

@@ -0,0 +1,47 @@
+from typing import Any
+
+import torch
+from transformers import DataCollatorForLanguageModeling
+
+from texteller.constants import MAX_TOKEN_SIZE, MIN_HEIGHT, MIN_WIDTH
+
+
+def _left_move(x: torch.Tensor, pad_val):
+    assert len(x.shape) == 2, "x should be 2-dimensional"
+    lefted_x = torch.ones_like(x)
+    lefted_x[:, :-1] = x[:, 1:]
+    lefted_x[:, -1] = pad_val
+    return lefted_x
+
+
+def tokenize_fn(samples: dict[str, list[Any]], tokenizer=None) -> dict[str, list[Any]]:
+    assert tokenizer is not None, "tokenizer should not be None"
+    tokenized_formula = tokenizer(samples["latex_formula"], return_special_tokens_mask=True)
+    tokenized_formula["pixel_values"] = samples["image"]
+    return tokenized_formula
+
+
+def collate_fn(samples: list[dict[str, Any]], tokenizer=None) -> dict[str, list[Any]]:
+    assert tokenizer is not None, "tokenizer should not be None"
+    pixel_values = [dic.pop("pixel_values") for dic in samples]
+
+    clm_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False)
+
+    batch = clm_collator(samples)
+    batch["pixel_values"] = pixel_values
+    batch["decoder_input_ids"] = batch.pop("input_ids")
+    batch["decoder_attention_mask"] = batch.pop("attention_mask")
+
+    batch["labels"] = _left_move(batch["labels"], -100)
+
+    # convert list of Image to a tensor with (B, C, H, W)
+    batch["pixel_values"] = torch.stack(batch["pixel_values"], dim=0)
+    return batch
+
+
+def filter_fn(sample, tokenizer=None) -> bool:
+    return (
+        sample["image"].height > MIN_HEIGHT
+        and sample["image"].width > MIN_WIDTH
+        and len(tokenizer(sample["latex_formula"])["input_ids"]) < MAX_TOKEN_SIZE - 10
+    )

examples/train_texteller/utils/transforms.py

@@ -0,0 +1,154 @@
+import torch
+import random
+import numpy as np
+import cv2
+
+from torchvision.transforms import v2
+from typing import Any
+from PIL import Image
+from collections import Counter
+
+from texteller.constants import (
+    IMG_CHANNELS,
+    MAX_RESIZE_RATIO,
+    MIN_RESIZE_RATIO,
+)
+from texteller.utils import transform as inference_transform
+from .augraphy_pipe import get_custom_augraphy
+
+augraphy_pipeline = get_custom_augraphy()
+
+
+def trim_white_border(image: np.ndarray):
+    if len(image.shape) != 3 or image.shape[2] != 3:
+        raise ValueError("Image is not in RGB format or channel is not in third dimension")
+
+    if image.dtype != np.uint8:
+        raise ValueError(f"Image should stored in uint8")
+
+    corners = [tuple(image[0, 0]), tuple(image[0, -1]), tuple(image[-1, 0]), tuple(image[-1, -1])]
+    bg_color = Counter(corners).most_common(1)[0][0]
+    bg_color_np = np.array(bg_color, dtype=np.uint8)
+
+    h, w = image.shape[:2]
+    bg = np.full((h, w, 3), bg_color_np, dtype=np.uint8)
+
+    diff = cv2.absdiff(image, bg)
+    mask = cv2.cvtColor(diff, cv2.COLOR_BGR2GRAY)
+
+    threshold = 15
+    _, diff = cv2.threshold(mask, threshold, 255, cv2.THRESH_BINARY)
+
+    x, y, w, h = cv2.boundingRect(diff)
+
+    trimmed_image = image[y : y + h, x : x + w]
+
+    return trimmed_image
+
+
+def add_white_border(image: np.ndarray, max_size: int) -> np.ndarray:
+    randi = [random.randint(0, max_size) for _ in range(4)]
+    pad_height_size = randi[1] + randi[3]
+    pad_width_size = randi[0] + randi[2]
+    if pad_height_size + image.shape[0] < 30:
+        compensate_height = int((30 - (pad_height_size + image.shape[0])) * 0.5) + 1
+        randi[1] += compensate_height
+        randi[3] += compensate_height
+    if pad_width_size + image.shape[1] < 30:
+        compensate_width = int((30 - (pad_width_size + image.shape[1])) * 0.5) + 1
+        randi[0] += compensate_width
+        randi[2] += compensate_width
+    return v2.functional.pad(
+        torch.from_numpy(image).permute(2, 0, 1),
+        padding=randi,
+        padding_mode="constant",
+        fill=(255, 255, 255),
+    )
+
+
+def padding(images: list[torch.Tensor], required_size: int) -> list[torch.Tensor]:
+    images = [
+        v2.functional.pad(
+            img, padding=[0, 0, required_size - img.shape[2], required_size - img.shape[1]]
+        )
+        for img in images
+    ]
+    return images
+
+
+def random_resize(images: list[np.ndarray], minr: float, maxr: float) -> list[np.ndarray]:
+    if len(images[0].shape) != 3 or images[0].shape[2] != 3:
+        raise ValueError("Image is not in RGB format or channel is not in third dimension")
+
+    ratios = [random.uniform(minr, maxr) for _ in range(len(images))]
+    return [
+        # Anti-aliasing
+        cv2.resize(
+            img, (int(img.shape[1] * r), int(img.shape[0] * r)), interpolation=cv2.INTER_LANCZOS4
+        )
+        for img, r in zip(images, ratios)
+    ]
+
+
+def rotate(image: np.ndarray, min_angle: int, max_angle: int) -> np.ndarray:
+    # Get the center of the image to define the point of rotation
+    image_center = tuple(np.array(image.shape[1::-1]) / 2)
+
+    # Generate a random angle within the specified range
+    angle = random.randint(min_angle, max_angle)
+
+    # Get the rotation matrix for rotating the image around its center
+    rotation_mat = cv2.getRotationMatrix2D(image_center, angle, 1.0)
+
+    # Determine the size of the rotated image
+    cos = np.abs(rotation_mat[0, 0])
+    sin = np.abs(rotation_mat[0, 1])
+    new_width = int((image.shape[0] * sin) + (image.shape[1] * cos))
+    new_height = int((image.shape[0] * cos) + (image.shape[1] * sin))
+
+    # Adjust the rotation matrix to take into account translation
+    rotation_mat[0, 2] += (new_width / 2) - image_center[0]
+    rotation_mat[1, 2] += (new_height / 2) - image_center[1]
+
+    # Rotate the image with the specified border color (white in this case)
+    rotated_image = cv2.warpAffine(
+        image, rotation_mat, (new_width, new_height), borderValue=(255, 255, 255)
+    )
+
+    return rotated_image
+
+
+def ocr_aug(image: np.ndarray) -> np.ndarray:
+    if random.random() < 0.2:
+        image = rotate(image, -5, 5)
+    image = add_white_border(image, max_size=25).permute(1, 2, 0).numpy()
+    image = augraphy_pipeline(image)
+    return image
+
+
+def train_transform(images: list[Image.Image]) -> list[torch.Tensor]:
+    assert IMG_CHANNELS == 1, "Only support grayscale images for now"
+
+    images = [np.array(img.convert("RGB")) for img in images]
+    # random resize first
+    images = random_resize(images, MIN_RESIZE_RATIO, MAX_RESIZE_RATIO)
+    images = [trim_white_border(image) for image in images]
+
+    # OCR augmentation
+    images = [ocr_aug(image) for image in images]
+
+    # general transform pipeline
+    images = inference_transform(images)
+    return images
+
+
+def img_train_transform(samples: dict[str, list[Any]]) -> dict[str, list[Any]]:
+    processed_img = train_transform(samples["pixel_values"])
+    samples["pixel_values"] = processed_img
+    return samples
+
+
+def img_inf_transform(samples: dict[str, list[Any]]) -> dict[str, list[Any]]:
+    processed_img = inference_transform(samples["pixel_values"])
+    samples["pixel_values"] = processed_img
+    return samples

