conv1d¶

torch.nn.functional.conv1d(input, weight, bias=None, stride=1, padding=0, dilation=1, groups=1) → Tensor¶

Applies a 1D convolution over an input signal composed of several input planes.

This operator supports TensorFloat32.

See Conv1d for details and output shape.

Note

In some circumstances when using the CUDA backend with CuDNN, this operator may select a nondeterministic algorithm to increase performance. If this is undesirable, you can try to make the operation deterministic (potentially at a performance cost) by setting torch.backends.cudnn.deterministic = True. Please see the notes on Reproducibility for background.

Parameters

• input – input tensor of shape $(\text{minibatch} , \text{in\_channels} , iW)$

• weight – filters of shape $(\text{out\_channels} , \frac{\text{in\_channels}}{\text{groups}} , kW)$

• bias – optional bias of shape $(\text{out\_channels})$ . Default: None

• stride – the stride of the convolving kernel. Can be a single number or a one-element tuple (sW,). Default: 1

• padding – implicit paddings on both sides of the input. Can be a single number or a one-element tuple (padW,). Default: 0

• dilation – the spacing between kernel elements. Can be a single number or a one-element tuple (dW,). Default: 1

• groups – split input into groups, $\text{in\_channels}$ should be divisible by the number of groups. Default: 1

Examples:

>>> filters = torch.randn(33, 16, 3)
>>> inputs = torch.randn(20, 16, 50)
>>> F.conv1d(inputs, filters)

conv2d¶

torch.nn.functional.conv2d(input, weight, bias=None, stride=1, padding=0, dilation=1, groups=1) → Tensor¶

Applies a 2D convolution over an input image composed of several input planes.

This operator supports TensorFloat32.

See Conv2d for details and output shape.

Note

In some circumstances when using the CUDA backend with CuDNN, this operator may select a nondeterministic algorithm to increase performance. If this is undesirable, you can try to make the operation deterministic (potentially at a performance cost) by setting torch.backends.cudnn.deterministic = True. Please see the notes on Reproducibility for background.

Parameters

• input – input tensor of shape $(\text{minibatch} , \text{in\_channels} , iH , iW)$

• weight – filters of shape $(\text{out\_channels} , \frac{\text{in\_channels}}{\text{groups}} , kH , kW)$

• bias – optional bias tensor of shape $(\text{out\_channels})$ . Default: None

• stride – the stride of the convolving kernel. Can be a single number or a tuple (sH, sW). Default: 1

• padding – implicit paddings on both sides of the input. Can be a single number or a tuple (padH, padW). Default: 0

• dilation – the spacing between kernel elements. Can be a single number or a tuple (dH, dW). Default: 1

• groups – split input into groups, $\text{in\_channels}$ should be divisible by the number of groups. Default: 1

Examples:

>>> # With square kernels and equal stride
>>> filters = torch.randn(8,4,3,3)
>>> inputs = torch.randn(1,4,5,5)
>>> F.conv2d(inputs, filters, padding=1)

conv3d¶

torch.nn.functional.conv3d(input, weight, bias=None, stride=1, padding=0, dilation=1, groups=1) → Tensor¶

Applies a 3D convolution over an input image composed of several input planes.

This operator supports TensorFloat32.

See Conv3d for details and output shape.

Note

In some circumstances when using the CUDA backend with CuDNN, this operator may select a nondeterministic algorithm to increase performance. If this is undesirable, you can try to make the operation deterministic (potentially at a performance cost) by setting torch.backends.cudnn.deterministic = True. Please see the notes on Reproducibility for background.

Parameters

• input – input tensor of shape $(\text{minibatch} , \text{in\_channels} , iT , iH , iW)$

• weight – filters of shape $(\text{out\_channels} , \frac{\text{in\_channels}}{\text{groups}} , kT , kH , kW)$

• bias – optional bias tensor of shape $(\text{out\_channels})$ . Default: None

• stride – the stride of the convolving kernel. Can be a single number or a tuple (sT, sH, sW). Default: 1

• padding – implicit paddings on both sides of the input. Can be a single number or a tuple (padT, padH, padW). Default: 0

• dilation – the spacing between kernel elements. Can be a single number or a tuple (dT, dH, dW). Default: 1

• groups – split input into groups, $\text{in\_channels}$ should be divisible by the number of groups. Default: 1

Examples:

>>> filters = torch.randn(33, 16, 3, 3, 3)
>>> inputs = torch.randn(20, 16, 50, 10, 20)
>>> F.conv3d(inputs, filters)

conv_transpose1d¶

torch.nn.functional.conv_transpose1d(input, weight, bias=None, stride=1, padding=0, output_padding=0, groups=1, dilation=1) → Tensor¶

Applies a 1D transposed convolution operator over an input signal composed of several input planes, sometimes also called “deconvolution”.

This operator supports TensorFloat32.

