pyMSDtorch documentation

Welcome to the pyMSDtorch documentation, a collection of routines for machine learning in scientific imaging. PyMSDtorch offers a friendly API for the easy and flexible deployment of custom deep neural networks. At its core, pyMSDtorch leverages PyTorch, an open source machine learning framework maintained by the Linux Foundation.

The pyMSDtorch libraries were originally intended to provide a platform to test the Mixed-Scale Dense Convolutional Neural Network architecture (MSDNet with accompanying PyCUDA implementation) and extend its usage to bioimaging and x-ray scattering applications. MSDNets are a parameter-lean alternative to U-Nets that remain robust to overfitting and have shown to provide a competitive performance over UNets, particularly when training data is limited.

Over the past years, these libraries have grown into a collection of modules for different interchangeable approaches. The pyMSDtorch libraries now contain, apart from MSDNets, Tuneable U-Nets, Tuneable U-Net3+, and randomized Sparse Mixed Scale networks (SMSNet). A uniform interface and a variety of utility scripts for training provides one with convenient options to change network and network architectures at will and find the one most suitable for your application. In addition to routines for segmentation or denoising, our SMSNets can be used in Autoencoders, image classification, and a variety of ensemble methods. We provide tutorial notebooks for most of these use cases. The development of pyMSDtorch is funded by a Laboratory Directed Research and Development grant and by DOE-ASCR/BES funds.

Getting Started:

• Welcome to pyMSDtorch’s documentation!
• Install pyMSDtorch
  • pip: Python package installer
  • From source
  • Tutorials only
• Network Initialization
  • Mixed-Scale dense networks (MSDNet)
  • Sparse mixed-scale dense network (SMSNet)
  • Tunable U-Nets
• Training
• Saving and loading models
  • Saving model weights (recommended)
  • Saving the entire model

Tutorials

MSDNet & TUNet Demo:

• Segmention demo in 2D using Mixed-scale Dense Networks and Tunable U-Nets

SMSNet Ensemble Learning:

• Ensemble Learning with Randomized Sparse Mixed-Scale Networks

Autoencoder Latent Space:

• Latent Space Exploration with UMap and Randomized Sparse Mixed Scale Autoencoders
• Latent Space Exploration with Randomized Sparse Mixed Scale Autoencoders, regularized by the availability of image labels
• (Newer) Latent Space Exploration with Randomized Sparse Mixed Scale Autoencoders, regularized by the availability of image labels

Indices and tables

• Index

• Module Index

• Search Page

License and legal stuff

This software has been developed from funds that originate from the US tax payer and is free for academics. Please have a look at the license agreement for more details. Commercial usage will require some extra steps. Please contact ipo@lbl.gov for more details.

Final thoughts

This documentation is far from complete, but have some notebooks as part of the codebase, which could provide a good entry point.

More to come!