Skip to content

Latest commit

 

History

History
116 lines (80 loc) · 3.89 KB

CHANGELOG.md

File metadata and controls

116 lines (80 loc) · 3.89 KB
Raw

Version 0.12.0

  • Added dropout option to all of the 2d conv blocks.
  • Added dropout to UNet, AutoEncoder2d and VAE2d.

Version 0.11.1

  • Updated the UNet's docstring.

Version 0.11.0

  • Made changes to VAE2d:
    • Added the option to have the mean and variance nets be CNNs, rather than MLPs.
    • Started returning the KL divergence during training and validation.
    • Added extra tests accordingly.

Version 0.10.3

  • Added the option to freeze the encoder in the forward pass of SimpleConvNet2d.

Version 0.10.2

  • Account for the attribution being initialised by normal_init being None.

Version 0.10.1

  • Made Scikit Image a requirement of torch_tools rather than just having it in the dev env; the shapes dataset needs it.

Version 0.10.0

  • Updated torch_tools.weight_init.normal_init with options for which attrs get initialised.

Version 0.9.0

  • Added total variational loss function total_image_variation.

Version 0.8.0

  • Add get_features method to SimpleConvNet2d for extracting encoded features.
  • Add get_features method to ConvNet2d for extracting encoded features.
  • Added a demo using a multiple-instance-learning attention model.

Version 0.7.0

  • Changed AutoEncoder2d demo to use ovarian histology images.

Version 0.6.1

  • Added residual blocks as optional block style to all relevant models.
  • Changed the UNet demo to use a nuclei segmentation data set.

Version 0.6.0

  • Added synthetic shapes dataset and demo.

Version 0.5.6

  • Changed the classification demo for FCNet to the penguin problem.

Version 0.5.5

  • Added py.typed package data to toml.

Version 0.5.4

  • Added missing docs for VAE2d.

Version 0.5.3

  • Added torch_tools.models._blocks_2d.ConvResBlock.

Version 0.5.2

  • Removed biases in residual block.

Version 0.5.1

  • Added py.typed file.

Version 0.5.0

  • Added torch_tools.VAE2d model.

Version 0.4.3

  • Updated the docstring in torch_tools.ConvNet2d.

Version 0.4.2

  • Fixed error in the doc-building caused by Torchvision (and possibly PyTorch).

Version 0.4.1

  • Added a demo for the patchify_img_batch function.

Version 0.4.0

Changes:

  • Added mixup augmentation option to ```DataSet``.

Version 0.3.2

Changes:

  • Fixed typo in FCNet docstring: DenseNetwork -> FCNet
  • Minor refactoring of GitHub workflows.
  • Move some of the dev requirements into the pip section.

Version 0.3.1

Updated the docstring in ConvNet2d to include the mobilenet options.

Version 0.3.0

This release introduces:

  • Mobilenet v3 as an encoder style for ConvNet2d.
  • Fixing a docstring typo in ConvNet2d.

Version 0.2.2

The only difference in this minor version update is a tiny patch to the README: the instructions for the pip install ... commands were missing the "git+" prefixes before the URL. They have now been added.

Version 0.2.1

The source-code links in docs weren't working. This ha now been fixed.

Version 0.2.0

New Features

Variable kernel size

For all of the models using 2D convolutional layers (except ConvNet2d, which uses predetermined architectures), the kernel size argument is now optional. Spoiler alert: it has to be an odd, positive, integer.

Documentation

  • Fixed a typo in the docstring for torch_tools.models._argument_processing.process_dropout_prob. The prob arg should be on [0.0, 1.0), and not (0.0, 1.0] as described. This was only a typo in the docstring and not a bug.
  • Made source link available in the docs.

Python package

  • The repo has moved from the old setup.py to use a pyproject.toml — hopefully correctly. The version imports and python dependencies have been updated accordingly.
  • The demos now use the requirements-dev.conda.yaml.
  • You can now install the package with pip from Github.