1. Added support for Python3.12 2. Added pre-compiled binaries:
- CPU binaries (macOS/Linux/Windows) - available via _pip install gbrl_ - CUDA 12 binaries (Linux/Windows) - available via _pip install gbrl-gpu_
1.0.11
Fixed bug with PyTorch tensors causing alternating prediction values
1.0.10
1. Fixed bug in training continuation. 2. GBRL now requires to specify the input dim when creating a model instance.
1.0.8
Changed observations to be optional for step function. Updating a model can now be down via model.step() instead of model.step(X).
1.0.7
- Fixed integration with MSVC. - Changed cosine score function in CPU and CUDA.
1.0.6
Improved PyTorch integration
The `step` and `predict/__call__` functions, GBRL functions in GBRL now work directly with PyTorch tensors, eliminating the need to clone data to the CPU and convert it to NumPy.