dnn 0.0.8
- Cox fitting now handles tied observed times with Breslow risk sets
in R and C++. Baseline hazards, survival curves, residuals, and RMST use
consistent grouped risk sets. Full-batch fitting caches tie groups.
- Corrected native tanh and Cox gradients, weighted MSE, binomial,
Poisson, and MAE loss calculations, Adam learning-rate scaling, and
optimizer-state initialization. Native fitting preserves the supplied
model.
- Observation weights default to 1 and accept finite nonnegative
vectors of sample size, including one-column matrices converted to
vectors. Weights are not normalized. Cox sample weights remain
unsupported.
- Corrected survival formula handling, single-layer and narrow-network
dimensions, censored-maximum handling, and deepAFT convergence
updates.
- Corrected survival residuals, exact integration of stepwise
model-based RMST, IBS ordering, small-sample IPCW grouping, and
uncensored imputation.
- Tuning rebuilds sampled architectures, returns the best model with
its matching controls, bounds failed attempts, and errors when no
candidate is selected. Iteration and fold counts are validated before
training.
- Rcpp fitting disables unsupported dropout with a warning. Final
dnnFit diagnostics are evaluated on all observations with dropout
disabled.
- Non-finite numerical results now produce an explicit error
recommending rescaling the data or adjusting model parameters.
- Added regression tests for the corrected loss functions, fitting
paths, survival calculations, tuning, and input contracts.