fix: properly accumulate gradient penalty gradients in WGANGP

- Restructure WGANGP to capture and combine gradients from all sources:
  real image gradients, fake image gradients, and gradient penalty gradients
- Create ComputeGradientPenaltyWithGradients to return both penalty value
  and parameter gradients
- Add UpdateCriticParametersWithGradients for combined gradient updates
- Fix DCGAN InputType to use ThreeDimensional for CNN compatibility
- Fix serialization in multiple GAN implementations

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

Author: franklinic

src/NeuralNetworks/ConditionalGAN.cs

@@ -369,7 +369,9 @@ private T TrainDiscriminatorBatch(Tensor<T> images, Tensor<T> labels)
     private T TrainGeneratorBatch(Tensor<T> generatorInput, Tensor<T> fakeImagesWithConditions, Tensor<T> targetLabels)
     {
         Generator.SetTrainingMode(true);
-        Discriminator.SetTrainingMode(false);
+        // Keep Discriminator in training mode for backpropagation (required for Backpropagate to work)
+        // We just won't update its parameters
+        Discriminator.SetTrainingMode(true);
 
         // Get discriminator output
         var discriminatorOutput = Discriminator.Predict(fakeImagesWithConditions);
@@ -380,7 +382,7 @@ private T TrainGeneratorBatch(Tensor<T> generatorInput, Tensor<T> fakeImagesWith
         // Calculate gradients
         var outputGradients = CalculateBatchGradients(discriminatorOutput, targetLabels);
 
-        // Backpropagate through discriminator
+        // Backpropagate through discriminator to get input gradients (but don't update discriminator weights)
         var discriminatorInputGradients = Discriminator.Backpropagate(outputGradients);
 
         // Extract gradients for the image part (not the condition part)

src/NeuralNetworks/DCGAN.cs

@@ -72,7 +72,7 @@ public DCGAN(
         : base(
             CreateDCGANGeneratorArchitecture(latentSize, imageChannels, imageHeight, imageWidth, generatorFeatureMaps),
             CreateDCGANDiscriminatorArchitecture(imageChannels, imageHeight, imageWidth, discriminatorFeatureMaps),
-            InputType.TwoDimensional,
+            InputType.ThreeDimensional,
             lossFunction,
             initialLearningRate)
     {
@@ -113,11 +113,20 @@ private static NeuralNetworkArchitecture<T> CreateDCGANGeneratorArchitecture(
         int imageWidth,
         int featureMaps)
     {
+        // For DCGAN generator, the latent vector is first projected and reshaped to an initial
+        // 3D feature map. The typical starting spatial size is 4x4 which gets upsampled through
+        // transposed convolutions. The depth represents the number of feature channels.
+        // Note: The actual latent vector (1D) handling is done by the first projection layer.
+        int initialSpatialSize = 4;
+        int initialChannels = featureMaps * 8;  // Standard DCGAN uses 8x feature maps initially
+
         return new NeuralNetworkArchitecture<T>(
-            InputType.OneDimensional,
+            InputType.ThreeDimensional,
             NeuralNetworkTaskType.Generative,
             NetworkComplexity.Medium,
-            inputSize: latentSize,
+            inputDepth: initialChannels,
+            inputHeight: initialSpatialSize,
+            inputWidth: initialSpatialSize,
             outputSize: imageChannels * imageHeight * imageWidth);
     }
 
@@ -154,12 +163,15 @@ private static NeuralNetworkArchitecture<T> CreateDCGANDiscriminatorArchitecture
         int imageWidth,
         int featureMaps)
     {
+        // DCGAN discriminator takes 3D images as input (channels x height x width)
+        // and outputs a single probability value for real/fake classification
         return new NeuralNetworkArchitecture<T>(
-            InputType.TwoDimensional,
+            InputType.ThreeDimensional,
             NeuralNetworkTaskType.BinaryClassification,
             NetworkComplexity.Medium,
+            inputDepth: imageChannels,
             inputHeight: imageHeight,
-            inputWidth: imageWidth * imageChannels,
+            inputWidth: imageWidth,
             outputSize: 1);
     }
 }