Add OpenVINO backend support for argmin and argmax #21060

Hmm-1224 · 2025-03-18T05:28:20Z

@rkazants please check out this PR

codecov-commenter · 2025-03-18T05:36:09Z

Codecov Report

Attention: Patch coverage is 77.77778% with 8 lines in your changes missing coverage. Please review.

Project coverage is 82.68%. Comparing base (1ba2012) to head (46e51b6).
Report is 25 commits behind head on master.

Files with missing lines	Patch %	Lines
keras/src/backend/openvino/numpy.py	77.77%	4 Missing and 4 partials ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master   #21060      +/-   ##
==========================================
+ Coverage   82.47%   82.68%   +0.21%     
==========================================
  Files         563      564       +1     
  Lines       53837    54168     +331     
  Branches     8359     8419      +60     
==========================================
+ Hits        44402    44789     +387     
+ Misses       7394     7300      -94     
- Partials     2041     2079      +38

Flag	Coverage Δ
keras	`82.49% <77.77%> (+0.20%)`	⬆️
keras-jax	`64.03% <0.00%> (+0.23%)`	⬆️
keras-numpy	`59.06% <0.00%> (+0.31%)`	⬆️
keras-openvino	`32.92% <77.77%> (+0.20%)`	⬆️
keras-tensorflow	`64.31% <0.00%> (+0.20%)`	⬆️
keras-torch	`64.07% <0.00%> (+0.24%)`	⬆️

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.

Hmm-1224 · 2025-03-22T20:01:00Z

hello @rkazants,
please check out this PR and guide me further .....

Hmm-1224 · 2025-03-24T18:25:31Z

Hey @rkazants,
As we are approaching the final stages of GSoC 2025, I’m truly passionate about contributing to OpenVINO and would love to successfully complete this PR as a prerequisite. I tried to reach out to you many times but may be those messages couldn't reach you. I’d really appreciate it if you could take a moment to review it.

Apart from this as you are a valuable member of this community, i want you to review my proposal and give me suggestions if needed....
OpenVINO (1).pdf

rkazants · 2025-03-27T07:04:05Z

keras/src/backend/openvino/excluded_concrete_tests.txt

+NumpyOneInputOpsCorrectnessTest::test_exp
+NumpyOneInputOpsCorrectnessTest::test_expand_dims


revert this change

rkazants · 2025-03-27T07:04:17Z

keras/src/backend/openvino/excluded_concrete_tests.txt

 NumpyOneInputOpsCorrectnessTest::test_exp2
 NumpyOneInputOpsCorrectnessTest::test_expm1
 NumpyOneInputOpsCorrectnessTest::test_flip
 NumpyOneInputOpsCorrectnessTest::test_floor_divide
-NumpyOneInputOpsCorrectnessTest::test_hstack


revert this change

rkazants · 2025-03-27T07:04:25Z

keras/src/backend/openvino/excluded_concrete_tests.txt

revert this change

rkazants · 2025-03-27T07:04:46Z

keras/src/backend/openvino/numpy.py

 def hstack(xs):
-    raise NotImplementedError("`hstack` is not supported with openvino backend")
+    if not isinstance(xs, (list, tuple)):
+        raise TypeError("Input to `hstack` must be a list or tuple of tensors.")
+    if len(xs) == 0:
+        raise ValueError("Input list to `hstack` cannot be empty.")
+    element_type = None
+    for x in xs:
+        if isinstance(x, OpenVINOKerasTensor):
+            element_type = x.output.get_element_type()
+            break
+    xs = [get_ov_output(x, element_type) for x in xs]
+    xs = _align_operand_types(xs[0], xs[1], "hstack()")
+    rank = len(xs[0].get_partial_shape())
+    axis = 1 if rank > 1 else 0
+    return OpenVINOKerasTensor(ov_opset.concat(xs, axis=axis).output(0))


it relates different PR. please remove it

oh yeah, sorry about it.i did remove it

@rkazants,
please check it out again..

rkazants · 2025-03-28T05:24:09Z

keras/src/backend/openvino/excluded_concrete_tests.txt

+NumpyOneInputOpsCorrectnessTest::test_log1p
+NumpyOneInputOpsCorrectnessTest::test_logaddexp


revert this change

rkazants · 2025-03-28T05:24:21Z

keras/src/backend/openvino/excluded_concrete_tests.txt

 NumpyDtypeTest::test_linspace
-NumpyDtypeTest::test_log10
 NumpyDtypeTest::test_log1p
-NumpyDtypeTest::test_log
+NumpyDtypeTest::test_logaddexp
 NumpyDtypeTest::test_logspace


revert these changes

Hello @rkazants ,
check out the if changes is made correctly...sorry if this frustrate you. Actually i just copy paste from keras-team:master excluded_conrete_test.txt and remove argmin and argmax but changes you told me to make was different from that...

Hmm-1224 · 2025-03-29T09:29:17Z

hey @rkazants,
please check it out. i guess i made all mandatory changes u suggested..

Hmm-1224 · 2025-03-29T09:32:16Z

Why is this coming ? since i removed argmax argmin the no. of lines reduced and last line come at 164 from 168..then why is it creating issue?

