Release v0.1.2 (2025-06-14)

Performance & Bug Fixes

• Fixed an error that caused full-batch updates of the validation set even when batch size was specified. This resolves the reported out-of-memory issues for large datasets (>100k cells).
• Added @torch.no_grad() decorator to the _eval function, reducing GPU memory usage by nearly 50%.

New Features

• Added noising_mode parameter to control denoising autoencoding schemes:
  • "default": Original permutation procedure (sampling without replacement)
  • "zero": Zero masking approach
  • "sampling": Sampling with replacement (used for any keyword other than "default" or "zero")
• Both new schemes ("zero" and "sampling") are significantly faster than the original implementation.
• The "sampling" option closely resembles the default procedure but allows replacement, making it more computationally efficient.

Performance Improvements

The new version has been tested on a 250k cells × 2000 genes dataset on a Linux server with an NVIDIA A6000 Ada GPU, completing 100-epoch training in <30 minutes using 16GB GPU memory.