[OpenVINO Backend] support ops.slice_update #21362

Mohamed-Ashraf273 · 2025-06-06T23:56:27Z

Hi @rkazants ,
I've supported ops,silce_update for my GSoC project, but I can't enable tests for it until getting __getitem__ merge.

It could be implemented more easily if we use opset15 instead of opset14.

codecov-commenter · 2025-06-07T00:02:42Z

Codecov Report

Attention: Patch coverage is 92.59259% with 4 lines in your changes missing coverage. Please review.

Project coverage is 82.75%. Comparing base (d16424e) to head (bc166fc).
Report is 15 commits behind head on master.

Files with missing lines	Patch %	Lines
keras/src/backend/openvino/core.py	92.59%	1 Missing and 3 partials ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master   #21362      +/-   ##
==========================================
+ Coverage   82.73%   82.75%   +0.01%     
==========================================
  Files         565      565              
  Lines       55215    55354     +139     
  Branches     8606     8631      +25     
==========================================
+ Hits        45682    45807     +125     
- Misses       7424     7440      +16     
+ Partials     2109     2107       -2

Flag	Coverage Δ
keras	`82.56% <92.59%> (+0.02%)`	⬆️
keras-jax	`63.27% <0.00%> (-0.07%)`	⬇️
keras-numpy	`58.48% <0.00%> (-0.05%)`	⬇️
keras-openvino	`33.84% <92.59%> (+0.05%)`	⬆️
keras-tensorflow	`63.70% <0.00%> (-0.05%)`	⬇️
keras-torch	`63.31% <0.00%> (-0.07%)`	⬇️

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.

rkazants · 2025-06-19T18:43:36Z

let us wait until slice_update tests are switched on

Mohamed-Ashraf273 · 2025-06-20T19:16:56Z

@rkazants
I've switched on tests for slice_update.

rkazants · 2025-06-24T07:02:44Z

keras/src/backend/openvino/core.py

@@ -810,9 +810,83 @@ def prepare_slice_index(val):


 def slice_update(inputs, start_indices, updates):


please add comments for each block

Mohamed-Ashraf273 · 2025-06-24T12:59:24Z

@rkazants
I can't change it to opest15 because I'm facing a lot of segfaults when compiling models. I tried the version 2024.5.0 as mentioned here: https://docs.openvino.ai/2025/documentation/openvino-ir-format/operation-sets/available-opsets.html, but it didn't work either.
So I added comments.

divyashreepathihalli · 2025-07-11T00:55:43Z

/gemini review

gemini-code-assist

Code Review

This pull request implements the slice_update operation for the OpenVINO backend. The implementation generates indices for all elements in the updates tensor and uses scatter_nd_update to perform the update. The changes also enable the corresponding tests for this operation. The implementation assumes a static rank for the input tensors, and the index generation function could be refactored for better readability.

gemini-code-assist · 2025-07-11T00:57:02Z

keras/src/backend/openvino/core.py

+    updates_shape = ov_opset.shape_of(updates_tensor, Type.i32).output(0)
+    rank = updates_tensor.get_partial_shape().rank.get_length()
+    num_updates = ov_opset.reduce_prod(
+        updates_shape, zero_tensor, keep_dims=False
+    ).output(0)


The implementation assumes that the rank of the updates tensor is statically known. The call to .get_length() on this line will raise an exception if the rank is dynamic. Keras layers and backends should ideally support dynamic shapes, including dynamic ranks. The logic within create_meshgrid_indices and process_index also relies on a static rank. Please consider how to support dynamic ranks, which might involve using OpenVINO's control flow operations (like ov.opset.loop) instead of a Python for loop over the rank. If supporting dynamic rank is out of scope for this PR, this limitation should be documented in a docstring for the function.

gemini-code-assist · 2025-07-11T00:57:02Z

keras/src/backend/openvino/core.py

+    def create_meshgrid_indices():
+        all_indices = []
+
+        for dim_idx in range(rank):
+            dim_size = ov_opset.gather(
+                updates_shape,
+                ov_opset.constant([dim_idx], Type.i32),
+                zero_scalar,
+            ).output(0)
+            dim_size_scalar = ov_opset.squeeze(dim_size, zero_tensor).output(0)
+
+            # Create range for this dimension
+            dim_range = ov_opset.range(
+                zero_scalar, dim_size_scalar, one_scalar, output_type=Type.i32
+            ).output(0)
+
+            # Calculate factors for meshgrid
+            if dim_idx < rank - 1:
+                # Product of dimensions after current
+                remaining_shape = ov_opset.slice(
+                    updates_shape,
+                    ov_opset.constant([dim_idx + 1], Type.i32),
+                    ov_opset.constant([rank], Type.i32),
+                    one_tensor,
+                    zero_tensor,
+                ).output(0)
+                repeat_factor = ov_opset.reduce_prod(
+                    remaining_shape, zero_tensor, keep_dims=False
+                ).output(0)
+            else:
+                repeat_factor = one_scalar
+
+            if dim_idx > 0:
+                # Product of dimensions before current
+                preceding_shape = ov_opset.slice(
+                    updates_shape,
+                    zero_tensor,
+                    ov_opset.constant([dim_idx], Type.i32),
+                    one_tensor,
+                    zero_tensor,
+                ).output(0)
+                tile_factor = ov_opset.reduce_prod(
+                    preceding_shape, zero_tensor, keep_dims=False
+                ).output(0)
+            else:
+                tile_factor = one_scalar
+
+            # Apply meshgrid transformations
+            # Step 1: Repeat elements
+            dim_expanded = ov_opset.unsqueeze(dim_range, one_scalar).output(0)
+            repeat_shape = ov_opset.concat(
+                [
+                    one_tensor,
+                    ov_opset.unsqueeze(repeat_factor, zero_scalar).output(0),
+                ],
+                axis=0,
+            ).output(0)
+            dim_repeated = ov_opset.tile(dim_expanded, repeat_shape).output(0)
+
+            # Step 2: Flatten
+            flat_size = ov_opset.multiply(
+                dim_size_scalar, repeat_factor
+            ).output(0)
+            dim_flat = ov_opset.reshape(
+                dim_repeated,
+                ov_opset.unsqueeze(flat_size, zero_scalar).output(0),
+                special_zero=False,
+            ).output(0)
+
+            # Step 3: Tile entire sequence
+            final_indices = ov_opset.tile(
+                dim_flat, ov_opset.unsqueeze(tile_factor, zero_scalar).output(0)
+            ).output(0)
+
+            all_indices.append(final_indices)
+
+        return all_indices


