search

apache / mxnet

Lightweight, Portable, Flexible Distributed/Mobile Deep Learning with Dynamic, Mutation-aware Dataflow Dep Scheduler; for Python, R, Julia, Scala, Go, Javascript and more
https://mxnet.apache.org
Apache License 2.0
20.77k stars 6.79k forks source link

Error in mx.io.internal.arrayiter()? #4386

Closed p2004r closed 7 years ago

p2004r commented 7 years ago

I rewrote an example available here (link to data in document) https://github.com/dmlc/mxnet/blob/master/example/recommenders/demo1-MF.ipynb

library(mxnet)

DF <- read.delim("./data/ml-100k/u.data", header=F)
names(DF) <- c("user", "item", "score", "time")

max_user<- max(DF$user)
max_item <- max(DF$item)

DF_mat <- data.matrix(DF[,1:3])

k <- 64

## model
user <- mx.symbol.Variable("user")
item <- mx.symbol.Variable("item")
score <- mx.symbol.Variable("score")
## user feature lookup
user1 <- mx.symbol.Embedding(data = user, input_dim = max_user, output_dim = k, name="user1") 
## item feature lookup
item1 <- mx.symbol.Embedding(data = item, input_dim = max_item, output_dim = k, name="item1")
## predict by the inner product, which is elementwise product and then sum
pred <- user1 * item1
pred1 <- mx.symbol.sum_axis(data = pred, axis = 1, name="pred1")
pred2 <- mx.symbol.Flatten(data = pred1, name="pred2")
## loss layer
pred3 <- mx.symbol.LinearRegressionOutput(data = pred2, label = score, name="pred3")

devices <- mx.cpu()
mx.set.seed(123)

mx.model.FeedForward.create(pred3, X = t(DF_mat[,1:2]), y = data.matrix(data.frame(score=DF$score)),
                            ctx=devices, num.round=10, array.batch.size=100,
                            verbose=verbose, array.layout="rowmajor",
                            initializer=mx.init.uniform(0.07),
                            eval.metric=mx.metric.rmse,
                            epoch.end.callback=mx.callback.log.train.metric(100))

But I get an error:

Error in mx.io.internal.arrayiter(as.array(data), as.array(label), unif.rnds,  : 
  basic_string::resize
> 
> traceback()
6: stop(list(message = "basic_string::resize", call = mx.io.internal.arrayiter(as.array(data), 
       as.array(label), unif.rnds, batch.size, shuffle), cppstack = NULL))
5: .External(list(name = "InternalFunction_invoke", address = <pointer: 0x1c7f300>, 
       dll = list(name = "Rcpp", path = "/usr/local/lib/R/site-library/Rcpp/libs/Rcpp.so", 
           dynamicLookup = TRUE, handle = <pointer: 0x2de8840>, 
           info = <pointer: 0x7f78e7f27d80>), numParameters = -1L), 
       <pointer: 0x4034e80>, ...)
4: mx.io.internal.arrayiter(as.array(data), as.array(label), unif.rnds, 
       batch.size, shuffle)
3: mx.io.arrayiter(X, y, batch.size = batch.size, shuffle = is.train)
2: mx.model.init.iter(X, y, batch.size = array.batch.size, is.train = TRUE)
1: mx.model.FeedForward.create(pred3, X = t(DF_mat[, 1:2]), y = data.matrix(data.frame(score = DF$score)), 
       ctx = devices, num.round = 10, array.batch.size = 100, verbose = verbose, 
       array.layout = "rowmajor", initializer = mx.init.uniform(0.07), 
       eval.metric = mx.metric.rmse, epoch.end.callback = mx.callback.log.train.metric(100))

Attempts to change the dimension of the data does not help.

