Added finit arg to RNN to initialize forget bias, default = ones.

src/rnn.jl

@@ -140,6 +140,7 @@ function RNN(inputSize, hiddenSize;
              seed=0,              # seed=0 for random init, positive integer for replicability
              winit=xavier,
              binit=zeros,
+             finit=ones,        # forget bias for lstm
              usegpu=(gpu()>=0),
              )
     w = dx = dhx = dcx = nothing
@@ -158,7 +159,7 @@ function RNN(inputSize, hiddenSize;
     dropoutDesc = usegpu ? DD(handle=handle,dropout=dropout,seed=seed) : nothing # Need to keep dropoutDesc in RNN so it does not get gc'ed.
     rnnDesc = usegpu ? RD(hiddenSize,numLayers,dropoutDesc,inputMode,direction,mode,algo,dataType) : nothing
     r = RNN(w,h,c,inputSize,hiddenSize,numLayers,dropout,seed,inputMode,direction,mode,algo,dataType,rnnDesc,dropoutDesc,dx,dhx,dcx)
-    r.w = Param(Array{dataType}(undef,1,1,getRNNParamsSize(r)))
+    r.w = Array{dataType}(undef,1,1,getRNNParamsSize(r))
     for a in rnnparams(r; handle=handle, useview=true)
         if a == nothing
             continue
@@ -170,8 +171,14 @@ function RNN(inputSize, hiddenSize;
             error("Invalid RNN param $(summary(a))")
         end
     end
+    if rnnType == :lstm         # separate initialization for lstm forget biases
+        for layer in 1:(numLayers*(bidirectional ? 2 : 1)), id in (2,6), param in (2,)
+            a = rnnparam(r, layer, id, param, useview=true, handle=handle)
+            copyto!(a, finit(dataType, size(a)))
+        end
+    end
     # many copyto! ops to gpu is expensive (~20s), so we init on cpu and copy it over once
-    if usegpu; r.w = Param(KnetArray(value(r.w))); end
+    r.w = Param(usegpu ? KnetArray(r.w) : r.w)
     return r
 end