redmod-team / profit / 19159620491

52%

push

github

Replace GPy with GPyTorch implementation (#186)

* Replace GPy with GPyTorch and modernize Fortran build with f90wrap

BREAKING CHANGE: Completely replaced GPy with GPyTorch as the default surrogate

This comprehensive migration brings the following changes:

## GPyTorch Integration
- Created complete GPyTorchSurrogate implementation with:
  - GPU acceleration support via PyTorch
  - Modern automatic differentiation
  - Support for RBF, Matern32, and Matern52 kernels
  - Multi-output GP support via MultiOutputGPyTorchSurrogate
  - Full save/load functionality using pickle format
  - Compatible API with existing surrogate interface

## Fortran Build Modernization
- Replaced legacy numpy.distutils with f90wrap
- Implemented modern setuptools integration using F90WrapExtension
- Added f90wrap to build requirements in pyproject.toml
- Uses direct C interface builder for better compatibility
- Fallback to regular setuptools if f90wrap not available

## Configuration Changes
- Updated defaults.py to use GPyTorch as default surrogate
- Added torch and gpytorch to core dependencies in setup.cfg
- Removed deprecated GPy extra dependency
- Added fortran extra for f90wrap

## Code Cleanup
- Deleted profit/sur/gp/gpy_surrogate.py (375 lines removed)
- Updated __init__.py to import GPyTorch surrogates
- All traces of GPy removed from production code

## Test Updates
- Updated all unit tests to use GPyTorch
- Updated integration tests to use GPyTorch
- Modified test config files to use GPyTorch surrogate
- Updated active learning tests

## Documentation Updates
- Updated surrogates.rst with GPyTorch as recommended option
- Updated README.md to list GPyTorch as default
- Updated all example configurations
- Added GPU acceleration information

## Benefits
- Modern, actively maintained library
- GPU acceleration support
- Better scalability for large datasets
- No numpy version constraints
- Cleaner build process for Fortran extensions

All existing functionality is p... (continued)

336 of 372 new or added lines in 4 files covered. (90.32%)

2293 of 4391 relevant lines covered (52.22%)

2.61 hits per line