Merge pull request #215 from jpuigcerver/PR

Dropout argument for recurrent modules & BGRU

8 files changed, 25 insertions, 16 deletions

diff --git a/BGRU.lua b/BGRU.lua
new file mode 100644
index 0000000..31d5d92
--- /dev/null
+++ b/BGRU.lua
@@ -0,0 +1,9 @@
+local BGRU, parent = torch.class('cudnn.BGRU', 'cudnn.RNN')
+
+function BGRU:__init(inputSize, hiddenSize, numLayers, batchFirst, dropout)
+    parent.__init(self, inputSize, hiddenSize, numLayers, batchFirst, dropout)
+    self.bidirectional = 'CUDNN_BIDIRECTIONAL'
+    self.mode = 'CUDNN_GRU'
+    self.numDirections = 2
+    self:reset()
+end
diff --git a/BLSTM.lua b/BLSTM.lua
index 8feebf1..d17dbca 100644
--- a/BLSTM.lua
+++ b/BLSTM.lua
@@ -1,7 +1,7 @@
 local BLSTM, parent = torch.class('cudnn.BLSTM', 'cudnn.RNN')
 
-function BLSTM:__init(inputSize, hiddenSize, numLayers, batchFirst)
-    parent.__init(self,inputSize, hiddenSize, numLayers, batchFirst)
+function BLSTM:__init(inputSize, hiddenSize, numLayers, batchFirst, dropout)
+    parent.__init(self, inputSize, hiddenSize, numLayers, batchFirst, dropout)
     self.bidirectional = 'CUDNN_BIDIRECTIONAL'
     self.mode = 'CUDNN_LSTM'
     self.numDirections = 2
diff --git a/GRU.lua b/GRU.lua
index 615bdf7..c41fc5d 100644
--- a/GRU.lua
+++ b/GRU.lua
@@ -1,7 +1,7 @@
 local GRU, parent = torch.class('cudnn.GRU', 'cudnn.RNN')
 
-function GRU:__init(inputSize, hiddenSize, numLayers, batchFirst)
-    parent.__init(self,inputSize, hiddenSize, numLayers, batchFirst)
+function GRU:__init(inputSize, hiddenSize, numLayers, batchFirst, dropout)
+    parent.__init(self,inputSize, hiddenSize, numLayers, batchFirst, dropout)
     self.mode = 'CUDNN_GRU'
     self:reset()
 end
diff --git a/LSTM.lua b/LSTM.lua
index 29c199c..d9f10a2 100644
--- a/LSTM.lua
+++ b/LSTM.lua
@@ -1,7 +1,7 @@
 local LSTM, parent = torch.class('cudnn.LSTM', 'cudnn.RNN')
 
-function LSTM:__init(inputSize, hiddenSize, numLayers, batchFirst)
-    parent.__init(self,inputSize, hiddenSize, numLayers, batchFirst)
+function LSTM:__init(inputSize, hiddenSize, numLayers, batchFirst, dropout)
+    parent.__init(self,inputSize, hiddenSize, numLayers, batchFirst, dropout)
     self.mode = 'CUDNN_LSTM'
     self:reset()
 end
diff --git a/RNN.lua b/RNN.lua
index 17faa3a..4871676 100644
--- a/RNN.lua
+++ b/RNN.lua
@@ -4,7 +4,7 @@ local errcheck = cudnn.errcheck
 
 local DESCS = {'rnnDesc', 'dropoutDesc', 'wDesc', 'xDescs', 'yDescs', 'hxDesc', 'hyDesc', 'cxDesc', 'cyDesc'}
 
-function RNN:__init(inputSize, hiddenSize, numLayers, batchFirst)
+function RNN:__init(inputSize, hiddenSize, numLayers, batchFirst, dropout)
    parent.__init(self)
 
    self.datatype = 'CUDNN_DATA_FLOAT'
@@ -17,7 +17,7 @@ function RNN:__init(inputSize, hiddenSize, numLayers, batchFirst)
    self.numDirections = 1 -- set to 2 for bi-directional.
    self.inputMode = 'CUDNN_LINEAR_INPUT'
    self.mode = 'CUDNN_RNN_RELU'
-   self.dropout = 0
+   self.dropout = dropout or 0
    self.seed = 0x01234567
    self.batchFirst = batchFirst or false -- Set to true for batch x time x inputdim.
 