See ConvTranspose1d for details and output shape.

Note

In some circumstances when using the CUDA backend with CuDNN, this operator may select a nondeterministic algorithm to increase performance. If this is undesirable, you can try to make the operation deterministic (potentially at a performance cost) by setting torch.backends.cudnn.deterministic = True. Please see the notes on Reproducibility for background.

Parameters

• input – input tensor of shape $(\text{minibatch} , \text{in\_channels} , iW)$

• weight – filters of shape $(\text{in\_channels} , \frac{\text{out\_channels}}{\text{groups}} , kW)$

• bias – optional bias of shape $(\text{out\_channels})$ . Default: None

• stride – the stride of the convolving kernel. Can be a single number or a tuple (sW,). Default: 1

• padding – dilation * (kernel_size - 1) - padding zero-padding will be added to both sides of each dimension in the input. Can be a single number or a tuple (padW,). Default: 0

• output_padding – additional size added to one side of each dimension in the output shape. Can be a single number or a tuple (out_padW). Default: 0

• groups – split input into groups, $\text{in\_channels}$ should be divisible by the number of groups. Default: 1

• dilation – the spacing between kernel elements. Can be a single number or a tuple (dW,). Default: 1

Examples:

>>> inputs = torch.randn(20, 16, 50)
>>> weights = torch.randn(16, 33, 5)
>>> F.conv_transpose1d(inputs, weights)

conv_transpose2d¶

torch.nn.functional.conv_transpose2d(input, weight, bias=None, stride=1, padding=0, output_padding=0, groups=1, dilation=1) → Tensor¶

Applies a 2D transposed convolution operator over an input image composed of several input planes, sometimes also called “deconvolution”.

This operator supports TensorFloat32.

See ConvTranspose2d for details and output shape.

Note

In some circumstances when using the CUDA backend with CuDNN, this operator may select a nondeterministic algorithm to increase performance. If this is undesirable, you can try to make the operation deterministic (potentially at a performance cost) by setting torch.backends.cudnn.deterministic = True. Please see the notes on Reproducibility for background.

Parameters

• input – input tensor of shape $(\text{minibatch} , \text{in\_channels} , iH , iW)$

• weight – filters of shape $(\text{in\_channels} , \frac{\text{out\_channels}}{\text{groups}} , kH , kW)$

• bias – optional bias of shape $(\text{out\_channels})$ . Default: None

• stride – the stride of the convolving kernel. Can be a single number or a tuple (sH, sW). Default: 1

• padding – dilation * (kernel_size - 1) - padding zero-padding will be added to both sides of each dimension in the input. Can be a single number or a tuple (padH, padW). Default: 0

• output_padding – additional size added to one side of each dimension in the output shape. Can be a single number or a tuple (out_padH, out_padW). Default: 0

• groups – split input into groups, $\text{in\_channels}$ should be divisible by the number of groups. Default: 1

• dilation – the spacing between kernel elements. Can be a single number or a tuple (dH, dW). Default: 1

Examples:

>>> # With square kernels and equal stride
>>> inputs = torch.randn(1, 4, 5, 5)
>>> weights = torch.randn(4, 8, 3, 3)
>>> F.conv_transpose2d(inputs, weights, padding=1)

conv_transpose3d¶

torch.nn.functional.conv_transpose3d(input, weight, bias=None, stride=1, padding=0, output_padding=0, groups=1, dilation=1) → Tensor¶

Applies a 3D transposed convolution operator over an input image composed of several input planes, sometimes also called “deconvolution”

This operator supports TensorFloat32.

See ConvTranspose3d for details and output shape.

Note

In some circumstances when using the CUDA backend with CuDNN, this operator may select a nondeterministic algorithm to increase performance. If this is undesirable, you can try to make the operation deterministic (potentially at a performance cost) by setting torch.backends.cudnn.deterministic = True. Please see the notes on Reproducibility for background.

Parameters

• input – input tensor of shape $(\text{minibatch} , \text{in\_channels} , iT , iH , iW)$

• weight – filters of shape $(\text{in\_channels} , \frac{\text{out\_channels}}{\text{groups}} , kT , kH , kW)$

• bias – optional bias of shape $(\text{out\_channels})$ . Default: None

• stride – the stride of the convolving kernel. Can be a single number or a tuple (sT, sH, sW). Default: 1

• padding – dilation * (kernel_size - 1) - padding zero-padding will be added to both sides of each dimension in the input. Can be a single number or a tuple (padT, padH, padW). Default: 0

• output_padding – additional size added to one side of each dimension in the output shape. Can be a single number or a tuple (out_padT, out_padH, out_padW). Default: 0

• groups – split input into groups, $\text{in\_channels}$ should be divisible by the number of groups. Default: 1

• dilation – the spacing between kernel elements. Can be a single number or a tuple (dT, dH, dW). Default: 1

Examples:

>>> inputs = torch.randn(20, 16, 50, 10, 20)
>>> weights = torch.randn(16, 33, 3, 3, 3)
>>> F.conv_transpose3d(inputs, weights)

unfold¶

torch.nn.functional.unfold(input, kernel_size, dilation=1, padding=0, stride=1)[source]¶

Extracts sliding local blocks from an batched input tensor.

