I'm trying to train a list of text datasets at the character level (for example, a cat => "a", " ", "c", "a", "t") so that I can classify them with great accuracy. I'm using mxnet package (CNN Network) in R and using crepe model. So to prepare for training, I need to do iterations for both training and test datasets. So the code is as follow:
train.iter <- CustomCSVIter$new(iter=NULL, data.csv=train.file.output,
batch.size=args$batch_size, alphabet=alphabet,
feature.len=feature.len)
test.iter <- CustomCSVIter$new(iter=NULL, data.csv=test.file.output,
batch.size=args$batch_size, alphabet=alphabet,
feature.len=feature.len)
data.csv where I have these datasets, batch.size is just an integer, feature.len is also just an integer, and alphabet is a vector of alphanumeric quotations (abcd...?!""). When I run the above code, I get a message saying I have a fatal error and Rstudio crashes and reloads. I don't know what I'm doing wrong. To run the above code, you need the following function:
CustomCSVIter <- setRefClass("CustomCSVIter",
fields=c("iter", "data.csv", "batch.size",
"alphabet","feature.len"),
contains = "Rcpp_MXArrayDataIter",
methods=list(
initialize=function(iter, data.csv, batch.size,
alphabet, feature.len){
csv_iter <- mx.io.CSVIter(data.csv=data.csv,
data.shape=feature.len+1, #=features + label
batch.size=batch.size)
.self$iter <- csv_iter
.self$data.csv <- data.csv
.self$batch.size <- batch.size
.self$alphabet <- alphabet
.self$feature.len <- feature.len
.self
},
value=function(){
val <- as.array(.self$iter$value()$data)
val.y <- val[1,]
val.x <- val[-1,]
val.x <- dict.decoder(data=val.x,
alphabet=.self$alphabet,
feature.len=.self$feature.len,
batch.size=.self$batch.size)
val.x <- mx.nd.array(val.x)
val.y <- mx.nd.array(val.y)
list(data=val.x, label=val.y)
},
iter.next=function(){
.self$iter$iter.next()
},
reset=function(){
.self$iter$reset()
},
num.pad=function(){
.self$iter$num.pad()
},
finalize=function(){
.self$iter$finalize()
}
)
)
