Skip to content

[OpenVINO backend] Support for numpy.det using Faddeev-LeVerrier algorithm #22355

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

Open
andersendsa wants to merge 2 commits into
base: master
Choose a base branch
from
Open

[OpenVINO backend] Support for numpy.det using Faddeev-LeVerrier algorithm #22355

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
e634240
Implement numpy.det in Openvino Backend
andersendsa Mar 5, 2026
5650c8f
formatting fix
andersendsa Mar 5, 2026
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
1 change: 0 additions & 1 deletion keras/src/backend/openvino/excluded_concrete_tests.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -346,7 +346,6 @@ LayerTest::test_end_to_end_masking
LayerTest::test_quantized_layer_with_remat
LayerTest::test_stateless_call
LinalgOpsCorrectnessTest::test_cholesky
LinalgOpsCorrectnessTest::test_det
LinalgOpsCorrectnessTest::test_eig
LinalgOpsCorrectnessTest::test_eigh
LinalgOpsCorrectnessTest::test_lstsq
Expand Down
118 changes: 117 additions & 1 deletion keras/src/backend/openvino/linalg.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,4 +1,5 @@
import openvino.opset15 as ov_opset
import openvino as ov
from openvino import Type

from keras.src.backend import config
Expand Down Expand Up @@ -34,7 +35,122 @@ def cholesky_inverse(a, upper=False):


def det(a):
raise NotImplementedError("`det` is not supported with openvino backend")
a = convert_to_tensor(a)
a_ov = get_ov_output(a)
original_type = a_ov.get_element_type()

# Avoid constant folding bug for f64 in OpenVINO CPU Loop evaluate
if original_type == Type.f64:
a_ov = ov_opset.convert(a_ov, Type.f32).output(0)

a_shape = ov_opset.shape_of(a_ov, output_type="i32").output(0)

rank = a_ov.get_partial_shape().rank.get_length()

minus_1 = ov_opset.constant([-1], Type.i32).output(0)
minus_2 = ov_opset.constant([-2], Type.i32).output(0)

N_node_1d = ov_opset.gather(
a_shape, minus_1, ov_opset.constant(0, Type.i32).output(0)
).output(0)
N_node_scalar = ov_opset.squeeze(
N_node_1d, ov_opset.constant([0], Type.i32).output(0)
).output(0)
Comment on lines +50 to +58
Copy link
Contributor

@gemini-code-assist gemini-code-assist bot Mar 5, 2026

Choose a reason for hiding this comment

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

medium

To improve readability and maintainability, it's good practice to define constants at the top of their scope. Here, you can define zero_i32 and zero_i32_vec and reuse them. This avoids creating the same constant inline multiple times later in the function.

    minus_1 = ov_opset.constant([-1], Type.i32).output(0)
    minus_2 = ov_opset.constant([-2], Type.i32).output(0)

    zero_i32 = ov_opset.constant(0, Type.i32).output(0)
    zero_i32_vec = ov_opset.constant([0], Type.i32).output(0)

    N_node_1d = ov_opset.gather(
        a_shape, minus_1, zero_i32
    ).output(0)
    N_node_scalar = ov_opset.squeeze(
        N_node_1d, zero_i32_vec
    ).output(0)


num_batch_dims = rank - 2
if num_batch_dims > 0:
batch_dims_shape = ov_opset.broadcast(
ov_opset.constant([1], Type.i32).output(0),
ov_opset.constant([num_batch_dims], Type.i32).output(0)
).output(0)
eye_shape = ov_opset.concat(
[batch_dims_shape, N_node_1d, N_node_1d], 0
).output(0)
else:
eye_shape = ov_opset.concat([N_node_1d, N_node_1d], 0).output(0)

eye = ov_opset.eye(
N_node_scalar,
N_node_scalar,
ov_opset.constant(0, Type.i32).output(0),
a_ov.get_element_type()
).output(0)
eye_reshaped = ov_opset.reshape(eye, eye_shape, False).output(0)

trip_count = N_node_scalar
loop = ov_opset.loop(
trip_count, ov_opset.constant(True, Type.boolean).output(0)
)

M_param = ov_opset.parameter(
[-1] * rank, a_ov.get_element_type(), "M"
)
k_param = ov_opset.parameter([], Type.i32, "k")
A_body_param = ov_opset.parameter(
[-1] * rank, a_ov.get_element_type(), "A_body"
)
eye_body_param = ov_opset.parameter(
[-1] * rank, a_ov.get_element_type(), "eye_body"
)

k_next = ov_opset.add(
k_param.output(0), ov_opset.constant(1, Type.i32).output(0)
).output(0)
k_f32 = ov_opset.convert(k_next, a_ov.get_element_type()).output(0)

M_diag = ov_opset.multiply(
M_param.output(0), eye_body_param.output(0)
).output(0)
trace_axes = ov_opset.concat([minus_2, minus_1], 0).output(0)
trace = ov_opset.reduce_sum(M_diag, trace_axes, keep_dims=True).output(0)

minus_one = ov_opset.constant(-1.0, a_ov.get_element_type()).output(0)
c_k_factor = ov_opset.divide(minus_one, k_f32).output(0)
c_k = ov_opset.multiply(c_k_factor, trace).output(0)

c_k_I = ov_opset.multiply(c_k, eye_body_param.output(0)).output(0)
M_plus_c_k_I = ov_opset.add(M_param.output(0), c_k_I).output(0)

M_next = ov_opset.matmul(
A_body_param.output(0), M_plus_c_k_I, False, False
).output(0)

cond_next = ov_opset.constant(True, Type.boolean).output(0)

body = ov.Model(
[M_next, k_next, c_k, cond_next],
[M_param, k_param, A_body_param, eye_body_param]
)
loop.set_function(body)
loop.set_special_body_ports([-1, 3])

loop.set_merged_input(M_param, a_ov, M_next)
loop.set_merged_input(
k_param, ov_opset.constant(0, Type.i32).output(0), k_next
)
loop.set_invariant_input(A_body_param, a_ov)
loop.set_invariant_input(eye_body_param, eye_reshaped)

out_c_k = loop.get_iter_value(c_k, -1)

det_c_k = ov_opset.squeeze(out_c_k, trace_axes).output(0)

N_mod_2 = ov_opset.mod(
N_node_scalar, ov_opset.constant(2, Type.i32).output(0)
).output(0)
N_mod_2_f32 = ov_opset.convert(N_mod_2, a_ov.get_element_type()).output(0)
one = ov_opset.constant(1.0, a_ov.get_element_type()).output(0)
two = ov_opset.constant(2.0, a_ov.get_element_type()).output(0)
sign = ov_opset.subtract(
one, ov_opset.multiply(two, N_mod_2_f32).output(0)
).output(0)

det = ov_opset.multiply(det_c_k, sign).output(0)

if original_type == Type.f64:
det = ov_opset.convert(det, Type.f64).output(0)

return OpenVINOKerasTensor(det)


def eig(a):
Expand Down
Loading