> d.it <- mx.io.arrayiter(data = t(data.matrix(DF[, -which(colnames(DF) %in% c("score"))])), label = t(data.matrix(data.frame(score=DF$score))))
> d.it
C++ object <0x4babf40> of class 'MXArrayDataIter' <0x4bb7850>
> mx.io.arrayiter(data = data.matrix(DF[, -which(colnames(DF) %in% c("score"))]), label = data.matrix(data.frame(score=DF$score)))
Error in mx.io.internal.arrayiter(as.array(data), as.array(label), unif.rnds,  : 
  basic_string::resize
> mx.model.FeedForward.create(pred3, X = d.it,
+                             ctx=devices, num.round=10,
+                             verbose=verbose, #array.layout="rowmajor",
+                             initializer=mx.init.uniform(0.07),
+                             epoch.end.callback=mx.callback.log.train.metric(100))
[22:13:20] /root/mxnet/dmlc-core/include/dmlc/logging.h:235: [22:13:20] src/symbol/symbol.cc:155: Symbol.InferShapeKeyword argument name data not found.
Candidate arguments:
    [0]user
    [1]user1_weight
    [2]item
    [3]item1_weight
    [4]score

Error in symbol$infer.shape(list(...)) : basic_string::resize
> traceback()
6: stop(list(message = "basic_string::resize", call = symbol$infer.shape(list(...)), 
       cppstack = NULL))
5: .External(list(name = "CppMethod__invoke_notvoid", address = <pointer: 0x28022e0>, 
       dll = list(name = "Rcpp", path = "/usr/local/lib/R/site-library/Rcpp/libs/Rcpp.so", 
           dynamicLookup = TRUE, handle = <pointer: 0x396b840>, 
           info = <pointer: 0x7f4f81b48d80>), numParameters = -1L), 
       <pointer: 0x4b9f060>, <pointer: 0x4af8300>, .pointer, ...)
4: symbol$infer.shape(list(...))
3: mx.symbol.infer.shape(symbol, data = input.shape)
2: mx.model.init.params(symbol, input.shape, initializer, mx.cpu())
1: mx.model.FeedForward.create(pred3, X = d.it, ctx = devices, num.round = 10, 
       verbose = verbose, initializer = mx.init.uniform(0.07), epoch.end.callback = mx.callback.log.train.metric(100))

Environment info

Operating System: Ubuntu 14.04

Package used R MXNet commit hash (git rev-parse HEAD): ~/mxnet# git rev-parse HEAD 9d066eebfb4f42f75e18ea8925c28d5526153576

R sessionInfo():

R version 3.3.2 (2016-10-31) Platform: x86_64-pc-linux-gnu (64-bit) Running under: Ubuntu 14.04.5 LTS

locale: [1] C

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

p2004r commented 7 years ago

And mx.io.CSVIter() not working ...

> system("head data/ml-100k/data.y.csv")
"score"
3
3
1
2
1
4
2
5
3
> system("head data/ml-100k/data.x.csv")
"user","item"
196,242
186,302
22,377
244,51
166,346
298,474
115,265
253,465
305,451
> mx.io.CSVIter("./data/ml-100k/data.x.csv", c(1,2), "./data/ml-100k/data.y.csv", c(1,1))
Error in mx.varg.io.CSVIter(list(...)) : basic_string::resize
> traceback()
4: stop(list(message = "basic_string::resize", call = mx.varg.io.CSVIter(list(...)), 
       cppstack = NULL))
3: .External(list(name = "InternalFunction_invoke", address = <pointer: 0x1347f10>, 
       dll = list(name = "Rcpp", path = "/usr/local/lib/R/site-library/Rcpp/libs/Rcpp.so", 
           dynamicLookup = TRUE, handle = <pointer: 0x1581520>, 
           info = <pointer: 0x7f4fba6a0d80>), numParameters = -1L), 
       <pointer: 0x39c9fb0>, alist)
2: mx.varg.io.CSVIter(list(...))
1: mx.io.CSVIter("./data/ml-100k/data.x.csv", c(1, 2), "./data/ml-100k/data.y.csv", 
       c(1, 1))
>
p2004r commented 7 years ago

