Update code

[henkb1987]

Hi there,

Thanks for this great package. Would you happen to know if I can fit a model on a subset of data and then update that model with further data at a later time?

Thanks again.

[DimitriF]

late answer,
there is a model argument in trainr that does exactly that, just supply the previously trained model and it will keep training

[gyubaekkim]

Hi there,

I want ask more about this issue.
I implemented update(re-train) logic but it may not work properly.
Please review the following code that excludes unnecessary part and tell me what's wrong.

require(rnn)
require(caret)

if (file.exists(filename)) {
:
previous_model <- readRDS(filename)
:
model <- trainr(Y=T_Y, X=T_X, model=previous_model, learningrate = 0.2, hidden_dim=5, numepochs=1)
:

} else {
:
model <- trainr(Y=T_Y, X=T_X, learningrate = 0.2, hidden_dim=5, numepochs=1)
:
}
:
P <- predictr(model, P_X)
:
saveRDS(model, filename)
:

The above code is run daily so re-training is done with small data.
I checked the size of saved model file and found the proportional increase.
The issue is that accuracy of prediction using re-trained model was bad.

Thanks in advance.

[DimitriF]

Hi,
Normally that would be the way to do it (without inputing again options which are ignored).
Could you make str(model) for 2 consecutive days?
Dimitri

[gyubaekkim]

Here are the description of 2 consecutive model (initial & re-trained)

• Inital
  List of 29
  $ input_dim : int 4
  $ hidden_dim : num 5
  $ output_dim : int 1
  $ synapse_dim : num [1:3] 4 5 1
  $ time_dim : int 32
  $ sigmoid : chr "logistic"
  $ network_type : chr "rnn"
  $ numepochs : num 1
  $ batch_size : num 1
  $ learningrate : num 0.2
  $ learningrate_decay : num 1
  $ momentum : num 0
  $ update_rule : chr "sgd"
  $ use_bias : logi FALSE
  $ seq_to_seq_unsync : logi FALSE
  $ epoch_function :List of 2
  ..$ :function (model)
  ..$ :function (model)
  $ loss_function :function (model)
  $ last_layer_error : num [1:8637, 1:32, 1] -0.501 0.568 -0.573 -0.446 0.437 ...
  $ last_layer_delta : num [1:8637, 1:32, 1] -0.125 0.139 -0.14 -0.11 0.107 ...
  $ store :List of 2
  ..$ : num [1:8637, 1:32, 1:5] 0.338 0.261 0.211 0.26 0.34 ...
  ..$ : num [1:8637, 1:32, 1] 0.501 0.432 0.573 0.446 0.563 ...
  $ time_synapse :List of 2
  ..$ : num [1:4, 1:5] -0.637 -0.1629 -0.0229 -0.225 -0.3313 ...
  ..$ : num [1:5, 1] -0.85 4.794 -0.744 -0.78 -2.605
  $ recurrent_synapse :List of 1
  ..$ : num [1:5, 1:5] 1.592 0.665 1.752 1.533 -0.28 ...
  $ bias_synapse :List of 2
  ..$ : num [1:5] -0.085244 -0.038063 0.043454 0.000909 -0.0694
  ..$ : num 5.89
  $ time_synapse_update :List of 2
  ..$ : num [1:4, 1:5] 0 0 0 0 0 0 0 0 0 0 ...
  ..$ : num [1:5, 1] 0 0 0 0 0
  $ bias_synapse_update :List of 2
  ..$ : num [1:5] 0 0 0 0 0
  ..$ : num 0
  $ recurrent_synapse_update:List of 1
  ..$ : num [1:5, 1:5] 0 0 0 0 0 0 0 0 0 0 ...
  $ error : num [1:8637, 1] 0.941 1.233 1.148 0.908 0.875 ...
  $ current_epoch : int 1
  $ store_best :List of 2
  ..$ : num [1:8637, 1:32, 1:5] 0.338 0.261 0.211 0.26 0.34 ...
  ..$ : num [1:8637, 1:32, 1] 0.501 0.432 0.573 0.446 0.563 ...

