Skip to contents

SimCLR pre-training on raster covariate patches (Pillar 4 scaffold)

Source: R/foundation_simclr.R
foundation_simclr_pretrain.Rd

Trains a small CNN encoder on unlabeled raster patches via the SimCLR contrastive objective. Each forward pass draws two independent augmented views of every patch in the mini-batch and enforces high embedding similarity between views of the same patch and low similarity otherwise.

Usage

foundation_simclr_pretrain(
  patches,
  feature_dim = 32L,
  proj_dim = 16L,
  batch_size = 8L,
  epochs = 30L,
  lr = 0.005,
  temperature = 0.2,
  noise_sd = 0.1,
  seed = NULL,
  verbose = FALSE
)

Arguments

patches

A 4-D R array shaped (N, C, H, W): N patches, C covariate channels, spatial H x W.

feature_dim, proj_dim

Integer; backbone and projection head widths.

batch_size

Integer; SimCLR mini-batch size. Each batch contributes 2 * batch_size - 2 negatives per anchor.

epochs, lr

Training hyperparameters for Adam.

temperature

Numeric; NT-Xent temperature.

noise_sd

Numeric; additive-noise strength during augmentation.

seed, verbose

As elsewhere.

Value

A edaphos_foundation_simclr object (S3).