Hmm-1224 · 2025-04-01T16:17:55Z

hey @rkazants,
please check out the modifications.8th April is deadline and i have to complete it before deadline. Please help me in competing this task on time..
Also i requested you to review my OpenVINO proposal. And i truly need your feedback on it.. If you could take out few minutes it would be great. Sorry i have to talk about that here but i personally mailed you but you probably missed it.

rkazants · 2025-04-03T10:47:27Z

keras/src/backend/openvino/numpy.py

+        if topk_indices_shape.rank.get_length() == rank - 1:
+            topk_indices = ov_opset.unsqueeze(topk_indices, [axis]).output(0)


I don't think we need this code. Please remove because TopK will preserve the required dimension

rkazants · 2025-04-03T10:58:51Z

keras/src/backend/openvino/numpy.py

+            axis = rank + axis
+        k = ov_opset.constant(1, Type.i32).output(0)
+    topk_outputs = ov_opset.topk(
+        x, k=k, axis=axis, mode="max", sort="none", index_element_type=Type.i32


According to numpy documentation "In case of multiple occurrences of the maximum values, the indices corresponding to the first occurrence are returned" in https://numpy.org/doc/2.2/reference/generated/numpy.argmax.html, we should use sort = value , mode = max , stable = true

rkazants · 2025-04-03T18:43:12Z

keras/src/backend/openvino/numpy.py

+    if keepdims:
+        new_shape = ov_opset.constant(
+            [1 if i == axis else -1 for i in range(rank)], Type.i32
+        ).output(0)
+        topk_indices = ov_opset.reshape(topk_indices, new_shape, True).output(0)


this branch looks not needed, please check

i tried without it first but it didn't work and were failing for two testcases and resulted in following error..
TypeError: reshape() missing 1 required positional argument: 'special_zero' .

To resolve this, I ensured that reshape() receives the correct arguments by explicitly setting the shape adjustment logic when keepdims=True. This avoids errors and ensures correct behavior. If you have alternate suggestions please do share i will work on it...

topk_indices = topk_outputs.output(1) if not keepdims: topk_indices = ov_opset.squeeze(topk_indices, [axis]).output(0)

this is how i firstly tried as you told TopK already preserves the shape when keepdims=True but didn't work for two testcases as i described above..

Hey @rkazants,
Please check it out. I request you to do it fast please...

@Hmm-1224,

please leave only branch if not keepdims: as you tried

and change flatten_shape above to:

flatten_shape = ov_opset.constant([-1] + [1] * (rank - 1), Type.i32).output(0)

axis is None case turned to be a bit tricky case.

rkazants · 2025-04-04T09:36:29Z

keras/src/backend/openvino/numpy.py

@@ -342,7 +394,7 @@ def argsort(x, axis=-1):
    if rank == 0:
        return OpenVINOKerasTensor(ov_opset.constant([0], Type.i32).output(0))
    if axis is None:
-        flatten_shape = ov_opset.constant([-1], Type.i32).output(0)
+        flatten_shape = ov_opset.constant([-1] + [1] * (rank - 1), Type.i32).output(0)


is something wrong? You told to change flatten_shape..

hey @rkazants, whats the isssue?

okay got it..sorry sir i mistakenly change argsort...

@rkazants please check i fixed that issue...

Hmm-1224 · 2025-04-05T06:03:52Z

hey @rkazants , please check..

Hmm-1224 added 6 commits March 13, 2025 17:57

Update numpy.py

5bce71d

Update excluded_concrete_tests.txt

6fb3be2

Update excluded_concrete_tests.txt

4d1d012

Update numpy.py

f967cad

Update numpy.py

200b0ef

Update excluded_concrete_tests.txt

2f30b59

google-ml-butler bot added the size:L label Mar 18, 2025

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

Hmm-1224 and others added 2 commits March 18, 2025 11:00

Merge branch 'master' into Hmm-1224-patch-1

a9ee5c9

numpy.py formatted

caf6990

Hmm-1224 mentioned this pull request Mar 18, 2025

[Good First Issue][Keras 3 OpenVINO Backend]: Support numpy.hstack operation openvinotoolkit/openvino#29361

Open

gbaned requested a review from mattdangerw March 18, 2025 11:55

google-ml-butler bot added the awaiting review label Mar 18, 2025

Hmm-1224 mentioned this pull request Mar 22, 2025

[Good First Issue][Keras 3 OpenVINO Backend]: Support numpy.argmax and numpy.argmin operations openvinotoolkit/openvino#29011

