perf: lazy init for VGG + CapsuleNetwork to unblock NeuralNetworks ModelFamily shard

[ooples]

Summary

Second PR in the #1136 series, targeting the cancelled Tests (net10.0) - ModelFamily - NeuralNetworks shard on CI. Baseline run 24398739627 cancelled that shard at 45 min with 14 VGG timeouts, 8 Capsule timeouts, 110 FastText OOMs, and several others stacked up in a shared xunit process.

What's in this PR

Applies the same InitializationStrategies<T>.Lazy pattern introduced in PR #1137 for diffusion noise predictors, but to the two biggest per-instance OOM offenders in the NeuralNetworks shard.

VGGNetwork (14 timeouts + 3 OOMs on baseline): VGG16BN at 224×224×3 with 1000 classes allocates ~1.1 GB of weights per instance. The three fully connected classifier layers alone are ~1 GB (FC1 at 25088×4096 = ~822 MB). Each ModelFamilyTests.NeuralNetworks.VGGNetworkTests.* was eagerly renting those tensors in new DenseLayer<T>(...).

CapsuleNetwork (8 timeouts + 1 OOM): Same eager-rent pattern on the initial 9×9 convolutional layer at production spatial sizes.

Changes LayerHelper<T>.CreateDefaultVGGLayers and LayerHelper<T>.CreateDefaultCapsuleNetworkLayers to thread InitializationStrategies<T>.Lazy into the relevant ConvolutionalLayer and DenseLayer constructions. Weight tensors stay at shape [0,…,0] until the first Forward() call.

Intentionally out of scope

• FastText's 110 OOMs were caused by per-test accumulation, not a single large alloc. PR perf: lazy init in diffusion noise predictors + GC between tests #1137's IAsyncLifetime.DisposeAsync hook on NeuralNetworkModelTestBase forces a full GC between test methods, which covers this.
• VoxelCNN's 3 OOMs go through Conv3DLayer, which doesn't have a lazy-init branch yet. Deferred to the follow-up MHA/LayerNorm/Conv3D lazy-init PR tracked in perf: fix 5 cancelled CI jobs — Diffusion models OOM/timeout from eager weight allocation #1136.
• NEAT's 12 failures are evolutionary-topology-driven and don't have a single eager-alloc hot path to fix the same way.

Verification

• Local build: dotnet build src/AiDotNet.csproj --framework net10.0 -c Release — 0 errors.
• CI target: Tests (net10.0) - ModelFamily - NeuralNetworks should transition CANCELLED → SUCCESS or FAILURE. Pre-existing test failures are acceptable at this stage — the goal is unblocking the CI signal.

Depends on: none (branched off master). Does not conflict with PR #1137 (different files).

Refs #1136

Summary by CodeRabbit

• Refactor

  • Layer initialization now defers heavy weight/tensor allocations until first use, reducing upfront memory and improving startup efficiency.
  • Parameter counting is initialization-aware to reflect deferred allocation.

• Bug Fixes

  • Serialization/deserialization and parameter access now reliably preserve or trigger initialization to avoid stale or mismatched parameters and gradients.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

2 Skipped Deployments

Project	Deployment	Actions	Updated (UTC)
aidotnet_website	Ignored	Preview	Apr 14, 2026 7:19pm
aidotnet-playground-api	Ignored	Preview	Apr 14, 2026 7:19pm

[coderabbitai]

Note

Reviews paused

It looks like this branch is under active development. To avoid overwhelming you with review comments due to an influx of new commits, CodeRabbit has automatically paused this review. You can configure this behavior by changing the reviews.auto_review.auto_pause_after_reviewed_commits setting.

Use the following commands to manage reviews:

• @coderabbitai resume to resume automatic reviews.
• @coderabbitai review to trigger a single review.

Use the checkboxes below for quick actions:

• ▶️ Resume reviews
• 🔍 Trigger review

Walkthrough

Layer construction now uses lazy initialization for selected convolutional and classifier dense layers (allocation deferred until first Forward()). ConvolutionalLayer parameter access, gradients, SetParameters and Deserialize were updated to respect and trigger deferred initialization when needed.

