TorchDefaultInitializationStrategy<T>
Models & Types · AiDotNet.Initialization
PyTorch’s default initialization of nn.Conv*d and nn.Linear: kaiming_uniform_ with
a = √5, which is U(±1/√fan_in) for the weight, and U(±1/√fan_in) for the bias.
Inheritance: InitializationStrategyBase<T> → TorchDefaultInitializationStrategy<T>
For Beginners
Use this when porting a PyTorch model whose code builds its layers and never calls an init function on them; it reproduces the weights such a model starts training from.
How It Works
Reference implementations that never re-initialize their modules train from this distribution, so a port that wants their starting point needs it rather than the gain-scaled Kaiming default of the convolution layers, whose bound is √3 times wider for an identity activation and whose biases start at zero.
Constructors
| Constructor | Summary |
|---|---|
TorchDefaultInitializationStrategy() |
Creates the strategy with the framework’s default thread-safe RNG. |
TorchDefaultInitializationStrategy(Random rng) |
Creates the strategy with a caller-supplied Random source. |
Properties
| Property | Summary |
|---|---|
bool IsLazy { get; } |
|
bool LoadFromExternal { get; } |
Methods
| Method | Summary |
|---|---|
void InitializeBiases(Tensor<T> biases) |
|
void InitializeWeights(Tensor<T> weights, int inputSize, int outputSize) |