pyproject.toml

@@ -0,0 +1,87 @@
+[build-system]
+requires = ["hatchling", "hatch-vcs"]
+build-backend = "hatchling.build"
+
+[project]
+name = "texteller"
+authors = [
+    { name="OleehyO", email="leehy0357@gmail.com" }
+]
+dynamic = ["version"]
+description = "Texteller is a tool for converting rendered image to original latex code"
+readme = "README.md"
+license = { file = "LICENSE" }
+requires-python = ">=3.10"
+dependencies = [
+    "click>=8.1.8",
+    "colorama>=0.4.6",
+    "opencv-python-headless>=4.11.0.86",
+    "pyclipper>=1.3.0.post6",
+    "shapely>=2.1.0",
+    "streamlit>=1.44.1",
+    "streamlit-paste-button>=0.1.2",
+    "torch>=2.6.0",
+    "torchvision>=0.21.0",
+    "transformers==4.45.2",
+    "wget>=3.2",
+    "optimum[onnxruntime]>=1.24.0",
+    "python-multipart>=0.0.20",
+    "ray[serve]>=2.44.1",
+]
+
+[tool.hatch.version]
+source = "vcs"
+
+[tool.ruff]
+exclude = [".git", ".mypy_cache", ".ruff_cache", ".venv", "dist"]
+target-version = "py310"
+line-length = 100
+
+[tool.ruff.format]
+line-ending = "lf"
+quote-style = "double"
+
+[tool.ruff.lint]
+select = ["E", "W"]
+ignore = [
+    "E999",
+    "EXE001",
+    "UP009",
+    "F401",
+    "TID252",
+    "F403",
+    "F841",
+    "E501",
+    "W291",
+    "W293",
+    "E741",
+    "E712",
+]
+
+[tool.hatch.build.targets.wheel]
+packages = ["texteller"]
+
+[project.scripts]
+texteller = "texteller.cli:cli"
+
+[project.optional-dependencies]
+onnxruntime-gpu = [
+    "onnxruntime-gpu>=1.21.0",
+]
+test = [
+    "pytest>=8.3.5",
+]
+train = [
+    "accelerate>=1.6.0",
+    "augraphy>=8.2.6",
+    "datasets>=3.5.0",
+    "tensorboardx>=2.6.2.2",
+]
+docs = [
+    "myst-parser>=4.0.1",
+    "nbsphinx>=0.9.7",
+    "sphinx>=8.1.3",
+    "sphinx-book-theme>=1.1.4",
+    "sphinx-copybutton>=0.5.2",
+    "sphinx-design>=0.6.1",
+]

requirements.txt

@@ -1,13 +0,0 @@
-transformers
-datasets
-evaluate
-streamlit
-opencv-python
-ray[serve]
-accelerate
-tensorboardX
-nltk
-python-multipart
-
-pdf2image
-augraphy

src/client_demo.py

@@ -1,16 +0,0 @@
-import requests
-
-# 服务的 URL
-url = "http://127.0.0.1:9900/predict"
-
-# 替换成你要预测的图像的路径
-img_path = "/home/lhy/code/TeXify/src/7.png"
-
-# 构造请求数据
-data = {"img_path": img_path}
-
-# 发送 POST 请求
-response = requests.post(url, json=data)
-
-# 打印响应
-print(response.text)

src/inference.py

@@ -1,40 +0,0 @@
-import os
-import argparse
-
-from pathlib import Path
-from models.ocr_model.utils.inference import inference
-from models.ocr_model.model.TexTeller import TexTeller
-
-
-if __name__ == '__main__':
-    parser = argparse.ArgumentParser()
-    parser.add_argument(
-        '-img', 
-        type=str, 
-        required=True,
-        help='path to the input image'
-    )
-    parser.add_argument(
-        '-cuda', 
-        default=False,
-        action='store_true',
-        help='use cuda or not'
-    )
-
-    args = parser.parse_args([
-        '-img', './models/ocr_model/test_img/1.png',
-        '-cuda'
-    ])
-
-    script_dirpath = Path(__file__).resolve().parent
-    os.chdir(script_dirpath)
-
-    model = TexTeller.from_pretrained('./models/ocr_model/model_checkpoint')
-    tokenizer = TexTeller.get_tokenizer('./models/tokenizer/roberta-tokenizer-550K')
-
-    # base = '/home/lhy/code/TeXify/src/models/ocr_model/test_img'
-    # img_path = [base + f'/{i}.png' for i in range(7, 12)]
-    img_path = [args.img]
-
-    res = inference(model, tokenizer, img_path, args.cuda)
-    print(res[0])

src/models/globals.py

@@ -1,60 +0,0 @@
-# 公式图片(灰度化后)的均值和方差
-IMAGE_MEAN = 0.9545467
-IMAGE_STD  = 0.15394445
-
-
-# =========================   ocr模型用的参数   ============================= #
-
-# 输入图片的最大最小的宽和高
-MIN_HEIGHT = 32
-MAX_HEIGHT = 512
-MIN_WIDTH  = 32
-MAX_WIDTH  = 1280
-# LaTex-OCR中分别是 32、192、32、672
-
-# ocr模型所用数据集，pdf转图片所用的Density值(dpi)
-TEXIFY_INPUT_DENSITY = 100
-
-# ocr模型的tokenizer中的词典数量
-VOCAB_SIZE = 15000
-
-# ocr模型是否固定输入图片的大小
-OCR_FIX_SIZE = True
-# ocr模型训练时，输入图片所固定的大小 (when OCR_FIX_SIZE is True)
-OCR_IMG_SIZE = 448
-# ocr模型训练时，输入图片最大的宽和高（when OCR_FIX_SIZE is False）
-OCR_IMG_MAX_HEIGHT = 512
-OCR_IMG_MAX_WIDTH  = 768
-
-# ocr模型输入图片的通道数
-OCR_IMG_CHANNELS = 1  # 灰度图
-
-# ocr模型训练数据集的最长token数
-MAX_TOKEN_SIZE = 1024     # 模型最长的embedding长度(默认512)
-# MAX_TOKEN_SIZE = 2048     # 模型最长的embedding长度(默认512)
-# MAX_TOKEN_SIZE = 600
-
-# ocr模型训练时随机缩放的比例
-MAX_RESIZE_RATIO = 1.15
-MIN_RESIZE_RATIO = 0.75
-
-# ocr模型输入的图片要求的最低宽和高(过滤垃圾数据)
-MIN_HEIGHT = 12
-MIN_WIDTH  = 30
-
-# ============================================================================= #
-
-
-# =========================   Resizer模型用的参数   ============================= #
-
-# Resizer模型所用数据集中，图片所用的Density渲染值
-RESIZER_INPUT_DENSITY  = 200   
-
-LABEL_RATIO   = 1.0 * TEXIFY_INPUT_DENSITY / RESIZER_INPUT_DENSITY
-
-NUM_CLASSES   = 1      # 模型使用回归预测
-NUM_CHANNELS  = 1      # 输入单通道图片（灰度图）
-
-# Resizer在训练时，图片所固定的的大小
-RESIZER_IMG_SIZE = 448    
-# ============================================================================= #

src/models/ocr_model/README.md

@@ -1,6 +0,0 @@
-* Encoder-Decoder架构
-
-* Encoder使用Deit_{BASE}
-
-* Decoder使用RoBERTa_{LARGE}
-    * Decoder的tokenizer也使用RoBERTa_{LARGE}的

src/models/ocr_model/model/TexTeller.py

@@ -1,65 +0,0 @@
-from pathlib import Path
-
-from models.globals import (
-    VOCAB_SIZE,
-    OCR_IMG_SIZE,
-    OCR_IMG_CHANNELS,
-    MAX_TOKEN_SIZE
-)
-
-from transformers import (
-    ViTConfig,
-    ViTModel,
-    TrOCRConfig,
-    TrOCRForCausalLM,
-    RobertaTokenizerFast,
-    VisionEncoderDecoderModel,
-)
-
-
-class TexTeller(VisionEncoderDecoderModel):
-    def __init__(self, decoder_path=None, tokenizer_path=None):
-        encoder = ViTModel(ViTConfig(
-            image_size=OCR_IMG_SIZE,
-            num_channels=OCR_IMG_CHANNELS
-        ))
-        decoder = TrOCRForCausalLM(TrOCRConfig(
-            vocab_size=VOCAB_SIZE,
-            max_position_embeddings=MAX_TOKEN_SIZE
-        ))
-        super().__init__(encoder=encoder, decoder=decoder)
-    
-    @classmethod
-    def from_pretrained(cls, model_path: str):
-        model_path = Path(model_path).resolve()
-        return VisionEncoderDecoderModel.from_pretrained(str(model_path))
-
-    @classmethod
-    def get_tokenizer(cls, tokenizer_path: str) -> RobertaTokenizerFast:
-        tokenizer_path = Path(tokenizer_path).resolve()
-        return RobertaTokenizerFast.from_pretrained(str(tokenizer_path))
-
-
-if __name__ == "__main__":
-    pause = 1
-    # texteller = TexTeller()
-    # from ..utils.inference import inference
-    # model = TexTeller.from_pretrained('/home/lhy/code/TexTeller/src/models/ocr_model/model/ckpt')
-    # model.save_pretrained('/home/lhy/code/TexTeller/src/models/ocr_model/model/ckpt2', safe_serialization=False)
-    # tokenizer = TexTeller.get_tokenizer('/home/lhy/code/TeXify/src/models/tokenizer/roberta-tokenizer-550Kformulas')
-
-    # base = '/home/lhy/code/TeXify/src/models/ocr_model/model'
-    # imgs_path = [
-    #     # base + '/1.jpg',
-    #     # base + '/2.jpg',
-    #     # base + '/3.jpg',
-    #     # base + '/4.jpg',
-    #     # base + '/5.jpg',
-    #     # base + '/6.jpg',
-    #     base + '/foo.jpg'
-    # ]
-
-    # # res = inference(model, [img1, img2, img3, img4, img5, img6, img7], tokenizer)
-    # res = inference(model, imgs_path, tokenizer)
-    # pause = 1
-