Warning

Currently, only 4-D input tensors (batched image-like tensors) are supported.

Warning

More than one element of the unfolded tensor may refer to a single memory location. As a result, in-place operations (especially ones that are vectorized) may result in incorrect behavior. If you need to write to the tensor, please clone it first.

See torch.nn.Unfold for details

fold¶

torch.nn.functional.fold(input, output_size, kernel_size, dilation=1, padding=0, stride=1)[source]¶

Combines an array of sliding local blocks into a large containing tensor.

Warning

Currently, only 3-D output tensors (unfolded batched image-like tensors) are supported.

See torch.nn.Fold for details

avg_pool1d¶

torch.nn.functional.avg_pool1d(input, kernel_size, stride=None, padding=0, ceil_mode=False, count_include_pad=True) → Tensor¶

Applies a 1D average pooling over an input signal composed of several input planes.

See AvgPool1d for details and output shape.

Parameters

• input – input tensor of shape $(\text{minibatch} , \text{in\_channels} , iW)$

• kernel_size – the size of the window. Can be a single number or a tuple (kW,)

• stride – the stride of the window. Can be a single number or a tuple (sW,). Default: kernel_size

• padding – implicit zero paddings on both sides of the input. Can be a single number or a tuple (padW,). Default: 0

• ceil_mode – when True, will use ceil instead of floor to compute the output shape. Default: False

• count_include_pad – when True, will include the zero-padding in the averaging calculation. Default: True

Examples:

>>> # pool of square window of size=3, stride=2
>>> input = torch.tensor([[[1, 2, 3, 4, 5, 6, 7]]], dtype=torch.float32)
>>> F.avg_pool1d(input, kernel_size=3, stride=2)
tensor([[[ 2.,  4.,  6.]]])

avg_pool2d¶

torch.nn.functional.avg_pool2d(input, kernel_size, stride=None, padding=0, ceil_mode=False, count_include_pad=True, divisor_override=None) → Tensor¶

Applies 2D average-pooling operation in $kH \times kW$ regions by step size $sH \times sW$ steps. The number of output features is equal to the number of input planes.

See AvgPool2d for details and output shape.

Parameters

• input – input tensor $(\text{minibatch} , \text{in\_channels} , iH , iW)$

• kernel_size – size of the pooling region. Can be a single number or a tuple (kH, kW)

• stride – stride of the pooling operation. Can be a single number or a tuple (sH, sW). Default: kernel_size

• padding – implicit zero paddings on both sides of the input. Can be a single number or a tuple (padH, padW). Default: 0

• ceil_mode – when True, will use ceil instead of floor in the formula to compute the output shape. Default: False

• count_include_pad – when True, will include the zero-padding in the averaging calculation. Default: True

• divisor_override – if specified, it will be used as divisor, otherwise size of the pooling region will be used. Default: None

avg_pool3d¶

torch.nn.functional.avg_pool3d(input, kernel_size, stride=None, padding=0, ceil_mode=False, count_include_pad=True, divisor_override=None) → Tensor¶

Applies 3D average-pooling operation in $kT \times kH \times kW$ regions by step size $sT \times sH \times sW$ steps. The number of output features is equal to $\lfloor\frac{\text{input planes}}{sT}\rfloor$ .

See AvgPool3d for details and output shape.

Parameters

• input – input tensor $(\text{minibatch} , \text{in\_channels} , iT \times iH , iW)$

• kernel_size – size of the pooling region. Can be a single number or a tuple (kT, kH, kW)

• stride – stride of the pooling operation. Can be a single number or a tuple (sT, sH, sW). Default: kernel_size

• padding – implicit zero paddings on both sides of the input. Can be a single number or a tuple (padT, padH, padW), Default: 0

• ceil_mode – when True, will use ceil instead of floor in the formula to compute the output shape

• count_include_pad – when True, will include the zero-padding in the averaging calculation

• divisor_override – if specified, it will be used as divisor, otherwise size of the pooling region will be used. Default: None

max_pool1d¶

torch.nn.functional.max_pool1d(*args, **kwargs)¶

Applies a 1D max pooling over an input signal composed of several input planes.

See MaxPool1d for details.

max_pool2d¶

torch.nn.functional.max_pool2d(*args, **kwargs)¶

Applies a 2D max pooling over an input signal composed of several input planes.

See MaxPool2d for details.

max_pool3d¶

torch.nn.functional.max_pool3d(*args, **kwargs)¶

Applies a 3D max pooling over an input signal composed of several input planes.

