GroupedInvertibleConvFlowLayer<T>
Layers · AiDotNet.NeuralNetworks.Layers
Glow-TTS’s invertible 1×1 convolution with channel grouping: the channels are split into groups of
nSplit and the same nSplit × nSplit matrix mixes each group.
Inheritance: LayerBase<T> → GroupedInvertibleConvFlowLayer<T>
Implements: IInvertibleFlowStep<T>, IShapeContract
For Beginners
Shuffles information between channels in a way that can be undone exactly.
How It Works
Kim et al. 2020, §3.3, Fig. 8c: a full invertible 1×1 convolution (Glow) over 160 channels costs a 160 × 160
determinant; Glow-TTS instead mixes groups of 4 channels with one shared 4 × 4 matrix, taking two channels from each
half so the following coupling layer sees a mix (reference implementation modules.InvConvNear, n_split=4).
The matrix starts as a random rotation (QR of a Gaussian matrix, with a positive determinant). The log-determinant is
(C / nSplit) · T · log|det W|.
Constructors
| Constructor | Summary |
|---|---|
GroupedInvertibleConvFlowLayer(int channels, int split) |
Creates the layer. |
Properties
| Property | Summary |
|---|---|
bool SupportsTraining { get; } |
Methods
| Method | Summary |
|---|---|
Dictionary<string, string> GetParameterRoles() |
Returns parameter roles for per-role optimizer configuration (e.g., weight decay exemption for biases). |
IReadOnlyList<OutputAxisContract> OutputAxesFor(int inputRank) |
Derived from this type’s [TensorLayout] declarations. |
IReadOnlyList<Tensor<T>> GetTrainableParameters() |
Returns all trainable parameter tensors marked with [TrainableParameter]. |
ValueTuple<Tensor<T>, Tensor<T>> Transform(Tensor<T> input, bool reverse) |
|
void ResetState() |
|
void SetTrainableParameters(IReadOnlyList<Tensor<T>> parameters) |
Replaces trainable parameter tensors (e.g., with ParameterBuffer views). |
void UpdateParameters(T learningRate) |
|
void ZeroGrad() |
Clears all gradient fields discovered by convention ({paramName}Gradient). |