Manually I get data one labels

> X <- mx.io.arrayiter(data=DF_mat_x, label=t(data.matrix(data.frame(score=DF_y))))
> Z <- mxnet:::mx.model.init.iter(X, NULL, batch.size = 120, is.train = TRUE)
> Z$iter.next()
[1] TRUE
> Z$value()
$data
     [,1] [,2] [,3] [,4] [,5] [,6] [,7] [,8] [,9] [,10] [,11] [,12] [,13] [,14]
[1,]  196  186   22  244  166  298  115  253  305     6    62   286   200   210
[2,]  242  302  377   51  346  474  265  465  451    86   257  1014   222    40
     [,15] [,16] [,17] [,18] [,19] [,20] [,21] [,22] [,23] [,24] [,25] [,26]
[1,]   224   303   122   194   291   234   119   167   299   291   308    95
[2,]    29   785   387   274  1042  1184   392   486   144   118     1   546
     [,27] [,28] [,29] [,30] [,31] [,32] [,33] [,34] [,35] [,36] [,37] [,38]
[1,]    38   102    63   160    50   301   225   290    97   157   181   278
[2,]    95   768   277   234   246    98   193    88   194   274  1081   603
     [,39] [,40] [,41] [,42] [,43] [,44] [,45] [,46] [,47] [,48] [,49] [,50]
[1,]   276     7    10   284   201   276   287   246   242   249    99   178
[2,]   796    32    16   304   979   564   327   201  1137   241     4   332
     [,51] [,52] [,53] [,54] [,55] [,56] [,57] [,58] [,59] [,60] [,61] [,62]
[1,]   251    81   260    25    59    72    87   290    42   292   115    20
[2,]   100   432   322   181   196   679   384   143   423   515    20   288
     [,63] [,64] [,65] [,66] [,67] [,68] [,69] [,70] [,71] [,72] [,73] [,74]
[1,]   201    13   246   138   167    60    57   223   189   243    92   246
[2,]   219   526   919    26   232   427   304   274   512    15  1049   416
     [,75] [,76] [,77] [,78] [,79] [,80] [,81] [,82] [,83] [,84] [,85] [,86]
[1,]   194   241   178   254   293   127   225   299   225   276   291   222
[2,]   165   690   248  1444     5   229   237   229   480    54   144   366
     [,87] [,88] [,89] [,90] [,91] [,92] [,93] [,94] [,95] [,96] [,97] [,98]
[1,]   267    42    11    95     8   162    87   279   145   119    62    62
[2,]   518   403   111   625   338    25  1016   154   275  1153   498   382
     [,99] [,100] [,101] [,102] [,103] [,104] [,105] [,106] [,107] [,108]
[1,]    28    135     32     90    286    293    216    166    250    271
[2,]   209     23    294    382    208    685    144    328    496    132
     [,109] [,110] [,111] [,112] [,113] [,114] [,115] [,116] [,117] [,118]
[1,]    160    265    198     42    168    110     58     90    271     62
[2,]    174    118    498     96    151    307    144    648    346     21
     [,119] [,120] [,121] [,122] [,123] [,124] [,125] [,126] [,127] [,128]
[1,]    279    237     94    128    298     44    264    194     72    222
[2,]    832    514    789    485    317    195    200    385    195    750

$label
     [,1] [,2] [,3] [,4] [,5] [,6] [,7] [,8] [,9] [,10] [,11] [,12] [,13] [,14]
[1,]    3    3    1    2    1    4    2    5    3     3     2     5     5     3
     [,15] [,16] [,17] [,18] [,19] [,20] [,21] [,22] [,23] [,24] [,25] [,26]
[1,]     3     3     5     2     4     2     4     4     4     2     4     2
     [,27] [,28] [,29] [,30] [,31] [,32] [,33] [,34] [,35] [,36] [,37] [,38]
[1,]     5     2     4     5     3     4     4     4     3     4     1     5
     [,39] [,40] [,41] [,42] [,43] [,44] [,45] [,46] [,47] [,48] [,49] [,50]
[1,]     1     4     4     4     2     3     5     5     5     5     5     3
     [,51] [,52] [,53] [,54] [,55] [,56] [,57] [,58] [,59] [,60] [,61] [,62]
[1,]     4     2     4     5     5     2     4     5     5     4     3     1
     [,63] [,64] [,65] [,66] [,67] [,68] [,69] [,70] [,71] [,72] [,73] [,74]
[1,]     4     3     4     5     1     5     5     4     4     3     1     3
     [,75] [,76] [,77] [,78] [,79] [,80] [,81] [,82] [,83] [,84] [,85] [,86]
[1,]     4     2     4     3     3     5     5     3     5     3     5     4
     [,87] [,88] [,89] [,90] [,91] [,92] [,93] [,94] [,95] [,96] [,97] [,98]
[1,]     5     3     4     4     4     4     4     5     2     5     4     3
     [,99] [,100] [,101] [,102] [,103] [,104] [,105] [,106] [,107] [,108]
[1,]     4      4      3      5      4      3      4      5      4      5
     [,109] [,110] [,111] [,112] [,113] [,114] [,115] [,116] [,117] [,118]
[1,]      5      4      3      5      5      4      4      4      4      3
     [,119] [,120] [,121] [,122] [,123] [,124] [,125] [,126] [,127] [,128]
[1,]      3      4      4      3      4      5      5      2      5      5