See MaxPool3d for details.

max_unpool1d¶

torch.nn.functional.max_unpool1d(input, indices, kernel_size, stride=None, padding=0, output_size=None)[source]¶

Computes a partial inverse of MaxPool1d.

See MaxUnpool1d for details.

max_unpool2d¶

torch.nn.functional.max_unpool2d(input, indices, kernel_size, stride=None, padding=0, output_size=None)[source]¶

Computes a partial inverse of MaxPool2d.

See MaxUnpool2d for details.

max_unpool3d¶

torch.nn.functional.max_unpool3d(input, indices, kernel_size, stride=None, padding=0, output_size=None)[source]¶

Computes a partial inverse of MaxPool3d.

See MaxUnpool3d for details.

lp_pool1d¶

torch.nn.functional.lp_pool1d(input, norm_type, kernel_size, stride=None, ceil_mode=False)[source]¶

Applies a 1D power-average pooling over an input signal composed of several input planes. If the sum of all inputs to the power of p is zero, the gradient is set to zero as well.

See LPPool1d for details.

lp_pool2d¶

torch.nn.functional.lp_pool2d(input, norm_type, kernel_size, stride=None, ceil_mode=False)[source]¶

Applies a 2D power-average pooling over an input signal composed of several input planes. If the sum of all inputs to the power of p is zero, the gradient is set to zero as well.

See LPPool2d for details.

adaptive_max_pool1d¶

torch.nn.functional.adaptive_max_pool1d(*args, **kwargs)¶

Applies a 1D adaptive max pooling over an input signal composed of several input planes.

See AdaptiveMaxPool1d for details and output shape.

Parameters

• output_size – the target output size (single integer)

• return_indices – whether to return pooling indices. Default: False

adaptive_max_pool2d¶

torch.nn.functional.adaptive_max_pool2d(*args, **kwargs)¶

Applies a 2D adaptive max pooling over an input signal composed of several input planes.

See AdaptiveMaxPool2d for details and output shape.

Parameters

• output_size – the target output size (single integer or double-integer tuple)

• return_indices – whether to return pooling indices. Default: False

adaptive_max_pool3d¶

torch.nn.functional.adaptive_max_pool3d(*args, **kwargs)¶

Applies a 3D adaptive max pooling over an input signal composed of several input planes.

See AdaptiveMaxPool3d for details and output shape.

Parameters

• output_size – the target output size (single integer or triple-integer tuple)

• return_indices – whether to return pooling indices. Default: False

adaptive_avg_pool1d¶

torch.nn.functional.adaptive_avg_pool1d(input, output_size) → Tensor¶

Applies a 1D adaptive average pooling over an input signal composed of several input planes.

See AdaptiveAvgPool1d for details and output shape.

Parameters

output_size – the target output size (single integer)

adaptive_avg_pool2d¶

torch.nn.functional.adaptive_avg_pool2d(input, output_size)[source]¶

Applies a 2D adaptive average pooling over an input signal composed of several input planes.

See AdaptiveAvgPool2d for details and output shape.

Parameters

output_size – the target output size (single integer or double-integer tuple)

adaptive_avg_pool3d¶

torch.nn.functional.adaptive_avg_pool3d(input, output_size)[source]¶

Applies a 3D adaptive average pooling over an input signal composed of several input planes.

See AdaptiveAvgPool3d for details and output shape.

Parameters

output_size – the target output size (single integer or triple-integer tuple)

threshold¶

torch.nn.functional.threshold(input, threshold, value, inplace=False)¶

Thresholds each element of the input Tensor.

See Threshold for more details.

torch.nn.functional.threshold_(input, threshold, value) → Tensor¶

In-place version of threshold().

relu¶

torch.nn.functional.relu(input, inplace=False) → Tensor[source]¶

Applies the rectified linear unit function element-wise. See ReLU for more details.

torch.nn.functional.relu_(input) → Tensor¶

In-place version of relu().

hardtanh¶

torch.nn.functional.hardtanh(input, min_val=-1., max_val=1., inplace=False) → Tensor[source]¶

Applies the HardTanh function element-wise. See Hardtanh for more details.

torch.nn.functional.hardtanh_(input, min_val=-1., max_val=1.) → Tensor¶

In-place version of hardtanh().

hardswish¶

torch.nn.functional.hardswish(input: torch.Tensor, inplace: bool = False) → torch.Tensor[source]¶

Applies the hardswish function, element-wise, as described in the paper:

Searching for MobileNetV3.

$$\text{Hardswish}(x) = \begin{cases} 0 & \text{if~} x \le -3, \\ x & \text{if~} x \ge +3, \\ x \cdot (x + 3) /6 & \text{otherwise} \end{cases}$$

See Hardswish for more details.

relu6¶

torch.nn.functional.relu6(input, inplace=False) → Tensor[source]¶

Applies the element-wise function $\text{ReLU6}(x) = \min(\max(0,x), 6)$ .

