daniel
/
NNSDR
2
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

add: big data simulation with p growing proportional to n

This commit is contained in:
Daniel Kapla 2021-09-01 12:44:05 +02:00
parent 0214823794
commit 69a008535b
3 changed files with 104 additions and 45 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
simulations/simulations.R
View File

@ -14,7 +14,7 @@ args <- parse.args(defaults = list(
dataset = '1', # Name (number) of the data set dataset = '1', # Name (number) of the data set
# Neuronal Net. structure/definitions # Neuronal Net. structure/definitions
hidden_units = 512L, hidden_units = 512L,
activation = 'relu', # or `relu` activation = 'relu',
trainable_reduction = TRUE, trainable_reduction = TRUE,
# Neuronal Net. training # Neuronal Net. training
epochs = c(200L, 400L), # Number of training epochs for (`OPG`, Refinement) epochs = c(200L, 400L), # Number of training epochs for (`OPG`, Refinement)

98
simulations/simulations_bigdata.R
View File

@ -12,6 +12,10 @@ args <- parse.args(defaults = list(
# Simulation configuration # Simulation configuration
reps = 10L, # Number of replications reps = 10L, # Number of replications
dataset = '6', # Name (number) of the data set dataset = '6', # Name (number) of the data set
# Sets if reference methods shall be evaluated
run_mave = TRUE,
run_cve = TRUE,
run_nn = TRUE,
# Neuronal Net. structure/definitions # Neuronal Net. structure/definitions
hidden_units = 512L, hidden_units = 512L,
activation = 'relu', activation = 'relu',
@ -34,7 +38,7 @@ ds <- dataset(args$dataset, n = 100L, p = args$p) # Generates a list with `X`, `
## Build Dimension Reduction Neuronal Network model (matching the data) ## Build Dimension Reduction Neuronal Network model (matching the data)
nn <- nnsdr$new( nn <- nnsdr$new(
input_shapes = list(x = ncol(ds$X)), input_shapes = list(x = ncol(ds$X)),
d = ncol(ds$B), d = ncol(ds$B), # depends on the dataset type
hidden_units = args$hidden_units, hidden_units = args$hidden_units,
activation = args$activation, activation = args$activation,
trainable_reduction = args$trainable_reduction trainable_reduction = args$trainable_reduction
@ -53,51 +57,57 @@ for (rep in seq_len(args$reps)) {
## Sample test dataset ## Sample test dataset
ds.test <- dataset(ds$name, n = 1000L, p = args$p) ds.test <- dataset(ds$name, n = 1000L, p = args$p)
## First the reference method `MAVE` if (args$run_mave) {
# To be fair for measuring the time, set `max.dim` to true reduction dimension ## First the reference method `MAVE`
# and with `screen = ncol(X)` screening is turned "off". # To be fair for measuring the time, set `max.dim` to true reduction
time <- system.time(dr <- mave.compute(X, Y, max.dim = ncol(B), # dimension and with `screen = ncol(X)` screening is turned "off".
method = "meanMAVE", screen = ncol(X))) time <- system.time(dr <- mave.compute(X, Y, max.dim = ncol(B),
d.sub <- dist.subspace(B, coef(dr, ncol(B)), normalize = TRUE) method = "meanMAVE", screen = ncol(X)))
d.gra <- dist.grassmann(B, coef(dr, ncol(B))) d.sub <- dist.subspace(B, coef(dr, ncol(B)), normalize = TRUE)
mse <- mean((predict(dr, ds.test$X, dim = ncol(B)) - ds.test$Y)^2) d.gra <- dist.grassmann(B, coef(dr, ncol(B)))
cat('"mave",', rep, ',', d.sub, ',', d.gra, ',', mse, ',', mse <- mean((predict(dr, ds.test$X, dim = ncol(B)) - ds.test$Y)^2)
time['user.self'], ',', time['sys.self'], ',', time['elapsed'], '\n', cat('"mave",', rep, ',', d.sub, ',', d.gra, ',', mse, ',',