Changes

Cohort / File(s)	Summary
Layer helper updates src/Helpers/LayerHelper.cs	Added using AiDotNet.Initialization;. VGG 3×3 conv blocks, initial Capsule conv, and classifier DenseLayer<T>s now constructed with initializationStrategy: InitializationStrategies<T>.Lazy, deferring kernel/bias allocation until first Forward().
Convolutional layer initialization & parameter access src/NeuralNetworks/Layers/ConvolutionalLayer.cs	Deserialize(BinaryReader) re-registers trainable tensors and sets _isInitialized = true. ParameterCount computes expected size when uninitialized; GetParameters()/SetParameters(...) call EnsureInitialized(); GetParameterGradients() returns a zero-vector sized by ParameterCount when gradient tensors are absent and only forces init when gradients exist. No public signatures changed.

Sequence Diagram(s)

sequenceDiagram
    participant Builder as LayerHelper/ModelBuilder
    participant Init as InitializationStrategies<T>
    participant Layer as ConvolutionalLayer<T>
    participant Runner as Runtime/Forward

    Builder->>Layer: construct(..., initializationStrategy: Init.Lazy)
    Note right of Layer: kernels & biases NOT allocated
    Runner->>Layer: Forward(input)
    Layer->>Layer: EnsureInitialized()
    Layer->>Init: allocate kernels & biases
    Init-->>Layer: return tensors
    Layer->>Runner: perform convolution -> output

Loading

Estimated code review effort

🎯 3 (Moderate) | ⏱️ ~20 minutes

Possibly related issues

• perf: fix 5 cancelled CI jobs — Diffusion models OOM/timeout from eager weight allocation #1136: Direct match — implements lazy/deferred initialization for layer weights and updates parameter/serialization behavior to avoid eager allocation.

No stubs, TODOs, or non-production placeholders were detected in the changed files; nothing flagged as BLOCKING.

Poem

Kernels snooze in quiet rows, until Forward calls them near,
Lazy threads yawn, then softly grow — weights bloom when inputs appear.
Small change, big relief, memory sits light — models cheer. 🚀

🚥 Pre-merge checks | ✅ 3

✅ Passed checks (3 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title accurately captures the main change: applying lazy initialization to VGG and CapsuleNetwork layers to resolve performance and OOM issues in CI testing.
Docstring Coverage	✅ Passed	Docstring coverage is 100.00% which is sufficient. The required threshold is 80.00%.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

📝 Generate docstrings

• ✅ Generated successfully - (🔄 Check to regenerate)
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch perf/nn-lazy-init-oom

---

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[Copilot]

Pull request overview

This PR applies the repository’s lazy-weight-initialization strategy to the largest allocation hot spots in the NeuralNetworks ModelFamily shard—VGG and CapsuleNetwork—so model construction doesn’t eagerly rent huge weight tensors and overwhelm CI memory/time budgets.

Changes:

• Thread InitializationStrategies<T>.Lazy into VGG’s ConvolutionalLayer<T> and classifier DenseLayer<T> creation in LayerHelper<T>.CreateDefaultVGGLayers.
• Thread InitializationStrategies<T>.Lazy into CapsuleNetwork’s initial ConvolutionalLayer<T> creation in LayerHelper<T>.CreateDefaultCapsuleNetworkLayers.
• Add the required AiDotNet.Initialization import in LayerHelper.cs.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[coderabbitai]

Actionable comments posted: 1

🤖 Prompt for all review comments with AI agents

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@src/NeuralNetworks/Layers/ConvolutionalLayer.cs`:
- Line 1325: The calls to EnsureInitialized() in the layer (e.g., at
EnsureInitialized() invocations) can re-randomize weights for layers that were
deserialized because Deserialize() doesn't mark the layer as initialized; update
the deserialization path to set the layer's initialized state instead of letting
EnsureInitialized() run later — either set the existing initialization flag
(e.g., _initialized or similar) to true inside Deserialize() or add a distinct
_isDeserialized flag and make EnsureInitialized() respect it (skip random init
if deserialized). Locate Deserialize(), EnsureInitialized(), and any
lazy-initialization fields in ConvolutionalLayer and ensure Deserialize() marks
the layer as initialized to prevent overwriting loaded parameters.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Pro

Run ID: 38a4b580-af56-4e96-896a-c231c8f6c8b1

📥 Commits

Reviewing files that changed from the base of the PR and between 41342c5 and dbbac2e.

📒 Files selected for processing (1)

• src/NeuralNetworks/Layers/ConvolutionalLayer.cs

[coderabbitai]

Caution

Some comments are outside the diff and can’t be posted inline due to platform limitations.

