Skip to content

xuguodong1999/pen-simulator

Repository files navigation

pen-simulator

PEN-SIMULATOR is a handwriting / svg synthesis system built for simulation of pen-based interaction in AI research. It converts many text representations, such as latex of mathematical expression or smiles of chemical structural formula, into sketching trajectories or vectorized elements.

samples.mp4

Feature Overview

More Samples

  • Mathematical expressions from im2latex.
im2latex-samples.mp4
  • Structural formulas from drugbank.
drugbank-samples.mp4
  • Draft playground of decimal multiplication
multiply-samples.mp4

API Usage

PEN-SIMULATOR is written in C++ and has Python bindings.

Visualization utilities:

Python

import pen_simulator as ps
import numpy as np

ps.global_init_couch('path/to/SCUT_IRAC/Couch', [
  'Couch_Digit_195', 'Couch_Letter_195', 'Couch_Symbol_130',
  'Couch_GB1_188', 'Couch_GB2_195',
])
ps.global_init_makemeahanzi('path/to/makemeahanzi/graphics.txt')

def example():
  def on_label(origin: str, fallback: str):
    pass  # record mathml element / atom of a molecule
  def on_hierarchy(hierarchies: list[str], label: str):
    pass  # record mathml stack / bond of a molecule
  def draw_line(x0: float, y0: float, x1: float, y1: float):
    pass  # do embedding or visualization for sketching trajectories
  def draw_path(d: str, transform: np.array):
    pass  # ... for svg path (maybe fallback)
  def draw_text(label: str, box: np.array, transform: np.array):
    pass  # ... for system font (maybe fallback)
  pen_context = ps.PenContext()
  pen_context.register_draw_line(draw_line)         # raise exception if missing
  pen_context.register_draw_path(draw_path)         # raise exception if missing
  pen_context.register_draw_text(draw_text)         # raise exception if missing
  pen_context.register_on_label(on_label)           # optional
  pen_context.register_on_hierarchy(on_hierarchy)   # optional

  pen_op = ps.generate(
    text='your-awesome-latex-code',
    source_type=ps.SourceType.HANDWRITING,
    text_type=ps.TextType.LATEX,
    traverse_order=ps.TraverseOrder.DEFAULT,
  )
  pen_op.fit_into_keep_ratio(720 * 0.8, 480 * 0.8)
  pen_op.move_center_to(np.array([720 / 2, 480 / 2]))
  pen_op.set_context(pen_context)
  pen_op.on_paint()  # will invoke all callbacks

C++ (a quick start is adding new cmake subprojects under apps/)

#include "az/data/synthesis_tex_generator.h"
#include "az/data/structural_formula_item.h"

#include "az/data/couch_reader.h"
#include "az/data/makemeahanzi_reader.h"

void example() {
    using namespace az::pen;
    using namespace az::data;

    CouchDatasetReader hw_reader;
    hw_reader.sync_load_all("path/to/SCUT_IRAC/Couch", {
        "Couch_Digit_195", "Couch_Letter_195", "Couch_Symbol_130",
        "Couch_GB1_188", "Couch_GB2_195",
    });
    const auto hw_shape_provider = [&](const UCharType &label) { return hw_reader.select(label); };

    UnicodeSvgReader svg_reader;
    svg_reader.sync_load_all("path/to/makemeahanzi/graphics.txt");
    const auto svg_shape_provider = [&](const UCharType &label) { return svg_reader.select(label); };

    const auto shape_provider = hw_shape_provider; /* or svg_shape_provider */
    /* 1. generate latex */
    auto pen_tex = SynthesisTexGenerator::generate_next("your-awesome-latex-code", shape_provider);
    /* 2. generate structural formula */
    StructuralFormulaItem item;
    item.create_from_format(shape_provider, "your-awesome-smiles", "smi");
    auto pen_smi = item.data;
}

You can make use of derivations of PenOp like render_with_skia function in apps/az_data_cli/skia_utils.cpp:

void example(az::pen::PenOp *data) {
    auto pen_context = std::make_shared<PenContext>();
    pen_context->register_draw_line([&](
            const ScalarType &x0, const ScalarType &y0,
            const ScalarType &x1, const ScalarType &y1
    ) { /* ... for most handwriting cases ... */ });
    pen_context->register_draw_path([&](
            const SkPath *path,
            const TransformMatrix2 &mat
    ) { /* ...for most svg cases, also work as a handwriting fallback... */ });
    pen_context->register_draw_text([&](
            const UCharType &label,
            const Box2 &box,
            const TransformMatrix2 &mat
    ) { /* ... if label is absent even in svg, use system font as a fallback */ });
    data->set_context(pen_context);
    data->on_paint();
    data->set_context(nullptr);
}

Build

Recommended system configurations:

CMake 3.27.x, Ninja 1.11.x, GCC 11.x / MSVC 14.38+ / Apple Clang 15.x, Node 20.x,
Python 3.11.x (optional), Qt 6.5.3 (optional)

Get latest source code:

git clone [email protected]:xuguodong1999/pen-simulator.git --branch main --single-branch --recursive

# if you did not perform recursive clone on the first time, update submodules before building
git submodule init
git submodule sync
git submodule update

Build mathjax with json output and generate js codes as C++ string buffer:

pushd www
pnpm install
pnpm --filter @xgd/mathjax-json run build
popd

Build root as a common CMake project:

mkdir build
cmake -G Ninja -B build -S . -DCMAKE_BUILD_TYPE=Release
# for development purpose, add -DBUILD_SHARED_LIBS=ON may be helpful
  • (optional) If you need opencv highgui for visualization, add flags like -DCMAKE_PREFIX_PATH=/path/to/Qt/6.5.3/gcc_64 to enable Qt integration.

Build python bindings (optional): add -DPYTHON_EXECUTABLE=path/to/python in cmake configurations.

We pack static version of python wheel by running python setup.py bdist_wheel under python/.

Reference

License

GPLv3 Clause

Citation

@misc{PENSIM2024,
  title = {pen simulator: data synthesis for simulation of pen-based interaction},
  author = {xuguodong1999},
  year = {2024},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/xuguodong1999/pen-simulator}}
}

Have fun training your LLM to use a pen !