tensorplay@1.0.0.dev20260909

Return the recommended gain value for the given nonlinearity.

Supported names: linear/conv{1,2,3}d/conv_transpose{1,2,3}d and sigmoid map to 1, tanh to 5/3, relu to sqrt(2), selu to 3/4, and leaky_relu to sqrt(2 / (1 + negative_slope**2)) where the slope is taken from param (default 0.01).

Fill tensor in-place with the scalar val.

Fill the {3, 4, 5}-D tensor in-place with Dirac delta kernels.

Preserves identity of inputs in convolutional layers; with groups > 1 each group of output channels preserves identity independently.

Fill the 2-D tensor in-place with the identity matrix.

As many inputs as possible are preserved through a Linear layer.

Fill tensor in-place with N(0, std^2) where std = gain / sqrt(fan).

Fill tensor in-place with U(-bound, bound) where bound = gain * sqrt(3 / fan).

Fill tensor in-place with samples from N(mean, std^2).

Fill tensor in-place with the scalar value 1.

Fill tensor in-place with a (semi) orthogonal matrix.

The tensor must have at least 2 dimensions; trailing dimensions are flattened. Rows (or columns, when narrower) are orthonormalized with a QR factorization of a standard-normal sample, and Q is rescaled by the diagonal signs of R so its distribution is uniform over the orthogonal group.

Fill the 2-D tensor in-place as a sparse matrix.

Each column gets sparsity * rows zeroed entries (a random row subset per column); the remaining entries come from N(0, std^2).

Fills the input Tensor with values drawn from a truncated normal distribution.

Method is based on the rejection-sampling scheme in https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf —

Fill tensor in-place with samples from U(a, b).

Fill tensor in-place with N(0, std^2) where std = gain * sqrt(2 / (fan_in + fan_out)).

Fill tensor in-place with U(-a, a) where a = gain * sqrt(6 / (fan_in + fan_out)).

Fill tensor in-place with the scalar value 0.