Validate input channels match kernel in depthwise/separable conv by rstar327 · Pull Request #22687 · keras-team/keras

rstar327 · 2026-04-17T01:53:27Z

Description

keras.ops.separable_conv and keras.ops.depthwise_conv previously gave confusing error messages when the input's channel dimension didn't match the kernel's expected input channels (e.g., passing channels-last shaped data with data_format="channels_first"):

Eager: TF backend error: input depth must be evenly divisible by filter depth: 5 vs 3
Symbolic: Kernel shape must have the same length as input

Neither message clearly identifies the root cause: a mismatch between the input's channel count and the kernel's expected input channels.

This PR adds explicit channel validation in both the symbolic path (compute_output_spec) and the eager path (before dispatching to the backend), producing a clear error:

ValueError: The number of input channels must match the kernel's input channels.
Received: input channels=5, kernel input channels=3, data_format='channels_first'.

Changes:

Added _check_input_channels_match_kernel helper in keras/src/ops/nn.py
Added channel validation in DepthwiseConv.compute_output_spec (symbolic path)
Added channel validation in depthwise_conv() and separable_conv() (eager path)

Contributor Agreement

I am a human, and not a bot.
I will be responsible for responding to review comments in a timely manner.
I will work with the maintainers to push this PR forward until submission.

…eparable_conv

gemini-code-assist

Code Review

This pull request introduces a _check_input_channels_match_kernel helper function to validate input channels for depthwise and separable convolutions. The review feedback suggests enhancing this helper to support symbolic tensors with None dimensions, which would allow for its reuse in DepthwiseConv.compute_output_spec and eliminate redundant validation logic.

gemini-code-assist · 2026-04-17T01:56:04Z

+def _check_input_channels_match_kernel(inputs, kernel, data_format):
+    input_channels = (
+        inputs.shape[-1] if data_format == "channels_last" else inputs.shape[1]
+    )
+    kernel_input_channels = kernel.shape[-2]
+    if input_channels != kernel_input_channels:
+        raise ValueError(
+            "The number of input channels must match the kernel's "
+            f"input channels. Received: input channels="
+            f"{input_channels}, kernel input channels="
+            f"{kernel_input_channels}, data_format='{data_format}'."
+        )


To improve reusability and reduce code duplication, you can modify this helper function to also handle symbolic tensors, where shape dimensions can be None. This would allow you to use it in DepthwiseConv.compute_output_spec as well.

By adding is not None checks, this function becomes safe for both eager and symbolic execution paths.

def _check_input_channels_match_kernel(inputs, kernel, data_format): input_channels = ( inputs.shape[-1] if data_format == "channels_last" else inputs.shape[1] ) kernel_input_channels = kernel.shape[-2] if ( input_channels is not None and kernel_input_channels is not None and input_channels != kernel_input_channels ): raise ValueError( "The number of input channels must match the kernel's " f"input channels. Received: input channels=" f"{input_channels}, kernel input channels=" f"{kernel_input_channels}, data_format='{data_format}'." )

gemini-code-assist · 2026-04-17T01:56:05Z

+        input_channels = (
+            inputs.shape[-1]
+            if self.data_format == "channels_last"
+            else inputs.shape[1]
+        )
+        kernel_input_channels = kernel.shape[-2]
+        if (
+            input_channels is not None
+            and kernel_input_channels is not None
+            and input_channels != kernel_input_channels
+        ):
+            raise ValueError(
+                "The number of input channels must match the kernel's "
+                f"input channels. Received: input channels="
+                f"{input_channels}, kernel input channels="
+                f"{kernel_input_channels}, data_format="
+                f"'{self.data_format}'."
+            )


Following the suggested change to _check_input_channels_match_kernel, you can replace this entire validation block with a single call to the updated helper function. This simplifies the code and removes duplication.

_check_input_channels_match_kernel(inputs, kernel, self.data_format)

codecov-commenter · 2026-04-17T02:07:23Z

Codecov Report

❌ Patch coverage is 94.73684% with 2 lines in your changes missing coverage. Please review.
✅ Project coverage is 83.03%. Comparing base (e94cb07) to head (2fae3a7).
⚠️ Report is 24 commits behind head on master.