src/models/ocr_model/train/fonts/Rotodesign License.txt

@@ -1,14 +0,0 @@
-Congratulations on your download of this fine Rotodesign brand font product. We hope it will bring you many hours of typesetting pleasure and riches beyond your wildest dreams. We DO NOT, however, guarantee either of these things. Your mileage may vary. 
-
-This font is freeware, and is provided with no warranties as to its quality or its utility. After all, how much did you pay? Anyway, this font can be copied and used as you wish provided all copies include this readme file. Don't lie to your friends and tell 'em you made it yourself. You only cheat yourself when you do that. In the unlikely event you use this font to design something really cool or that makes you a ton of cash money, that's okay with me, just send me a copy or two of the finished item, and remember me when you get rich and famous. Enjoy!
-
-©2006 
-Patrick Broderick
-Rotodesign
-
-http://www.rotodesign.com
-roto@rotodesign.net
-
-Rotodesign
-1288 Columbus Ave. #176
-San Francisco, CA 94133

src/models/ocr_model/train/google_bleu/google_bleu.py

@@ -1,168 +0,0 @@
-# Copyright 2020 The HuggingFace Evaluate Authors.
-#
-# 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.
-""" Google BLEU (aka GLEU) metric. """
-
-from typing import Dict, List
-
-import datasets
-from nltk.translate import gleu_score
-
-import evaluate
-from evaluate import MetricInfo
-
-from .tokenizer_13a import Tokenizer13a
-
-
-_CITATION = """\
-@misc{wu2016googles,
-      title={Google's Neural Machine Translation System: Bridging the Gap between Human and Machine Translation},
-      author={Yonghui Wu and Mike Schuster and Zhifeng Chen and Quoc V. Le and Mohammad Norouzi and Wolfgang Macherey
-              and Maxim Krikun and Yuan Cao and Qin Gao and Klaus Macherey and Jeff Klingner and Apurva Shah and Melvin
-              Johnson and Xiaobing Liu and Łukasz Kaiser and Stephan Gouws and Yoshikiyo Kato and Taku Kudo and Hideto
-              Kazawa and Keith Stevens and George Kurian and Nishant Patil and Wei Wang and Cliff Young and
-              Jason Smith and Jason Riesa and Alex Rudnick and Oriol Vinyals and Greg Corrado and Macduff Hughes
-              and Jeffrey Dean},
-      year={2016},
-      eprint={1609.08144},
-      archivePrefix={arXiv},
-      primaryClass={cs.CL}
-}
-"""
-
-_DESCRIPTION = """\
-The BLEU score has some undesirable properties when used for single
-sentences, as it was designed to be a corpus measure. We therefore
-use a slightly different score for our RL experiments which we call
-the 'GLEU score'. For the GLEU score, we record all sub-sequences of
-1, 2, 3 or 4 tokens in output and target sequence (n-grams). We then
-compute a recall, which is the ratio of the number of matching n-grams
-to the number of total n-grams in the target (ground truth) sequence,
-and a precision, which is the ratio of the number of matching n-grams
-to the number of total n-grams in the generated output sequence. Then
-GLEU score is simply the minimum of recall and precision. This GLEU
-score's range is always between 0 (no matches) and 1 (all match) and
-it is symmetrical when switching output and target. According to
-our experiments, GLEU score correlates quite well with the BLEU
-metric on a corpus level but does not have its drawbacks for our per
-sentence reward objective.
-"""
-
-_KWARGS_DESCRIPTION = """\
-Computes corpus-level Google BLEU (GLEU) score of translated segments against one or more references.
-Instead of averaging the sentence level GLEU scores (i.e. macro-average precision), Wu et al. (2016) sum up the matching
-tokens and the max of hypothesis and reference tokens for each sentence, then compute using the aggregate values.
-
-Args:
-    predictions (list of str): list of translations to score.
-    references (list of list of str): list of lists of references for each translation.
-    tokenizer : approach used for tokenizing `predictions` and `references`.
-        The default tokenizer is `tokenizer_13a`, a minimal tokenization approach that is equivalent to `mteval-v13a`, used by WMT.
-        This can be replaced by any function that takes a string as input and returns a list of tokens as output.
-    min_len (int): The minimum order of n-gram this function should extract. Defaults to 1.
-    max_len (int): The maximum order of n-gram this function should extract. Defaults to 4.
-
-Returns:
-    'google_bleu': google_bleu score
-
-Examples:
-    Example 1:
-        >>> predictions = ['It is a guide to action which ensures that the rubber duck always disobeys the commands of the cat', \
-        'he read the book because he was interested in world history']
-        >>> references = [['It is the guiding principle which guarantees the rubber duck forces never being under the command of the cat'], \
-        ['he was interested in world history because he read the book']]
-        >>> google_bleu = evaluate.load("google_bleu")
-        >>> results = google_bleu.compute(predictions=predictions, references=references)
-        >>> print(round(results["google_bleu"], 2))
-        0.44
-
-    Example 2:
-        >>> predictions = ['It is a guide to action which ensures that the rubber duck always disobeys the commands of the cat', \
-        'he read the book because he was interested in world history']
-        >>> references = [['It is the guiding principle which guarantees the rubber duck forces never being under the command of the cat', \
-        'It is a guide to action that ensures that the rubber duck will never heed the cat commands', \
-        'It is the practical guide for the rubber duck army never to heed the directions of the cat'], \
-        ['he was interested in world history because he read the book']]
-        >>> google_bleu = evaluate.load("google_bleu")
-        >>> results = google_bleu.compute(predictions=predictions, references=references)
-        >>> print(round(results["google_bleu"], 2))
-        0.61
-
-    Example 3:
-        >>> predictions = ['It is a guide to action which ensures that the rubber duck always disobeys the commands of the cat', \
-        'he read the book because he was interested in world history']
-        >>> references = [['It is the guiding principle which guarantees the rubber duck forces never being under the command of the cat', \
-        'It is a guide to action that ensures that the rubber duck will never heed the cat commands', \
-        'It is the practical guide for the rubber duck army never to heed the directions of the cat'], \
-        ['he was interested in world history because he read the book']]
-        >>> google_bleu = evaluate.load("google_bleu")
-        >>> results = google_bleu.compute(predictions=predictions, references=references, min_len=2)
-        >>> print(round(results["google_bleu"], 2))
-        0.53
-
-    Example 4:
-        >>> predictions = ['It is a guide to action which ensures that the rubber duck always disobeys the commands of the cat', \
-        'he read the book because he was interested in world history']
-        >>> references = [['It is the guiding principle which guarantees the rubber duck forces never being under the command of the cat', \
-        'It is a guide to action that ensures that the rubber duck will never heed the cat commands', \
-        'It is the practical guide for the rubber duck army never to heed the directions of the cat'], \
-        ['he was interested in world history because he read the book']]
-        >>> google_bleu = evaluate.load("google_bleu")
-        >>> results = google_bleu.compute(predictions=predictions,references=references, min_len=2, max_len=6)
-        >>> print(round(results["google_bleu"], 2))
-        0.4
-"""
-
-
-@evaluate.utils.file_utils.add_start_docstrings(_DESCRIPTION, _KWARGS_DESCRIPTION)
-class GoogleBleu(evaluate.Metric):
-    def _info(self) -> MetricInfo:
-        return evaluate.MetricInfo(
-            description=_DESCRIPTION,
-            citation=_CITATION,
-            inputs_description=_KWARGS_DESCRIPTION,
-            features=[
-                datasets.Features(
-                    {
-                        "predictions": datasets.Value("string", id="sequence"),
-                        "references": datasets.Sequence(datasets.Value("string", id="sequence"), id="references"),
-                    }
-                ),
-                datasets.Features(
-                    {
-                        "predictions": datasets.Value("string", id="sequence"),
-                        "references": datasets.Value("string", id="sequence"),
-                    }
-                ),
-            ],
-        )
-
-    def _compute(
-        self,
-        predictions: List[str],
-        references: List[List[str]],
-        tokenizer=Tokenizer13a(),
-        min_len: int = 1,
-        max_len: int = 4,
-    ) -> Dict[str, float]:
-        # if only one reference is provided make sure we still use list of lists
-        if isinstance(references[0], str):
-            references = [[ref] for ref in references]
-
-        references = [[tokenizer(r) for r in ref] for ref in references]
-        predictions = [tokenizer(p) for p in predictions]
-        return {
-            "google_bleu": gleu_score.corpus_gleu(
-                list_of_references=references, hypotheses=predictions, min_len=min_len, max_len=max_len
-            )
-        }

