Merge pull request #2853 from levskaya/topkjvp

Add top_k jvp and batching rules and tests

docs/jax.lax.rst

@@ -124,6 +124,7 @@ Operators
     sub
     tan
     tie_in
+    top_k
     transpose
 
 

jax/lax/lax.py

@@ -1187,7 +1187,8 @@ def sort_key_val(keys: Array, values: Array, dimension: int = -1) -> Array:
   sorted_keys, sorted_values = result
   return sorted_keys, sorted_values
 
-def top_k(operand: Array, k: int) -> Array:
+def top_k(operand: Array, k: int) -> Tuple[Array, Array]:
+  """Returns top ``k`` values and their indices along the last axis of ``operand``."""
   k = int(k)
   if k < 0:
     raise ValueError("k argument to top_k must be nonnegative, got {}".format(k))
@@ -4624,12 +4625,53 @@ def _top_k_abstract_eval(operand, *, k):
   return (ShapedArray(shape, operand.dtype),
           ShapedArray(shape, onp.dtype(onp.int32)))
 
+def _top_k_jvp(primals, tangents, *, k):
+  operand, = primals
+  tangent, = tangents
+  primals_out = top_k(operand, k)
+  if tangent is ad_util.zero:
+    tangents_out = (ad_util.zero, ad_util.zero)
+  else:
+    _, k_idxs = primals_out
+    idx_shape = k_idxs.shape
+    rank = len(idx_shape)
+    gather_index_shape = idx_shape + (1,)
+    gather_indices = []
+    for i in range(rank-1):
+      _iota = iota(k_idxs.dtype, idx_shape[i])
+      _iota = tie_in(operand, _iota)
+      _iota = broadcast_in_dim(_iota, gather_index_shape, (i,))
+      gather_indices.append(_iota)
+    gather_indices.append(reshape(k_idxs, gather_index_shape))
+    gather_indices = concatenate(gather_indices, dimension=rank)
+    slice_sizes = (1,) * rank
+    dnums = GatherDimensionNumbers(
+      offset_dims=(),
+      collapsed_slice_dims=tuple(range(rank)),
+      start_index_map=tuple(range(rank)))
+    tangents_out = (gather(tangent, gather_indices, dnums, slice_sizes),
+                    ad_util.zero)
+  return primals_out, tangents_out
+
+def _top_k_batch_rule(batched_args, batch_dims, *, k):
+  operand, = batched_args
+  bdim, = batch_dims
+  if bdim == operand.ndim-1:
+    perm = onp.arange(operand.ndim)
+    perm[bdim-1], perm[bdim] = perm[bdim], perm[bdim-1]
+    top_k_v, top_k_i = top_k(transpose(operand, perm), k=k)
+    return (transpose(top_k_v, perm),
+            transpose(top_k_i, perm)), (bdim, bdim)
+  else:
+    return top_k(operand, k=k), (bdim, bdim)
+
 top_k_p = Primitive('top_k')
 top_k_p.multiple_results = True
 top_k_p.def_impl(partial(xla.apply_primitive, top_k_p))
 top_k_p.def_abstract_eval(_top_k_abstract_eval)
 xla.translations[top_k_p] = partial(standard_translate, 'top_k')
-
+ad.primitive_jvps[top_k_p] = _top_k_jvp
+batching.primitive_batchers[top_k_p] = _top_k_batch_rule
 
 def _tie_in_transpose_rule(t):
   return [ad_util.zero, t]

jax/test_util.py

@@ -634,6 +634,13 @@ def fn(shape, dtype):
     return randint(low, high=high, size=shape, dtype=dtype)
   return fn
 
+def rand_unique_int():
+  randchoice = npr.RandomState(0).choice
+  def fn(shape, dtype):
+    return randchoice(onp.arange(onp.prod(shape), dtype=dtype),
+                      size=shape, replace=False)
+  return fn
+
 def rand_bool():
   rng = npr.RandomState(0)
   def generator(shape, dtype):

tests/lax_test.py

@@ -2484,6 +2484,22 @@ def args_maker():
     fun = lambda keys, values: lax.sort_key_val(keys, values, axis)
     check_grads(fun, (keys, values), 2, ["fwd", "rev"], 1e-2, 1e-2, 1e-2)
 
+  @parameterized.named_parameters(jtu.cases_from_list(
+      {"testcase_name": "_shape={}_k={}".format(
+          jtu.format_shape_dtype_string(shape, dtype), k),
+       "rng_factory": rng_factory, "shape": shape, "dtype": dtype, "k": k}
+      for dtype in [onp.float32,]
+      for shape in [(4,), (5, 5), (2, 1, 4)]
+      for k in [1, 3]
+      for rng_factory in [jtu.rand_default]))
+  def testTopKGrad(self, shape, dtype, k, rng_factory):
+    rng = rng_factory()
+    perm_rng = onp.random.RandomState(0)
+    flat_values = onp.arange(onp.prod(shape, dtype=int), dtype=dtype)
+    values = perm_rng.permutation(flat_values).reshape(shape)
+    fun = lambda vs: lax.top_k(vs, k=k)[0]
+    check_grads(fun, (values,), 2, ["fwd", "rev"], eps=1e-2)
+
   @parameterized.named_parameters(jtu.cases_from_list(
       {"testcase_name": "_shape={}_idxs={}_axes={}".format(
           jtu.format_shape_dtype_string(shape, dtype), idxs, axes),
@@ -3213,6 +3229,27 @@ def testBroadcastShapesReturnsPythonInts(self):
     out_shape = lax.broadcast_shapes(shape1, shape2)
     self.assertTrue(all(type(s) is int for s in out_shape))
 
+  @parameterized.named_parameters(jtu.cases_from_list(
+      {"testcase_name": "_shape={}_k={}_bdims={}".format(
+          jtu.format_shape_dtype_string(shape, dtype), k, bdims),
+       "shape": shape, "dtype": dtype, "k": k, "bdims": bdims, "rng_factory": rng_factory}
+      for shape in [(4,), (3, 4, 5)]
+      for k in [1, 3]
+      for bdims in all_bdims(shape)
+      # TODO(b/155170120): test with repeats once the XLA:CPU stable top_k bug is fixed:
+      # The top_k indices for integer arrays with identical entries won't match between
+      # vmap'd version and manual reference, so only test unique integer arrays for int_dtypes.
+      for dtype, rng_factory in itertools.chain(
+        zip(float_dtypes, itertools.repeat(jtu.rand_default)),
+        zip(int_dtypes, itertools.repeat(jtu.rand_unique_int)))))
+  def testTopK(self, shape, dtype, k, bdims, rng_factory):
+    rng = rng_factory()
+    # _CheckBatching doesn't work with tuple outputs, so test outputs separately.
+    op1 = lambda x: lax.top_k(x, k=k)[0]
+    self._CheckBatching(op1, 5, bdims, (shape,), (dtype,), rng)
+    op2 = lambda x: lax.top_k(x, k=k)[1]
+    self._CheckBatching(op2, 5, bdims, (shape,), (dtype,), rng)
+
   # TODO Concatenate
   # TODO Reverse
   # TODO DynamicSlice