Open

rkazants reviewed Mar 27, 2025

View reviewed changes

keras/src/backend/openvino/excluded_concrete_tests.txt

NumpyTwoInputOpsCorrectnessTest::test_where

Copy link

Contributor

rkazants Mar 27, 2025

Choose a reason for hiding this comment

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

revert this change

rkazants reviewed Mar 27, 2025

View reviewed changes

Hmm-1224 added 2 commits March 27, 2025 15:30

Update numpy.py

1e8a155

Update excluded_concrete_tests.txt

7d18391

Hmm-1224 requested a review from rkazants March 27, 2025 10:21

rkazants reviewed Mar 28, 2025

View reviewed changes

Hmm-1224 and others added 2 commits March 28, 2025 06:03

Revert changes to excluded_concrete_tests.txt

7e304bb

Update excluded_concrete_tests.txt

e8d44c3

Hmm-1224 requested a review from rkazants March 28, 2025 06:19

Update excluded_concrete_tests.txt

5f0c1cb

Hmm-1224 added 4 commits March 28, 2025 12:16

Update excluded_concrete_tests.txt

0924558

Update excluded_concrete_tests.txt

ab431a8

Update excluded_concrete_tests.txt

47a83db

Update excluded_concrete_tests.txt

d9a6d8a

rkazants reviewed Apr 3, 2025

View reviewed changes

Hmm-1224 and others added 6 commits April 3, 2025 19:58

Update numpy.py

bb6e9ec

Update numpy.py

b96856b

file reformatted

dffb76a

Update numpy.py

93bb579

Update numpy.py

3b5b5a6

numpy.py formatted

41195ab

rkazants reviewed Apr 3, 2025

View reviewed changes

Hmm-1224 requested a review from rkazants April 3, 2025 19:20

Hmm-1224 and others added 2 commits April 3, 2025 19:39

removed virtual env folder

0d8d1b2

Update numpy.py

617f984

rkazants reviewed Apr 4, 2025

View reviewed changes

Hmm-1224 added 2 commits April 4, 2025 15:06

Update numpy.py

23370ed

Update numpy.py

46e51b6

Hmm-1224 requested a review from rkazants April 4, 2025 09:46

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add OpenVINO backend support for argmin and argmax #21060

Add OpenVINO backend support for argmin and argmax #21060

Hmm-1224 commented Mar 18, 2025

codecov-commenter commented Mar 18, 2025 •

edited

Loading

Hmm-1224 commented Mar 22, 2025

Hmm-1224 commented Mar 24, 2025 •

edited

Loading

rkazants Mar 27, 2025

rkazants Mar 27, 2025

rkazants Mar 27, 2025

rkazants Mar 27, 2025

Hmm-1224 Mar 27, 2025

Hmm-1224 Mar 27, 2025 •

edited

Loading

rkazants Mar 28, 2025

rkazants Mar 28, 2025

Hmm-1224 Mar 28, 2025 •

edited

Loading

Hmm-1224 commented Mar 29, 2025

Hmm-1224 commented Mar 29, 2025 •

edited

Loading

Hmm-1224 commented Apr 1, 2025 •

edited

Loading

rkazants Apr 3, 2025

rkazants Apr 3, 2025

rkazants Apr 3, 2025

Hmm-1224 Apr 3, 2025 •

edited

Loading

Hmm-1224 Apr 3, 2025 •

edited

Loading

Hmm-1224 Apr 4, 2025

rkazants Apr 4, 2025

rkazants Apr 4, 2025

Hmm-1224 Apr 4, 2025

Hmm-1224 Apr 4, 2025

Hmm-1224 Apr 4, 2025

Hmm-1224 Apr 4, 2025

Hmm-1224 commented Apr 5, 2025

		NumpyOneInputOpsCorrectnessTest::test_exp
		NumpyOneInputOpsCorrectnessTest::test_expand_dims

		NumpyOneInputOpsCorrectnessTest::test_log1p
		NumpyOneInputOpsCorrectnessTest::test_logaddexp

		if topk_indices_shape.rank.get_length() == rank - 1:
		topk_indices = ov_opset.unsqueeze(topk_indices, [axis]).output(0)

Add OpenVINO backend support for argmin and argmax #21060

Are you sure you want to change the base?

Add OpenVINO backend support for argmin and argmax #21060

Conversation

Hmm-1224 commented Mar 18, 2025

codecov-commenter commented Mar 18, 2025 • edited Loading

Codecov Report

Hmm-1224 commented Mar 22, 2025

Hmm-1224 commented Mar 24, 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

Choose a reason for hiding this comment

Hmm-1224 Mar 27, 2025 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Hmm-1224 Mar 28, 2025 • edited Loading

Choose a reason for hiding this comment

Hmm-1224 commented Mar 29, 2025

Hmm-1224 commented Mar 29, 2025 • edited Loading

Hmm-1224 commented Apr 1, 2025 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Hmm-1224 Apr 3, 2025 • edited Loading

Choose a reason for hiding this comment

Hmm-1224 Apr 3, 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

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

Hmm-1224 commented Apr 5, 2025

codecov-commenter commented Mar 18, 2025 •

edited

Loading

Hmm-1224 commented Mar 24, 2025 •

edited

Loading

Hmm-1224 Mar 27, 2025 •

edited

Loading

Hmm-1224 Mar 28, 2025 •

edited

Loading

Hmm-1224 commented Mar 29, 2025 •

edited

Loading

Hmm-1224 commented Apr 1, 2025 •

edited

Loading

Hmm-1224 Apr 3, 2025 •

edited

Loading

Hmm-1224 Apr 3, 2025 •

edited

Loading