Changelog

Release history for foreblocks and foretools.

v0.1.15 (current)

Added mHC (multi-Hyper-Connections) residual stream mixing to TransformerEncoder and TransformerDecoder
Added paper-style Attention Residuals (use_attention_residual, attn_residual_type, attention_residual_block_size)
Added Mixture-of-Depths token routing (use_mod, mod_mode, mod_budget_scheduler)
Added latent-MoE path (moe_use_latent, moe_latent_dim, moe_latent_d_ff) for higher expert count at lower cost
Added adaptive_noisy_topk, hash_topk, and multi_hash_topk router families
Added afocp conformal method with attention-based feature network
Added cptc (state-aware rolling conformal) method
Removed classic dense load-balancing auxiliary loss; expert utilization now handled via router expert-bias adaptation
label_len <= 0 in Informer-like mode no longer generates a full-decoder Informer padding mask
forward_one_step(...) is now incompatible with use_mod=True and use_mhc=True (explicitly guarded)

Added tsgen synthetic series generator with AR, seasonal, trend, and noise components
Added tsaug AutoDA augmentation search
Added foreminer changepoint detection, cluster analysis, and stationarity diagnostics
bohb plotter and observation store improvements

Code importing TimeSeriesSeq2Seq directly should switch to ForecastingModel
load_balance_weight no longer drives the primary balancing mechanism in MoE; tune z_loss_weight and router type instead

Initial public release of foretools/bohb Bayesian hyperparameter optimization
Added ConformalPredictionEngine with split, rolling, agaci, and enbpi methods
Added GateSkip sublayer-level residual gating
Added CT-PatchTST encoder tokenization path (ct_patchtst=True)
Added foretools/fengineer feature engineering pipeline with RFECV and mutual-information selection
Added foretools/vmd decomposition toolkit (VMD, EMD-family, hierarchical VMD)
Stabilized MLTracker local FastAPI server and mltracker-tui TUI

See git log for commit-level history.