Fit the GASF-image CNN single-sample discriminator

Maps the TRAINING matrix to the per-sample robust CLR over the frozen feature universe (colnames(X_train)), computes the FROZEN per-feature GASF bounds, renders each training specimen to a GASF image, and trains a small Conv-BN-ReLU-MaxPool CNN end-to-end by full-batch BCEWithLogitsLoss (Adam, no shuffle, seed-before-build) via reticulate-python torch + torchvision. After training the net is put in eval() mode (frozen BatchNorm) and its FULL state_dict is EXPORTED to R as named float64 numeric arrays (incl. BN running stats); the python module is DISCARDED. The fitted model holds NO external pointer – score rebuilds the CNN from the R state.

Usage

fit_img_gasfcnn(X_train, y_train, meta_train = NULL, hp = list())

Arguments

X_train

Numeric matrix (samples \(\times\) features) of non-negative abundances with unique, non-empty feature names (colnames).

y_train

Numeric / integer 0/1 labels (1 = case), length nrow(X_train), with at least one case and one control.

meta_train

Optional per-sample metadata. Accepted for interface uniformity and otherwise ignored.

hp

Optional list of hyperparameters. Allowed fields: channels (positive-integer vector of per-block conv channel widths; default c(8L, 16L) – two blocks), epochs (full-batch training epochs, positive integer; default 200L), lr (positive Adam learning rate; default 1e-3), weight_decay (non-negative Adam L2; default 1e-4), min_features (feature-overlap floor at scoring, positive integer; default 3L), device ("cpu" (default), "cuda", or "auto"; "cuda"/"auto" fall back to CPU with no GPU), seed (integer; default 42L), and image_size (fixed GASF image side length, integer \(\ge 2\); default 128L). The frozen-order rCLR profile is PAA-reduced to length \(L = \min(p, image\_size)\) before imaging, so the GASF image is at most image_size x image_size regardless of the feature count \(p\); when \(p \le image\_size\) no reduction is applied.

Value

Object of class img_gasfcnn_model: a list with feature_universe, state_dict (named float64 arrays + BN counters), channels, gasf_lo, gasf_hi (frozen per-coordinate GASF bounds, length paa_len), image_size (the PAA target side length) and paa_len (= min(p, image_size), the actual GASF image side), device (resolved), seed, and hp. No python pointer.

References

Wang Z, Oates T. (2015) Imaging Time-Series to Improve Classification and Imputation. IJCAI; arXiv:1506.00327.

Examples

if (FALSE) { # \dontrun{
set.seed(1)
n <- 80; p <- 16; k <- 6
L <- matrix(stats::rnorm(n * p, 4, 0.5), nrow = n,
            dimnames = list(NULL, paste0("miR-", seq_len(p))))
y <- rep(c(0, 1), each = n / 2)
L[y == 1, seq_len(k)] <- L[y == 1, seq_len(k)] + 1.2
X <- exp(L)
model <- fit_img_gasfcnn(X, y)
score_img_gasfcnn(model, X[1, , drop = FALSE])
} # }