[OpenVINO Backend] : add support for numpy.nan_to_num #21186

11happy · 2025-04-18T17:57:55Z

Overview:

fixes [Good First Issue][Keras 3 OpenVINO Backend]: Support numpy.nan_to_num operation openvinotoolkit/openvino#30119

Testing:

Currently passsing all tests with this approach but having issues with isinf approach . I have mentioned in detail in comments below.

CC:

@rkazants

Signed-off-by: 11happy <[email protected]>

11happy · 2025-04-18T18:00:32Z

current approach is passing all tests but when I use is_inf to create masks for neg_inf & pos_inf somehow i think it is not able to create those masks properly any insights on that is appreciated.

previous approach to create masks:

posinf_mask = ov_opset.is_inf(x,{'detect_negative': False,'detect_positive': True}).output(0)
neginf_mask = ov_opset.is_inf(x,{'detect_negative': True,'detect_positive': False}).output(0)

error faced(infs are not detected here I checked them ):

Thank you

11happy · 2025-04-18T18:02:28Z

as numpy documentation mentions to use dtype max I create new dictionaries is there any other approach to directly get dtype maxes & mins please let me know.
Thank you

codecov-commenter · 2025-04-18T18:05:19Z

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 82.61%. Comparing base (50dae30) to head (1ed1b37).
Report is 18 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master   #21186      +/-   ##
==========================================
+ Coverage   82.59%   82.61%   +0.02%     
==========================================
  Files         564      564              
  Lines       54407    54474      +67     
  Branches     8449     8469      +20     
==========================================
+ Hits        44936    45003      +67     
+ Misses       7396     7395       -1     
- Partials     2075     2076       +1

Flag	Coverage Δ
keras	`82.42% <100.00%> (+0.02%)`	⬆️
keras-jax	`63.71% <0.00%> (-0.10%)`	⬇️
keras-numpy	`58.83% <0.00%> (-0.10%)`	⬇️
keras-openvino	`32.99% <100.00%> (+0.04%)`	⬆️
keras-tensorflow	`64.11% <0.00%> (-0.11%)`	⬇️
keras-torch	`63.79% <0.00%> (-0.11%)`	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

Signed-off-by: 11happy <[email protected]>

rkazants · 2025-04-19T11:13:38Z

keras/src/backend/openvino/excluded_concrete_tests.txt

@@ -96,7 +96,6 @@ NumpyOneInputOpsCorrectnessTest::test_median
 NumpyOneInputOpsCorrectnessTest::test_meshgrid


please also remove NumpyDtypeTest::test_nan

rkazants · 2025-04-19T11:15:12Z

keras/src/backend/openvino/numpy.py

+    nan_vector = ov_opset.broadcast(
+        ov_opset.constant(nan, dtype), shape_x
+    ).output(0)


it can be just a scalar. select operation will broadcast it automatically. No need broadcast by yourself.

the same comment for pos_inf and neg_inf:)

rkazants · 2025-04-19T11:17:00Z

current approach is passing all tests but when I use is_inf to create masks for neg_inf & pos_inf somehow i think it is not able to create those masks properly any insights on that is appreciated.

previous approach to create masks:
posinf_mask = ov_opset.is_inf(x,{'detect_negative': False,'detect_positive': True}).output(0)
neginf_mask = ov_opset.is_inf(x,{'detect_negative': True,'detect_positive': False}).output(0)
error faced(infs are not detected here I checked them ):

Thank you

Can you please create standalone code that demonstrates this issue and create issue in openvino github? It will be greatly to fix it on our side.

Best regards,
Roman

Signed-off-by: 11happy <[email protected]>

11happy · 2025-04-19T13:03:54Z

Can you please create standalone code that demonstrates this issue and create issue in openvino github? It will be greatly to fix it on our side.

Best regards, Roman

Sure I will create a new issue there , also I have correctec the code based on new tests & previous reviews.
Thank you

Signed-off-by: 11happy <[email protected]>

11happy · 2025-04-19T13:41:34Z

when I tried to reproduce the same error in a colab

now the is_inf are detecting it properly however still the error persists when I try to do it in numpy.nan_to_num implementation .. not sure why, but am debugging it , any insights on this is appreciated.
Thank you

rkazants · 2025-04-19T16:35:52Z

when I tried to reproduce the same error in a colab now the is_inf are detecting it properly however still the error persists when I try to do it in numpy.nan_to_num implementation .. not sure why, but am debugging it , any insights on this is appreciated. Thank you

try to create constant node instead of parameter, i.e. no params and one result node.

