ConstantInitializationStrategy<T> |
Constant initialization strategy that sets every weight to one value and every bias to zero. |
EagerInitializationStrategy<T> |
Eager initialization strategy that initializes weights immediately on construction. |
FromFileInitializationStrategy<T> |
Initialization strategy that loads weights from an external file. |
GlorotUniformInitializationStrategy<T> |
Glorot and Bengio’s uniform initialization (TensorFlow’s and Keras’s default): weights from U(−√(6 / (fan_in + fan_out)), √(6 / (fan_in + fan_out))) and zero biases. |
HeInitializationStrategy<T> |
He/Kaiming initialization strategy for ReLU-family activations. |
LazyInitializationStrategy<T> |
Lazy initialization strategy that defers weight allocation until first Forward() call. |
LeCunInitializationStrategy<T> |
LeCun initialization strategy for SELU activations and self-normalizing networks. |
NormalInitializationStrategy<T> |
Initializes weights from a normal distribution with a configurable mean and standard deviation. |
NormalizedColumnsInitializationStrategy<T> |
OpenAI Baselines’ normc_initializer: standard-normal weights, each output unit’s incoming column rescaled to a fixed norm, and zero biases. |
OrthogonalInitializationStrategy<T> |
Orthogonal initialization strategy for RNNs, LSTMs, and deep networks. |
TorchDefaultInitializationStrategy<T> |
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. |
UniformInitializationStrategy<T> |
Uniform random initialization with configurable range. |
ZeroInitializationStrategy<T> |
Zero initialization strategy that sets all weights to zero. |