use new session and store params back in the model

src/backend/tensorflow/graph.jl

@@ -76,3 +76,9 @@ function makesession(model, n)
   run(sess, initialize_all_variables())
   sess, params, inputs, output
 end
+
+function storeparams!(sess, params)
+  for (p, t) in params
+    p.x = run(sess, t)
+  end
+end

src/backend/tensorflow/recurrent.jl

@@ -47,19 +47,20 @@ end
 (m::SeqModel)(x::Seq) = first(m(batchone(x)))
 
 function Flux.train!(m::SeqModel, Xs, Ys; epoch = 1, η = 0.1,
-                     loss = (y, ŷ) -> -reduce_sum(y .* log2(ŷ)),
-                     opt = TensorFlow.train.GradientDescentOptimizer(η))
+                     loss = (y, ŷ) -> -reduce_sum(y .* log(ŷ)),
+                     opt = () -> TensorFlow.train.GradientDescentOptimizer(η))
   batchlen, seqlen = length(first(Xs)), length(first(Xs)[1])
   state = batchone.(m.m.model.state)
+  sess, params, (instates, input), (outstates, output) = makesession(m.m.model)
   Y = placeholder(Float32)
-  Loss = loss(Y, m.m.output[end])/batchlen/seqlen
-  minimize_op = TensorFlow.train.minimize(opt, Loss)
+  Loss = loss(Y, output)/batchlen/seqlen
+  minimize_op = TensorFlow.train.minimize(opt(), Loss)
   for e in 1:epoch
     info("Epoch $e\n")
     @progress for (i, (x, y)) in enumerate(zip(Xs,Ys))
-      out = run(m.m.session, vcat(m.m.output..., Loss, minimize_op),
-                merge(Dict(m.m.inputs[end]=>batchone(x), Y=>batchone(y)),
-                      Dict(zip(m.m.inputs[1:end-1], state))))
+      out = run(sess, vcat(outstates..., output, Loss, minimize_op),
+                merge(Dict(input=>batchone(x), Y=>batchone(y)),
+                      Dict(zip(instates, state))))
       state = out[1:length(state)]
       loss = out[end-1]
       isnan(loss) && error("Loss is NaN")
@@ -67,4 +68,6 @@ function Flux.train!(m::SeqModel, Xs, Ys; epoch = 1, η = 0.1,
       (i-1) % 10 == 0 && @show loss
     end
   end
+  storeparams!(sess, params)
+  return
 end