NormalizedColumnsInitializationStrategy<T>
Models & Types · AiDotNet.Initialization
OpenAI Baselines’ normc_initializer: standard-normal weights, each output unit’s incoming column rescaled
to a fixed norm, and zero biases.
Inheritance: InitializationStrategyBase<T> → NormalizedColumnsInitializationStrategy<T>
For Beginners
Random starting weights, scaled so every neuron starts equally “loud”. A small scale on the last layer makes a new policy start out undecided between its actions.
How It Works
out = randn(in, out); out *= std / sqrt(sum(out², axis=0)). Every unit starts with the same incoming
weight norm, std, whatever the fan-in. Baselines’ PPO (the code of Schulman et al. 2017)
builds its policy and value networks with it: 1.0 on the hidden layers and the value head, 0.01 on the policy
head, so the initial policy is close to uniform (or, for a Gaussian, centred on zero).
Constructors
| Constructor | Summary |
|---|---|
NormalizedColumnsInitializationStrategy(double std, Random random) |
Creates the initializer. |
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) |