Skip to content

Add VectorizedMap op class. #21516

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Jul 28, 2025
Merged

Add VectorizedMap op class. #21516

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
39 changes: 34 additions & 5 deletions keras/src/ops/core.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,13 +16,16 @@ def call(self, f, xs):
return backend.core.map(f, xs)

def compute_output_spec(self, f, xs):
x = xs[0]
n = xs.shape[0]
x = tree.map_structure(lambda t: t[0], xs)
n = tree.flatten(xs)[0].shape[0]
y = backend.compute_output_spec(f, x)

def append_batch_axis(x):
def append_batch_axis(t):
return KerasTensor(
shape=(n,) + x.shape, dtype=x.dtype, sparse=x.sparse
shape=(n,) + t.shape,
dtype=t.dtype,
sparse=t.sparse,
ragged=t.ragged,
)

y = tree.map_structure(append_batch_axis, y)
Expand Down Expand Up @@ -1078,7 +1081,31 @@ def cond(pred, true_fn, false_fn):
return Cond()(pred, true_fn, false_fn)


# TODO: also create an Op subclass VectorizedMap.
class VectorizedMap(Operation):
def __init__(self, function, *, name=None):
super().__init__(name=name)
self.function = function

def call(self, elements):
return backend.core.vectorized_map(self.function, elements)

def compute_output_spec(self, elements):
x = tree.map_structure(lambda t: t[0], elements)
n = tree.flatten(elements)[0].shape[0]
y = backend.compute_output_spec(self.function, x)

def append_batch_axis(t):
return KerasTensor(
shape=(n,) + t.shape,
dtype=t.dtype,
sparse=t.sparse,
ragged=t.ragged,
)

y = tree.map_structure(append_batch_axis, y)
return y
Comment on lines +1092 to +1106
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

medium

The logic in this compute_output_spec method is identical to the one in the Map operation. To improve maintainability and reduce code duplication, you can reuse the implementation from Map.

def compute_output_spec(self, elements):
    # Reuse the implementation from `Map` to avoid code duplication.
    return Map().compute_output_spec(self.function, elements)



@keras_export("keras.ops.vectorized_map")
def vectorized_map(function, elements):
"""Parallel map of `function` on axis 0 of tensor(s) `elements`.
Expand Down Expand Up @@ -1109,6 +1136,8 @@ def vectorized_map(function, elements):
In this case, `function` is expected to take as input
a single list of tensor arguments.
"""
if any_symbolic_tensors((elements,)):
return VectorizedMap(function)(elements)
return backend.core.vectorized_map(function, elements)


Expand Down
81 changes: 71 additions & 10 deletions keras/src/ops/core_test.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -56,17 +56,24 @@ def test_map(self):
def f(x):
return x**2

xs = KerasTensor((None,))
self.assertEqual(core.map(f, xs).shape, (None,))
xs = KerasTensor((None, 5))
self.assertEqual(core.map(f, xs).shape, (None, 5))

# Test nested output
def f2(x):
return {"a": x**2, "b": x * 10}

xs = KerasTensor((None,))
xs = KerasTensor((None, 5))
ys = core.map(f2, xs)
self.assertEqual(ys["a"].shape, (None,))
self.assertEqual(ys["b"].shape, (None,))
self.assertEqual(ys["a"].shape, (None, 5))
self.assertEqual(ys["b"].shape, (None, 5))

# Test nested input
def f3(x):
return x[0] + x[1]

xs = (KerasTensor((None, 5)), KerasTensor((None, 5)))
self.assertEqual(core.map(f3, xs).shape, (None, 5))

def test_saturate_cast(self):
x = KerasTensor((3, 5, None), dtype="float32")
Expand Down Expand Up @@ -125,6 +132,29 @@ def fn(x, y):
self.assertEqual(result[0].shape, (None,))
self.assertEqual(result[1].shape, (None,))

def test_vectorized_map(self):
def f(x):
return x**2

xs = KerasTensor((None, 5))
self.assertEqual(core.vectorized_map(f, xs).shape, (None, 5))

# Test nested output
def f2(x):
return {"a": x**2, "b": x * 10}

xs = KerasTensor((None, 5))
ys = core.vectorized_map(f2, xs)
self.assertEqual(ys["a"].shape, (None, 5))
self.assertEqual(ys["b"].shape, (None, 5))

