Version 2.3

Adds the ability to train across multiple nodes with multiple GPUs, some Jupyter Notebooks to analyse results, and some minor bug fixes/clean-ups.

Changes

• Added the ability to train using PyTorch's DistributedDataParallel module, speeding up training by around 5x
• Changed Generator architecture to only use 4x4 kernels
• Added more control over saving a model during training; an interval in epochs can be specific at which to save the model
• Added lists of clean and stripe indices to PatchVisualiser class
• Models are now saved with the --name argument, rather than the --mode argument
• Added parameter for the Centre of Rotation in the reconstruct() function
• Added a script to generate a mask of stripe locations
• Added a script to apply the model to arbitrary tomographic scans
• Added Jupyter Notebooks for visualising results, RMSEs, residuals and graphs
• Added a tutorial on how to use the repository
• Updated documentation and removed old/unused code
• Minor bug fixes

Version 2.2

Adds the ability to generate "patches" of sinograms, rather than whole sinograms.
Patches are of size (1801, 256) by default, and are not downsampled.
The motivation behind this is two-fold:

1. The network only sees a smaller area around each stripe, allowing it to focus on important local information
2. By splitting one sinogram into many patches, the network has more data to train on

Changes

• Added a new mode of generating data: "patch"
• Added a new mode for training & testing: "patch"
• Added a new visualizer class to display patch-based data
• Added wandb support
• Updated Dataset classes so that they can process synthetic, real-life and patch-based directory structures
• Datasets are now shuffled before the train/validate/test split, as well as after
• Increased width of simulated stripes, and removed the expansion of mask widths
• Minor bug fixes

Version 2.1

This version includes a series of changes to make the project more readable and understandable.

Changes

• Added documentation to most functions
• Added installation instructions to README.md and moved old content to PROJECT_DESCRIPTION.md
• Improved the code in 'simulator/realdata_loader.py' to make it neater, more modular, and easier to understand
• Added a description of parameters & options for both data generation and training/testing
• Improved & updated command-line option descriptions for data generation and training/testing
• Removed from ... import * from utils/__init__.py to hopefully speed up importing from utils
• Updated simulateStripe() from simulator/data_simulator.py to use a stripe/ directory (rather than shift00/)
• Changed name of real mode in data generation to paired to make its function more clear & distinguish from other methods of processing real data
• loadHDF() from utils/data_io.py now only returns data, rather than the tuple (data, angles)
• Made all lines less than 80 characters long

Version 2.0

New Features

• Real-life data from hdf5 and nexus files can now be processed and saved as tiffs, so that a model can be trained on them
• There are currently three different ways of saving data:
  • Raw: apply no post-processing methods, save the data exactly as it is on disk
  • Paired: create an input/target pair for each sinogram in the data
  • Dynamic (for dynamic experiments only): save each "frame" of a dynamic tomography scan
• New Stripe Detection method featuring Morphological Processing and Clustering

Changes

• New utils directory contains functions shared between different sub-modules
• Some small improvements to utils functions, including test metrics, rescaling, and plotting
• Some old dataset code has been deprecated; a new, more scalable dataset class has been created

Version 1.0

Features

• Simulated data can be generated using either simple stripe artifacts or more realistic flat fields
• A neural network can be trained on this dataset with the following modes:
  • base: input whole sinograms, generate whole sinograms
  • window: input windowed sinograms, generate windowed sinograms; windows are automatically created from inputs
  • full: input whole sinograms, generate whole sinograms; whole sinograms are created by concatenating windows
  • mask: input masked sinograms, generate masked sinograms; masks are automatically created from inputs
  • simple: same as mask, but masks are created from both inputs and targets
• Neural networks can also be trained using LSGANs (rather than cGANs) with any of the above modes
• A trained network can then be tested, outputting the following test statistics:
  • MAE, L2 norm, MSE, Gradient Difference, Dice Coefficient, IoU, Histogram Intersection, Structural Similarity and PSNR