discovery-unicamp / Minerva

Minerva is a framework for training machine learning models for researchers.

https://discovery-unicamp.github.io/Minerva/

MIT License

3 stars 7 forks source link

Hyperparameter Optimization Pipelines and Supporting Tools #74

Open GabrielBG0 opened 2 weeks ago

GabrielBG0 commented 2 weeks ago

Change Log

This PR implements a series of new features:

Transforms

new Crop
new Normalize
new Transpose
new Resize
new EmptyTransform
new _RandomSyncedTransform
new RandomResize
new RandomFlip
new ClassRatioCrop

Metrics

new CroppedMetric
new ResizedMetric

Callbacks

new TrainerReportKeepOnlyLastCallback
new TrainerReportOnIntervalCallback

Datasets

SupervisedReconstructionDataset now supports support_context_transforms

Engines

new _Engine
new PatchInferencer

Models

SETR_PUP now supports optimizers as parameters
SETR_PUP now supports selection of the aux output layers and aux weights
SETR_PUP now supports loading backbone weights and freezing backbone weights
SETR_PUP now supports weighted losses
SETR_PUP now supports interpolation of loaded weights input size
SETR_PUP now supports different learning rates for backbone and head
SETR_PUP now supports different prediction engines for val and test
_VisionTransformerBackbone now supports selection of the aux output layers
_VisionTransformerBackbone now supports loading backbone weights and freezing backbone weights
_VisionTransformerBackbone now supports interpolation of loaded weights input size

Pipelines

new RayHyperParameterSearch
new HyperoptHyperParameterSearch

Other Improvements

General typo fixes
General typing fixes

Issue ticket number and link (If apply)

48

Checklist before requesting a review

[ ] I have performed a self-review of my code.
[ ] I have added tests to this functionality.
[ ] All tests are passing.
[ ] I have updated the documentation as needed.
[ ] I have followed the project's coding guidelines.