Incubator-mxnet: [R] mx.io.CSVIter not loading labels in version 1.3.0

[some-guy1]

The code given below used to work fine in older versions of mxnet (older than 1.2.0). In version 1.3.0, it seems data.csv is loaded, but label.csv is skipped.

### sessionInfo()
R version 3.5.0 (2018-04-23)
Platform: x86_64-w64-mingw32/x64 (64-bit)
Running under: Windows >= 8 x64 (build 9200)

Matrix products: default

locale:
[1] LC_COLLATE=English_United States.1252 LC_CTYPE=English_United States.1252
[3] LC_MONETARY=English_United States.1252 LC_NUMERIC=C
[5] LC_TIME=English_United States.1252

attached base packages:
[1] stats graphics grDevices utils datasets methods base

other attached packages:
[1] mxnet_1.3.0

loaded via a namespace (and not attached):
[1] Rcpp_0.12.16 pillar_1.2.2 compiler_3.5.0 RColorBrewer_1.1-2 influenceR_0.1.0 plyr_1.8.4
[7] bindr_0.1.1 viridis_0.5.1 tools_3.5.0 digest_0.6.15 jsonlite_1.5 viridisLite_0.3.0
[13] tibble_1.4.2 gtable_0.2.0 rgexf_0.15.3 pkgconfig_2.0.1 rlang_0.2.0 igraph_1.2.1
[19] rstudioapi_0.7 yaml_2.1.19 bindrcpp_0.2.2 gridExtra_2.3 downloader_0.4 DiagrammeR_1.0.0
[25] dplyr_0.7.4 stringr_1.3.1 htmlwidgets_1.2 hms_0.4.2 grid_3.5.0 glue_1.2.0
[31] R6_2.2.2 Rook_1.1-1 XML_3.98-1.11 readr_1.1.1 purrr_0.2.4 tidyr_0.8.0
[37] ggplot2_2.2.1 magrittr_1.5 codetools_0.2-15 scales_0.5.0 htmltools_0.3.6 assertthat_0.2.0
[43] colorspace_1.3-2 brew_1.0-6 stringi_1.2.2 visNetwork_2.0.3 lazyeval_0.2.1 munsell_0.4.3

Build info

https://github.com/apache/incubator-mxnet/tree/master/R-package

cran <- getOption("repos")
cran["dmlc"] <- "https://apache-mxnet.s3-accelerate.dualstack.amazonaws.com/R/CRAN/"
options(repos = cran)
install.packages("mxnet")

Error Message:

Error in symbol$infer.shape(list(...)) :
Error in operator lro: Shape inconsistent, Provided=[1], inferred shape=[1,1,3,3]

Minimum reproducible example

batch_size = 1

train_iter <- mx.io.CSVIter(
  data.csv = "./matty_inv/A.csv", data.shape = c(3,3,1),
  label.csv = "./matty_inv/A.csv", label.shape =  c(3, 3, 1), 
  batch.size = batch_size
)

data <- mx.symbol.Variable('data')
label <- mx.symbol.Variable('label')

conv_1 <- mx.symbol.Convolution(data= data, kernel = c(1,1), num_filter = 4, name="conv_1")
conv_act_1 <- mx.symbol.Activation(data= conv_1, act_type = "relu", name="conv_act_1")
flat <- mx.symbol.flatten(data = conv_act_1,  name="flatten")
fcl_1 <- mx.symbol.FullyConnected(data = flat, num_hidden = 9, name="fc_1")
fcl_2 <- mx.symbol.reshape(fcl_1, shape=c(3,3, 1, batch_size))
NN_Model <- mx.symbol.LinearRegressionOutput(data=fcl_2 , label=label, name="lro")

mx.set.seed(99)
  autoencoder <- mx.model.FeedForward.create(
    NN_Model, X=train_iter, initializer = mx.init.uniform(0.01),
    ctx=mx.cpu(), num.round=n.rounds, array.batch.size=batch_size,
    learning.rate=8e-3, array.layout = "rowmajor",
    eval.metric = mx.metric.rmse, optimizer = "adam",
    verbose = TRUE)

[anirudhacharya]

@mxnet-label-bot [R, Bug, Data-loading]

[ankkhedia]

Hi @some-guy1
Could you please follow CSVIterator example mentioned over here https://mxnet.incubator.apache.org/tutorials/r/CustomIterator.html. Over here , it assumes label to be a part of the input csv file. This is the specification for csviter API
https://mxnet.apache.org/api/python/io/io.html#mxnet.io.CSVIter

[jeremiedb]

@some-guy1 How did you install the package? I suspect it's a corrupt build you're have. I'm running on 1.3.0 (GPU build on Windows) and the following did work fine:

library(mxnet)

train_iter <- mx.io.CSVIter(
  data_csv = "train-data.csv", data.shape = c(2),
  label_csv = "train-label.csv", label.shape =  c(1),
  batch_size = 2
)

train_iter$reset()
train_iter$iter.next()
train_iter$value()

train-data.csv contains two columns while train-label.csv contains a single column.

[some-guy1]

@ankkhedia the csviter API describes using two different files for data csv and label csv.