src/models/ocr_model/train/google_bleu/tokenizer_13a.py

@@ -1,100 +0,0 @@
-# Source: https://github.com/mjpost/sacrebleu/blob/master/sacrebleu/tokenizers/tokenizer_13a.py
-# Copyright 2020 SacreBLEU Authors.
-#
-# 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 re
-from functools import lru_cache
-
-
-class BaseTokenizer:
-    """A base dummy tokenizer to derive from."""
-
-    def signature(self):
-        """
-        Returns a signature for the tokenizer.
-        :return: signature string
-        """
-        return "none"
-
-    def __call__(self, line):
-        """
-        Tokenizes an input line with the tokenizer.
-        :param line: a segment to tokenize
-        :return: the tokenized line
-        """
-        return line
-
-
-class TokenizerRegexp(BaseTokenizer):
-    def signature(self):
-        return "re"
-
-    def __init__(self):
-        self._re = [
-            # language-dependent part (assuming Western languages)
-            (re.compile(r"([\{-\~\[-\` -\&\(-\+\:-\@\/])"), r" \1 "),
-            # tokenize period and comma unless preceded by a digit
-            (re.compile(r"([^0-9])([\.,])"), r"\1 \2 "),
-            # tokenize period and comma unless followed by a digit
-            (re.compile(r"([\.,])([^0-9])"), r" \1 \2"),
-            # tokenize dash when preceded by a digit
-            (re.compile(r"([0-9])(-)"), r"\1 \2 "),
-            # one space only between words
-            # NOTE: Doing this in Python (below) is faster
-            # (re.compile(r'\s+'), r' '),
-        ]
-
-    @lru_cache(maxsize=2**16)
-    def __call__(self, line):
-        """Common post-processing tokenizer for `13a` and `zh` tokenizers.
-        :param line: a segment to tokenize
-        :return: the tokenized line
-        """
-        for (_re, repl) in self._re:
-            line = _re.sub(repl, line)
-
-        # no leading or trailing spaces, single space within words
-        # return ' '.join(line.split())
-        # This line is changed with regards to the original tokenizer (seen above) to return individual words
-        return line.split()
-
-
-class Tokenizer13a(BaseTokenizer):
-    def signature(self):
-        return "13a"
-
-    def __init__(self):
-        self._post_tokenizer = TokenizerRegexp()
-
-    @lru_cache(maxsize=2**16)
-    def __call__(self, line):
-        """Tokenizes an input line using a relatively minimal tokenization
-        that is however equivalent to mteval-v13a, used by WMT.
-
-        :param line: a segment to tokenize
-        :return: the tokenized line
-        """
-
-        # language-independent part:
-        line = line.replace("<skipped>", "")
-        line = line.replace("-\n", "")
-        line = line.replace("\n", " ")
-
-        if "&" in line:
-            line = line.replace("&quot;", '"')
-            line = line.replace("&amp;", "&")
-            line = line.replace("&lt;", "<")
-            line = line.replace("&gt;", ">")
-
-        return self._post_tokenizer(f" {line} ")

src/models/ocr_model/train/train.py

@@ -1,114 +0,0 @@
-import os
-
-from functools import partial
-from pathlib import Path
-
-from datasets import load_dataset
-from transformers import Trainer, TrainingArguments, Seq2SeqTrainer, Seq2SeqTrainingArguments, GenerationConfig
-
-from .training_args import CONFIG
-from ..model.TexTeller import TexTeller
-from ..utils.functional import tokenize_fn, collate_fn, img_train_transform, img_inf_transform, filter_fn
-from ..utils.metrics import bleu_metric
-from ...globals import MAX_TOKEN_SIZE, MIN_WIDTH, MIN_HEIGHT   
-
-
-def train(model, tokenizer, train_dataset, eval_dataset, collate_fn_with_tokenizer):
-    training_args = TrainingArguments(**CONFIG)
-    debug_mode = False
-    if debug_mode:
-        training_args.auto_find_batch_size = False
-        training_args.num_train_epochs = 2
-        # training_args.per_device_train_batch_size = 3
-        training_args.per_device_train_batch_size = 2
-        training_args.per_device_eval_batch_size = 2 * training_args.per_device_train_batch_size
-        training_args.jit_mode_eval = False
-        training_args.torch_compile = False
-        training_args.dataloader_num_workers = 1
-    
-    trainer = Trainer(
-        model,
-        training_args,
-
-        train_dataset=train_dataset,
-        eval_dataset=eval_dataset,
-
-        tokenizer=tokenizer, 
-        data_collator=collate_fn_with_tokenizer,
-    )
-
-    trainer.train(resume_from_checkpoint=None)
-    # trainer.train(resume_from_checkpoint='/home/lhy/code/TexTeller/src/models/ocr_model/train/train_result/TexTellerv2/checkpoint-288000')
-
-
-def evaluate(model, tokenizer, eval_dataset, collate_fn):
-    eval_config = CONFIG.copy()
-    eval_config['predict_with_generate'] = True
-    generate_config = GenerationConfig(
-        max_length=MAX_TOKEN_SIZE-100,
-        num_beams=1,
-        do_sample=False,
-        pad_token_id=tokenizer.pad_token_id,
-        eos_token_id=tokenizer.eos_token_id,
-        bos_token_id=tokenizer.bos_token_id,
-    )
-    eval_config['generation_config'] = generate_config
-    eval_config['auto_find_batch_size'] = False
-    seq2seq_config = Seq2SeqTrainingArguments(**eval_config)
-
-    trainer = Seq2SeqTrainer(
-        model,
-        seq2seq_config,
-
-        eval_dataset=eval_dataset,
-        tokenizer=tokenizer, 
-        data_collator=collate_fn,
-        compute_metrics=partial(bleu_metric, tokenizer=tokenizer)
-    )
-
-    res = trainer.evaluate()
-    print(res)
-    
-
-if __name__ == '__main__':
-    cur_path = os.getcwd()
-    script_dirpath = Path(__file__).resolve().parent
-    os.chdir(script_dirpath)
-
-    dataset = load_dataset(
-        '/home/lhy/code/TexTeller/src/models/ocr_model/train/data/loader.py'
-    )['train']
-    tokenizer = TexTeller.get_tokenizer('/home/lhy/code/TexTeller/src/models/tokenizer/roberta-tokenizer-7Mformulas')
-    filter_fn_with_tokenizer = partial(filter_fn, tokenizer=tokenizer)
-
-    # dataset = dataset.filter(lambda x: x['image'].height > MIN_HEIGHT and x['image'].width > MIN_WIDTH)
-    dataset = dataset.filter(filter_fn_with_tokenizer, num_proc=16)
-    dataset = dataset.shuffle(seed=42)
-    dataset = dataset.flatten_indices()
-
-    map_fn = partial(tokenize_fn, tokenizer=tokenizer)
-    tokenized_dataset = dataset.map(map_fn, batched=True, remove_columns=dataset.column_names, num_proc=8, load_from_cache_file=True)
-
-    split_dataset = tokenized_dataset.train_test_split(test_size=0.005, seed=42)
-    train_dataset, eval_dataset = split_dataset['train'], split_dataset['test']
-
-    train_dataset = train_dataset.with_transform(img_train_transform)
-    eval_dataset  = eval_dataset.with_transform(img_inf_transform)
-
-    collate_fn_with_tokenizer = partial(collate_fn, tokenizer=tokenizer)
-    # model = TexTeller()
-    model = TexTeller.from_pretrained('/home/lhy/code/TexTeller/src/models/ocr_model/model/ckpt')
-
-    # =================  debug  =======================
-    # foo = train_dataset[:50]
-    # bar = eval_dataset[:50]
-    # =================  debug  =======================
-
-    enable_train    = True
-    enable_evaluate = True
-    if enable_train:
-        train(model, tokenizer, train_dataset, eval_dataset, collate_fn_with_tokenizer)
-    if enable_evaluate:
-        evaluate(model, tokenizer, eval_dataset, collate_fn_with_tokenizer)
-
-    os.chdir(cur_path)

