A PyTorch-based library for working with 3D and 2D convolutional neural networks, with focus on semantic segmentation of volumetric biomedical image data.

Quick overview of elektronn3's code structure:

• elektronn3.training: Utilities for training, monitoring, visualization and model evaluation. Provides a flexible Trainer class that can be used for arbitrary PyTorch models and Data sets.
• elektronn3.data: Data loading and augmentation code for semantic segmentation and other dense prediction tasks. The main focus is on 3D (volumetric) biomedical image data stored as HDF5 files, but most of the code also supports 2D and n-dimensional data.
• elektronn3.inference: Code for deployment of trained models and for efficient tiled inference on large input volumes.
• elektronn3.models: Neural network architectures for segmentation and other pixel-wise prediction tasks. models.unet.UNet provides a highly flexible PyTorch model class inspired by 3D U-Net that works in 2D and 3D and supports custom depths, data anisotropy handling, batch normalization and many more configurable features.
• elektronn3.modules: Modules (in the sense of torch.nn.Module) for building neural networks and loss functions.
• examples: Scripts that demonstrate how the library can be used for biomedical image segmentation.

elektronn3's modular codebase makes it easy to extend/replace parts of it with your own code: For example, you can use the training tools included in elektronn3.training with your own data sets, augmentation methods, network models etc. or use the data loading and augmentation code of elektronn3.data with your own training code. The neural network architectures in elektronn3.models can also be freely used with custom training and/or data loading code.

Documentation can be found at elektronn3.readthedocs.io.

For a roadmap of planned features, see the "enhancement" issues on the tracker.

• Linux (support for Windows, MacOS and other systems is not planned)
• Python 3.6 or later
• PyTorch 1.1
• For other requirements see requirements.txt

Ensure that all of the requirements listed above are installed. We recommend using conda or a virtualenv for that. To install elektronn3 in development mode, run

git clone https://github.com/ELEKTRONN/elektronn3 elektronn3-dev
pip install -e elektronn3-dev

To update your installation, just git pull in your clone directory.

If you are not familiar with virtualenv and conda or are not sure about some of the required steps, you can find a more detailed setup guide here

For a quick test run, first ensure that the neuro_data_cdhw data set is in the expected path:

wget https://github.com/ELEKTRONN/elektronn.github.io/releases/download/neuro_data_cdhw/neuro_data_cdhw.zip
unzip neuro_data_cdhw.zip -d ~/neuro_data_cdhw

To test training with our custom U-Net-inspired architecture in elektronn3, you can run:

python3 train_unet_neurodata.py

• Hitting Ctrl-C anytime during the training will drop you to the IPython training shell where you can access training data and make interactive changes.
• To continue training, hit Ctrl-D twice.
• If you want the process to terminate after leaving the shell, set self.terminate = True inside it and then hit Ctrl-D twice.

Tensorboard logs are saved in ~/e3training/ by default, so you can track training progress by running a tensorboard server there:

tensorboard --logdir ~/e3training/

Then you can view the visualizations at http://localhost:6006.

The elektronn3 project is being developed by the ELEKTRONN team at the Max Planck Institute of Neurobiology and is funded by Winfried Denk's lab.

Jörgen Kornfeld is academic advisor to this project.