2
0
Fork 0
CVE/CVE_R/R/cve_momentum.R

140 lines
4.6 KiB
R
Raw Normal View History

2019-09-25 12:49:12 +00:00
#' Implementation of the CVE method as a Riemann Conjugated Gradient method.
#'
2019-09-25 12:49:12 +00:00
#' @references A Riemannian Conjugate Gradient Algorithm with Implicit Vector
#' Transport for Optimization on the Stiefel Manifold
#' @keywords internal
#' @export
2019-09-25 12:49:12 +00:00
cve_momentum <- function(X, Y, k,
nObs = sqrt(nrow(X)),
h = NULL,
tau = 1.0,
tol = 1e-4,
rho = 0.1, # Momentum update.
slack = 0,
epochs = 50L,
attempts = 10L,
logger = NULL
) {
# Set `grad` functions environment to enable if to find this environments
# local variabels, needed to enable the manipulation of this local variables
# from within `grad`.
environment(grad) <- environment()
# Get dimensions.
n <- nrow(X) # Number of samples.
p <- ncol(X) # Data dimensions
q <- p - k # Complement dimension of the SDR space.
# Save initial learning rate `tau`.
tau.init <- tau
# Addapt tolearance for break condition.
tol <- sqrt(2 * q) * tol
# Estaimate bandwidth if not given.
if (missing(h) || !is.numeric(h)) {
h <- estimate.bandwidth(X, k, nObs)
}
# Compute persistent data.
# Compute lookup indexes for symmetrie, lower/upper
# triangular parts and vectorization.
pair.index <- elem.pairs(seq(n))
i <- pair.index[1, ] # `i` indices of `(i, j)` pairs
j <- pair.index[2, ] # `j` indices of `(i, j)` pairs
# Index of vectorized matrix, for lower and upper triangular part.
lower <- ((i - 1) * n) + j
upper <- ((j - 1) * n) + i
# Create all pairewise differences of rows of `X`.
X_diff <- X[i, , drop = F] - X[j, , drop = F]
# Identity matrix.
I_p <- diag(1, p)
# Init tracking of current best (according multiple attempts).
V.best <- NULL
loss.best <- Inf
# Start loop for multiple attempts.
for (attempt in 1:attempts) {
# Reset learning rate `tau`.
tau <- tau.init
# Sample a `(p, q)` dimensional matrix from the stiefel manifold as
# optimization start value.
V <- rStiefl(p, q)
# Initial loss and gradient.
loss <- Inf
G <- grad(X, Y, V, h, loss.out = TRUE, persistent = TRUE)
# Set last loss (aka, loss after applying the step).
loss.last <- loss
# Call logger with initial values before starting optimization.
if (is.function(logger)) {
epoch <- 0 # Set epoch count to 0 (only relevant for logging).
error <- NA
logger(environment())
}
2019-09-25 12:49:12 +00:00
M <- matrix(0, p, q)
## Start optimization loop.
for (epoch in 1:epochs) {
# Apply learning rate `tau`.
2019-09-25 12:49:12 +00:00
A <- projTangentStiefl(V, G)
# Momentum update.
M <- A + rho * projTangentStiefl(V, M)
# Parallet transport (on Stiefl manifold) into direction of `G`.
2019-09-25 12:49:12 +00:00
V.tau <- retractStiefl(V - tau * M)
# Loss at position after a step.
loss <- grad(X, Y, V.tau, h, loss.only = TRUE, persistent = TRUE)
# Check if step is appropriate, iff not reduce learning rate.
if ((loss - loss.last) > slack * loss.last) {
tau <- tau / 2
next() # Keep position and try with smaller `tau`.
}
# Compute error.
error <- norm(V %*% t(V) - V.tau %*% t(V.tau), type = "F")
# Check break condition (epoch check to skip ignored gradient calc).
# Note: the devision by `sqrt(2 * k)` is included in `tol`.
if (error < tol || epoch >= epochs) {
# take last step and stop optimization.
V <- V.tau
# Call logger last time befor stoping.
if (is.function(logger)) {
logger(environment())
}
break()
}
# Perform the step and remember previous loss.
V <- V.tau
loss.last <- loss
# Call logger after taking a step.
if (is.function(logger)) {
logger(environment())
}
# Compute gradient at new position.
G <- grad(X, Y, V, h, persistent = TRUE)
}
# Check if current attempt improved previous ones
if (loss < loss.best) {
loss.best <- loss
V.best <- V
}
}
return(list(
loss = loss.best,
V = V.best,
B = null(V.best),
h = h
))
}