@jeremiedb I just copied and pasted the following into RStudio:

cran <- getOption("repos")
cran["dmlc"] <- "https://apache-mxnet.s3-accelerate.dualstack.amazonaws.com/R/CRAN/"
options(repos = cran)
install.packages("mxnet")

Have you tried a label csv with more than a single column? It appears that csv iter is ONLY reading the first column of the csv file even though I said the shape was c(3,3,1)

[jeremiedb]

I tried the following with a label.csv having 6 columns and it returns proper array of shape (3,2,1). Note that batch size is ignored in defining the shape of data and label.

train_iter <- mx.io.CSVIter(
  data_csv = "train-data.csv", data.shape = c(2),
  label_csv = "train-label.csv", label.shape =  c(3, 2),
  batch_size = 1
)

[anirudhacharya]

@some-guy1 can you please share a small snapshot of the data file you are trying to read - ./matty_inv/A.csv

[some-guy1]

[jeremiedb]

Can you confirm whether the following is working:

cran <- getOption("repos")
cran["dmlc"] <- "https://apache-mxnet.s3-accelerate.dualstack.amazonaws.com/R/CRAN/"
options(repos = cran)
install.packages("mxnet")


library(data.table)
library(mxnet)

data <- matrix(rnorm(9*5), ncol = 9)
fwrite(x = as.data.table(data), file = "data.csv", col.names = F)

labels <- matrix(rnorm(9*5), ncol = 9)
fwrite(x = as.data.table(data), file = "label.csv", col.names = F)

iter <- mx.io.CSVIter(data_csv = "data.csv",
                      data_shape = c(9),
                      label_csv = "label.csv",
                      label_shape = c(3,1,3),
                      batch.size = 2)

iter$reset()
iter$iter.next()
iter$value()

This has been run on Windows 10 and R 3.5.1 and worked fine.
From the error message provided symbol$infer.shape(list(...)), issue seems more related to the network definition.

[some-guy1]

Ok this is confusing living heck out of me. I ran your code above and iter$value() results in the dimensions I would expect. Hence, this appears correct.

BUT, I tried again to simply run only my code below

train_iter <- mx.io.CSVIter(
  data_csv = "A.csv", data_shape = c(3, 3, 1),
  label_csv = "A.csv", label_shape =  c(3, 3, 1),
  batch_size = 1
)

train_iter$reset()
train_iter$iter.next()
train_iter$value()

and I still only get one value in the label. I am expecting $label to be identical to $data

$label
[1] 1.373546

Now instead if I run:

train_iter <- mx.io.CSVIter(
  data_csv = "A.csv", data_shape = c(3, 3, 1),
  label_csv = "A.csv", label_shape =  c(3, 1, 3),
  batch_size = 1
)

train_iter$reset()
train_iter$iter.next()
train_iter$value()

I get

$`data`
, , 1, 1

          [,1]     [,2]     [,3]
[1,] 1.1643714 2.689276 2.152658
[2,] 0.8696243 1.527834 1.233204
[3,] 2.1957829 2.738622 1.815747


$label
, , 1, 1

          [,1]
[1,] 1.1643714
[2,] 0.8696243
[3,] 2.1957829

, , 2, 1

         [,1]
[1,] 2.689276
[2,] 1.527834
[3,] 2.738622

, , 3, 1

         [,1]
[1,] 2.152658
[2,] 1.233204
[3,] 1.815747

Is it possible that the label axis is rotated relative to the data axis? Are we supposed to transpose the label axis until is matches the data axis?

[jeremiedb]

Interesting! Really looks like a bug when the last label axis is set to 1.
As a quick turnaround, what about dropping the 1 axis for all data and label in the csv.iter and perform a reshape within the model architecture?

Something like:

iter <- mx.io.CSVIter(data_csv = "data.csv",
                      data_shape = c(3,3),
                      label_csv = "label.csv",
                      label_shape = c(3,3),
                      batch.size = 2)

data <- mx.symbol.Variable("data")
data <- mx.symbol.reshape(data, shape = c(3,1,3,0))

label <- mx.symbol.Variable("label")
label <- mx.symbol.reshape(label, shape = c(3,3,1,0))
...
loss <- mx.symbol.LinearRegressionOutput(data = model, label = label)

[some-guy1]

Ok so I think we are now moving on to a new problem. I would think the workaround you have posted above should work just fine....but it doesnt. It causes the entire R environment to crash.

I am getting the exact same symptoms when I setup an iterator based on arrays in memory. See the following separate issue I submitted:

https://github.com/apache/incubator-mxnet/issues/12431

[piyushghai]

@some-guy1 The other issue that you opened was fixed. (#12431) .
Do you think the fix for that fixed this as well ?

[karan6181]

@some-guy1 Could you please re-run your script again with the latest MXNet master, the R environment crash issue has been fixed. If you are able to run your script without error then we can close this issue. Thanks!

[piyushghai]

@some-guy1 I suggest this issue be closed as the snippet used by @jeremiedb loads the labels.
Can you upgrade MXNet R package to V1.4.0 and try it out again and let us know ?
Please feel free to re-open if closed in error.

@Roshrini Can you close this issue ?

[Roshrini]

As the issue is fixed in v1.4.0, I am closing this issue. Please reopen if you still encounter it or if closed in error. Thanks!