diff --git a/src/data.jl b/src/data.jl
new file mode 100644
index 000000000..86f50c14d
--- /dev/null
+++ b/src/data.jl
@@ -0,0 +1,55 @@
+"""
+Data is an abstract type for generating data in minibatches.
+Its subtypes should implement the Iterator interface:
+
+* `start(iter) => state`
+* `next(iter,state) => (item,state)`
+* `done(iter,state) => Bool`
+"""
+abstract Data
+
+import Base: start, next, done
+
+type ItemTensor <: Data; x; rng; epochsize; batchsize; bootstrap; shuffle; batch;
+    function ItemTensor(x...; rng=MersenneTwister(), epochsize=ccount(x[1]), batchsize=128, bootstrap=false, shuffle=false)
+        nx = ccount(x[1])
+        all(xi->ccount(xi)==nx, x) || error("Item count mismatch")
+        shuffle = (shuffle ? (1:nx) : nothing)
+        batch = map(xi->KUdense(cget(xi,1:batchsize)), x)
+        new(x, rng, epochsize, batchsize, bootstrap, shuffle, batch)
+    end
+end
+
+function start(d::ItemTensor)
+    if d.shuffle != nothing
+        d.shuffle=randperm(d.rng,length(d.shuffle))
+    end
+    return 0
+end
+
+function done(d::ItemTensor, n)
+    n >= d.epochsize
+end
+
+function next(d::ItemTensor, n)
+    nx = ccount(d.x[1])
+    nb = min(d.batchsize, d.epochsize-n)
+    if d.bootstrap
+        ix = rand(d.rng, 1:nx, nb)
+    else
+        error("This is not working yet")
+        i1 = mod1(n+1, nx)
+        i2 = mod1(n+nb, nx)
+        if d.shuffle == nothing
+            ix = (i1 < i2 ? (i1:i2) : [i1:nx; 1:i2])
+            # nb>nx?
+        else
+            # i1==1?
+        end
+    end
+    length(ix) == nb || error()
+    for i=1:length(d.x)
+        cslice!(d.batch[i], d.x[i], ix)
+    end
+    (d.batch, n+nb)
+end