src/models/ocr_model/train/training_args.py

@@ -1,84 +0,0 @@
-CONFIG = {
-    "seed": 42,                            # 随机种子，用于确保实验的可重复性
-    "use_cpu": False,                      # 是否使用cpu（刚开始测试代码的时候先用cpu跑会更容易debug）
-    # "data_seed": 42,                     # data sampler的采样也固定
-    # "full_determinism": True,            # 使整个训练完全固定（这个设置会有害于模型训练，只用于debug）
-
-    "output_dir": "train_result/TexTellerv3",          # 输出目录
-    "overwrite_output_dir": False,         # 如果输出目录存在，不删除原先的内容
-    "report_to": ["tensorboard"],          # 输出日志到TensorBoard，
-                                           #+通过在命令行：tensorboard --logdir ./logs 来查看日志
-
-    "logging_dir": None,                   # TensorBoard日志文件的存储目录(使用默认值)
-    "log_level": "warning",                   # 其他可选:‘debug’, ‘info’, ‘warning’, ‘error’ and ‘critical’（由低级别到高级别）
-    "logging_strategy": "steps",           # 每隔一定步数记录一次日志
-    "logging_steps": 4000,                  # 记录日志的步数间隔，可以是int也可以是(0~1)的float，当是float时表示总的训练步数的ratio(比方说可以设置成1.0 / 2000)
-                                           #+通常与eval_steps一致
-    "logging_nan_inf_filter": False,       # 对loss=nan或inf进行记录
-
-    "num_train_epochs": 4,                # 总的训练轮数
-    # "max_steps": 3,                      # 训练的最大步骤数。如果设置了这个参数，
-                                           #+那么num_train_epochs将被忽略（通常用于调试）
-
-    # "label_names": ['your_label_name'],  # 指定data_loader中的标签名，如果不指定则默认为'labels'
-
-    "per_device_train_batch_size": 3,    # 每个GPU的batch size
-    "per_device_eval_batch_size": 6,      # 每个GPU的evaluation batch size
-    # "auto_find_batch_size": True,          # 自动搜索合适的batch size（指数decay）
-    "auto_find_batch_size": False,          # 自动搜索合适的batch size（指数decay）
-
-    "optim": "adamw_torch",                # 还提供了很多AdamW的变体（相较于经典的AdamW更加高效）
-                                           #+当设置了optim后，就不需要在Trainer中传入optimizer
-    "lr_scheduler_type": "cosine",         # 设置lr_scheduler
-    "warmup_ratio": 0.1,                   # warmup占整个训练steps的比例(假如训练1000步，那么前100步就是从lr=0慢慢长到参数设定的lr)
-    # "warmup_steps": 500,                 # 预热步数, 这个参数与warmup_ratio是矛盾的
-    "weight_decay": 0,                     # 权重衰减
-    "learning_rate": 5e-5,                 # 学习率
-    "max_grad_norm": 1.0,                  # 用于梯度裁剪，确保梯度的范数不超过1.0（默认1.0）
-    "fp16": False,                         # 是否使用16位浮点数进行训练（一般不推荐，loss很容易炸）
-    "bf16": False,                         # 是否使用16位宽浮点数进行训练（如果架构支持的话推荐使用）
-    "gradient_accumulation_steps": 2,      # 梯度累积步数，当batch size无法开很大时，可以考虑这个参数来实现大batch size的效果
-    "gradient_checkpointing": False,       # 当为True时，会在forward时适当丢弃一些中间量（用于backward），从而减轻显存压力（但会增加forward的时间）
-    "label_smoothing_factor": 0.0,         # softlabel，等于0时表示未开启
-    # "debug": "underflow_overflow",       # 训练时检查溢出，如果发生，则会发出警告。（该模式通常用于debug）
-    "jit_mode_eval": False,                 # 是否在eval的时候使用PyTorch jit trace（可以加速模型，但模型必须是静态的，否则会报错）
-    "torch_compile": False,                 # 是否使用torch.compile来编译模型（从而获得更好的训练和推理性能）
-                                           #+ 要求torch > 2.0，这个功能很好使，当模型跑通的时候可以开起来
-    # "deepspeed": "your_json_path",       #  使用deepspeed来训练，需要指定ds_config.json的路径
-                                           #+ 在Trainer中使用Deepspeed时一定要注意ds_config.json中的配置是否与Trainer的一致（如学习率，batch size，梯度累积步数等）
-                                           #+ 如果不一致，会出现很奇怪的bug（而且一般还很难发现）													
-
-    "dataloader_pin_memory": True,         # 可以加快数据在cpu和gpu之间转移的速度
-    "dataloader_num_workers": 16,          # 默认不会使用多进程来加载数据，通常设成4*所用的显卡数
-    "dataloader_drop_last": True,          # 丢掉最后一个minibatch，保证训练的梯度稳定
-
-    "evaluation_strategy": "steps",        # 评估策略，可以是"steps"或"epoch"
-    "eval_steps": 4000,                     # if evaluation_strategy="step"
-                                           #+默认情况下与logging_steps一样，可以是int也可以是(0~1)的float，当是float时表示总的训练步数的ratio(比方说可以设置成1.0 / 2000)
-
-    "save_strategy": "steps",              # 保存checkpoint的策略
-    "save_steps": 4000,                     # checkpoint保存的步数间隔，可以是int也可以是(0~1)的float，当是float时表示总的训练步数的ratio(比方说可以设置成1.0 / 2000)
-    "save_total_limit": 10,                 # 保存的模型的最大数量。如果超过这个数量，最旧的模型将被删除
-
-    "load_best_model_at_end": True,        # 训练结束时是否加载最佳模型
-                                           #+当设置True时，会保存训练时评估结果最好的checkpoint
-                                           #+当设置True时，evaluation_strategy必须与save_strategy一样，并且save_steps必须是eval_steps的整数倍
-    "metric_for_best_model": "eval_loss",  # 用于选择最佳模型的指标(必须与load_best_model_at_end一起用)
-                                           #+可以使用compute_metrics输出的evaluation的结果中（一个字典）的某个值
-                                           #+注意：Trainer会在compute_metrics输出的字典的键前面加上一个prefix，默认就是“eval_”
-    "greater_is_better": False,            # 指标值越小越好(必须与metric_for_best_model一起用)
-
-    "do_train": True,                      # 是否进行训练，通常用于调试
-    "do_eval": True,                       # 是否进行评估，通常用于调试
-
-    "remove_unused_columns": False,        # 是否删除没有用到的列（特征），默认为True
-                                           #+当删除了没用到的列后，making it easier to unpack inputs into the model’s call function
-    #+注意：remove_unused_columns去除列的操作会把传入的dataset的columns_names与模型forward方法中的参数名进行配对，对于不存在forward方法中的列名就会直接删掉整个feature
-    #+因此如果在dataset.with_transform(..)中给数据进行改名，那么这个remove操作会直接把原始的数据直接删掉，从而导致之后会拿到一个空的dataset，导致在对dataset进行切片取值时出问题
-    #+例如读进来的dataset图片对应的feature name叫"images"，而模型forward方法中对应的参数名叫“pixel_values”，
-    #+此时如果是在data.withtransfrom(..)中根据这个"images"生成其他模型forward方法中需要的参数，然后再把"images"改名成“pixel_values”，那么整个过程就会出问题
-    #+因为设置了remove_unused_columns=True后，会先给dataset进行列名检查，然后“images”这个feature会直接被删掉（导致with_transform的transform_fn拿不到“images”这个feature）
-    #+所以一个good practice就是：对于要改名的特征，先提前使用dataset.rename_column进行改名
-
-    "push_to_hub": False,                  # 是否训练完后上传hub，需要先在命令行：huggingface-cli login进行登录认证的配置，配置完后，认证信息会存到cache文件夹里
-}

src/models/ocr_model/utils/functional.py

@@ -1,59 +0,0 @@
-import torch
-
-from transformers import DataCollatorForLanguageModeling
-from typing import List, Dict, Any
-from .transforms import train_transform, inference_transform
-from ...globals import MIN_HEIGHT, MIN_WIDTH, MAX_TOKEN_SIZE
-
-
-def left_move(x: torch.Tensor, pad_val):
-    assert len(x.shape) == 2, 'x should be 2-dimensional'
-    lefted_x = torch.ones_like(x)
-    lefted_x[:, :-1] = x[:, 1:]
-    lefted_x[:, -1] = pad_val
-    return lefted_x
-
-
-def tokenize_fn(samples: Dict[str, List[Any]], tokenizer=None) -> Dict[str, List[Any]]:
-    assert tokenizer is not None, 'tokenizer should not be None'
-    tokenized_formula = tokenizer(samples['latex_formula'], return_special_tokens_mask=True)
-    tokenized_formula['pixel_values'] = samples['image']
-    return tokenized_formula
-
-
-def collate_fn(samples: List[Dict[str, Any]], tokenizer=None) -> Dict[str, List[Any]]:
-    assert tokenizer is not None, 'tokenizer should not be None'
-    pixel_values = [dic.pop('pixel_values') for dic in samples]
-
-    clm_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False)
-    
-    batch = clm_collator(samples)
-    batch['pixel_values'] = pixel_values
-    batch['decoder_input_ids'] = batch.pop('input_ids')
-    batch['decoder_attention_mask'] = batch.pop('attention_mask')
-
-    # 左移labels和decoder_attention_mask
-    batch['labels'] = left_move(batch['labels'], -100)
-
-    # 把list of Image转成一个tensor with (B, C, H, W)
-    batch['pixel_values'] = torch.stack(batch['pixel_values'], dim=0)
-    return batch
-
-
-def img_train_transform(samples: Dict[str, List[Any]]) -> Dict[str, List[Any]]:
-    processed_img = train_transform(samples['pixel_values'])
-    samples['pixel_values'] = processed_img
-    return samples
-
-
-def img_inf_transform(samples: Dict[str, List[Any]]) -> Dict[str, List[Any]]:
-    processed_img = inference_transform(samples['pixel_values'])
-    samples['pixel_values'] = processed_img
-    return samples
-
-
-def filter_fn(sample, tokenizer=None) -> bool:
-    return (
-        sample['image'].height > MIN_HEIGHT and sample['image'].width > MIN_WIDTH
-        and len(tokenizer(sample['latex_formula'])['input_ids']) < MAX_TOKEN_SIZE - 10
-    )

src/models/ocr_model/utils/helpers.py

@@ -1,39 +0,0 @@
-import cv2
-import numpy as np
-from typing import List
-from PIL import Image
-
-
-def convert2rgb(image_paths: List[str]) -> List[np.ndarray]:
-    # 输出的np.ndarray的格式为：[H, W, C]（通道在第三维），通道的排列顺序为RGB
-    processed_images = []
-
-    for path in image_paths:
-        # 读取图片
-        image = cv2.imread(path, cv2.IMREAD_UNCHANGED)
-
-        if image is None:
-            print(f"Image at {path} could not be read.")
-            continue
-
-        # 检查图片是否使用 uint16 类型
-        if image.dtype == np.uint16:
-            raise ValueError(f"Image at {path} is stored in uint16, which is not supported.")
-
-        # 获取图片通道数
-        channels = 1 if len(image.shape) == 2 else image.shape[2]
-
-        # 如果是 RGBA (4通道), 转换为 RGB
-        if channels == 4:
-            image = cv2.cvtColor(image, cv2.COLOR_BGRA2RGB)
-
-        # 如果是 I 模式 (单通道灰度图), 转换为 RGB
-        elif channels == 1:
-            image = cv2.cvtColor(image, cv2.COLOR_GRAY2RGB)
-
-        # 如果是 BGR (3通道), 转换为 RGB
-        elif channels == 3:
-            image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
-        processed_images.append(image)
-
-    return processed_images

