Skip to contents

Changelog

Source: NEWS.md

DeepPatientLevelPrediction 2.2.0

  • [Feature] Add positional encodings that use timeId (PR #154):
    • SinusoidalPE: Absolute sinusoidal positional embeddings using
    • LearnablePE: Learnable absolute positional embeddings
    • TapePE: Time-absolute Positional Encodings (tAPE)
    • RotaryPE: Rotary position embeddings (RoPE)
    • RelativePE: Relative positional embeddings (RPE)
    • EfficientRPE: Efficient relative positional embeddings (eRPE)
    • TemporalPE: Combines an absolute temporal embedding and a semantic one
    • StochasticConvPE: Relative learnable positional embedding (convSPE)
    • HybridRoPEconvPE: Combines RoPE and convSPE
    • TUPE: Transformer with learnable positional encoding (TUPE)
    • ALiBiPE: Attention with linear biases (ALiBi)
  • [Feature] Use py_require from reticulate to manage python dependencies and update min requirements (PR #150)
  • [Internal] Refactor transformer/dataset/embedding classes to use same code wether temporal or not (PR #147)
  • [Feature] Use train/validation split for model selection instead of cross validation (PR #145)
  • [Feature] Temporal transformer added which supports RopE and time tokens (PR #147)
  • [Feature] Temporal data processing added (PR #147)
  • [CI] Use uv for python in github actions (PR #136)
  • [Feature] Add an option to use torch compile (PR #133)
  • [Feature] More efficient conversions from polars to torch in dataset processing (PR #133)
  • [CI] Automatically detect broken links in docs using github actions
  • [Feature] Initialize models with embeddings, including Poincare (PR #125 and PR #130)
  • [Feature] Model initialization made more flexible with classes (PR #124)

DeepPatientLevelPrediction 2.1.0

  • Added basic transfer learning functionality. See vignette(“TransferLearning”)
  • Add a gpu memory cleaner to clean cached memory after out of memory error
  • The python module torch is now accessed through an exported function instead of loading the module at package load
  • Added gradient accumulation. Studies running at different sites using different hardware can now use same effective batch size by accumulating gradients.
  • Refactored out the cross validation from the hyperparameter tuning
  • Remove predictions from non-optimal hyperparameter combinations to save space
  • Only use html vignettes
  • Rename MLP to MultiLayerPerceptron

DeepPatientLevelPrediction 2.0.3

  • Hotfix: Fix count for polars v0.20.x

DeepPatientLevelPrediction 2.0.2

  • Ensure output from predict_proba is numeric instead of 1d array
  • Refactoring: Move cross-validation to a separate function
  • Refactoring: Move paramsToTune to a separate function
  • linting: Enforcing HADES style
  • Calculate AUC ourselves with torch, get rid of scikit-learn dependancy
  • added Andromeda to dev dependencies

DeepPatientLevelPrediction 2.0.1

  • Connection parameter fixed to be in line with newest polars
  • Fixed a bug where LRFinder used a hardcoded batch size
  • Seed is now used in LRFinder so it’s reproducible
  • Fixed a bug in NumericalEmbedding
  • Fixed a bug for Transformer and numerical features
  • Fixed a bug when resuming from a full TrainingCache (thanks Zoey Jiang and Linying Zhang )
  • Updated installation documentation after feedback from HADES hackathon
  • Fixed a bug where order of numeric features wasn’t conserved between training and test set
  • TrainingCache now only saves prediction dataframe for the best performing model

DeepPatientLevelPrediction 2.0.0

  • New backend which uses pytorch through reticulate instead of torch in R
  • All models ported over to python
  • Dataset class now in python
  • Estimator class in python
  • Learning rate finder in python
  • Added input checks and tests for wrong inputs
  • Training-cache for single hyperparameter combination added
  • Fixed empty test for training-cache

DeepPatientLevelPrediction 1.1.6

  • Caching and resuming of hyperparameter iterations

DeepPatientLevelPrediction 1.1.5

  • Fix bug where device function was not working for LRFinder

DeepPatientLevelPrediction 1.1.4

  • Remove torchopt dependancy since adamw is now in torch
  • Update torch dependency to >=0.10.0
  • Allow device to be a function that resolves during Estimator initialization

DeepPatientLevelPrediction 1.1.3

  • Fix actions after torch updated to v0.10 (#65)

DeepPatientLevelPrediction 1.1.2

  • Fix bug introduced by removing modelType from attributes (#59)

DeepPatientLevelPrediction 1.1

  • Check for if number of heads is compatible with embedding dimension fixed (#55)
  • Now transformer width can be specified as a ratio of the embedding dimensions (dimToken), (#53)
  • A custom metric can now be defined for earlyStopping and learning rate schedule (#51)
  • Added a setEstimator function to configure the estimator (#51)
  • Seed added for model weight initialization to improve reproducibility (#51)
  • Added a learning rate finder for automatic calculatio of learning rate (#51)
  • Add seed for sampling hyperparameters (#50)
  • used vectorised torch operations to speed up data conversion in torch dataset

DeepPatientLevelPrediction 1.0.2

  • Fix torch binaries issue when running tests from other github actions
  • Fix link on website
  • Fix tidyselect to silence warnings.

DeepPatientLevelPrediction 1.0.1

  • Added changelog to website
  • Added a first model tutorial
  • Fixed small bug in default ResNet and Transformer

DeepPatientLevelPrediction 1.0.0

  • created an Estimator R6 class to handle the model fitting
  • Added three non-temporal models. An MLP, a ResNet and a Transformer
  • ResNet and Transformer have default versions of hyperparameters
  • Created tests and documentation for the package