v5.1.0

Improvements

General

• Update readme to include architectural guidelines

Core

• Refactored classes to have one class per file for better visibility
• Refactored tests to fit with new structure

v5.0.1

Bug Fixes

• Fix issue with pipeline in core that wasn't setting hashes properly

v5.0.0

Improvements

General

• Rename epochalyst to epochlib
• Integrate agogos code into package

CI/CD

• Configure main merging workflow to use rye

v4.1.0

Improvements

General

• Make numpy a required dependency (not optional)

CI/CD

• Configure pre-commit.ci to cache environments
• Dependabot merges to version branch workflow

Torch Trainer

• Unprivatise functions
• Support for mixed precision models
• Updated logging messages
• Support for custom dataset input

Data

• New dataset that can take training pipeline
• New data class based on enum retrieval

Cacher

• Reduced complexity
• Paths are now Path instead of str

v4.0.0

Improvements

General

• Shorten import path lengths
• Add default logger
• Setup dependabot to manage dependencies
• Deploy documentation on GitHub Pages

CI/CD

• Only run tests on python version 3.10 every PR and on every version for release PRs

Documentation

• Add optional dependencies description to README.md

TIMM

• Refactor class to include

Torch Trainer

• Add ONNX/Openvino compile option to trainer

Augmentations

• Add BirdCLEF audio augmentations

Bug Fixes

General

• Fix PyPi logo

v0.3.7

Improvements

General

• Allow project to be managed with Rye
• Update pre-commit configuration
• Add CITATION.cff file to allow project to be cited

Torch trainer

• Option for prefix and suffix for logging in torch trainer

Bug Fixes

Torch Trainer

• Make patience have default value -1 in trainer
• to_predict and tm folder are not considered in torch trainer hash

v0.3.6

Torch trainer Bugs

• Fix collate_fn changing hash of torch trainer

v0.3.5

Pyproject.toml Bugs

• Remove setup tools

v0.3.4

Improvements

Torch Trainer

• Implement Checkpoints (checkpointing_enabled, checkpointing_keep_every, checkpointing_resume_if_exists). After each epoch, the model is saved to disk. By default only the current checkpoint is saved on disk, while old ones are deleted. This can also be changed.
• A custom collate_fn for the DataLoader class is can now be passed into TorchTrainer. This is useful for e.g. GNN's

Models/timm Improvements

• It is now possible to specify whether to use the pre-trained model or not.

PyTest

• Refactored Tests to use fixtures
• Resolve most warnings

Bug Fixes

Torch Trainer

• Now raises an error if model_name is not specified.
• Now check if the specified trained_models_directory exists, and creates it.
• Fix a bug in _train_one_epoch: Copied x into the y_batch, instead of y.

v0.3.3

Improvements

Torch Trainer

• Fix bug in scheduler step
• Add ability to specify dataloader arguments
• Add ability to specify type of the x and y tensors

Logging

• When cache is stored a message is logged to terminal, fixing issue with large datasets taking a long time to store with no indicator what is happening within the pipeline