fix accuracy for batches

src/training.jl

@@ -2,12 +2,14 @@ tobatch(xs::Batch) = rawbatch(xs)
 tobatch(xs) = tobatch(batchone(xs))
 
 function accuracy(m, data)
+  n = 0
   correct = 0
   for (x, y) in data
     x, y = tobatch.((x, y))
+    n += size(x, 1)
     correct += sum(onecold(m(x)) .== onecold(y))
   end
-  return correct/length(data)
+  return correct/n
 end
 
 function train!(m, train, test = [];