torch.nn.init — PyTorch 2.7 documentation (original) (raw)

Warning

All the functions in this module are intended to be used to initialize neural network parameters, so they all run in torch.no_grad() mode and will not be taken into account by autograd.

torch.nn.init.calculate_gain(nonlinearity, param=None)[source][source]

Return the recommended gain value for the given nonlinearity function.

The values are as follows:

nonlinearity gain
Linear / Identity 11
Conv{1,2,3}D 11
Sigmoid 11
Tanh 53\frac{5}{3}
ReLU 2\sqrt{2}
Leaky Relu 21+negative_slope2\sqrt{\frac{2}{1 + \text{negative\_slope}^2}}
SELU 34\frac{3}{4}

Warning

In order to implement Self-Normalizing Neural Networks , you should use nonlinearity='linear' instead of nonlinearity='selu'. This gives the initial weights a variance of 1 / N, which is necessary to induce a stable fixed point in the forward pass. In contrast, the default gain for SELU sacrifices the normalization effect for more stable gradient flow in rectangular layers.

Parameters

Examples

gain = nn.init.calculate_gain('leaky_relu', 0.2) # leaky_relu with negative_slope=0.2

torch.nn.init.uniform_(tensor, a=0.0, b=1.0, generator=None)[source][source]

Fill the input Tensor with values drawn from the uniform distribution.

U(a,b)\mathcal{U}(a, b).

Parameters

Return type

Tensor

Examples

w = torch.empty(3, 5) nn.init.uniform_(w)

torch.nn.init.normal_(tensor, mean=0.0, std=1.0, generator=None)[source][source]

Fill the input Tensor with values drawn from the normal distribution.

N(mean,std2)\mathcal{N}(\text{mean}, \text{std}^2).

Parameters

Return type

Tensor

Examples

w = torch.empty(3, 5) nn.init.normal_(w)

torch.nn.init.constant_(tensor, val)[source][source]

Fill the input Tensor with the value val\text{val}.

Parameters

Return type

Tensor

Examples

w = torch.empty(3, 5) nn.init.constant_(w, 0.3)

torch.nn.init.ones_(tensor)[source][source]

Fill the input Tensor with the scalar value 1.

Parameters

tensor (Tensor) – an n-dimensional torch.Tensor

Return type

Tensor

Examples

w = torch.empty(3, 5) nn.init.ones_(w)

torch.nn.init.zeros_(tensor)[source][source]

Fill the input Tensor with the scalar value 0.

Parameters

tensor (Tensor) – an n-dimensional torch.Tensor

Return type

Tensor

Examples

w = torch.empty(3, 5) nn.init.zeros_(w)

torch.nn.init.eye_(tensor)[source][source]

Fill the 2-dimensional input Tensor with the identity matrix.

Preserves the identity of the inputs in Linear layers, where as many inputs are preserved as possible.

Parameters

tensor – a 2-dimensional torch.Tensor

Examples

w = torch.empty(3, 5) nn.init.eye_(w)

torch.nn.init.dirac_(tensor, groups=1)[source][source]

Fill the {3, 4, 5}-dimensional input Tensor with the Dirac delta function.

Preserves the identity of the inputs in Convolutionallayers, where as many input channels are preserved as possible. In case of groups>1, each group of channels preserves identity

Parameters

Examples

w = torch.empty(3, 16, 5, 5) nn.init.dirac_(w) w = torch.empty(3, 24, 5, 5) nn.init.dirac_(w, 3)

torch.nn.init.xavier_uniform_(tensor, gain=1.0, generator=None)[source][source]

Fill the input Tensor with values using a Xavier uniform distribution.

The method is described in Understanding the difficulty of training deep feedforward neural networks - Glorot, X. & Bengio, Y. (2010). The resulting tensor will have values sampled fromU(−a,a)\mathcal{U}(-a, a) where

a=gain×6fan_in+fan_outa = \text{gain} \times \sqrt{\frac{6}{\text{fan\_in} + \text{fan\_out}}}

Also known as Glorot initialization.

Parameters

Return type

Tensor

Examples

w = torch.empty(3, 5) nn.init.xavier_uniform_(w, gain=nn.init.calculate_gain('relu'))

Note

Be aware that fan_in and fan_out are calculated assuming that the weight matrix is used in a transposed manner, (i.e., x @ w.T in Linear layers, where w.shape = [fan_out, fan_in]). This is important for correct initialization. If you plan to use x @ w, where w.shape = [fan_in, fan_out], pass in a transposed weight matrix, i.e. nn.init.xavier_uniform_(w.T, ...).

torch.nn.init.xavier_normal_(tensor, gain=1.0, generator=None)[source][source]

Fill the input Tensor with values using a Xavier normal distribution.

