Skip to content

Latest commit

 

History

History
178 lines (148 loc) · 6.48 KB

HISTORY.md

File metadata and controls

178 lines (148 loc) · 6.48 KB
Raw

0.0.17

  • Added a CLI for metrics computation
  • Added metrics specific for image-to-image comparison
  • Allow penalty and class_weights defined in the config to be used instead of a re-computation
  • Resume functionality now includes penalty calculation

0.0.16

  • Added a script "gandlf_deploy", allowing deployment of models into MLCubes (currently requires Docker)
  • ImageNet pre-trained models for UNet with variable encoders is now available
  • ACS/Soft conversion is available for ImageNet-pretrained UNet
  • Updated links, copyright and email to MLCommons
  • Allowing provision for user to generate multiple configurations for experimentation
  • Added ability to combine classification inference results from different architectures
  • Added ability to save initial and latest models in addition to best
  • Added ability to specify testing data csv in main cli
  • Normalized surface dice has been added
  • Added dedicated script to perform post-training model optimization
  • Added CI and documentation for OpenFL integration
  • Added getting started guide
  • Added documentation for all loss functions and updated guideline
  • Added ability to save resized images instead of loading them directly

0.0.15

  • Updated setup.py for python>=3.8
  • stride_size is now handled internally for histology data
  • Probability maps are now saved overlaid with original WSI
  • Added ability to print model size and summary at run-time
  • Improved error checking added for WSI inference
  • VIPS has been removed from dependencies
  • Failed unit test cases are now recorded
  • Per class accuracy has been added as a metric
  • Dedicated rescaling preprocessing function added for increased flexibility
  • Largest Connected Component Analysis is now added
  • Included metrics using overall predictions and ground truths

0.0.14

  • Add an option ("save_training") to save training patches
  • Add option to save per-label segmentation metrics
  • Separate "motion" artifact
  • DenseNet now supports InstanceNorm
  • Updated implementations of VGG and DenseNet to use ModelBase for consistency
  • Model saving now includes the git commit hash
  • Added FAQ in documentation
  • Accuracy is now standardized from torchmetrics
  • New post-processing module added
  • Anonymization module has been added
  • More progress bars added for better feedback
  • NIfTI conversion added in anonymization
  • Using TiffSlide instead of OpenSlide
  • Minimum resampling resolution is now available
  • Adding option to resize images and resize patches separately
  • Reverse one-hot logic is now updated to output unique labels
  • User can now resume previous training with and without parameter/data updates
  • Docker images are now getting built
  • Inference works without having access to ground truth labels
  • Map output labels using post-processing before saving
  • Enable customized histology classification output via heatmaps
  • Added ImageNet pre-trained models
  • Added RGBA to RGB conversion for preprocessing
  • Model can now be saved at every epoch
  • Different options for final inference
  • Added submodule to handle template-based normalization
  • RGB conversion submodule added to handle alpha channel conversions
  • Sigmoid multiplier option has been added
  • Compute objects can now be requested using developer-level functions
  • Transformer-based networks, TransUNet and UNetR are now available
  • Can now perform histology computation using microns

0.0.13

  • Deep supervision added
  • Documentation updated
  • Model IO is now standardized

0.0.12

  • Misc bugfixes
  • Automatic check-pointing of the model has been added
  • Extending the codebase has been simplified
  • New optimizers added
  • New metrics added
  • Affine augmentation can now be significantly fine-tuned
  • Update logic for penalty calculation
  • RGB-specific augmentation added
  • Cropping added

0.0.11

  • Misc bugfixes for segmentation and classification
  • DFU 2021 parameter file added
  • Added SDNet for supervised learning - https://doi.org/10.1016/j.media.2019.101535
  • Added option to re-orient all images to canonical
  • Preprocessing and augmentation made into separate submodules

0.0.10

  • Half-time epoch loss and metric output added for increased information
  • Gradient clipping added
  • Per-epoch details in validation output added
  • Different types of normalization layer options added
  • Hausdorff as a validation metric has been added
  • New option to save preprocessed data before the training starts

0.0.9

  • Refactoring the training and inference code
  • Added offline mechanism to generate padded images to improve training RAM requirements

0.0.8

  • Pre-split training/validation data can now be provided
  • Major code refactoring to make extensions easier
  • Added a way to ignore a label during validation dice calculation
  • Added more options for VGG
  • Tests can now be run on GPU
  • New scheduling options added

0.0.7

  • New modality switch added for rad/path
  • Class list can now be defined as a range
  • Added option to train and infer on fused labels
  • Rotation 90 and 180 augmentation added
  • Cropping zero planes added for preprocessing
  • Normalization options added
  • Added option to save generated masks on validation and (if applicable) testing data

0.0.6

  • Added PyVIPS support
  • SubjectID-based split added

0.0.5

  • 2D support added
  • Pre-processing module added
    • Added option to threshold or clip the input image
  • Code consolidation
  • Added generic DenseNet
  • Added option to switch between Uniform and Label samplers
  • Added histopathology input (patch-based extraction)

0.0.4

  • Added full image validation for generating loss and dice scores
  • Nested cross-validation added
    • Collect statistics and plot them
  • Weighted DICE computation for handling class imbalances in segmentation

0.0.3

  • Added detailed documentation
  • Added MSE from Torch
  • Added option to parameterize model properties
    • Final convolution layer (softmax/sigmoid/none)
  • Added option to resize input dataset
  • Added new regression architecture (VGG)
  • Version checking in config file

0.0.2

  • More scheduling options
  • Automatic mixed precision training is now enabled by default
  • Subject-based shuffle for training queue construction is now enabled by default
  • Single place to parse and pass around parameters to make training/inference API easier to handle
  • Configuration file mechanism switched to YAML

0.0.1 (2020/08/25)

  • First tag of GaNDLF
  • Initial feature list:
    • Supports multiple
      • Deep Learning model architectures
      • Channels/modalities
      • Prediction classes
    • Data augmentation
    • Built-in cross validation