src/models/ocr_model/utils/inference.py

@@ -1,39 +0,0 @@
-import torch
-
-from transformers import RobertaTokenizerFast, GenerationConfig
-from typing import List
-
-from models.ocr_model.model.TexTeller import TexTeller
-from models.ocr_model.utils.transforms import inference_transform
-from models.ocr_model.utils.helpers import convert2rgb
-from models.globals import MAX_TOKEN_SIZE
-
-
-def inference(
-    model: TexTeller, 
-    tokenizer: RobertaTokenizerFast,
-    imgs_path: List[str], 
-    use_cuda: bool,
-    num_beams: int = 1,
-) -> List[str]:
-    model.eval()
-    imgs = convert2rgb(imgs_path)
-    imgs = inference_transform(imgs)
-    pixel_values = torch.stack(imgs)
-
-    if use_cuda:
-        model = model.to('cuda')
-        pixel_values = pixel_values.to('cuda')
-
-
-    generate_config = GenerationConfig(
-        max_new_tokens=MAX_TOKEN_SIZE,
-        num_beams=num_beams,
-        do_sample=False,
-        pad_token_id=tokenizer.pad_token_id,
-        eos_token_id=tokenizer.eos_token_id,
-        bos_token_id=tokenizer.bos_token_id,
-    )
-    pred = model.generate(pixel_values, generation_config=generate_config)
-    res = tokenizer.batch_decode(pred, skip_special_tokens=True)
-    return res

src/models/ocr_model/utils/metrics.py

@@ -1,17 +0,0 @@
-import evaluate
-import numpy as np
-from transformers import EvalPrediction, RobertaTokenizer
-from typing import Dict
-
-def bleu_metric(eval_preds:EvalPrediction, tokenizer:RobertaTokenizer) -> Dict:
-    metric = evaluate.load('/home/lhy/code/TexTeller/src/models/ocr_model/train/google_bleu')  # 这里需要联网，所以会卡住
-    
-    logits, labels = eval_preds.predictions, eval_preds.label_ids
-    preds = logits
-    # preds = np.argmax(logits, axis=1)  # 把logits转成对应的预测标签
-
-    labels = np.where(labels == -100, 1, labels)
-
-    preds = tokenizer.batch_decode(preds, skip_special_tokens=True)
-    labels = tokenizer.batch_decode(labels, skip_special_tokens=True)
-    return metric.compute(predictions=preds, references=labels)

src/models/ocr_model/utils/transforms.py

@@ -1,221 +0,0 @@
-import torch
-import random
-import numpy as np
-import cv2
-
-from torchvision.transforms import v2
-from typing import List
-from PIL import Image
-
-from ...globals import (
-    OCR_IMG_CHANNELS,
-    OCR_IMG_SIZE,
-    OCR_FIX_SIZE,
-    IMAGE_MEAN, IMAGE_STD,
-    MAX_RESIZE_RATIO, MIN_RESIZE_RATIO
-)
-from .ocr_aug import ocr_augmentation_pipeline
-
-# train_pipeline = default_augraphy_pipeline(scan_only=True)
-train_pipeline = ocr_augmentation_pipeline()
-
-general_transform_pipeline = v2.Compose([
-    v2.ToImage(),    # Convert to tensor, only needed if you had a PIL image
-                     #+返回一个List of torchvision.Image，list的长度就是batch_size
-                     #+因此在整个Compose pipeline的最后，输出的也是一个List of torchvision.Image
-                     #+注意：不是返回一整个torchvision.Image，batch_size的维度是拿出来的
-    v2.ToDtype(torch.uint8, scale=True),  # optional, most input are already uint8 at this point
-    v2.Grayscale(),  # 转灰度图（视具体任务而定）
-
-    v2.Resize(       # 固定resize到一个正方形上
-        size=OCR_IMG_SIZE - 1,  # size必须小于max_size 
-        interpolation=v2.InterpolationMode.BICUBIC,
-        max_size=OCR_IMG_SIZE,
-        antialias=True
-    ),
-
-    v2.ToDtype(torch.float32, scale=True),  # Normalize expects float input
-    v2.Normalize(mean=[IMAGE_MEAN], std=[IMAGE_STD]),
-
-    # v2.ToPILImage()  # 用于观察转换后的结果是否正确（debug用）
-])
-
-
-def trim_white_border(image: np.ndarray):
-    # image是一个3维的ndarray，RGB格式，维度分布为[H, W, C]（通道维在第三维上）
-
-    # # 检查images中的第一个元素是否是嵌套的列表结构
-    # if isinstance(image, list):
-    #     image = np.array(image, dtype=np.uint8)
-
-    # 检查图像是否为RGB格式，同时检查通道维是不是在第三维上
-    if len(image.shape) != 3 or image.shape[2] != 3:
-        raise ValueError("Image is not in RGB format or channel is not in third dimension")
-
-    # 检查图片是否使用 uint8 类型
-    if image.dtype != np.uint8:
-        raise ValueError(f"Image should stored in uint8")
-
-    # 创建与原图像同样大小的纯白背景图像
-    h, w = image.shape[:2]
-    bg = np.full((h, w, 3), 255, dtype=np.uint8)
-
-    # 计算差异
-    diff = cv2.absdiff(image, bg)
-
-    # 只要差值大于1，就全部转化为255
-    _, diff = cv2.threshold(diff, 1, 255, cv2.THRESH_BINARY)
-
-    # 把差值转灰度图
-    gray_diff = cv2.cvtColor(diff, cv2.COLOR_RGB2GRAY)
-    # 计算图像中非零像素点的最小外接矩阵
-    x, y, w, h = cv2.boundingRect(gray_diff) 
-
-    # 裁剪图像
-    trimmed_image = image[y:y+h, x:x+w]
-
-    return trimmed_image
-
-
-def add_white_border(image: np.ndarray, max_size: int) -> np.ndarray:
-    randi = [random.randint(0, max_size) for _ in range(4)]
-    pad_height_size = randi[1] + randi[3]
-    pad_width_size  = randi[0] + randi[2]
-    if (pad_height_size + image.shape[0] < 30):
-        compensate_height = int((30 - (pad_height_size + image.shape[0])) * 0.5) + 1
-        randi[1] += compensate_height
-        randi[3] += compensate_height
-    if (pad_width_size + image.shape[1] < 30):
-        compensate_width = int((30 - (pad_width_size + image.shape[1])) * 0.5) + 1
-        randi[0] += compensate_width
-        randi[2] += compensate_width
-    return v2.functional.pad(
-        torch.from_numpy(image).permute(2, 0, 1),
-        padding=randi,
-        padding_mode='constant',
-        fill=(255, 255, 255)
-    )
-
-
-def padding(images: List[torch.Tensor], required_size: int) -> List[torch.Tensor]:
-    images = [  
-        v2.functional.pad(
-            img,
-            padding=[0, 0, required_size - img.shape[2], required_size - img.shape[1]]
-        )
-        for img in images
-    ]
-    return images
-
-
-def random_resize(
-    images: List[np.ndarray], 
-    minr: float, 
-    maxr: float
-) -> List[np.ndarray]:
-    # np.ndarray的格式：3维，RGB格式，维度分布为[H, W, C]（通道维在第三维上）
-
-    # # 检查images中的第一个元素是否是嵌套的列表结构
-    # if isinstance(images[0], list):
-    #     # 将嵌套的列表结构转换为np.ndarray
-    #     images = [np.array(img, dtype=np.uint8) for img in images]
-
-    if len(images[0].shape) != 3 or images[0].shape[2] != 3:
-        raise ValueError("Image is not in RGB format or channel is not in third dimension")
-
-    ratios = [random.uniform(minr, maxr) for _ in range(len(images))]
-    return [
-        cv2.resize(img, (int(img.shape[1] * r), int(img.shape[0] * r)), interpolation=cv2.INTER_LANCZOS4)  # 抗锯齿
-        for img, r in zip(images, ratios)
-    ]
-
-
-def rotate(image: np.ndarray, min_angle: int, max_angle: int) -> np.ndarray:
-    # Get the center of the image to define the point of rotation
-    image_center = tuple(np.array(image.shape[1::-1]) / 2)
-
-    # Generate a random angle within the specified range
-    angle = random.randint(min_angle, max_angle)
-
-    # Get the rotation matrix for rotating the image around its center
-    rotation_mat = cv2.getRotationMatrix2D(image_center, angle, 1.0)
-
-    # Determine the size of the rotated image
-    cos = np.abs(rotation_mat[0, 0])
-    sin = np.abs(rotation_mat[0, 1])
-    new_width = int((image.shape[0] * sin) + (image.shape[1] * cos))
-    new_height = int((image.shape[0] * cos) + (image.shape[1] * sin))
-
-    # Adjust the rotation matrix to take into account translation
-    rotation_mat[0, 2] += (new_width / 2) - image_center[0]
-    rotation_mat[1, 2] += (new_height / 2) - image_center[1]
-
-    # Rotate the image with the specified border color (white in this case)
-    rotated_image = cv2.warpAffine(image, rotation_mat, (new_width, new_height), borderValue=(255, 255, 255))
-
-    return rotated_image
-
-
-def ocr_aug(image: np.ndarray) -> np.ndarray:
-    # 20%的概率进行随机旋转
-    if random.random() < 0.2:
-        image = rotate(image, -5, 5)
-    # 增加白边
-    image = add_white_border(image, max_size=25).permute(1, 2, 0).numpy()
-    # 数据增强
-    image = train_pipeline(image)
-    return image
-
-
-def train_transform(images: List[Image.Image]) -> List[torch.Tensor]:
-    assert OCR_IMG_CHANNELS == 1 , "Only support grayscale images for now"
-    assert OCR_FIX_SIZE == True, "Only support fixed size images for now"
-
-    images = [np.array(img.convert('RGB')) for img in images]
-    # random resize first
-    images = random_resize(images, MIN_RESIZE_RATIO, MAX_RESIZE_RATIO)
-    # 裁剪掉白边
-    images = [trim_white_border(image) for image in images]
-
-    # OCR augmentation
-    images = [ocr_aug(image) for image in images]
-
-    # general transform pipeline
-    images = [general_transform_pipeline(image) for image in  images]
-    # padding to fixed size
-    images = padding(images, OCR_IMG_SIZE)
-    return images
-
-
-def inference_transform(images: List[np.ndarray]) -> List[torch.Tensor]:
-    assert OCR_IMG_CHANNELS == 1 , "Only support grayscale images for now"
-    assert OCR_FIX_SIZE == True, "Only support fixed size images for now"
-    images = [np.array(img.convert('RGB')) for img in images]
-    # 裁剪掉白边
-    images = [trim_white_border(image) for image in images]
-    # general transform pipeline
-    images = [general_transform_pipeline(image) for image in  images]  # imgs: List[PIL.Image.Image]
-    # padding to fixed size
-    images = padding(images, OCR_IMG_SIZE)
-
-    return images
-
-
-if __name__ == '__main__':
-    from pathlib import Path
-    from .helpers import convert2rgb
-    base_dir = Path('/home/lhy/code/TeXify/src/models/ocr_model/model')
-    imgs_path = [
-        base_dir / '1.jpg',
-        base_dir / '2.jpg',
-        base_dir / '3.jpg',
-        base_dir / '4.jpg',
-        base_dir / '5.jpg',
-        base_dir / '6.jpg',
-        base_dir / '7.jpg',
-    ]
-    imgs_path = [str(img_path) for img_path in imgs_path]
-    imgs = convert2rgb(imgs_path)
-    res = random_resize(imgs, 0.5, 1.5)
-    pause = 1
-