@@ -46,7 +46,7 @@ function RNN:reset(stdv)
    errcheck('cudnnGetRNNParamsSize',
             cudnn.getHandle(),
             self.rnnDesc[0],
-            self.xDescs[0],	    
+            self.xDescs[0],
             weightSize:data(),
 	    self.datatype)
    weightSize[1] = (weightSize[1] + 3) / 4 -- sizeof(float)
@@ -119,7 +119,7 @@ end
 function RNN:resetRNNDescriptor()
    if not self.rnnDesc then
       self.rnnDesc = self:createRNNDescriptors(1)
-   end   
+   end
    errcheck('cudnnSetRNNDescriptor',
             self.rnnDesc[0],
             self.hiddenSize,
@@ -279,7 +279,7 @@ function RNN:updateOutput(input)
    local oSize = torch.LongStorage({self.seqLength, self.miniBatch, self.hiddenSize * self.numDirections})
    local oStride = torch.LongStorage({self.miniBatch * self.hiddenSize * self.numDirections, self.hiddenSize * self.numDirections, 1})
    self.output:resize(oSize, oStride)
-   local y = self.output 
+   local y = self.output
    local w = self.weight
    local hy = self:resizeHidden(self.hiddenOutput):zero()
    local cy = self:resizeHidden(self.cellOutput):zero()
diff --git a/RNNReLU.lua b/RNNReLU.lua
index 3aa8ee9..fc262e2 100644
--- a/RNNReLU.lua
+++ b/RNNReLU.lua
@@ -1,7 +1,7 @@
 local RNNReLU, parent = torch.class('cudnn.RNNReLU', 'cudnn.RNN')
 
-function RNNReLU:__init(inputSize, hiddenSize, numLayers, batchFirst)
-    parent.__init(self,inputSize, hiddenSize, numLayers, batchFirst)
+function RNNReLU:__init(inputSize, hiddenSize, numLayers, batchFirst, dropout)
+    parent.__init(self,inputSize, hiddenSize, numLayers, batchFirst, dropout)
     self.mode = 'CUDNN_RNN_RELU'
     self:reset()
 end
diff --git a/RNNTanh.lua b/RNNTanh.lua
index 98fa87c..3382a52 100644
--- a/RNNTanh.lua
+++ b/RNNTanh.lua
@@ -1,7 +1,7 @@
 local RNNTanh, parent = torch.class('cudnn.RNNTanh', 'cudnn.RNN')
 
-function RNNTanh:__init(inputSize, hiddenSize, numLayers, batchFirst)
-    parent.__init(self,inputSize, hiddenSize, numLayers, batchFirst)
+function RNNTanh:__init(inputSize, hiddenSize, numLayers, batchFirst, dropout)
+    parent.__init(self,inputSize, hiddenSize, numLayers, batchFirst, dropout)
     self.mode = 'CUDNN_RNN_TANH'
     self:reset()
 end
diff --git a/cudnn-scm-1.rockspec b/cudnn-scm-1.rockspec
index bc36117..44c5f28 100644
--- a/cudnn-scm-1.rockspec
+++ b/cudnn-scm-1.rockspec
@@ -8,7 +8,7 @@ source = {
 description = {
    summary = "Torch7 FFI bindings for NVIDIA CuDNN kernels!",
    detailed = [[
-   All CuDNN modules exposed as nn.Module derivatives so 
+   All CuDNN modules exposed as nn.Module derivatives so
    that they can be used with torch's neural network package
    ]],
    homepage = "https://github.com/soumith/cudnn.torch",