Changelog¶
v0.1.0¶
Model¶
Replaced the learnable channel-wise scaling with a fixed constant.
Added an exponential moving average (EMA) of model weights during training.
Changed the final profile convolution to
kernel_width=1.Set the default model size to 96 filters.
Tuned the Muon and AdamW learning rates and weight decay values for improved convergence.
Best-model selection now monitors the validation profile loss rather than the total validation loss.
Training¶
Default
max_jitterfor fitting lowered from 500 to 50 (applied across the package’s fitting defaults).
Packaging and tooling¶
Migrated from
setup.pytopyproject.tomlwithuvsupport.Refactored the CLI from a monolithic script into the
cherimoya_climodular package.Raised the minimum Python version to 3.10 and minimum PyTorch to 2.9.
Added
macs3,bam2bw,bpnet-lite,triton, andjoblibas dependencies.Added a Sphinx documentation site hosted on Read the Docs.
v0.0.1¶
Initial release of the Cherimoya model and pipeline.
Includes the CheriBlock architecture and custom kernels.
Features a dual-optimizer training strategy (AdamW + Muon).
Implements a full end-to-end processing and modeling pipeline.