basic back pass

src/backend/tensorflow/model.jl

@@ -1,18 +1,21 @@
 using Flux: mapt, collectt, shapecheckt
 
 struct Exec
-  session::Session
+  session ::Session
-  input::Any
+  input   ::Any
-  output::Any
+  output  ::Any
-  params::Dict{Flux.Param,Tensor}
+  grads   ::Any
-  stacks::Dict{Any,Any}
+  params  ::Dict{Flux.Param,Tensor}
+  stacks  ::Dict{Any,Any}
 end
 
 function makesession(model, inputs; session = Session(Graph()))
   inputs = mapt(_ -> placeholder(Float32), inputs)
   params, stacks, output = tograph(model, inputs...)
+  # grads = gradients(output, [collectt(inputs)..., values(params)...])
+  grads = placeholder(Float32)
   run(session, global_variables_initializer())
-  Exec(session, inputs, output, params, stacks)
+  Exec(session, inputs, output, grads, params, stacks)
 end
 
 retuple(xs) = xs
@@ -20,11 +23,37 @@ retuple(xs::AbstractArray{<:AbstractArray}) = (retuple.(xs)...,)
 
 dictt(xs, ys) = Dict(zip(collectt(xs), collectt(ys)))
 
-function (m::Exec)(args...)
+function params(m::Exec, args...)
   shapecheckt(m.input, args)
   idict = dictt(m.input, args)
   pdict = Dict(t => p.x for (p, t) in m.params)
-  retuple(run(m.session, m.output, merge(idict, pdict)))
+  merge(idict, pdict)
+end
+
+function (m::Exec)(args...)
+  retuple(run(m.session, m.output, params(m, args...)))
+end
+
+pullt!(_, xs) = shift!(xs)
+pullt!(x::Tuple, xs) = map(x -> pullt!(x, xs), x)
+
+# TODO: gradients don't work yet
+# `gradients` lacks support for `grad_y`s and multiple `y`s
+
+function Flux.back!(m::Exec, Δ, args...)
+  Δps = run(m.session, m.grads, params(m, args...))
+  Δin = pullt!(m.input, Δps)
+  for (p, Δ) in zip(keys(m.params), Δps)
+    p.Δx .+= Δ
+  end
+  Δin
+end
+
+function Flux.update!(m::Exec, η)
+  for p in keys(m.params)
+    update!(p, η)
+  end
+  return m
 end
 
 mutable struct Model
@@ -41,11 +70,8 @@ function (m::Model)(args...)
   @tferr m.exec.stacks m.exec(args...)
 end
 
-for f in :[back!, update!].args
+Flux.back!(m::Model, Δ, args...) = back!(m.exec, Δ, args...)
-  @eval function Flux.$f(m::Model, args...)
+Flux.update!(m::Model, η) = (update!(m.exec, η); m)
-    error($(string(f)) * " is not yet supported on TensorFlow models")
-  end
-end
 
 # Recurrent Models