sep = '', file = log, append = TRUE) time['user.self'], ',', time['sys.self'], ',', time['elapsed'],
## and the `OPG` method '\n', sep = '', file = log, append = TRUE)
time <- system.time(dr <- mave.compute(X, Y, max.dim = ncol(B), ## and the `OPG` method
method = "meanOPG", screen = ncol(X))) time <- system.time(dr <- mave.compute(X, Y, max.dim = ncol(B),
d.sub <- dist.subspace(B, coef(dr, ncol(B)), normalize = TRUE) method = "meanOPG", screen = ncol(X)))
d.gra <- dist.grassmann(B, coef(dr, ncol(B))) d.sub <- dist.subspace(B, coef(dr, ncol(B)), normalize = TRUE)
mse <- mean((predict(dr, ds.test$X, dim = ncol(B)) - ds.test$Y)^2) d.gra <- dist.grassmann(B, coef(dr, ncol(B)))
cat('"opg",', rep, ',', d.sub, ',', d.gra, ',', mse, ',', mse <- mean((predict(dr, ds.test$X, dim = ncol(B)) - ds.test$Y)^2)
time['user.self'], ',', time['sys.self'], ',', time['elapsed'], '\n', cat('"opg",', rep, ',', d.sub, ',', d.gra, ',', mse, ',',
sep = '', file = log, append = TRUE) time['user.self'], ',', time['sys.self'], ',', time['elapsed'],
'\n', sep = '', file = log, append = TRUE)
}
## Next the CVE method if (args$run_cve) {
time <- system.time(dr <- cve.call(X, Y, k = ncol(B))) ## Next the CVE method
d.sub <- dist.subspace(B, coef(dr, ncol(B)), normalize = TRUE) time <- system.time(dr <- cve.call(X, Y, k = ncol(B)))
d.gra <- dist.grassmann(B, coef(dr, ncol(B))) d.sub <- dist.subspace(B, coef(dr, ncol(B)), normalize = TRUE)
mse <- mean((predict(dr, ds.test$X, k = ncol(B)) - ds.test$Y)^2) d.gra <- dist.grassmann(B, coef(dr, ncol(B)))
cat('"cve",', rep, ',', d.sub, ',', d.gra, ',', mse, ',', mse <- mean((predict(dr, ds.test$X, k = ncol(B)) - ds.test$Y)^2)
time['user.self'], ',', time['sys.self'], ',', time['elapsed'], '\n', cat('"cve",', rep, ',', d.sub, ',', d.gra, ',', mse, ',',
sep = '', file = log, append = TRUE) time['user.self'], ',', time['sys.self'], ',', time['elapsed'],
'\n', sep = '', file = log, append = TRUE)
}
## Fit `DR` Neuronal Network model if (args$run_nn) {
time <- system.time(nn$fit(X, Y, epochs = args$epochs, ## Fit `DR` Neuronal Network model
batch_size = args$batch_size, initializer = args$initializer)) time <- system.time(nn$fit(X, Y, epochs = args$epochs,
# OPG estimate batch_size = args$batch_size, initializer = args$initializer))
d.sub <- dist.subspace(B, coef(nn, 'OPG'), normalize = TRUE) # OPG estimate
d.gra <- dist.grassmann(B, coef(nn, 'OPG')) d.sub <- dist.subspace(B, coef(nn, 'OPG'), normalize = TRUE)
cat('"nn.opg",', rep, ',', d.sub, ',', d.gra, ',NA,NA,NA,NA\n', d.gra <- dist.grassmann(B, coef(nn, 'OPG'))
sep = '', file = log, append = TRUE) cat('"nn.opg",', rep, ',', d.sub, ',', d.gra, ',NA,NA,NA,NA\n',
# Refinement estimate sep = '', file = log, append = TRUE)
d.sub <- dist.subspace(B, coef(nn), normalize = TRUE) # Refinement estimate
d.gra <- dist.grassmann(B, coef(nn)) d.sub <- dist.subspace(B, coef(nn), normalize = TRUE)
mse <- mean((nn$predict(ds.test$X) - ds.test$Y)^2) d.gra <- dist.grassmann(B, coef(nn))
cat('"nn.ref",', rep, ',', d.sub, ',', d.gra, ',', mse, ',', mse <- mean((nn$predict(ds.test$X) - ds.test$Y)^2)
time['user.self'], ',', time['sys.self'], ',', time['elapsed'], '\n', cat('"nn.ref",', rep, ',', d.sub, ',', d.gra, ',', mse, ',',
sep = '', file = log, append = TRUE) time['user.self'], ',', time['sys.self'], ',', time['elapsed'],
'\n', sep = '', file = log, append = TRUE)
}
}) })
## Invoke the garbage collector ## Invoke the garbage collector

