Make data.Serial return a generator function instead of a generator so it can compose, add default arg for easy calling.

PiperOrigin-RevId: 323283358

Author: Lukasz Kaiser

trax/data/inputs.py

@@ -82,11 +82,12 @@
 
 
 def Serial(*fns):  # pylint: disable=invalid-name
-  """Creates an input pipeline by running all functions one after another."""
-  generator = None
-  for f in fastmath.tree_flatten(fns):
-    generator = f(generator)
-  return generator
+  """Combines generator functions into one that runs them in turn."""
+  def composed_fns(generator=None):
+    for f in fastmath.tree_flatten(fns):
+      generator = f(generator)
+    return generator
+  return composed_fns
 
 
 def Log(n_steps_per_example=1, only_shapes=True):  # pylint: disable=invalid-name

trax/data/inputs_test.py

@@ -76,7 +76,16 @@ def test_batch_data(self):
   def test_serial(self):
     dataset = lambda _: ((i, i+1) for i in range(10))
     batches = data.Serial(dataset, data.Shuffle(3), data.Batch(10))
-    batch = next(batches)
+    batch = next(batches())
+    self.assertLen(batch, 2)
+    self.assertEqual(batch[0].shape, (10,))
+
+  def test_serial_composes(self):
+    """Check that data.Serial works inside another data.Serial."""
+    dataset = lambda _: ((i, i+1) for i in range(10))
+    serial1 = data.Serial(dataset, data.Shuffle(3))
+    batches = data.Serial(serial1, data.Batch(10))
+    batch = next(batches())
     self.assertLen(batch, 2)
     self.assertEqual(batch[0].shape, (10,))
 
@@ -88,7 +97,7 @@ def test_serial_with_python(self):
         lambda g: filter(lambda x: x[0] % 2 == 1, g),
         data.Batch(2)
     )
-    batch = next(batches)
+    batch = next(batches())
     self.assertLen(batch, 2)
     (xs, ys) = batch
     # First tuple after filtering is (1, 3) = (1, 2+1).

trax/data/tf_inputs.py

@@ -242,7 +242,8 @@ def select_from(example):
   dataset = dataset.map(select_from)
   dataset = dataset.repeat()
 
-  def gen(unused_arg):
+  def gen(generator=None):
+    del generator
     for example in fastmath.dataset_as_numpy(dataset):
       yield example
   return gen