Files with missing lines	Patch %	Lines
keras/src/backend/common/backend_utils.py	60.00%	1 Missing and 1 partial ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master   #22687      +/-   ##
==========================================
+ Coverage   82.95%   83.03%   +0.08%     
==========================================
  Files         596      596              
  Lines       69252    69858     +606     
  Branches    10814    10884      +70     
==========================================
+ Hits        57451    58010     +559     
- Misses       8973     8999      +26     
- Partials     2828     2849      +21

Flag	Coverage Δ
keras	`82.86% <94.73%> (+0.08%)`	⬆️
keras-jax	`58.35% <42.10%> (-0.37%)`	⬇️
keras-numpy	`54.23% <44.73%> (-0.34%)`	⬇️
keras-openvino	`59.77% <23.68%> (+0.34%)`	⬆️
keras-tensorflow	`59.90% <34.21%> (-0.38%)`	⬇️
keras-torch	`58.69% <44.73%> (-0.37%)`	⬇️

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.

hertschuh

Thanks for looking into this!

hertschuh · 2026-04-23T01:31:37Z

        return DepthwiseConv(
            strides, padding, data_format, dilation_rate
        ).symbolic_call(inputs, kernel)
+    _check_input_channels_match_kernel(inputs, kernel, data_format)


The reason why we don't do any validation involving the inputs in the backend-independent code is that there are scenarios where this won't work:

inputs can be a nested Python array and won't have a shape

inputs can be a tracer / placeholder tensor and have dynamic dimensions

I would check which backend don't handle this correctly, then add this check for the backends that need it after the convert_to_tensor(inputs).

…per review feedback

rstar327 · 2026-04-23T19:15:49Z

Thanks for the review! I've pushed a refactor addressing the feedback:

Removed the check from keras/src/ops/nn.py (both the standalone helper and the DepthwiseConv.compute_output_spec check), so the backend-independent path no longer touches inputs.shape.
Added a single shared helper check_depthwise_conv_input_channels in keras/src/backend/common/backend_utils.py.
Called it from each of the four backends (TF / JAX / Torch / NumPy) after convert_to_tensor(inputs), so we only run it against a concrete tensor. The JAX / Torch / NumPy separable_conv already delegate to their own depthwise_conv, so the check fires once per call.

Verified all four backends now raise the same clear ValueError on the #22516 repro, and the existing nn_test.py + separable_conv_test.py + depthwise_conv_test.py suites pass on all four backends. Let me know if you'd like any adjustments.

hertschuh · 2026-04-24T00:12:15Z

+    if (
+        input_channels is None
+        or kernel_input_channels is None
+        or input_channels == kernel_input_channels
+    ):
+        return
+    raise ValueError(


Write it like this:

if ( isinstance(input_channels, int) and isinstance(kernel_input_channels, int) and input_channels != kernel_input_channels ): raise ValueError(...)

The reason is that dynamic dimensions can come in different forms, not just None, during tracing.

hertschuh · 2026-04-24T00:15:17Z

+    inputs = convert_to_tensor(inputs)
+    kernel = convert_to_tensor(kernel)
+    check_depthwise_conv_input_channels(inputs, kernel, data_format)
    num_spatial_dims = len(inputs.shape) - 2
    if num_spatial_dims > 2:
        raise ValueError(
            "`inputs` rank must be 3 (1D conv) or 4 (2D conv). Received: "
            f"{inputs.ndim}."
        )


Can you move the check_depthwise_conv_input_channels after the check on num_spatial_dims.

The reason is that if the number of dimensions is incorrect, the math in check_depthwise_conv_input_channels will be incorrect and meaningless and may raise a confusing error.

hertschuh · 2026-04-24T00:15:59Z

+    inputs = convert_to_tensor(inputs)
+    kernel = convert_to_tensor(kernel)
+    check_depthwise_conv_input_channels(inputs, kernel, data_format)


Any reason why you didn't do separable_conv?

hertschuh · 2026-04-24T00:16:16Z

+    inputs = convert_to_tensor(inputs)
+    depthwise_kernel = convert_to_tensor(depthwise_kernel)
+    pointwise_kernel = convert_to_tensor(pointwise_kernel)
+    check_depthwise_conv_input_channels(inputs, depthwise_kernel, data_format)
    num_spatial_dims = len(inputs.shape) - 2
    if num_spatial_dims > 2:
        raise ValueError(
            "`num_spatial_dims` must be 1 or 2. Received: "
            f"num_spatial_dims={num_spatial_dims}."
        )


Can you move the check_depthwise_conv_input_channels after the check on num_spatial_dims.

The reason is that if the number of dimensions is incorrect, the math in check_depthwise_conv_input_channels will be incorrect and meaningless and may raise a confusing error.

hertschuh · 2026-04-24T00:17:13Z

+    inputs = convert_to_tensor(inputs)
+    kernel = convert_to_tensor(kernel)
+    check_depthwise_conv_input_channels(inputs, kernel, data_format)


Any reason why you didn't do separable_conv?

hertschuh · 2026-04-24T00:17:33Z

+    data_format = backend.standardize_data_format(data_format)
+    inputs = convert_to_tensor(inputs)
    kernel = convert_to_tensor(kernel)
+    check_depthwise_conv_input_channels(inputs, kernel, data_format)


Any reason why you didn't do separable_conv?