# Test nested input
def f3(x):
return x[0] + x[1]

xs = (KerasTensor((None, 5)), KerasTensor((None, 5)))
self.assertEqual(core.vectorized_map(f3, xs).shape, (None, 5))

def test_while_loop(self):
def cond(args):
return tree.flatten(args)[0] < 10
Expand Down Expand Up @@ -203,18 +233,25 @@ def test_map(self):
def f(x):
return x**2

xs = KerasTensor((6,))
xs = KerasTensor((6, 5))
ys = core.map(f, xs)
self.assertEqual(ys.shape, (6,))
self.assertEqual(ys.shape, (6, 5))

# Test nested output
def f2(x):
return {"a": x**2, "b": x * 10}

xs = KerasTensor((6,))
xs = KerasTensor((6, 5))
ys = core.map(f2, xs)
self.assertEqual(ys["a"].shape, (6,))
self.assertEqual(ys["b"].shape, (6,))
self.assertEqual(ys["a"].shape, (6, 5))
self.assertEqual(ys["b"].shape, (6, 5))

# Test nested input
def f3(x):
return x[0] + x[1]

xs = (KerasTensor((6, 5)), KerasTensor((6, 5)))
self.assertEqual(core.map(f3, xs).shape, (6, 5))

def test_saturate_cast(self):
x = KerasTensor((3, 5, 7), dtype="float32")
Expand Down Expand Up @@ -307,6 +344,30 @@ def fn(x, y):
self.assertEqual(core.switch(index, [fn], x, y)[0].shape, (5,))
self.assertEqual(core.switch(index, [fn], x, y)[1].shape, (2,))

def test_vectorized_map(self):
def f(x):
return x**2

xs = KerasTensor((6, 5))
ys = core.vectorized_map(f, xs)
self.assertEqual(ys.shape, (6, 5))

# Test nested output
def f2(x):
return {"a": x**2, "b": x * 10}

xs = KerasTensor((6, 5))
ys = core.vectorized_map(f2, xs)
self.assertEqual(ys["a"].shape, (6, 5))
self.assertEqual(ys["b"].shape, (6, 5))

# Test nested input
def f3(x):
return x[0] + x[1]

xs = (KerasTensor((6, 5)), KerasTensor((6, 5)))
self.assertEqual(core.vectorized_map(f3, xs).shape, (6, 5))

def test_while_loop(self):
def cond(args):
return tree.flatten(args)[0] < 10
Expand Down
8 changes: 8 additions & 0 deletions keras/src/ops/numpy.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2340,6 +2340,14 @@ class Deg2rad(Operation):
def call(self, x):
return backend.numpy.deg2rad(x)

def compute_output_spec(self, x):
dtype = backend.standardize_dtype(x.dtype)
if dtype in ["int64", "float64"]:
dtype = "float64"
elif dtype not in ["bfloat16", "float16"]:
dtype = backend.floatx()
return KerasTensor(x.shape, dtype)


@keras_export(["keras.ops.deg2rad", "keras.ops.numpy.deg2rad"])
def deg2rad(x):
Expand Down
11 changes: 11 additions & 0 deletions keras/src/ops/ops_test.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -182,6 +182,7 @@ def test_class_function_consistency(self, module_name):
# Check order of parameters.
if name in (
"fori_loop",
"vectorized_map",
"while_loop",
"batch_normalization",
"dot_product_attention",
Expand Down Expand Up @@ -224,6 +225,16 @@ def test_class_function_consistency(self, module_name):
f"function `{name}` and op class `{op_class.__name__}`",
)

# ==== Check compute_output_spec is implement ====
# - op class should override Operation's `compute_output_spec`
self.assertTrue(
hasattr(op_class, "compute_output_spec")
and op_class.compute_output_spec
is not Operation.compute_output_spec,
f"Op class `{op_class.__name__}` should override "
"`compute_output_spec`",
)

@parameterized.named_parameters(named_product(module_name=OPS_MODULES))
def test_backend_consistency(self, module_name):
ops_module = getattr(ops, module_name)
Expand Down
Loading