TorchSketch is an open source software library for free-hand sketch oriented deep learning research, which is built on the top of PyTorch.

The project is under continuous update!

TorchSketch is developed based on Python 3.7.

To avoid any conflicts with your existing Python setup, it's better to install TorchSketch into a standalone environment, e.g., an Anaconda virtual environment.

Assume that you have installed Anaconda. Please create a virtual environment before installation of TorchSketch, as follows.

# Create a virtual environment in Anaconda.
conda create --name ${CUSTOMIZED_ENVIRONMENT_NAME} python=3.7

# Activate it.
conda activate ${CUSTOMIZED_ENVIRONMENT_NAME}

Please use the following command to install TorchSketch.

pip install torchsketch

Then, TorchSketch can be imported into your Python console as follows.

import torchsketch

If you are using MacOS, you may need cairo and pango installed. You can install them with homebrew

brew install cairo
brew install pango

In addition, TorchSketch also can be installed from source.

# Choose your workspace and download this repository.
cd ${CUSTOMIZED_WORKSPACE}
git clone https://github.com/PengBoXiangShang/torchsketch

# Enter the folder of TorchSketch.
cd torchsketch

# Install.
python setup.py install

TorchSketch has three main modules, including data, networks, utils, as shown in follows. The documents and example codes are provided in docs.

• torchsketch
  • data
    • dataloaders: provides the dataloader class files for the frequently-used sketch datasets, e.g., TU-Berlin, Sketchy, QuickDraw.
    • datasets: provides the specific API for each dataset, which integrates a series of functions including downloading, extraction, cleaning, MD5 checksum, and other preprocessings.
  • networks
    • cnn: provides all the SOTA CNNs.
    • gnn: provides the sketch-applicable implementations of GNNs, including GCN, GAT, graph transformer, etc.
    • rnn: provides the sketch-applicable implementations of RNNs.
    • tcn: provides the sketch-applicable implementations of TCNs.
  • utils
    • data_augmentation_utils
    • general_utils
    • metric_utils
    • self_supervised_utils
    • svg_specific_utils
  • docs
    • api_reference
    • examples

These modules and sub-modules can be imported as follows.

import torchsketch.data.dataloaders as dataloaders
import torchsketch.data.datasets as datasets

import torchsketch.networks.cnn as cnns
import torchsketch.networks.gnn as gnns
import torchsketch.networks.rnn as rnns
import torchsketch.networks.tcn as tcns

import torchsketch.utils.data_augmentation_utils as data_augmentation_utils
import torchsketch.utils.general_utils as general_utils
import torchsketch.utils.metric_utils as metric_utils
import torchsketch.utils.self_supervised_utils as self_supervised_utils
import torchsketch.utils.svg_specific_utils as svg_specific_utils

As shown in the following figure, a general PyTorch-based code project mainly includes four blocks, i.e., preparing data, preparing data loader, creating network/model, and training. The functions/APIs built-in torchsketch.utils are designed orienting at all four blocks. When researchers would prepare data and data loader, they could select functions/APIs from torchsketch.data. When researchers would create a network, they could select network classes from torchsketch.networks.

• TorchSketch supports both GPU based and Python built-in multi-processing acceleration.
• TorchSketch is modular, flexible, and extensible, without overly complex design patterns and excessive encapsulation.
• TorchSketch provides four kinds of network architectures that are applicable to sketch, i.e., CNN, RNN, GNN, TCN.
• TorchSketch is compatible to not only numerous datasets but also various formats of free-hand sketch, e.g., SVG, NumPy, PNG, JPEG, by providing numerous format-convert APIs, format-specific APIs, etc.
• TorchSketch supports self-supervised learning study for sketch.
• TorchSketch, beyond free-hand sketch research, also has some universal components that are applicable to the studies for other deep learning topics.

If you find this code useful, please cite our paper "Deep Learning for Free-Hand Sketch: A Survey" (https://arxiv.org/abs/2001.02600):

This project is licensed under the MIT License