⚠️ Outside diff range comments (1)

src/NeuralNetworks/Layers/ConvolutionalLayer.cs (1)

1349-1354: ⚠️ Potential issue | 🟠 Major

Don't allocate full weights just to return zero gradients.

Line 1349 eagerly initializes the layer even when no backward pass has run and this method immediately returns new Vector<T>(ParameterCount). That defeats lazy init on any code path that probes gradients before training and can bring back the same cold-start pressure this PR is trying to remove.

Proposed fix

 public override Vector<T> GetParameterGradients()
 {
-    EnsureInitialized();
     // If gradients haven't been computed yet, return zero gradients
-    if (_kernelsGradient == null || _biasesGradient == null)
+    if (!_isInitialized || _kernelsGradient == null || _biasesGradient == null)
     {
         return new Vector<T>(ParameterCount);
     }

     // Bulk copy from contiguous tensor storage — replaces 4-nested scalar loops

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@src/NeuralNetworks/Layers/ConvolutionalLayer.cs` around lines 1349 - 1354,
Don't call EnsureInitialized() before checking gradient fields; move the
null-check for _kernelsGradient and _biasesGradient to the top of the method and
return new Vector<T>(ParameterCount) immediately if either is null so you avoid
forcing layer initialization. Only call EnsureInitialized() when you actually
need to access or assemble gradients from the initialized fields (e.g., when
aggregating _kernelsGradient/_biasesGradient into the returned Vector),
preserving lazy init and avoiding the cold-start allocation.

♻️ Duplicate comments (1)

src/NeuralNetworks/Layers/ConvolutionalLayer.cs (1)

728-731: ⚠️ Potential issue | 🔴 Critical

Blocking: Deserialize() still leaves stale tensor state behind.

Setting _isInitialized = true fixes the re-randomization path, but Line 728 still swaps in new _kernels/_biases without clearing or rebuilding the state that was derived from the old tensors. On a reused instance, _biasReshaped4D, optimizer state, and persistent registrations can keep pointing at pre-deserialize tensors, so the loaded model can run or update with stale data.

Proposed fix

 public override void Deserialize(BinaryReader reader)
 {
     base.Deserialize(reader);
     InputDepth = reader.ReadInt32();
     OutputDepth = reader.ReadInt32();
     KernelSize = reader.ReadInt32();
     Stride = reader.ReadInt32();
     Padding = reader.ReadInt32();

+    Engine.InvalidatePersistentTensor(_kernels);
+    Engine.InvalidatePersistentTensor(_biases);
+
     // Deserialize _kernels — flat span iteration replaces 4-nested indexing loops
     _kernels = TensorAllocator.RentUninitialized<T>([OutputDepth, InputDepth, KernelSize, KernelSize]);
     var kernelSpan = _kernels.Data.Span;
     for (int i = 0; i < kernelSpan.Length; i++)
         kernelSpan[i] = NumOps.FromDouble(reader.ReadDouble());

     // Deserialize _biases — flat span iteration
     _biases = new Tensor<T>([OutputDepth]);
     var biasSpan = _biases.Data.Span;
     for (int i = 0; i < biasSpan.Length; i++)
         biasSpan[i] = NumOps.FromDouble(reader.ReadDouble());

     // Reinitialize _lastInput and _lastOutput
     _lastInput = new Tensor<T>([OutputDepth, InputDepth, KernelSize, KernelSize]);
     _lastOutput = new Tensor<T>([OutputDepth, InputDepth, KernelSize, KernelSize]);
+    RegisterTrainableParameter(_kernels, PersistentTensorRole.Weights);
+    RegisterTrainableParameter(_biases, PersistentTensorRole.Biases);
+    _biasReshaped4D = null;
+    _preAllocatedOutput = null;
+    _kernelsGradient = null;
+    _biasesGradient = null;
+    _kernelsVelocity = null;
+    _biasesVelocity = null;

     // Mark as initialized so EnsureInitialized() doesn't re-randomize the
     // just-deserialized weights on the next Forward/GetParameters call.
     _isInitialized = true;
 }

As per coding guidelines: “Production Readiness (CRITICAL)… Incomplete features: Half-implemented patterns where some code paths work but others silently do nothing.”