And the model does not see the data


## model
user <- mx.symbol.Variable("user")
item <- mx.symbol.Variable("item")
score <- mx.symbol.Variable("label")
## user feature lookup
user1 <- mx.symbol.Embedding(user, input_dim = max_user, output_dim = k, name="user1") 
## item feature lookup
item1 <- mx.symbol.Embedding(item, input_dim = max_item, output_dim = k, name="item1")
## predict by the inner product, which is elementwise product and then sum
pred <- user1 * item1
pred1 <- mx.symbol.sum_axis(pred, axis = 1, name="pred1")
pred2 <- mx.symbol.Flatten(pred1, name="pred2")
## loss layer
pred3 <- mx.symbol.LinearRegressionOutput(data=pred2, label=score, name="pred3")

devices <- mx.cpu()
mx.set.seed(123)

mx.model.FeedForward.create(pred3, X =  mx.io.arrayiter(data=DF_mat_x, label=t(data.matrix(data.frame(score=DF_y)))),
                            ctx=devices, num.round=10, array.batch.size=10,
                            verbose=T, #array.layout="rowmajor",
                            initializer=mx.init.uniform(0.07), learning.rate=0.07,
                            eval.metric=mx.metric.rmse, momentum = 0.9,
                            epoch.end.callback=mx.callback.log.train.metric(1))

[18:47:02] /root/mxnet/dmlc-core/include/dmlc/logging.h:235: [18:47:02] src/symbol/symbol.cc:155: Symbol.InferShapeKeyword argument name data not found.
Candidate arguments:
    [0]user
    [1]user1_weight
    [2]item
    [3]item1_weight
    [4]label

Error in symbol$infer.shape(list(...)) : basic_string::resize
> traceback()
6: stop(list(message = "basic_string::resize", call = symbol$infer.shape(list(...)), 
       cppstack = NULL))
5: .External(list(name = "CppMethod__invoke_notvoid", address = <pointer: 0x11caef0>, 
       dll = list(name = "Rcpp", path = "/usr/local/lib/R/site-library/Rcpp/libs/Rcpp.so", 
           dynamicLookup = TRUE, handle = <pointer: 0x1404520>, 
           info = <pointer: 0x7f348a598d80>), numParameters = -1L), 
       <pointer: 0x3835820>, <pointer: 0x3836540>, .pointer, ...)
4: symbol$infer.shape(list(...))
3: mx.symbol.infer.shape(symbol, data = input.shape)
2: mx.model.init.params(symbol, input.shape, initializer, mx.cpu())
1: mx.model.FeedForward.create(pred3, X = mx.io.arrayiter(data = DF_mat_x, 
       label = t(data.matrix(data.frame(score = DF_y)))), ctx = devices, 
       num.round = 10, array.batch.size = 10, verbose = T, initializer = mx.init.uniform(0.07), 
       learning.rate = 0.07, eval.metric = mx.metric.rmse, momentum = 0.9, 
       epoch.end.callback = mx.callback.log.train.metric(1))
p2004r commented 7 years ago

Original example uses 

    user = mx.nd.array(user)
    item = mx.nd.array(item)
    score = mx.nd.array(score)
    return mx.io.NDArrayIter(data={'user':user,'item':item},label={'score':score}, 
                             batch_size=batch_size, shuffle=True)

How to do it properly in R?

p2004r commented 7 years ago

The workaround does not work. Can you fix my solution?

How to use the operator mx.symbol.slice_axis() correctly?


k <- 64

## model

data <- mx.symbol.Variable("data")
###user <- mx.symbol.Variable("user")
user <- mx.symbol.slice_axis(data=data, axis=0, begin=0, end=0, name="user")
###item <- mx.symbol.Variable("item")
item <- mx.symbol.slice_axis(data=data, axis=0, begin=1, end=1, name="item")
score <- mx.symbol.Variable("label")
## user feature lookup
user1 <- mx.symbol.Embedding(data=user, input_dim = max_user, output_dim = k, name="user1")
## item feature lookup
item1 <- mx.symbol.Embedding(data=item, input_dim = max_item, output_dim = k, name="item1")
## predict by the inner product, which is elementwise product and then sum
pred <- user1 * item1
pred1 <- mx.symbol.sum_axis(pred, axis = 1, name="pred1")
pred2 <- mx.symbol.Flatten(pred1, name="pred2")
## loss layer
pred3 <- mx.symbol.LinearRegressionOutput(data=pred2, label=score, name="pred3")