The method is described in Understanding the difficulty of training deep feedforward neural networks - Glorot, X. & Bengio, Y. (2010). The resulting tensor will have values sampled from N(0,std2)\mathcal{N}(0, \text{std}^2) where

std=gain×2fan_in+fan_out\text{std} = \text{gain} \times \sqrt{\frac{2}{\text{fan\_in} + \text{fan\_out}}}

Also known as Glorot initialization.

Parameters

Return type

Tensor

Examples

w = torch.empty(3, 5) nn.init.xavier_normal_(w)

Note

Be aware that fan_in and fan_out are calculated assuming that the weight matrix is used in a transposed manner, (i.e., x @ w.T in Linear layers, where w.shape = [fan_out, fan_in]). This is important for correct initialization. If you plan to use x @ w, where w.shape = [fan_in, fan_out], pass in a transposed weight matrix, i.e. nn.init.xavier_normal_(w.T, ...).

torch.nn.init.kaiming_uniform_(tensor, a=0, mode='fan_in', nonlinearity='leaky_relu', generator=None)[source][source]

Fill the input Tensor with values using a Kaiming uniform distribution.

The method is described in Delving deep into rectifiers: Surpassing human-level performance on ImageNet classification - He, K. et al. (2015). The resulting tensor will have values sampled fromU(−bound,bound)\mathcal{U}(-\text{bound}, \text{bound}) where

bound=gain×3fan_mode\text{bound} = \text{gain} \times \sqrt{\frac{3}{\text{fan\_mode}}}

Also known as He initialization.

Parameters

Examples

w = torch.empty(3, 5) nn.init.kaiming_uniform_(w, mode='fan_in', nonlinearity='relu')

Note

Be aware that fan_in and fan_out are calculated assuming that the weight matrix is used in a transposed manner, (i.e., x @ w.T in Linear layers, where w.shape = [fan_out, fan_in]). This is important for correct initialization. If you plan to use x @ w, where w.shape = [fan_in, fan_out], pass in a transposed weight matrix, i.e. nn.init.kaiming_uniform_(w.T, ...).

torch.nn.init.kaiming_normal_(tensor, a=0, mode='fan_in', nonlinearity='leaky_relu', generator=None)[source][source]

Fill the input Tensor with values using a Kaiming normal distribution.

The method is described in Delving deep into rectifiers: Surpassing human-level performance on ImageNet classification - He, K. et al. (2015). The resulting tensor will have values sampled fromN(0,std2)\mathcal{N}(0, \text{std}^2) where

std=gainfan_mode\text{std} = \frac{\text{gain}}{\sqrt{\text{fan\_mode}}}

Also known as He initialization.

Parameters

Examples

w = torch.empty(3, 5) nn.init.kaiming_normal_(w, mode='fan_out', nonlinearity='relu')

Note

Be aware that fan_in and fan_out are calculated assuming that the weight matrix is used in a transposed manner, (i.e., x @ w.T in Linear layers, where w.shape = [fan_out, fan_in]). This is important for correct initialization. If you plan to use x @ w, where w.shape = [fan_in, fan_out], pass in a transposed weight matrix, i.e. nn.init.kaiming_normal_(w.T, ...).

torch.nn.init.trunc_normal_(tensor, mean=0.0, std=1.0, a=-2.0, b=2.0, generator=None)[source][source]

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

The values are effectively drawn from the normal distribution N(mean,std2)\mathcal{N}(\text{mean}, \text{std}^2)with values outside [a,b][a, b] redrawn until they are within the bounds. The method used for generating the random values works best when a≤mean≤ba \leq \text{mean} \leq b.

Parameters

Return type

Tensor

Examples

w = torch.empty(3, 5) nn.init.trunc_normal_(w)

torch.nn.init.orthogonal_(tensor, gain=1, generator=None)[source][source]

Fill the input Tensor with a (semi) orthogonal matrix.

Described in Exact solutions to the nonlinear dynamics of learning in deep linear neural networks - Saxe, A. et al. (2013). The input tensor must have at least 2 dimensions, and for tensors with more than 2 dimensions the trailing dimensions are flattened.

Parameters

Examples

w = torch.empty(3, 5) nn.init.orthogonal_(w)

torch.nn.init.sparse_(tensor, sparsity, std=0.01, generator=None)[source][source]

Fill the 2D input Tensor as a sparse matrix.

The non-zero elements will be drawn from the normal distributionN(0,0.01)\mathcal{N}(0, 0.01), as described in Deep learning via Hessian-free optimization - Martens, J. (2010).

Parameters

Examples

w = torch.empty(3, 5) nn.init.sparse_(w, sparsity=0.1)