suryakalia / pytorch_connectomics Goto Github PK

The field of connectomics aims to reconstruct the wiring diagram of the brain by mapping the neural connections at the level of individual synapses. Recent advances in electronic microscopy (EM) have enabled the collection of a large number of image stacks at nanometer resolution, but the annotation requires expertise and is super time-consuming. Here we provide a deep learning framework powered by PyTorch for automatic and semi-automatic semantic and instance segmentation in connectomics, which is called PyTorch Connectomics (PyTC). This repository is mainly maintained by the Visual Computing Group (VCG) at Harvard University.

PyTorch Connectomics is currently under active development!

• Multi-task, active and semi-supervised learning
• Distributed and mixed-precision optimization
• Scalability for handling large datasets
• Comprehensive augmentations for volumetric data

If you want new features that are relatively easy to implement (e.g., loss functions, models), please open a feature requirement discussion in issues or implement by yourself and submit a pull request. For other features that requires substantial amount of design and coding, please contact the author directly.

The code is developed and tested under the following configurations.

• Hardware: 1-8 Nvidia GPUs with at least 12G GPU memory (change SYSTEM.NUM_GPU accordingly based on the configuration of your machine)
• Software: CentOS Linux 7.4 (Core), CUDA>=11.1, Python>=3.8, PyTorch>=1.9.0, YACS>=0.1.8

Create a new conda environment and install PyTorch:

conda create -n py3_torch python=3.8
source activate py3_torch
conda install pytorch torchvision cudatoolkit=11.3 -c pytorch

Please note that this package is mainly developed on the Harvard FASRC cluster. More information about GPU computing on the FASRC cluster can be found here.

Download and install the package:

git clone https://github.com/zudi-lin/pytorch_connectomics.git
cd pytorch_connectomics
pip install --editable .

Since the codebase is under active development, the editable installation will allow any changes to the original package to reflect directly in the environment. For more information and frequently asked questions about installation, please check the installation guide.

Besides the installation guidance above, we also push a PyTC Docker image to the public docker registry (03/12/2022) to improve usability. Additionally, we provide the corresponding Dockerfile to enable individual modifications. Pleas refer to our PyTC Docker Guidance for more information.

We provide a data augmentation interface several different kinds of commonly used augmentation method for EM images. The interface is pure-python, and operate on and output only numpy arrays, so it can be easily incorporated into any kinds of python-based deep learning frameworks (e.g., TensorFlow). For more details about the design of the data augmentation module, please check the documentation.

We use the Yet Another Configuration System (YACS) library to manage the settings and hyperparameters in model training and inference. The configuration files for tutorial examples can be found here. All available configuration options can be found at connectomics/config/defaults.py. Please note that the default value of several options is None, which is only supported after YACS v0.1.8.

We provide several encoder-decoder architectures, which are customized 3D UNet and Feature Pyramid Network (FPN) models with various blocks and backbones. Those models can be applied for both semantic segmentation and bottom-up instance segmentation of 3D image stacks. Those models can also be constructed specifically for isotropic and anisotropic datasets. Please check the documentation for more details.

This project is built upon numerous previous projects. Especially, we'd like to thank the contributors of the following github repositories:

• pyGreenTea: HHMI Janelia FlyEM Team
• DataProvider: Princeton SeungLab
• Detectron2: Facebook AI Reserach

We gratefully acknowledge the support from NSF awards IIS-1835231 and IIS-2124179.

This project is licensed under the MIT License and the copyright belongs to all PyTorch Connectomics contributors - see the LICENSE file for details.

For a detailed description of our framework, please read this technical report. If you find PyTorch Connectomics (PyTC) useful in your research, please cite:

@article{lin2021pytorch,
  title={PyTorch Connectomics: A Scalable and Flexible Segmentation Framework for EM Connectomics},
  author={Lin, Zudi and Wei, Donglai and Lichtman, Jeff and Pfister, Hanspeter},
  journal={arXiv preprint arXiv:2112.05754},
  year={2021}
}