devices <- mx.cpu()
mx.set.seed(123)

mx.model.FeedForward.create(pred3, X =  mx.io.arrayiter(data=DF_mat_x, label=t(data.matrix(data.frame(score=DF_y)))),
                            ctx=devices, num.round=10, array.batch.size=10,
                            verbose=T, array.layout="rowmajor",
                            initializer=mx.init.uniform(0.07), learning.rate=0.07,
                            eval.metric=mx.metric.rmse, momentum = 0.9,
                            epoch.end.callback=mx.callback.log.train.metric(1))

Error

> mx.model.FeedForward.create(pred3, X =  mx.io.arrayiter(data=DF_mat_x, label=t(data.matrix(data.frame(score=DF_y)))),
+                             ctx=devices, num.round=10, array.batch.size=10,
+                             verbose=T, array.layout="rowmajor",
+                             initializer=mx.init.uniform(0.07), learning.rate=0.07,
+                             eval.metric=mx.metric.rmse, momentum = 0.9,
+                             epoch.end.callback=mx.callback.log.train.metric(1))
[11:52:15] /root/mxnet/dmlc-core/include/dmlc/logging.h:235: [11:52:15] src/operator/./matrix_op-inl.h:620: Check failed: (param.begin) < (param.end) 
Error in symbol$infer.shape(list(...)) : basic_string::resize
> traceback()
6: stop(list(message = "basic_string::resize", call = symbol$infer.shape(list(...)), 
       cppstack = NULL))
5: .External(list(name = "CppMethod__invoke_notvoid", address = <pointer: 0x182b480>, 
       dll = list(name = "Rcpp", path = "/usr/local/lib/R/site-library/Rcpp/libs/Rcpp.so", 
           dynamicLookup = TRUE, handle = <pointer: 0x2412840>, 
           info = <pointer: 0x7f177895cd80>), numParameters = -1L), 
       <pointer: 0x3646130>, <pointer: 0x359f480>, .pointer, ...)
4: symbol$infer.shape(list(...))
3: mx.symbol.infer.shape(symbol, data = input.shape)
2: mx.model.init.params(symbol, input.shape, initializer, mx.cpu())
1: mx.model.FeedForward.create(pred3, X = mx.io.arrayiter(data = DF_mat_x, 
       label = t(data.matrix(data.frame(score = DF_y)))), ctx = devices, 
       num.round = 10, array.batch.size = 10, verbose = T, array.layout = "rowmajor", 
       initializer = mx.init.uniform(0.07), learning.rate = 0.07, 
       eval.metric = mx.metric.rmse, momentum = 0.9, epoch.end.callback = mx.callback.log.train.metric(1))
thirdwing commented 7 years ago

@p2004r The proper way is to define your own iterator like below. You can find a small demo in #6673

CustomIter <- setRefClass("CustomIter", fields = c("iter1", "iter2"),
  contains = "Rcpp_MXArrayDataIter",
  methods = list(
    initialize = function(iter1, iter2) {
      .self$iter1 <- iter1
      .self$iter2 <- iter2
      .self
    },
    value = function() {
      user <- .self$iter1$value()$data
      item <- .self$iter2$value()$data
      label <- .self$iter1$value()$label
      list(user = user,
           item = item,
           label = label)
    },
    iter.next = function() {
      .self$iter1$iter.next()
      .self$iter2$iter.next()
    },
    reset = function() {
      .self$iter1$reset()
      .self$iter2$reset()
    },
    num.pad = function() {
      .self$iter1$num.pad()
    },
    finalize = function() {
      .self$iter1$finalize()
      .self$iter2$finalize()
    }
  )
)

user_iter = mx.io.arrayiter(data = DF[, 1], label = DF[, 3], batch.size = k)

item_iter = mx.io.arrayiter(data = DF[, 2], label = DF[, 3], batch.size = k)

train_iter <- CustomIter$new(user_iter, item_iter)
p2004r commented 7 years ago

Thank you.

thirdwing commented 7 years ago

Fixed in #6673