src/models/resizer/inference.py

@@ -1,44 +0,0 @@
-#!/usr/bin/env python3
-import os
-import argparse
-import torch
-
-from pathlib import Path
-from PIL import Image
-from .model.Resizer import Resizer
-from .utils import preprocess_fn
-
-from munch import Munch
-
-
-def inference(args):
-    img = Image.open(args.image)
-    img = img.convert('RGB') if img.format == 'PNG' else img
-    processed_img = preprocess_fn({"pixel_values": [img]})
-
-    ckt_path = Path(args.checkpoint).resolve()
-    model = Resizer.from_pretrained(ckt_path)
-    model.eval()
-    inpu = torch.stack(processed_img['pixel_values'])
-    pred = model(inpu) * 1.25
-    print(pred)
-
-    ...
-
-
-if __name__ == "__main__":
-    cur_dirpath = os.getcwd()
-    script_dirpath = Path(__file__).resolve().parent
-    os.chdir(script_dirpath)
-
-    parser = argparse.ArgumentParser()
-    parser.add_argument('-img', '--image', type=str, required=True)
-    parser.add_argument('-ckt', '--checkpoint', type=str, required=True)
-
-    args = parser.parse_args([
-        '-img', '/home/lhy/code/TeXify/src/models/resizer/foo5_140h.jpg',
-        '-ckt', '/home/lhy/code/TeXify/src/models/resizer/train/train_result_pred_height_v5'
-    ])
-    inference(args)
-
-    os.chdir(cur_dirpath)

src/models/resizer/model/Resizer.py

@@ -1,5 +0,0 @@
-from transformers import ResNetForImageClassification
-
-class Resizer(ResNetForImageClassification):
-    def __init__(self, config):
-        super().__init__(config)

src/models/resizer/train/train.py

@@ -1,122 +0,0 @@
-import os
-import datasets
-
-from pathlib import Path
-from transformers import (
-    ResNetConfig,
-    TrainingArguments,
-    Trainer
-)
-
-from ..utils import preprocess_fn
-from ..model.Resizer import Resizer
-from ...globals import NUM_CHANNELS, NUM_CLASSES, RESIZER_IMG_SIZE
-
-
-def train():
-    cur_dirpath = os.getcwd()
-    script_dirpath = Path(__file__).resolve().parent
-    os.chdir(script_dirpath)
-
-    data = datasets.load_dataset("./dataset").shuffle(seed=42)
-    data = data.rename_column("images", "pixel_values")
-    data.flatten_indices()
-    data = data.with_transform(preprocess_fn)
-    train_data, test_data = data['train'], data['test']
-
-    config = ResNetConfig(
-        num_channels=NUM_CHANNELS,
-        num_labels=NUM_CLASSES,
-        img_size=RESIZER_IMG_SIZE
-    )
-    model = Resizer(config)
-    model = Resizer.from_pretrained("/home/lhy/code/TeXify/src/models/resizer/train/train_result_pred_height_v4/checkpoint-213000")
-
-    training_args = TrainingArguments(
-        # resume_from_checkpoint="/home/lhy/code/TeXify/src/models/resizer/train/train_result_pred_height_v3/checkpoint-94500",
-        max_grad_norm=1.0,
-        # use_cpu=True,
-        seed=42,                            # 随机种子，用于确保实验的可重复性
-        # data_seed=42,                     # data sampler的采样也固定
-        # full_determinism=True,            # 使整个训练完全固定（这个设置会有害于模型训练，只用于debug）
-
-        output_dir='./train_result_pred_height_v5',        # 输出目录
-        overwrite_output_dir=False,         # 如果输出目录存在，不删除原先的内容
-        report_to=["tensorboard"],          # 输出日志到TensorBoard，
-                                            #+通过在命令行：tensorboard --logdir ./logs 来查看日志
-
-        logging_dir=None,               # TensorBoard日志文件的存储目录
-        log_level="info",
-        logging_strategy="steps",           # 每隔一定步数记录一次日志
-        logging_steps=500,                  # 记录日志的步数间隔
-        logging_nan_inf_filter=False,       # 对loss=nan或inf进行记录
-
-        num_train_epochs=50,                 # 总的训练轮数
-        # max_steps=3,                      # 训练的最大步骤数。如果设置了这个参数，
-                                            #+那么num_train_epochs将被忽略（通常用于调试）
-
-        # label_names = ['your_label_name'],    # 指定data_loader中的标签名，如果不指定则默认为'labels'
-
-        per_device_train_batch_size=55,     # 每个GPU的batch size
-        per_device_eval_batch_size=48*2,      # 每个GPU的evaluation batch size
-        auto_find_batch_size=False,         # 自动搜索合适的batch size（指数decay）
-
-        optim = 'adamw_torch',              # 还提供了很多AdamW的变体（相较于经典的AdamW更加高效）
-                                            #+当设置了optim后，就不需要在Trainer中传入optimizer
-        lr_scheduler_type="cosine",         # 设置lr_scheduler
-        warmup_ratio=0.1,                   # warmup占整个训练steps的比例
-        # warmup_steps=500,                 # 预热步数
-        weight_decay=0,                     # 权重衰减
-        learning_rate=5e-5,                 # 学习率
-        fp16=False,                         # 是否使用16位浮点数进行训练
-        gradient_accumulation_steps=1,      # 梯度累积步数，当batch size无法开很大时，可以考虑这个参数来实现大batch size的效果
-        gradient_checkpointing=False,       # 当为True时，会在forward时适当丢弃一些中间量（用于backward），从而减轻显存压力（但会增加forward的时间）
-        label_smoothing_factor=0.0,         # softlabel，等于0时表示未开启
-        # debug='underflow_overflow',       # 训练时检查溢出，如果发生，则会发出警告。（该模式通常用于debug）
-        torch_compile=True,                # 是否使用torch.compile来编译模型（从而获得更好的训练和推理性能）
-                                            #+ 要求torch > 2.0，并且这个功能现在还不是很稳定
-        # deepspeed='your_json_path',       #  使用deepspeed来训练，需要指定ds_config.json的路径
-                                            #+ 在Trainer中使用Deepspeed时一定要注意ds_config.json中的配置是否与Trainer的一致（如学习率，batch size，梯度累积步数等）
-                                            #+ 如果不一致，会出现很奇怪的bug（而且一般还很难发现）													
-
-        dataloader_pin_memory=True,         # 可以加快数据在cpu和gpu之间转移的速度
-        dataloader_num_workers=16,           # 默认不会使用多进程来加载数据
-        dataloader_drop_last=True,          # 丢掉最后一个minibatch
-
-        evaluation_strategy="steps",        # 评估策略，可以是"steps"或"epoch"
-        eval_steps=500,                       # if evaluation_strategy="step"
-        # eval_steps=10,                     # if evaluation_strategy="step"
-
-        save_strategy="steps",              # 保存checkpoint的策略
-        save_steps=1500,                    # 模型保存的步数间隔
-        save_total_limit=5,                 # 保存的模型的最大数量。如果超过这个数量，最旧的模型将被删除
-
-        load_best_model_at_end=True,        # 训练结束时是否加载最佳模型
-        metric_for_best_model="eval_loss",  # 用于选择最佳模型的指标
-        greater_is_better=False,            # 指标值越小越好
-
-        do_train=True,                      # 是否进行训练，通常用于调试
-        do_eval=True,                       # 是否进行评估，通常用于调试
-
-        remove_unused_columns=True,         # 是否删除没有用到的列（特征），默认为True
-                                            #+当删除了没用到的列后，making it easier to unpack inputs into the model’s call function
-
-        push_to_hub=False,                  # 是否训练完后上传hub，需要先在命令行：huggingface-cli login进行登录认证的配置，配置完后，认证信息会存到cache文件夹里
-        hub_model_id="a_different_name",    # 模型的名字
-                                            #+每次保存模型时，都会上传到hub，
-                                            #+训练完后，记得trainer.push_to_hub()，会将模型使用的参数以及验证集上的结果传到hub上 
-    )
-
-    trainer = Trainer(
-        model,
-        training_args,
-        train_dataset=train_data,
-        eval_dataset=test_data,
-    )
-    trainer.train()
-
-    os.chdir(cur_dirpath)
-
-
-if __name__ == '__main__':
-    train()