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@src/NeuralNetworks/Layers/ConvolutionalLayer.cs` around lines 728 - 731,
Deserialize currently replaces _kernels/_biases and sets _isInitialized but
leaves derived state (e.g., _biasReshaped4D), optimizer state entries, and any
persistent registrations pointing at old tensors, causing stale references;
update Deserialize (the method that assigns _kernels and _biases) to: clear or
rebuild derived fields like _biasReshaped4D, detach or reset optimizer state
entries associated with the previous tensors, and re-register persistent
resources so they point to the newly deserialized tensors (and ensure
EnsureInitialized logic is consistent with these rebuilt state fields).

🤖 Prompt for all review comments with AI agents

Verify each finding against the current code and only fix it if needed.

Outside diff comments:
In `@src/NeuralNetworks/Layers/ConvolutionalLayer.cs`:
- Around line 1349-1354: Don't call EnsureInitialized() before checking gradient
fields; move the null-check for _kernelsGradient and _biasesGradient to the top
of the method and return new Vector<T>(ParameterCount) immediately if either is
null so you avoid forcing layer initialization. Only call EnsureInitialized()
when you actually need to access or assemble gradients from the initialized
fields (e.g., when aggregating _kernelsGradient/_biasesGradient into the
returned Vector), preserving lazy init and avoiding the cold-start allocation.

---

Duplicate comments:
In `@src/NeuralNetworks/Layers/ConvolutionalLayer.cs`:
- Around line 728-731: Deserialize currently replaces _kernels/_biases and sets
_isInitialized but leaves derived state (e.g., _biasReshaped4D), optimizer state
entries, and any persistent registrations pointing at old tensors, causing stale
references; update Deserialize (the method that assigns _kernels and _biases)
to: clear or rebuild derived fields like _biasReshaped4D, detach or reset
optimizer state entries associated with the previous tensors, and re-register
persistent resources so they point to the newly deserialized tensors (and ensure
EnsureInitialized logic is consistent with these rebuilt state fields).

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Pro

Run ID: 69cdb026-6b5f-496a-be8f-bcf82e9a6b83

📥 Commits

Reviewing files that changed from the base of the PR and between dbbac2e and 93041b4.

📒 Files selected for processing (1)

• src/NeuralNetworks/Layers/ConvolutionalLayer.cs

[Copilot]

Pull request overview

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[coderabbitai]

♻️ Duplicate comments (2)

src/NeuralNetworks/Layers/ConvolutionalLayer.cs (2)

729-740: ⚠️ Potential issue | 🔴 Critical

Blocking: clear all caches/state that still reference the pre-deserialize tensors.

Deserialize() swaps in new _kernels and _biases, but _biasReshaped4D, _preAllocatedOutput, and the cached gradient/optimizer tensors still point at the old objects. On a reused instance, Forward() can keep adding the stale bias view, and UpdateParameters() can apply stale gradients/velocity to the newly loaded weights.

Suggested fix

         ClearRegisteredParameters();
         RegisterTrainableParameter(_kernels, PersistentTensorRole.Weights);
         RegisterTrainableParameter(_biases, PersistentTensorRole.Biases);
+        _biasReshaped4D = null;
+        _preAllocatedOutput = null;
+        _kernelsGradient = null;
+        _biasesGradient = null;
+        _kernelsVelocity = null;
+        _biasesVelocity = null;
 
         // Mark as initialized so EnsureInitialized() doesn't re-randomize the
         // just-deserialized weights on the next Forward/GetParameters call.
         _isInitialized = true;

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@src/NeuralNetworks/Layers/ConvolutionalLayer.cs` around lines 729 - 740,
Deserialize() replaces _kernels and _biases but leaves other fields and caches
pointing at the old tensors; clear or reset any cached state that references
pre-deserialize tensors (e.g., _biasReshaped4D, _preAllocatedOutput, any cached
gradient/velocity/optimizer tensors or views) immediately after swapping in the
new tensors and before calling
RegisterTrainableParameter/ClearRegisteredParameters, so Forward() and
UpdateParameters() cannot use stale references; also ensure these caches are
reallocated or nullified and that _isInitialized is set only after caches are
refreshed.

---

704-740: ⚠️ Potential issue | 🔴 Critical

Blocking: save/load is still broken for never-initialized lazy layers.

A lazy layer serialized before any Forward(), SetParameters(), or GetParameters() call still writes zero kernel/bias values from the [0,…,0] placeholders. This method always reads a full parameter payload, so loading that file will fail with EndOfStreamException.

