scAlignOptions: Set training options

In scAlign: An alignment and integration method for single cell genomics

Description

Defines parameters for optimizer and training procedure.

Usage

scAlignOptions(steps = 15000, batch.size = 150,
  learning.rate = 1e-04, log.every = 5000, architecture = "large",
  batch.norm.layer = FALSE, dropout.layer = TRUE, num.dim = 32,
  perplexity = 30, betas = 0, norm = TRUE, full.norm = FALSE,
  early.stop = FALSE, walker.loss = TRUE, reconc.loss = FALSE,
  walker.weight = 1, classifier.weight = 1, classifier.delay = NA,
  gpu.device = "0", seed = 1234)

Arguments

steps	(default: 15000) Number of training iterations for neural networks.
batch.size	(default: 150) Number of input samples per training batch.
learning.rate	(default: 1e-4) Initial learning rate for ADAM.
log.every	(default: 5000) Number of steps before saving results.
architecture	(default: "small") Network function name for scAlign.
batch.norm.layer	(default: FALSE) Include batch normalization in the network structure.
dropout.layer	(default: TRUE) Include dropout in the network.
num.dim	(default: 32) Number of dimensions for joint embedding space.
perplexity	(default: 30) Determines the neighborhood size for each sample.
betas	(default: 0) Sets the bandwidth of the gaussians to be the same if > 0. Otherwise per cell beta is computed.
norm	(default: TRUE) Normalize the data mini batches while training scAlign (repeated).
full.norm	(default: FALSE) Normalize the data matrix prior to scAlign (done once).
early.stop	(default: TRUE) Early stopping during network training.
walker.loss	(default: TRUE) Add walker loss to model.
reconc.loss	(default: FALSE) Add reconstruction loss to model during alignment.
walker.weight	(default: 1.0) Weight on walker loss component
classifier.weight	(default: 1.0) Weight on classifier loss component
classifier.delay	(default: NULL) Delay classifier component of loss function until specific training step. Defaults to (2/3)*steps.
gpu.device	(default: '0') Which gpu to use.
seed	(default: 1245) Sets graph level random seed in tensorflow.

Value

Options data.frame

Examples

options=scAlignOptions(steps=15000,
                       log.every=5000,
                       early.stop=FALSE,
                       architecture="large")

scAlign documentation built on April 28, 2020, 6:10 p.m.