src/models/resizer/utils/__init__.py

@@ -1 +0,0 @@
-from .preprocess import preprocess_fn

src/models/resizer/utils/preprocess.py

@@ -1,75 +0,0 @@
-import torch
-from torchvision.transforms import v2
-
-from PIL import Image, ImageChops
-from ...globals import (
-    IMAGE_MEAN, IMAGE_STD, 
-    LABEL_RATIO,
-    RESIZER_IMG_SIZE,
-    NUM_CHANNELS
-)
-
-from typing import (
-    Any,
-    List,
-    Dict,
-)
-
-
-def trim_white_border(image: Image):
-    if image.mode == 'RGB':
-        bg_color = (255, 255, 255)
-    elif image.mode == 'RGBA':
-        bg_color = (255, 255, 255, 255)
-    elif image.mode == 'L':
-        bg_color = 255
-    else:
-        raise ValueError("Unsupported image mode")
-    bg = Image.new(image.mode, image.size, bg_color)
-    diff = ImageChops.difference(image, bg)
-    diff = ImageChops.add(diff, diff, 2.0, -100)
-    bbox = diff.getbbox()
-    if bbox:
-        return image.crop(bbox)
-
-
-def preprocess_fn(samples: Dict[str, List[Any]]) -> Dict[str, List[Any]]:
-    imgs = samples['pixel_values']
-    imgs = [trim_white_border(img) for img in imgs]
-    labels = [float(img.height * LABEL_RATIO) for img in imgs]
-
-    assert NUM_CHANNELS == 1, "Only support grayscale images"
-    transform = v2.Compose([
-        v2.ToImage(),
-        v2.ToDtype(torch.uint8, scale=True),
-        v2.Grayscale(),
-        v2.Resize(
-            size=RESIZER_IMG_SIZE - 1,  # size必须小于max_size 
-            interpolation=v2.InterpolationMode.BICUBIC,
-            max_size=RESIZER_IMG_SIZE,
-            antialias=True
-        ),
-        v2.ToDtype(torch.float32, scale=True),
-        v2.Normalize(mean=[IMAGE_MEAN], std=[IMAGE_STD]),
-    ])
-    imgs = transform(imgs)
-    imgs = [
-        v2.functional.pad(
-            img,
-            padding=[0, 0, RESIZER_IMG_SIZE - img.shape[2], RESIZER_IMG_SIZE - img.shape[1]]
-        )
-        for img in imgs
-    ]
-
-    res = {'pixel_values': imgs, 'labels': labels}
-    return res
-
-
-if __name__ == "__main__":  # unit test
-    import datasets
-    data = datasets.load_dataset("/home/lhy/code/TeXify/src/models/resizer/train/dataset/dataset.py").shuffle(seed=42)
-    data = data.with_transform(preprocess_fn)
-    train_data, test_data = data['train'], data['test']
-
-    inpu = train_data[:10]
-    pause = 1

src/models/tokenizer/roberta-tokenizer-550K/special_tokens_map.json

@@ -1,15 +0,0 @@
-{
-  "bos_token": "<s>",
-  "cls_token": "<s>",
-  "eos_token": "</s>",
-  "mask_token": {
-    "content": "<mask>",
-    "lstrip": true,
-    "normalized": false,
-    "rstrip": false,
-    "single_word": false
-  },
-  "pad_token": "<pad>",
-  "sep_token": "</s>",
-  "unk_token": "<unk>"
-}

src/models/tokenizer/roberta-tokenizer-550K/tokenizer_config.json

@@ -1,57 +0,0 @@
-{
-  "add_prefix_space": false,
-  "added_tokens_decoder": {
-    "0": {
-      "content": "<s>",
-      "lstrip": false,
-      "normalized": true,
-      "rstrip": false,
-      "single_word": false,
-      "special": true
-    },
-    "1": {
-      "content": "<pad>",
-      "lstrip": false,
-      "normalized": true,
-      "rstrip": false,
-      "single_word": false,
-      "special": true
-    },
-    "2": {
-      "content": "</s>",
-      "lstrip": false,
-      "normalized": true,
-      "rstrip": false,
-      "single_word": false,
-      "special": true
-    },
-    "3": {
-      "content": "<unk>",
-      "lstrip": false,
-      "normalized": true,
-      "rstrip": false,
-      "single_word": false,
-      "special": true
-    },
-    "4": {
-      "content": "<mask>",
-      "lstrip": true,
-      "normalized": false,
-      "rstrip": false,
-      "single_word": false,
-      "special": true
-    }
-  },
-  "bos_token": "<s>",
-  "clean_up_tokenization_spaces": true,
-  "cls_token": "<s>",
-  "eos_token": "</s>",
-  "errors": "replace",
-  "mask_token": "<mask>",
-  "model_max_length": 1000000000000000019884624838656,
-  "pad_token": "<pad>",
-  "sep_token": "</s>",
-  "tokenizer_class": "RobertaTokenizer",
-  "trim_offsets": true,
-  "unk_token": "<unk>"
-}

src/models/tokenizer/roberta-tokenizer-7Mformulas/special_tokens_map.json

@@ -1,15 +0,0 @@
-{
-  "bos_token": "<s>",
-  "cls_token": "<s>",
-  "eos_token": "</s>",
-  "mask_token": {
-    "content": "<mask>",
-    "lstrip": true,
-    "normalized": false,
-    "rstrip": false,
-    "single_word": false
-  },
-  "pad_token": "<pad>",
-  "sep_token": "</s>",
-  "unk_token": "<unk>"
-}

src/models/tokenizer/roberta-tokenizer-7Mformulas/tokenizer_config.json

@@ -1,57 +0,0 @@
-{
-  "add_prefix_space": false,
-  "added_tokens_decoder": {
-    "0": {
-      "content": "<s>",
-      "lstrip": false,
-      "normalized": true,
-      "rstrip": false,
-      "single_word": false,
-      "special": true
-    },
-    "1": {
-      "content": "<pad>",
-      "lstrip": false,
-      "normalized": true,
-      "rstrip": false,
-      "single_word": false,
-      "special": true
-    },
-    "2": {
-      "content": "</s>",
-      "lstrip": false,
-      "normalized": true,
-      "rstrip": false,
-      "single_word": false,
-      "special": true
-    },
-    "3": {
-      "content": "<unk>",
-      "lstrip": false,
-      "normalized": true,
-      "rstrip": false,
-      "single_word": false,
-      "special": true
-    },
-    "4": {
-      "content": "<mask>",
-      "lstrip": true,
-      "normalized": false,
-      "rstrip": false,
-      "single_word": false,
-      "special": true
-    }
-  },
-  "bos_token": "<s>",
-  "clean_up_tokenization_spaces": true,
-  "cls_token": "<s>",
-  "eos_token": "</s>",
-  "errors": "replace",
-  "mask_token": "<mask>",
-  "model_max_length": 1000000000000000019884624838656,
-  "pad_token": "<pad>",
-  "sep_token": "</s>",
-  "tokenizer_class": "RobertaTokenizer",
-  "trim_offsets": true,
-  "unk_token": "<unk>"
-}