Suggested fix

 public override void Serialize(BinaryWriter writer)
 {
+    EnsureInitialized();
     base.Serialize(writer);
     writer.Write(InputDepth);
     writer.Write(OutputDepth);

🤖 Prompt for all review comments with AI agents

Verify each finding against the current code and only fix it if needed.

Duplicate comments:
In `@src/NeuralNetworks/Layers/ConvolutionalLayer.cs`:
- Around line 729-740: Deserialize() replaces _kernels and _biases but leaves
other fields and caches pointing at the old tensors; clear or reset any cached
state that references pre-deserialize tensors (e.g., _biasReshaped4D,
_preAllocatedOutput, any cached gradient/velocity/optimizer tensors or views)
immediately after swapping in the new tensors and before calling
RegisterTrainableParameter/ClearRegisteredParameters, so Forward() and
UpdateParameters() cannot use stale references; also ensure these caches are
reallocated or nullified and that _isInitialized is set only after caches are
refreshed.

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Pro

Run ID: b775940e-c998-4154-bf18-fb6ba95efa39

📥 Commits

Reviewing files that changed from the base of the PR and between 93041b4 and 02b532d.

📒 Files selected for processing (1)

• src/NeuralNetworks/Layers/ConvolutionalLayer.cs

[vercel]

Deployment failed with the following error:

Resource is limited - try again in 24 hours (more than 100, code: "api-deployments-free-per-day").

Learn More: https://vercel.com/franklins-projects-02a0b5a0?upgradeToPro=build-rate-limit

[coderabbitai]

Caution

Some comments are outside the diff and can’t be posted inline due to platform limitations.

⚠️ Outside diff range comments (1)

src/NeuralNetworks/Layers/ConvolutionalLayer.cs (1)

1399-1407: ⚠️ Potential issue | 🟠 Major

Avoid eager random initialization in SetParameters().

Calling EnsureInitialized() first allocates and fills the full tensor set before immediately overwriting it. For the large VGG/Capsule cases, that reintroduces the memory/CPU spike this PR is trying to remove, and even invalid-length inputs now pay that cost before throwing. Validate against ParameterCount first, then allocate backing storage without InitializeWeights() when the layer is still lazy.

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@src/NeuralNetworks/Layers/ConvolutionalLayer.cs` around lines 1399 - 1407,
Validate the incoming parameter length against the layer's ParameterCount before
calling EnsureInitialized so you don't eagerly allocate; in SetParameters
compare parameters.Length to ParameterCount (not kernelLen + biasLen) and throw
if mismatched, and only after validation allocate or ensure backing storage for
_kernels and _biases—but if allocation is needed do so without calling
InitializeWeights (avoid calling EnsureInitialized/InitializeWeights before
copying) so you don't pre-fill tensors unnecessarily; then copy values into
_kernels and _biases and set any initialized flags.

♻️ Duplicate comments (2)

src/NeuralNetworks/Layers/ConvolutionalLayer.cs (2)

734-740: ⚠️ Potential issue | 🔴 Critical

Blocking: reset cached bias views when deserializing new tensors.

Deserialize() replaces _biases, but _biasReshaped4D is left intact. The fused and inference fast paths reuse that cache with ??=, so a reused layer instance can keep adding the old bias tensor after load.

Suggested fix

         ClearRegisteredParameters();
         RegisterTrainableParameter(_kernels, PersistentTensorRole.Weights);
         RegisterTrainableParameter(_biases, PersistentTensorRole.Biases);
+        _biasReshaped4D = null;
+        _preAllocatedOutput = null;
 
         // Mark as initialized so EnsureInitialized() doesn't re-randomize the
         // just-deserialized weights on the next Forward/GetParameters call.
         _isInitialized = true;

As per coding guidelines: “Production Readiness (CRITICAL)… Incomplete features: Half-implemented patterns where some code paths work but others silently do nothing.”

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@src/NeuralNetworks/Layers/ConvolutionalLayer.cs` around lines 734 - 740, When
deserializing and replacing the layer's bias tensor in Deserialize(), reset any
cached bias views so old views aren't reused: after assigning the new _biases
set _biasReshaped4D to null (or its default/empty state) so the fused/inference
fast paths that use the ??= cache will recreate the correct view; also ensure
this happens before setting _isInitialized = true so
EnsureInitialized()/Forward/GetParameters won't re-use stale cached bias views.

---

729-740: ⚠️ Potential issue | 🔴 Critical

Blocking: lazy layers still cannot round-trip through serialization.