Signed-off-by: 11happy <[email protected]>

11happy · 2025-04-19T18:35:58Z

in is_inf we need to send the input by convert it to float inputs then only it works.

I first tried running the tests for is_inf by excluding it & saw they were also failing

def isinf(x):
    x = get_ov_output(x)
    return OpenVINOKerasTensor(ov_opset.is_inf(x).output(0))

correct usage would be & this change passes all tests same thing goes for is_finite I cheecked that as well.

def isinf(x):
    x = get_ov_output(x)
    return OpenVINOKerasTensor(ov_opset.is_inf(ov_opset.convert(x, Type.f32)).output(0))

also on colab it was always correctly workin also with constant node & single result with float dtype. but when I changed dtype to int they started failing & giving error

I suppose it is intentional to pass them as float
Thank you

rkazants

@fchollet, looks good to me. Recommend to merge.

rkazants · 2025-04-21T18:19:34Z

keras/src/backend/openvino/numpy.py

+        neginf if neginf is not None else DTYPES_MIN[dtype], dtype
+    ).output(0)
+    posinf_mask = ov_opset.is_inf(
+        ov_opset.convert(x, Type.f32),


probably it is a bug on OpenVINO side. Please create issue for non-f32 types in OpenVINO repo. For now, this WA absolutely works for us

Done , I have created the issue.
Thank you

rkazants · 2025-04-23T07:55:39Z

keras/src/backend/openvino/numpy.py

+    x = get_ov_output(x)
+    dtype = x.get_element_type()
+    nan_val = ov_opset.constant(nan, dtype).output(0)
+    posinf_val = ov_opset.constant(


@11happy, let us do not call is_inf and is_nan for integer types. You can return input as is as I understand.
For floating point types you need to call but convert to fp32 is not needed.
Can you fix it please in this PR?

I added a check for int type & early retunr & removed the dtype conversion to float ,

if dtype.is_integral(): return OpenVINOKerasTensor(x)

removed this ov_opset.convert(x, Type.f32)) conversion from is_inf, with these changes test is failing , it is not able to detect the infs.

@11happy, please clarify what cases exactly are failing. For int types, it should pass because there is no inf and nan values.

so the test case that is failing is previous one only

> self.assertAllClose( knp.nan_to_num(x, nan=2, posinf=3, neginf=4), [1.0, 2.0, 3.0, 4.0] ) keras/src/ops/numpy_test.py:4806: _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ keras/src/testing/test_case.py:48: in assertAllClose np.testing.assert_allclose(x1, x2, atol=atol, rtol=rtol, err_msg=msg) _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ args = (<function assert_allclose.<locals>.compare at 0x722684c8ed40>, array([ 1.00000000e+00, 2.00000000e+00, 3.40282347e+38, -3.40282347e+38]), array([1., 2., 3., 4.])) kwds = {'equal_nan': True, 'err_msg': 'None', 'header': 'Not equal to tolerance rtol=1e-06, atol=1e-06', 'verbose': True} @wraps(func) def inner(*args, **kwds): with self._recreate_cm(): > return func(*args, **kwds) E AssertionError: E Not equal to tolerance rtol=1e-06, atol=1e-06 E None E Mismatched elements: 2 / 4 (50%) E Max absolute difference: 3.40282347e+38 E Max relative difference: 1.13427449e+38 E x: array([ 1.000000e+00, 2.000000e+00, 3.402823e+38, -3.402823e+38]) E y: array([1., 2., 3., 4.]) /usr/lib/python3.12/contextlib.py:81: AssertionError

I tried to debug it , my observations:
this is failing for f64 dtype, here the test case with dtype f64 is failing , also in conversion logic if I intentionally convert it to f64 type instead of f32 the above test fails. here specifically it passes with f16,f32, but fails with f64.

