Train a Scaden ensemble in torch

Trains a three-model multilayer perceptron ensemble corresponding to the Scaden architectures `m256`, `m512`, and `m1024`. The function expects a processed training matrix with samples in rows and genes in columns, and a matching target matrix with cell-type proportions.

Usage

scaden(
  train_x,
  train_y,
  lr = 1e-04,
  batch_size = 128,
  epochs = 20,
  seed = 123,
  device = NULL
)

Arguments

train_x: Numeric matrix with samples in rows and genes in columns.
train_y: Numeric matrix with samples in rows and cell types in columns.
lr: Numeric scalar. Learning rate passed to Adam.
batch_size: Integer scalar. Minibatch size.
epochs: Integer scalar. Number of training epochs for each ensemble model.
seed: Integer scalar random seed.
device: Optional torch device. If `NULL`, CUDA is used when available, otherwise CPU is used.

Value

A named list with components:

model256: Trained `m256` torch model.
model512: Trained `m512` torch model.
model1024: Trained `m1024` torch model.
loss256: Numeric vector of batch-level losses for the `m256` model.
loss512: Numeric vector of batch-level losses for the `m512` model.
loss1024: Numeric vector of batch-level losses for the `m1024` model.
epoch_loss256: Numeric vector of mean epoch losses for the `m256` model.
epoch_loss512: Numeric vector of mean epoch losses for the `m512` model.
epoch_loss1024: Numeric vector of mean epoch losses for the `m1024` model.
total_training_time_sec: Numeric scalar giving total ensemble training time in seconds.
architectures: List of model architectures used.
lr: Learning rate used for training.
batch_size: Batch size used for training.
epochs: Number of training epochs.
inputdim: Number of input genes.
outputdim: Number of output cell types.
celltypes: Character vector of output cell-type names.
train_samples: Character vector of training sample names.
device: Torch device used for training.

Details

The ensemble consists of three feed-forward neural networks with softmax output. Predictions are averaged across the three trained models in [scaden_predict()].

Examples

if (FALSE) { # \dontrun{
fit <- scaden(
  train_x = proc$train_x,
  train_y = proc$train_y,
  lr = 1e-4,
  batch_size = 128,
  epochs = 20
)
} # }