See ReLU6 for more details.

elu¶

torch.nn.functional.elu(input, alpha=1.0, inplace=False)[source]¶

Applies element-wise, $\text{ELU}(x) = \max(0,x) + \min(0, \alpha * (\exp(x) - 1))$ .

See ELU for more details.

torch.nn.functional.elu_(input, alpha=1.) → Tensor¶

In-place version of elu().

selu¶

torch.nn.functional.selu(input, inplace=False) → Tensor[source]¶

Applies element-wise, $\text{SELU}(x) = scale * (\max(0,x) + \min(0, \alpha * (\exp(x) - 1)))$ , with $\alpha=1.6732632423543772848170429916717$ and $scale=1.0507009873554804934193349852946$ .

See SELU for more details.

celu¶

torch.nn.functional.celu(input, alpha=1., inplace=False) → Tensor[source]¶

Applies element-wise, $\text{CELU}(x) = \max(0,x) + \min(0, \alpha * (\exp(x/\alpha) - 1))$ .

See CELU for more details.

leaky_relu¶

torch.nn.functional.leaky_relu(input, negative_slope=0.01, inplace=False) → Tensor[source]¶

Applies element-wise, $\text{LeakyReLU}(x) = \max(0, x) + \text{negative\_slope} * \min(0, x)$

See LeakyReLU for more details.

torch.nn.functional.leaky_relu_(input, negative_slope=0.01) → Tensor¶

In-place version of leaky_relu().

prelu¶

torch.nn.functional.prelu(input, weight) → Tensor[source]¶

Applies element-wise the function $\text{PReLU}(x) = \max(0,x) + \text{weight} * \min(0,x)$ where weight is a learnable parameter.

See PReLU for more details.

rrelu¶

torch.nn.functional.rrelu(input, lower=1./8, upper=1./3, training=False, inplace=False) → Tensor[source]¶

Randomized leaky ReLU.

See RReLU for more details.

torch.nn.functional.rrelu_(input, lower=1./8, upper=1./3, training=False) → Tensor¶

In-place version of rrelu().

glu¶

torch.nn.functional.glu(input, dim=-1) → Tensor[source]¶

The gated linear unit. Computes:

$$\text{GLU}(a, b) = a \otimes \sigma(b)$$

where input is split in half along dim to form a and b, $\sigma$ is the sigmoid function and $\otimes$ is the element-wise product between matrices.

See Language Modeling with Gated Convolutional Networks.

Parameters

• input (Tensor) – input tensor

• dim (int) – dimension on which to split the input. Default: -1

gelu¶

torch.nn.functional.gelu(input) → Tensor[source]¶

Applies element-wise the function $\text{GELU}(x) = x * \Phi(x)$

where $\Phi(x)$ is the Cumulative Distribution Function for Gaussian Distribution.

See Gaussian Error Linear Units (GELUs).

logsigmoid¶

torch.nn.functional.logsigmoid(input) → Tensor¶

Applies element-wise $\text{LogSigmoid}(x_i) = \log \left(\frac{1}{1 + \exp(-x_i)}\right)$

See LogSigmoid for more details.

hardshrink¶

torch.nn.functional.hardshrink(input, lambd=0.5) → Tensor[source]¶

Applies the hard shrinkage function element-wise

See Hardshrink for more details.

tanhshrink¶

torch.nn.functional.tanhshrink(input) → Tensor[source]¶

Applies element-wise, $\text{Tanhshrink}(x) = x - \text{Tanh}(x)$

See Tanhshrink for more details.

softsign¶

torch.nn.functional.softsign(input) → Tensor[source]¶

Applies element-wise, the function $\text{SoftSign}(x) = \frac{x}{1 + |x|}$

See Softsign for more details.

softplus¶

torch.nn.functional.softplus(input, beta=1, threshold=20) → Tensor¶

Applies element-wise, the function $\text{Softplus}(x) = \frac{1}{\beta} * \log(1 + \exp(\beta * x))$ .

For numerical stability the implementation reverts to the linear function when $input \times \beta > threshold$ .

See Softplus for more details.

softmin¶

torch.nn.functional.softmin(input, dim=None, _stacklevel=3, dtype=None)[source]¶

Applies a softmin function.

Note that $\text{Softmin}(x) = \text{Softmax}(-x)$ . See softmax definition for mathematical formula.

See Softmin for more details.

Parameters

• input (Tensor) – input

• dim (int) – A dimension along which softmin will be computed (so every slice along dim will sum to 1).

• dtype (torch.dtype, optional) – the desired data type of returned tensor. If specified, the input tensor is casted to dtype before the operation is performed. This is useful for preventing data type overflows. Default: None.

softmax¶

torch.nn.functional.softmax(input, dim=None, _stacklevel=3, dtype=None)[source]¶

Applies a softmax function.

