Version 0.3.6

Changes

• fixed reverse strand padding for overlap
• fixed resume-training loading models in a frozen state so they couldn't learn; did not affect fine-tuning
• fixed Helixer not working with models trained on a different pool size than the default of 9 (all released models use that default, so they were not affected)
• bulk data read-in is now subsequence length dependent
• improved h5 file documentation

Version 0.3.5

Changes

• added multi-GPU training
• added the option to resume fine-tuning runs (for normal fine-tuning and fine-tuning with coverage) to address issue #175
• added the option to auto-download Helixer models to a different directory than the default, that can also be checked by Helixer.py to address issue #173

Version 0.3.4

Changes

• move package metadata to pyproject.toml
• only support Python 3.10
• improve documentation, both README.md and in docs/ folder
• change TensorFlow GPU support installation instructions in manual_install.md
• add deprecation warnings for unused parameters

v0.3.3

Version 0.3.3

Changes

• version parameter
• installation maintainability and versatility improvements
• issue templates
• simplifications in coverage handling

v0.3.2

release v0.3.2

Changes

• improve documentation, both README.md and in docs/ folder
• implement experimental transfer learning options from large generalizable models, to be
  able to tune a re-initalized final layer with the rest of the model frozen (with or without adding extrinsic information).
• improve flexibility in add_ngs_coverage.py script for extrinsic NGS information

version 0.3.1

CLI changes

• Helixer.py now sets reasonable default for --subsequence-length whenever --lineage is specified
• overlapping parameters can now be automatically derived from --subsequence-length

other changes

• check if output directory exists and is writable at start, instead of after data generation and prediction
• fetch_helixer_models.py is now a stand-alone script in the PATH (instead of having to call python fetch_helixer_models.py)
• user specified compression is now used for all steps in Helixer.py
• improvements to multi-threading performance during data generation
• improvements to clarity and accuracy of help function and error messages
• fixed corner-case bug causing overlapping to fail for very-small but still technically-possible batch-sizes (i.e. 5 w/ overlapping defaults).

Version 0.3.0

Context:

• release associated with training and testing of the paper: Helixer - ab initio Prediction of Primary Eukaryotic Gene Models Combining Deep Learning and a Hidden Markov Model

Changes:

• phase is now optional (select --predict-phase for the better predictions)
• added phase metric calculations
• cleanup and earlier error messages for missing HelixerPost binary