• attr(*, "error")= num 1.24
  NULL

• Re-trained
  List of 30
  $ model :List of 29
  ..$ input_dim : int 4
  ..$ hidden_dim : num 5
  ..$ output_dim : int 1
  ..$ synapse_dim : num [1:3] 4 5 1
  ..$ time_dim : int 32
  ..$ sigmoid : chr "logistic"
  ..$ network_type : chr "rnn"
  ..$ numepochs : num 1
  ..$ batch_size : num 1
  ..$ learningrate : num 0.2
  ..$ learningrate_decay : num 1
  ..$ momentum : num 0
  ..$ update_rule : chr "sgd"
  ..$ use_bias : logi FALSE
  ..$ seq_to_seq_unsync : logi FALSE
  ..$ epoch_function :List of 2
  .. ..$ :function (model)
  .. ..$ :function (model)
  ..$ loss_function :function (model)
  ..$ last_layer_error : num [1:8637, 1:32, 1] -0.501 0.568 -0.573 -0.446 0.437 ...
  ..$ last_layer_delta : num [1:8637, 1:32, 1] -0.125 0.139 -0.14 -0.11 0.107 ...
  ..$ store :List of 2
  .. ..$ : num [1:8637, 1:32, 1:5] 0.338 0.261 0.211 0.26 0.34 ...
  .. ..$ : num [1:8637, 1:32, 1] 0.501 0.432 0.573 0.446 0.563 ...
  ..$ time_synapse :List of 2
  .. ..$ : num [1:4, 1:5] -0.637 -0.1629 -0.0229 -0.225 -0.3313 ...
  .. ..$ : num [1:5, 1] -0.85 4.794 -0.744 -0.78 -2.605
  ..$ recurrent_synapse :List of 1
  .. ..$ : num [1:5, 1:5] 1.592 0.665 1.752 1.533 -0.28 ...
  ..$ bias_synapse :List of 2
  .. ..$ : num [1:5] -0.085244 -0.038063 0.043454 0.000909 -0.0694
  .. ..$ : num 5.89
  ..$ time_synapse_update :List of 2
  .. ..$ : num [1:4, 1:5] 0 0 0 0 0 0 0 0 0 0 ...
  .. ..$ : num [1:5, 1] 0 0 0 0 0
  ..$ bias_synapse_update :List of 2
  .. ..$ : num [1:5] 0 0 0 0 0
  .. ..$ : num 0
  ..$ recurrent_synapse_update:List of 1
  .. ..$ : num [1:5, 1:5] 0 0 0 0 0 0 0 0 0 0 ...
  ..$ error : num [1:8637, 1] 0.941 1.233 1.148 0.908 0.875 ...
  ..$ current_epoch : int 1
  ..$ store_best :List of 2
  .. ..$ : num [1:8637, 1:32, 1:5] 0.338 0.261 0.211 0.26 0.34 ...
  .. ..$ : num [1:8637, 1:32, 1] 0.501 0.432 0.573 0.446 0.563 ...
  ..- attr(*, "error")= num 1.24
  $ input_dim : int 4
  $ hidden_dim : num 5
  $ output_dim : int 1
  $ synapse_dim : num [1:3] 4 5 1
  $ time_dim : int 32
  $ sigmoid : chr "logistic"
  $ network_type : chr "rnn"
  $ numepochs : num 1
  $ batch_size : num 1
  $ learningrate : num 0.2
  $ learningrate_decay : num 1
  $ momentum : num 0
  $ update_rule : chr "sgd"
  $ use_bias : logi FALSE
  $ seq_to_seq_unsync : logi FALSE
  $ epoch_function :List of 2
  ..$ :function (model)
  ..$ :function (model)
  $ loss_function :function (model)
  $ last_layer_error : num [1:143, 1:32, 1] 0.9279 0.8914 0.9227 0.8843 -0.0786 ...
  $ last_layer_delta : num [1:143, 1:32, 1] 0.06211 0.08631 0.06582 0.09048 -0.00569 ...
  $ store :List of 2
  ..$ : num [1:143, 1:32, 1:5] 0.5 0.529 0.5 0.525 0.5 ...
  ..$ : num [1:143, 1:32, 1] 0.0721 0.1086 0.0773 0.1157 0.0786 ...
  $ time_synapse :List of 2
  ..$ : num [1:4, 1:5] -0.5844 0.1364 0.0783 0.2251 0.8466 ...
  ..$ : num [1:5, 1] -0.854 -0.669 -1.231 -1.201 -1.055
  $ recurrent_synapse :List of 1
  ..$ : num [1:5, 1:5] 0.853 0.194 1.079 0.848 -0.734 ...
  $ bias_synapse :List of 2
  ..$ : num [1:5] -0.085244 -0.038063 0.043454 0.000909 -0.0694
  ..$ : num -0.352
  $ time_synapse_update :List of 2
  ..$ : num [1:4, 1:5] 0 0 0 0 0 0 0 0 0 0 ...
  ..$ : num [1:5, 1] 0 0 0 0 0
  $ bias_synapse_update :List of 2
  ..$ : num [1:5] 0 0 0 0 0
  ..$ : num 0
  $ recurrent_synapse_update:List of 1
  ..$ : num [1:5, 1:5] 0 0 0 0 0 0 0 0 0 0 ...
  $ error : num [1:143, 1] 1.88 1.78 2.23 1.84 1.02 ...
  $ current_epoch : int 1
  $ store_best :List of 2
  ..$ : num [1:143, 1:32, 1:5] 0.5 0.529 0.5 0.525 0.5 ...
  ..$ : num [1:143, 1:32, 1] 0.0721 0.1086 0.0773 0.1157 0.0786 ...

• attr(*, "error")= num 1.92
  NULL

[DimitriF]

The second model is a list of length 30, something is wrong.

If I do:

X1 = sample(0:127, 10000, replace=TRUE)
X2 = sample(0:127, 10000, replace=TRUE)
Y <- X1 + X2
X1 <- int2bin(X1, length=8)
X2 <- int2bin(X2, length=8)
Y  <- int2bin(Y,  length=8)
X <- array( c(X1,X2), dim=c(dim(X1),2) )
model <- trainr(Y=Y, X=X,learningrate   =  0.1,momentum = 0.5,
hidden_dim     = 16  ,numepochs = 1,batch_size = 20)
length(model)
model <- trainr(Y=Y,X=X,model=model)
length(model)

Both length are 29.

[gyubaekkim]

learningrate parameter for re-training is missing in your sample.
But, in my case, it's not permitted so I found that library version was different.
I re-installed the latest version from github. After that, the issue that I reported does not happen again.
I don't know why. Anyway DimitriF 's comment was helpful. I appreciate.