Softmax is defined as:

$\text{Softmax}(x_{i}) = \frac{\exp(x_i)}{\sum_j \exp(x_j)}$

It is applied to all slices along dim, and will re-scale them so that the elements lie in the range [0, 1] and sum to 1.

See Softmax for more details.

Parameters

• input (Tensor) – input

• dim (int) – A dimension along which softmax will be computed.

• dtype (torch.dtype, optional) – the desired data type of returned tensor. If specified, the input tensor is casted to dtype before the operation is performed. This is useful for preventing data type overflows. Default: None.

Note

This function doesn’t work directly with NLLLoss, which expects the Log to be computed between the Softmax and itself. Use log_softmax instead (it’s faster and has better numerical properties).

softshrink¶

torch.nn.functional.softshrink(input, lambd=0.5) → Tensor¶

Applies the soft shrinkage function elementwise

See Softshrink for more details.

gumbel_softmax¶

torch.nn.functional.gumbel_softmax(logits, tau=1, hard=False, eps=1e-10, dim=-1)[source]¶

Samples from the Gumbel-Softmax distribution (Link 1 Link 2) and optionally discretizes.

Parameters

• logits – […, num_features] unnormalized log probabilities

• tau – non-negative scalar temperature

• hard – if True, the returned samples will be discretized as one-hot vectors, but will be differentiated as if it is the soft sample in autograd

• dim (int) – A dimension along which softmax will be computed. Default: -1.

Returns

Sampled tensor of same shape as logits from the Gumbel-Softmax distribution. If hard=True, the returned samples will be one-hot, otherwise they will be probability distributions that sum to 1 across dim.

Note

This function is here for legacy reasons, may be removed from nn.Functional in the future.

Note

The main trick for hard is to do y_hard - y_soft.detach() + y_soft

It achieves two things: - makes the output value exactly one-hot (since we add then subtract y_soft value) - makes the gradient equal to y_soft gradient (since we strip all other gradients)

Examples::

>>> logits = torch.randn(20, 32)
>>> # Sample soft categorical using reparametrization trick:
>>> F.gumbel_softmax(logits, tau=1, hard=False)
>>> # Sample hard categorical using "Straight-through" trick:
>>> F.gumbel_softmax(logits, tau=1, hard=True)

log_softmax¶

torch.nn.functional.log_softmax(input, dim=None, _stacklevel=3, dtype=None)[source]¶

Applies a softmax followed by a logarithm.

While mathematically equivalent to log(softmax(x)), doing these two operations separately is slower, and numerically unstable. This function uses an alternative formulation to compute the output and gradient correctly.

See LogSoftmax for more details.

Parameters

• input (Tensor) – input

• dim (int) – A dimension along which log_softmax will be computed.

• dtype (torch.dtype, optional) – the desired data type of returned tensor. If specified, the input tensor is casted to dtype before the operation is performed. This is useful for preventing data type overflows. Default: None.

tanh¶

torch.nn.functional.tanh(input) → Tensor[source]¶

Applies element-wise, $\text{Tanh}(x) = \tanh(x) = \frac{\exp(x) - \exp(-x)}{\exp(x) + \exp(-x)}$

See Tanh for more details.

sigmoid¶

torch.nn.functional.sigmoid(input) → Tensor[source]¶

Applies the element-wise function $\text{Sigmoid}(x) = \frac{1}{1 + \exp(-x)}$

See Sigmoid for more details.

hardsigmoid¶

torch.nn.functional.hardsigmoid(input) → Tensor[source]¶

Applies the element-wise function

$$\text{Hardsigmoid}(x) = \begin{cases} 0 & \text{if~} x \le -3, \\ 1 & \text{if~} x \ge +3, \\ x / 6 + 1 / 2 & \text{otherwise} \end{cases}$$

Parameters

inplace – If set to True, will do this operation in-place. Default: False

See Hardsigmoid for more details.

silu¶

torch.nn.functional.silu(input, inplace=False)[source]¶

Applies the silu function, element-wise.

$$\text{silu}(x) = x * \sigma(x), \text{where } \sigma(x) \text{ is the logistic sigmoid.}$$

Note

See Gaussian Error Linear Units (GELUs) where the SiLU (Sigmoid Linear Unit) was originally coined, and see Sigmoid-Weighted Linear Units for Neural Network Function Approximation in Reinforcement Learning and Swish: a Self-Gated Activation Function where the SiLU was experimented with later.

See SiLU for more details.

batch_norm¶

torch.nn.functional.batch_norm(input, running_mean, running_var, weight=None, bias=None, training=False, momentum=0.1, eps=1e-05)[source]¶

Applies Batch Normalization for each channel across a batch of data.