The create_meshgrid_indices function is complex, which can make it difficult to read and maintain. Consider refactoring this function by extracting parts of the logic into smaller helper functions. For example, the logic for calculating repeat_factor and tile_factor, and the meshgrid transformation steps (repeating, flattening, tiling) could each be in their own helper functions. This would make the main loop in create_meshgrid_indices more declarative and easier to understand.

google-ml-butler bot added the size:M label Jun 6, 2025

google-ml-butler bot assigned gbaned Jun 6, 2025

Mohamed-Ashraf273 force-pushed the support_slice_update branch 5 times, most recently from c968548 to 004626b Compare June 7, 2025 01:37

[OpenVINO Backend] support slice_update

aef0743

Mohamed-Ashraf273 force-pushed the support_slice_update branch from 004626b to aef0743 Compare June 7, 2025 01:38

gbaned added this to PR Queue Jun 11, 2025

github-project-automation bot moved this to Assigned Reviewer in PR Queue Jun 11, 2025

gbaned requested a review from mattdangerw June 11, 2025 02:26

google-ml-butler bot added the awaiting review label Jun 11, 2025

Mohamed-Ashraf273 and others added 2 commits June 20, 2025 22:10

Merge branch 'keras-team:master' into support_slice_update

0f78f8e

enable tests for slice_update

a285aa7

rkazants reviewed Jun 24, 2025

View reviewed changes

Mohamed-Ashraf273 force-pushed the support_slice_update branch 2 times, most recently from 799ecfa to 7ef40d5 Compare June 24, 2025 12:58

Mohamed-Ashraf273 force-pushed the support_slice_update branch 2 times, most recently from 307110e to a4188f1 Compare June 24, 2025 19:29

[OpenVINO backend] support slice_update

62d02fb

Mohamed-Ashraf273 force-pushed the support_slice_update branch from a4188f1 to 62d02fb Compare July 4, 2025 18:34

remove broadcast operations

b7c27dd

Mohamed-Ashraf273 force-pushed the support_slice_update branch from 9a2a8d8 to b7c27dd Compare July 4, 2025 20:34

add some comments for clarity

bc166fc

gemini-code-assist bot reviewed Jul 11, 2025

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[OpenVINO Backend] support ops.slice_update #21362

[OpenVINO Backend] support ops.slice_update #21362

Uh oh!

Mohamed-Ashraf273 commented Jun 6, 2025 •

edited

Loading

Uh oh!

codecov-commenter commented Jun 7, 2025 •

edited

Loading

Uh oh!

rkazants commented Jun 19, 2025

Uh oh!

Mohamed-Ashraf273 commented Jun 20, 2025

Uh oh!

rkazants Jun 24, 2025

Uh oh!

Mohamed-Ashraf273 commented Jun 24, 2025 •

edited

Loading

Uh oh!

divyashreepathihalli commented Jul 11, 2025

Uh oh!

gemini-code-assist bot left a comment

Uh oh!

gemini-code-assist bot Jul 11, 2025

Uh oh!

gemini-code-assist bot Jul 11, 2025

Uh oh!

Uh oh!

		@@ -810,9 +810,83 @@ def prepare_slice_index(val):


		def slice_update(inputs, start_indices, updates):

[OpenVINO Backend] support ops.slice_update #21362

Are you sure you want to change the base?

[OpenVINO Backend] support ops.slice_update #21362

Uh oh!

Conversation

Mohamed-Ashraf273 commented Jun 6, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

codecov-commenter commented Jun 7, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

rkazants commented Jun 19, 2025

Uh oh!

Mohamed-Ashraf273 commented Jun 20, 2025

Uh oh!

rkazants Jun 24, 2025

Choose a reason for hiding this comment

Uh oh!

Mohamed-Ashraf273 commented Jun 24, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

divyashreepathihalli commented Jul 11, 2025

Uh oh!

gemini-code-assist bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

gemini-code-assist bot Jul 11, 2025

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Jul 11, 2025

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Mohamed-Ashraf273 commented Jun 6, 2025 •

edited

Loading

codecov-commenter commented Jun 7, 2025 •

edited

Loading

Mohamed-Ashraf273 commented Jun 24, 2025 •

edited

Loading