49
simulations/simulations_bigdata.sh Normal file
View File

@ -0,0 +1,49 @@
#!/bin/bash
# Catch termination signal `SIGINT` and invoke `user_interupt`
trap user_interupt SIGINT
# Reports an user interupt and exits the simulation script (do not continue next
# statement, allows to exit bash loop with `^C`)
user_interupt() {
echo -e '\nUser Interrupt -> stopping simulation\n'
exit
}
# Simulation for big data with `p` proportional to `sqrt(n)`
for ds in 6 8; do
command="Rscript simulations_bigdata.R --reps=10 --run_mave=FALSE --run_cve=FALSE --dataset=$ds --n=1000 --p=32 --epochs=200,400"
echo -e "\n$command"
time eval "$command"
command="Rscript simulations_bigdata.R --reps=10 --run_mave=FALSE --run_cve=FALSE --dataset=$ds --n=4000 --p=63 --epochs=100,200"
echo -e "\n$command"
time eval "$command"
command="Rscript simulations_bigdata.R --reps=10 --run_mave=FALSE --run_cve=FALSE --dataset=$ds --n=16000 --p=126 --epochs=50,100"
echo -e "\n$command"
time eval "$command"
command="Rscript simulations_bigdata.R --reps=10 --run_mave=FALSE --run_cve=FALSE --dataset=$ds --n=64000 --p=253 --epochs=25,50"
echo -e "\n$command"
time eval "$command"
command="Rscript simulations_bigdata.R --reps=10 --run_mave=FALSE --run_cve=FALSE --dataset=$ds --n=256000 --p=506 --epochs=12,25"
echo -e "\n$command"
time eval "$command"
done
# Simulation for big data with `p` proportional to `n` (note: for the base case
# of `n = 1000`, `p = 32` see above)
# For i = 1, ..., 4 the sample size `n = 1000 * 4^i`, number of predictors
# `p = 32 * 4^i` and the training epochs `epochs ~ (200, 400) * 1.5^(-i)`
for ds in 6 8; do
command="Rscript simulations_bigdata.R --reps=10 --run_mave=TRUE --run_cve=TRUE --dataset=$ds --n=4000 --p=128 --epochs=133,266"
echo -e "\n$command"
time eval "$command"
command="Rscript simulations_bigdata.R --reps=10 --run_mave=FALSE --run_cve=FALSE --dataset=$ds --n=16000 --p=512 --epochs=88,176"
echo -e "\n$command"
time eval "$command"
command="Rscript simulations_bigdata.R --reps=10 --run_mave=FALSE --run_cve=FALSE --dataset=$ds --n=64000 --p=2048 --epochs=59,118"
echo -e "\n$command"
time eval "$command"
command="Rscript simulations_bigdata.R --reps=10 --run_mave=FALSE --run_cve=FALSE --dataset=$ds --n=256000 --p=8192 --epochs=39,78"
echo -e "\n$command"
time eval "$command"
done