…isinstance check

rstar327 · 2026-04-24T13:26:15Z

Thanks for the review! Pushed 2fae3a72d addressing all four points:

check_depthwise_conv_input_channels now uses isinstance(..., int) instead of an explicit None check, so dynamic/tracing dims in other forms fall through to the no-op path instead of accidentally passing the validation.
TF depthwise_conv / separable_conv: moved the channel check after the num_spatial_dims > 2 guard so we don't run wrong-axis math on a misranked input.
JAX / Torch / NumPy separable_conv: added the check explicitly. My original reasoning was that these three backends' separable_conv delegates to their own depthwise_conv (which already has the check), but you're right that making it explicit is cleaner and safer against future refactors.

Re-verified the #22516 repro raises the clear error on all four backends for both depthwise_conv and separable_conv, and the existing nn_test.py depthwise/separable tests (24 each) pass on TF / JAX / Torch / NumPy.

hertschuh · 2026-04-24T18:06:02Z

JAX / Torch / NumPy separable_conv: added the check explicitly. My original reasoning was that these three backends' separable_conv delegates to their own depthwise_conv (which already has the check), but you're right that making it explicit is cleaner and safer against future refactors.

Oh, I did not realize that.

hertschuh

Thanks for your contribution!

fix(nn): validate input channels match kernel in depthwise_conv and s…

709fab6

…eparable_conv

google-ml-butler Bot added the size:S label Apr 17, 2026

google-ml-butler Bot assigned gbaned Apr 17, 2026

gemini-code-assist Bot reviewed Apr 17, 2026

View reviewed changes

keerthanakadiri added the stat:awaiting keras-eng Awaiting response from Keras engineer label Apr 17, 2026

rstar327 added 3 commits April 17, 2026 14:41

chore: trigger CI re-run

5ba2dc3

fix(test): use PartitionSpec in distribution_lib_test assertions

58f418f

chore: merge upstream/master into fix-separable-conv-channels-first

3b676f4

hertschuh requested changes Apr 23, 2026

View reviewed changes

hertschuh added stat:awaiting response from contributor and removed stat:awaiting keras-eng Awaiting response from Keras engineer labels Apr 23, 2026

refactor(nn): move depthwise conv channel check from ops to backends …

c86cd2f

…per review feedback

google-ml-butler Bot removed the stat:awaiting response from contributor label Apr 23, 2026

hertschuh reviewed Apr 24, 2026

View reviewed changes

hertschuh added the stat:awaiting response from contributor label Apr 24, 2026

refactor(nn): address review feedback on channel check placement and …

2fae3a7

…isinstance check

google-ml-butler Bot removed the stat:awaiting response from contributor label Apr 24, 2026

hertschuh approved these changes Apr 24, 2026

View reviewed changes

google-ml-butler Bot added kokoro:force-run ready to pull Ready to be merged into the codebase labels Apr 24, 2026

hertschuh merged commit 634e895 into keras-team:master Apr 24, 2026
11 checks passed

Conversation

rstar327 commented Apr 17, 2026

Description

Contributor Agreement

Uh oh!

gemini-code-assist Bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

gemini-code-assist Bot Apr 17, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Apr 17, 2026

Choose a reason for hiding this comment

Uh oh!

codecov-commenter commented Apr 17, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

hertschuh left a comment

Choose a reason for hiding this comment

Uh oh!

hertschuh Apr 23, 2026

Choose a reason for hiding this comment

Uh oh!

rstar327 commented Apr 23, 2026

Uh oh!

hertschuh Apr 24, 2026

Choose a reason for hiding this comment

Uh oh!

hertschuh Apr 24, 2026

Choose a reason for hiding this comment

Uh oh!

hertschuh Apr 24, 2026

Choose a reason for hiding this comment

Uh oh!

hertschuh Apr 24, 2026

Choose a reason for hiding this comment

Uh oh!

hertschuh Apr 24, 2026

Choose a reason for hiding this comment

Uh oh!

hertschuh Apr 24, 2026

Choose a reason for hiding this comment

Uh oh!

rstar327 commented Apr 24, 2026

Uh oh!

hertschuh commented Apr 24, 2026

Uh oh!

hertschuh left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

codecov-commenter commented Apr 17, 2026 •

edited

Loading