I am also attaching a ss of PDB :

I have now a reproducible code on f64 dtype for the same error
I have posted that on the previous ticket openvinotoolkit/openvino#30264

Well, can you please do the implementation as follows:

return input as is for integer types

use is_inf and is_nan mask for floating point types. In addition, use WA with convert for fp64, for other types it is not needed.
Best regards,
Roman

I have also added a comment referencing the issue

Signed-off-by: 11happy <[email protected]>

rkazants

@fchollet, looks good to me. Recommend to merge.

feat: add support for numpy.nan_to_num

9497736

Signed-off-by: 11happy <[email protected]>

google-ml-butler bot added the size:M label Apr 18, 2025

google-ml-butler bot assigned gbaned Apr 18, 2025

use np.inf

eb04128

Signed-off-by: 11happy <[email protected]>

rkazants reviewed Apr 19, 2025

View reviewed changes

correct implementation based on new tests

4d72bbf

Signed-off-by: 11happy <[email protected]>

use np only torch having import errors

e8ad543

Signed-off-by: 11happy <[email protected]>

use inf approach

cff09be

Signed-off-by: 11happy <[email protected]>

rkazants reviewed Apr 21, 2025

View reviewed changes

rkazants reviewed Apr 23, 2025

View reviewed changes

refactor code

1ed1b37

Signed-off-by: 11happy <[email protected]>

rkazants reviewed Apr 27, 2025

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[OpenVINO Backend] : add support for numpy.nan_to_num #21186

[OpenVINO Backend] : add support for numpy.nan_to_num #21186

11happy commented Apr 18, 2025

11happy commented Apr 18, 2025 •

edited

Loading

11happy commented Apr 18, 2025

codecov-commenter commented Apr 18, 2025 •

edited

Loading

rkazants Apr 19, 2025

11happy Apr 19, 2025

rkazants Apr 19, 2025

rkazants Apr 19, 2025

11happy Apr 19, 2025

rkazants commented Apr 19, 2025

11happy commented Apr 19, 2025 •

edited

Loading

11happy commented Apr 19, 2025 •

edited

Loading

rkazants commented Apr 19, 2025

11happy commented Apr 19, 2025 •

edited

Loading

rkazants left a comment

rkazants Apr 21, 2025

11happy Apr 21, 2025

rkazants Apr 23, 2025

11happy Apr 23, 2025 •

edited

Loading

rkazants Apr 24, 2025

11happy Apr 25, 2025 •

edited

Loading

11happy Apr 25, 2025 •

edited

Loading

rkazants Apr 26, 2025

11happy Apr 27, 2025

11happy Apr 27, 2025

rkazants left a comment

		@@ -96,7 +96,6 @@ NumpyOneInputOpsCorrectnessTest::test_median
		NumpyOneInputOpsCorrectnessTest::test_meshgrid

[OpenVINO Backend] : add support for numpy.nan_to_num #21186

Are you sure you want to change the base?

[OpenVINO Backend] : add support for numpy.nan_to_num #21186

Conversation

11happy commented Apr 18, 2025

11happy commented Apr 18, 2025 • edited Loading

11happy commented Apr 18, 2025

codecov-commenter commented Apr 18, 2025 • edited Loading

Codecov Report

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

rkazants commented Apr 19, 2025

11happy commented Apr 19, 2025 • edited Loading

11happy commented Apr 19, 2025 • edited Loading

rkazants commented Apr 19, 2025

11happy commented Apr 19, 2025 • edited Loading

rkazants left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

11happy Apr 23, 2025 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

11happy Apr 25, 2025 • edited Loading

Choose a reason for hiding this comment

11happy Apr 25, 2025 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

rkazants left a comment

Choose a reason for hiding this comment

11happy commented Apr 18, 2025 •

edited

Loading

codecov-commenter commented Apr 18, 2025 •

edited

Loading

11happy commented Apr 19, 2025 •

edited

Loading

11happy commented Apr 19, 2025 •

edited

Loading

11happy commented Apr 19, 2025 •

edited

Loading

11happy Apr 23, 2025 •

edited

Loading

11happy Apr 25, 2025 •

edited

Loading

11happy Apr 25, 2025 •

edited

Loading