Train a tiny DDPM on a collection of soil-map patches

Train a tiny DDPM on a collection of soil-map patches

Usage

dm_fit(
  stack,
  conditioning = NULL,
  T = 50L,
  epochs = 100L,
  hidden = 32L,
  lr = 0.01,
  seed = NULL,
  backend = c("r", "torch"),
  device = c("cpu", "mps", "cuda")
)

Arguments

stack

A 3-D array of shape (n_patches, H, W) of soil- property patches (e.g. SOC at each pixel, already standardised to zero mean / unit variance).

conditioning

Optional matrix of shape (n_patches, cond_dim) giving a per-patch covariate summary fed to the denoising network as conditioning. Default NULL.

T

Integer; number of diffusion timesteps.

epochs

Integer; training epochs.

hidden

Integer; hidden width of the denoising MLP.

lr

Numeric; learning rate.

seed

Optional RNG seed.

backend

"r" (default, ELM-style MLP denoiser) or "torch" (autograd U-Net denoiser via torch::optim_adam; requires the torch Suggests dependency). v2.7.0 upgrade.

device

"cpu" (default), "mps", or "cuda" when backend = "torch".

Value

An edaphos_dm_fit fit.