e3x.nn.initializers

Custom initializers.

This module contains implementations of common initializers for the Tensor layer. Standard initialzers cannot be used, because the number of inputs and outputs cannot be inferred from the shape of the kernel (instead, all valid coupling paths are checked and the number of inputs and outputs determined accordingly).

Functions

fused_tensor_normal([scale, mask, dtype])

Initializer for fused tensor product kernels with normal distribution.

fused_tensor_uniform([scale, mask, dtype])

Initializer for fused tensor product kernels with uniform distribution.

tensor_glorot_normal([dtype])

Glorot normal initializer for tensor product kernels.

tensor_glorot_uniform([dtype])

Glorot uniform initializer for tensor product kernels.

tensor_he_normal([dtype])

He normal initializer for tensor product kernels.

tensor_he_uniform([dtype])

He uniform initializer for tensor product kernels.

tensor_kaiming_normal([dtype])

He normal initializer for tensor product kernels.

tensor_kaiming_uniform([dtype])

He uniform initializer for tensor product kernels.

tensor_lecun_normal([dtype])

Lecun normal initializer for tensor product kernels.

tensor_lecun_uniform([dtype])

Lecun uniform initializer for tensor product kernels.

tensor_variance_scaling(scale, mode, ...[, ...])

Variance scaling initializer for tensor product kernels.

tensor_xavier_normal([dtype])

Glorot normal initializer for tensor product kernels.

tensor_xavier_uniform([dtype])

Glorot uniform initializer for tensor product kernels.

Classes

FusedTensorInitializerFn(*args, **kwargs)

Protocol for fused tensor initializer functions.

InitializerFn(*args, **kwargs)

Protocol for all standard initializer functions.