See BatchNorm1d, BatchNorm2d, BatchNorm3d for details.

instance_norm¶

torch.nn.functional.instance_norm(input, running_mean=None, running_var=None, weight=None, bias=None, use_input_stats=True, momentum=0.1, eps=1e-05)[source]¶

Applies Instance Normalization for each channel in each data sample in a batch.

See InstanceNorm1d, InstanceNorm2d, InstanceNorm3d for details.

layer_norm¶

torch.nn.functional.layer_norm(input, normalized_shape, weight=None, bias=None, eps=1e-05)[source]¶

Applies Layer Normalization for last certain number of dimensions.

See LayerNorm for details.

local_response_norm¶

torch.nn.functional.local_response_norm(input, size, alpha=0.0001, beta=0.75, k=1.0)[source]¶

Applies local response normalization over an input signal composed of several input planes, where channels occupy the second dimension. Applies normalization across channels.

See LocalResponseNorm for details.

normalize¶

torch.nn.functional.normalize(input, p=2, dim=1, eps=1e-12, out=None)[source]¶

Performs $L_p$ normalization of inputs over specified dimension.

For a tensor input of sizes $(n_0, ..., n_{dim}, ..., n_k)$ , each $n_{dim}$ -element vector $v$ along dimension dim is transformed as

$$v = \frac{v}{\max(\lVert v \rVert_p, \epsilon)}.$$

With the default arguments it uses the Euclidean norm over vectors along dimension $1$ for normalization.

Parameters

• input – input tensor of any shape

• p (float) – the exponent value in the norm formulation. Default: 2

• dim (int) – the dimension to reduce. Default: 1

• eps (float) – small value to avoid division by zero. Default: 1e-12

• out (Tensor, optional) – the output tensor. If out is used, this operation won’t be differentiable.

linear¶

torch.nn.functional.linear(input, weight, bias=None)[source]¶

Applies a linear transformation to the incoming data: $y = xA^T + b$ .

This operator supports TensorFloat32.

Shape:

• Input: $(N, *, in\_features)$ N is the batch size, * means any number of additional dimensions

• Weight: $(out\_features, in\_features)$

• Bias: $(out\_features)$

• Output: $(N, *, out\_features)$

bilinear¶

torch.nn.functional.bilinear(input1, input2, weight, bias=None)[source]¶

Applies a bilinear transformation to the incoming data: $y = x_1^T A x_2 + b$

Shape:

• input1: $(N, *, H_{in1})$ where $H_{in1}=\text{in1\_features}$ and $*$ means any number of additional dimensions. All but the last dimension of the inputs should be the same.

• input2: $(N, *, H_{in2})$ where $H_{in2}=\text{in2\_features}$

• weight: $(\text{out\_features}, \text{in1\_features}, \text{in2\_features})$

• bias: $(\text{out\_features})$

• output: $(N, *, H_{out})$ where $H_{out}=\text{out\_features}$ and all but the last dimension are the same shape as the input.

dropout¶

torch.nn.functional.dropout(input, p=0.5, training=True, inplace=False)[source]¶

During training, randomly zeroes some of the elements of the input tensor with probability p using samples from a Bernoulli distribution.

See Dropout for details.

Parameters

• p – probability of an element to be zeroed. Default: 0.5

• training – apply dropout if is True. Default: True

• inplace – If set to True, will do this operation in-place. Default: False

alpha_dropout¶

torch.nn.functional.alpha_dropout(input, p=0.5, training=False, inplace=False)[source]¶

Applies alpha dropout to the input.

See AlphaDropout for details.

feature_alpha_dropout¶

torch.nn.functional.feature_alpha_dropout(input, p=0.5, training=False, inplace=False)[source]¶

Randomly masks out entire channels (a channel is a feature map, e.g. the $j$ -th channel of the $i$ -th sample in the batch input is a tensor $\text{input}[i, j]$ ) of the input tensor). Instead of setting activations to zero, as in regular Dropout, the activations are set to the negative saturation value of the SELU activation function.

Each element will be masked independently on every forward call with probability p using samples from a Bernoulli distribution. The elements to be masked are randomized on every forward call, and scaled and shifted to maintain zero mean and unit variance.

See FeatureAlphaDropout for details.

Parameters

• p – dropout probability of a channel to be zeroed. Default: 0.5

• training – apply dropout if is True. Default: True

• inplace – If set to True, will do this operation in-place. Default: False

dropout2d¶

torch.nn.functional.dropout2d(input, p=0.5, training=True, inplace=False)[source]¶

Randomly zero out entire channels (a channel is a 2D feature map, e.g., the $j$ -th channel of the $i$ -th sample in the batched input is a 2D tensor $\text{input}[i, j]$ ) of the input tensor). Each channel will be zeroed out independently on every forward call with probability p using samples from a Bernoulli distribution.

See Dropout2d for details.

