Help with Neural predict -> "subscript out of bounds" error

Hi,
Does anyone has any idea why this is happening during the NEURAL testing after training ?
Getting "subscript out of bounds" from the R script when there are multiple assets or NumWFOCycles >60.
If I remove 1 of the 2 asset, or set NumWFOCycles=50 , this issue is not happening.


/////////////////////////// Zorro script
#include <r.h>

var change(int n)
{
return scale((priceClose(0) - priceClose(n))/priceClose(0),100)/100;
}

function run()
{
NumCores = -1;
Script = "DeepLearnMXDEBUG";
Verbose = 31|DIAG;
StartDate = 2016;//2016;
BarPeriod = 1440; // 1 hour
LookBack = 100;
Weekend=0;

NumWFOCycles = 60;

assetList("History\AssetsBitmex.csv");
while(loop(Assets))
{
asset(Loop1);
set(RULES);

LifeTime = 3;
if(Train) Hedge = 2;
var Threshold = 0.5;

set(LOGFILE|PLOTNOW);
if(adviseLong(NEURAL+BALANCED,0,
change(1),change(2),change(3),change(4)) > 0.5)
enterLong();
if(adviseShort() > 0.5)
enterShort();

PlotWidth = 800;
PlotHeight1 = 340;

}
}

////////////////////////// Zorro assets

Name,Price,Spread,RollLong,RollShort,PIP,PIPCost,MarginCost,Leverage,LotAmount,Commission,Symbol,Type
ETH/BTC,1,0,0,0,0.00000001,0.00000001,0,1,1,-0.25,ETH/BTC,
BCH/BTC,1,0,0,0,0.00000001,0.00000001,0,1,1,-0.25,BCH/BTC,

/////////////////////////// R File

# how to install
#cran <- getOption("repos")
#cran["dmlc"] <- "https://s3-us-west-2.amazonaws.com/apache-mxnet/R/CRAN/"
#options(repos = cran)
#install.packages('mxnet')
# - or (if not yet available for current version) -
#install.packages("https://s3.ca-central-1.amazonaws.com/jeremiedb/share/mxnet/CPU/mxnet.zip", repos = NULL)
library('mxnet', quietly = T)
library('caret', quietly = T)

neural.train = function(model,XY)
{
X <- data.matrix(XY[,-ncol(XY)])
Y <- XY[,ncol(XY)]
Y <- ifelse(Y > 0,1,0)
Models[[model]] <<- mx.mlp(X,Y,
hidden_node = c(30),
out_node = 2,
activation = "sigmoid",
out_activation = "softmax",
num.round = 20,
array.batch.size = 20,
learning.rate = 0.05,
momentum = 0.9,
eval.metric = mx.metric.accuracy)
}

neural.predict = function(model,X)
{
if(is.vector(X)) X <- t(X)
X <- data.matrix(X)
Y <- predict(Models[[model]],X)
return(ifelse(Y[1,] > Y[2,],0,1))
}

neural.save = function(name)
{
for(i in c(1:length(Models)))
Models[[i]] <<- mx.serialize(Models[[i]])
save(Models,file=name)
}

neural.load <- function(name)
{
load(name,.GlobalEnv)
for(i in c(1:length(Models)))
Models[[i]] <<- mx.unserialize(Models[[i]])
}

neural.init = function()
{
mx.set.seed(365)
Models <<- vector("list")
}

neural.test = function()
{
neural.init()
XY <<- read.csv('MyZorroPath/Data/DeepLearnMX.csv',header = F)
splits <- nrow(XY)*0.8
XY.tr <<- head(XY,splits)
XY.ts <<- tail(XY,-splits)
neural.train(1,XY.tr)

X <<- XY.ts[,-ncol(XY.ts)]
Y <<- XY.ts[,ncol(XY.ts)]
Y.ob <<- ifelse(Y > 0,1,0)
Y.pr <<- neural.predict(1,X)
confusionMatrix(as.factor(Y.pr),as.factor(Y.ob))
}

///////////////////////////// Zorro output

Error in Models[[model]] : subscript out of bounds
Calls: neural.predict -> predict
Execution halted

I would interpret this error as if not all models are trained, so the Models vector is too small.

Look into the training log: has it error messages?

Hi jcl,
No errors in the training log.

I run the Debug View, and there I can see all input/output to R.
All logs attached (training/testing/diag/debugview), if can help.

Thanks

Just want to add that the error is in the debug-view log when testing (attached previour post).
Input/Output to R looks normal in this file.

The traininig looks like might be incomplete from the screenshot attached, as the purple bar didn't go to the end,and 57 runs instead of 60, something might be happening during training, that actually causes the error when testing.

But there are no errors during training at all...

Attaching a screenshot, the debug-view training log (the previous one was for testing) and the train log again if can help

Hi, you have day bars and only 3 years, so thats far too few bars for 60 wfo cycles. I think you cannot train a cycle that has only a few bars, maybe you get no result for some cycles and thats the reason of the error. Try 20 years instead of 3, or 10 wfo cycles instead of 60.

Yes I thought so, but then if remove either one of the 2 asset, then there are no errors and training complete successfully and testing afterwards has no issue

Thats exactly the problem, the first asset determines the bars, so you probably have some cycles with no candles of the second asset where they dont overlap. Especially with cryptos that have lots of gaps anyway.

I think the issue in training has been fixed with adding array.layout = "rowmajor" in the training function of the R script.
There is still a warning in the predict function as from the log below, but not sure how to add the same code to the predict function.
Anyone has any idea ?

Thanks

log:
[27312] <2> ExecuteCode: in >>> Y <- neural.predict(2,X)
[27312] <2> ExecuteCode: out <<< arning message:
[27312] In mx.model.select.layout.predict(X, model) :
[27312] Auto detect layout of input matrix, use rowmajor..


neural.train = function(model,XY)
{
X <- data.matrix(XY[,-ncol(XY)])
Y <- XY[,ncol(XY)]
Y <- ifelse(Y > 0,1,0)
Models[[model]] <<- mx.mlp(X,Y,
hidden_node = c(60),
out_node = 2,
activation = "sigmoid",
out_activation = "softmax",
num.round = 10,
array.batch.size = 40,
learning.rate = 0.02,
momentum = 1.5,
eval.metric = mx.metric.accuracy,
array.layout = "rowmajor")
}