mx batch semantics

src/backend/mxnet/model.jl

@@ -1,3 +1,4 @@
+using Flux: batchone, rebatch
 
 type Model <: Flux.Model
   model::Any
@@ -50,6 +51,10 @@ function runmodel(model::Model, input)
   copy(model.exec.outputs[1])
 end
 
+(m::Model)(x::Batch) = rebatch(runmodel(m, rawbatch(x)))
+
+(m::Model)(x) = first(m(batchone(x)))
+
 function Flux.back!(model::Model, Δ, x)
   ndzero!(model.grads[:input])
   mx.backward(model.exec, tond(Δ))

test/backend.jl

@@ -5,9 +5,8 @@ let dt = tf(d)
   @test d(xs) ≈ dt(xs)
 end
 
-# TODO: batch semantics
 let dm = mxnet(d, (1, 20))
-  @test d(xs)' ≈ dm(xs')
+  @test d(xs) ≈ dm(xs)
 end
 
 # TensorFlow native integration