Deserialize() now always reads a full kernel/bias payload and marks the layer initialized, but Serialize() still writes zero scalars when the layer stayed lazy and _kernels/_biases are still [0,…,0]. Saving a never-forwarded lazy layer will truncate the stream and fail on load or corrupt the next payload. Either materialize before writing or persist an explicit “uninitialized” state and read it symmetrically.

Suggested fix

 public override void Serialize(BinaryWriter writer)
 {
+    EnsureInitialized();
     base.Serialize(writer);
     writer.Write(InputDepth);
     writer.Write(OutputDepth);
     writer.Write(KernelSize);

As per coding guidelines: “Production Readiness (CRITICAL)… Incomplete features: Half-implemented patterns where some code paths work but others silently do nothing.”

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@src/NeuralNetworks/Layers/ConvolutionalLayer.cs` around lines 729 - 740,
Serialize() currently writes raw kernel/bias payloads even when the layer is
lazy, causing truncated/corrupt loads because Deserialize() always expects full
payloads and marks the layer initialized; fix this by making serialization
symmetric: in Serialize(), write an explicit initialized flag (e.g., a boolean
before weight data) and if _isInitialized == false write the flag and skip
writing _kernels/_biases payloads, and in Deserialize() first read that flag
and, if false, set _isInitialized = false and avoid reading/deserializing the
kernel/bias raw payloads (do not mark initialized or re-register parameters);
adjust places that call
EnsureInitialized()/RegisterTrainableParameter/_kernels/_biases handling so the
runtime behavior matches the new flag (or alternatively call EnsureInitialized()
from Serialize() if you prefer auto-materialization), but implement the explicit
flag approach in Serialize() and Deserialize() to guarantee symmetric, safe
round-trip behavior.

🤖 Prompt for all review comments with AI agents

Verify each finding against the current code and only fix it if needed.

Outside diff comments:
In `@src/NeuralNetworks/Layers/ConvolutionalLayer.cs`:
- Around line 1399-1407: Validate the incoming parameter length against the
layer's ParameterCount before calling EnsureInitialized so you don't eagerly
allocate; in SetParameters compare parameters.Length to ParameterCount (not
kernelLen + biasLen) and throw if mismatched, and only after validation allocate
or ensure backing storage for _kernels and _biases—but if allocation is needed
do so without calling InitializeWeights (avoid calling
EnsureInitialized/InitializeWeights before copying) so you don't pre-fill
tensors unnecessarily; then copy values into _kernels and _biases and set any
initialized flags.

---

Duplicate comments:
In `@src/NeuralNetworks/Layers/ConvolutionalLayer.cs`:
- Around line 734-740: When deserializing and replacing the layer's bias tensor
in Deserialize(), reset any cached bias views so old views aren't reused: after
assigning the new _biases set _biasReshaped4D to null (or its default/empty
state) so the fused/inference fast paths that use the ??= cache will recreate
the correct view; also ensure this happens before setting _isInitialized = true
so EnsureInitialized()/Forward/GetParameters won't re-use stale cached bias
views.
- Around line 729-740: Serialize() currently writes raw kernel/bias payloads
even when the layer is lazy, causing truncated/corrupt loads because
Deserialize() always expects full payloads and marks the layer initialized; fix
this by making serialization symmetric: in Serialize(), write an explicit
initialized flag (e.g., a boolean before weight data) and if _isInitialized ==
false write the flag and skip writing _kernels/_biases payloads, and in
Deserialize() first read that flag and, if false, set _isInitialized = false and
avoid reading/deserializing the kernel/bias raw payloads (do not mark
initialized or re-register parameters); adjust places that call
EnsureInitialized()/RegisterTrainableParameter/_kernels/_biases handling so the
runtime behavior matches the new flag (or alternatively call EnsureInitialized()
from Serialize() if you prefer auto-materialization), but implement the explicit
flag approach in Serialize() and Deserialize() to guarantee symmetric, safe
round-trip behavior.

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Pro

Run ID: b0018e32-b218-4d63-a99d-95159b9e06ce

📥 Commits

Reviewing files that changed from the base of the PR and between 02b532d and b7d9e8a.

📒 Files selected for processing (2)

• src/Helpers/LayerHelper.cs
• src/NeuralNetworks/Layers/ConvolutionalLayer.cs

[coderabbitai]

Note

Docstrings generation - SUCCESS
Generated docstrings for this pull request at #1150