Skip to content

NEWS.md

Latest commit

 

History

History
43 lines (29 loc) · 4.12 KB

NEWS.md

File metadata and controls

43 lines (29 loc) · 4.12 KB
Raw

RegDiffusion News / Changelog

0.2.3 [unreleased]

New Features

  • Added a seed argument to RegDiffusionTrainer for fully reproducible runs. When set, Python, NumPy, and PyTorch RNGs are seeded and cuDNN is put in deterministic mode. Two runs with the same seed now produce bit-identical adjacency matrices, verified across separate processes and under AMP, torch.compile, gradient accumulation, and train_split < 1.0. Without a seed (the default), runs remain non-deterministic.

Bug Fixes

  • Fixed RegDiffusionTrainer.train(n_steps=...) silently ignoring its argument. Both the normal and gradient-accumulation branches hardcoded n_steps=None, so training always ran for the n_steps given at initialization.
  • Fixed the L1 sparsity penalty being computed on signed adjacency values in the base RegDiffusion model. Soft thresholding preserves sign, so the penalty could in principle be reduced by making inhibitory (negative) edges more negative. Both models now share a single get_sparse_loss() that takes the mean absolute value over off-diagonal entries. In practice this does not change results — negative entries land in the soft-threshold dead zone, where both the output and its gradient are zero — and BEELINE metrics are unchanged; the penalty is now correct by construction rather than by accident.
  • RegDiffusionME.get_sampled_sparse_loss() is renamed to get_sparse_loss() and now excludes diagonal samples, matching the base model, whose get_adj_() masks the diagonal.
  • Registered the sampled adjacency index tensors (sampled_adj_row/col_nonparam) as buffers instead of nn.Parameter. These are integer indices, not learnable weights, and they no longer appear in model.parameters(). state_dict keys are unchanged, so existing checkpoints still load.
  • The sparse loss is no longer computed during the warmup window, where it is discarded.
  • Extended the tqdm.auto progress-bar fix from 0.2.2 (which covered trainer.py) to the remaining progress bars in grn.py and the data module, so dataset downloads and edgelist extraction also render properly in notebooks.

0.2.2

  • Avoid large progress printing in terminal

0.2.1

New Features

  • Added sparse matrix support for RegDiffusionTrainer. The exp_array argument now accepts scipy sparse matrices (e.g. adata.X) directly. Normalization statistics are computed in chunks and each sample is normalized on-the-fly during training, so the full dense matrix is never materialized. This enables training on datasets with 1M+ cells without excessive memory usage.
  • Updated CLI (regdiffusion command) to handle sparse adata.X from H5AD files, including sparse-safe data validation and log-transform via log1p on non-zero entries.

0.2.0

Memory Optimization

  • Removed three (n_gene, n_gene) helper matrices (zeros_nonparam, eye_nonparam, mask_nonparam) from RegDiffusion, replacing them with inline computations. Saves ~4.8 GB persistent GPU memory for 20K genes with no impact on model accuracy.

New Features

  • Added RegDiffusionME, a memory-efficient model variant enabled via memory_efficient=True in RegDiffusionTrainer. Reduces peak GPU memory by ~45% with no impact on accuracy. Uses a custom autograd function for soft thresholding (boolean masks instead of float32 tensors) and sampled sparse loss (avoids materializing full adjacency matrix for L1 regularization). Benchmarked on all 7 BEELINE datasets with identical AUROC/AUPRR/EPR.
  • Added automatic mixed precision (AMP) support via use_amp=True in RegDiffusionTrainer. Uses bfloat16 for forward pass and loss computation, reducing memory for autograd-saved activations while keeping model parameters in float32. Requires Ampere or newer GPU.

Bug Fixes

  • Fixed I_minus_A() using self.train instead of self.training, causing dropout to always apply during inference
  • Fixed missing self parameter in GRN.remove_weak_edges()
  • Fixed GRN.get_edgelist() incorrectly passing self to extract_edgelist()
  • Fixed external logger not being assigned in RegDiffusionTrainer.__init__()
  • Fixed forward() crash when n_celltype=None by adding conditional cell type embedding