Parameters

• p – probability of a channel to be zeroed. Default: 0.5

• training – apply dropout if is True. Default: True

• inplace – If set to True, will do this operation in-place. Default: False

dropout3d¶

torch.nn.functional.dropout3d(input, p=0.5, training=True, inplace=False)[source]¶

Randomly zero out entire channels (a channel is a 3D feature map, e.g., the $j$ -th channel of the $i$ -th sample in the batched input is a 3D tensor $\text{input}[i, j]$ ) of the input tensor). Each channel will be zeroed out independently on every forward call with probability p using samples from a Bernoulli distribution.

See Dropout3d for details.

Parameters

• p – probability of a channel to be zeroed. Default: 0.5

• training – apply dropout if is True. Default: True

• inplace – If set to True, will do this operation in-place. Default: False

embedding¶

torch.nn.functional.embedding(input, weight, padding_idx=None, max_norm=None, norm_type=2.0, scale_grad_by_freq=False, sparse=False)[source]¶

A simple lookup table that looks up embeddings in a fixed dictionary and size.

This module is often used to retrieve word embeddings using indices. The input to the module is a list of indices, and the embedding matrix, and the output is the corresponding word embeddings.

See torch.nn.Embedding for more details.

Parameters

• input (LongTensor) – Tensor containing indices into the embedding matrix

• weight (Tensor) – The embedding matrix with number of rows equal to the maximum possible index + 1, and number of columns equal to the embedding size

• padding_idx (int, optional) – If given, pads the output with the embedding vector at padding_idx (initialized to zeros) whenever it encounters the index.

• max_norm (float, optional) – If given, each embedding vector with norm larger than max_norm is renormalized to have norm max_norm. Note: this will modify weight in-place.

• norm_type (float, optional) – The p of the p-norm to compute for the max_norm option. Default 2.

• scale_grad_by_freq (boolean, optional) – If given, this will scale gradients by the inverse of frequency of the words in the mini-batch. Default False.

• sparse (bool, optional) – If True, gradient w.r.t. weight will be a sparse tensor. See Notes under torch.nn.Embedding for more details regarding sparse gradients.

Shape:

• Input: LongTensor of arbitrary shape containing the indices to extract

• Weight: Embedding matrix of floating point type with shape (V, embedding_dim),

  where V = maximum index + 1 and embedding_dim = the embedding size

• Output: (*, embedding_dim), where * is the input shape

Examples:

>>> # a batch of 2 samples of 4 indices each
>>> input = torch.tensor([[1,2,4,5],[4,3,2,9]])
>>> # an embedding matrix containing 10 tensors of size 3
>>> embedding_matrix = torch.rand(10, 3)
>>> F.embedding(input, embedding_matrix)
tensor([[[ 0.8490,  0.9625,  0.6753],
         [ 0.9666,  0.7761,  0.6108],
         [ 0.6246,  0.9751,  0.3618],
         [ 0.4161,  0.2419,  0.7383]],

        [[ 0.6246,  0.9751,  0.3618],
         [ 0.0237,  0.7794,  0.0528],
         [ 0.9666,  0.7761,  0.6108],
         [ 0.3385,  0.8612,  0.1867]]])

>>> # example with padding_idx
>>> weights = torch.rand(10, 3)
>>> weights[0, :].zero_()
>>> embedding_matrix = weights
>>> input = torch.tensor([[0,2,0,5]])
>>> F.embedding(input, embedding_matrix, padding_idx=0)
tensor([[[ 0.0000,  0.0000,  0.0000],
         [ 0.5609,  0.5384,  0.8720],
         [ 0.0000,  0.0000,  0.0000],
         [ 0.6262,  0.2438,  0.7471]]])

embedding_bag¶

torch.nn.functional.embedding_bag(input, weight, offsets=None, max_norm=None, norm_type=2, scale_grad_by_freq=False, mode='mean', sparse=False, per_sample_weights=None, include_last_offset=False)[source]¶

Computes sums, means or maxes of bags of embeddings, without instantiating the intermediate embeddings.

See torch.nn.EmbeddingBag for more details.

Note

When using the CUDA backend, this operation may induce nondeterministic behaviour in its backward pass that is not easily switched off. Please see the notes on Reproducibility for background.

Parameters

• input (LongTensor) – Tensor containing bags of indices into the embedding matrix

• weight (Tensor) – The embedding matrix with number of rows equal to the maximum possible index + 1, and number of columns equal to the embedding size

• offsets (LongTensor, optional) – Only used when input is 1D. offsets determines the starting index position of each bag (sequence) in input.

• max_norm (float, optional) – If given, each embedding vector with norm larger than max_norm is renormalized to have norm max_norm. Note: this will modify weight in-place.

• norm_type (float, optional) – The p in the p-norm to compute for the max_norm option. Default 2.

• scale_grad_by_freq (boolean, optional) – if given, this will scale gradients by the inverse of frequency of the words in the mini-batch. Default False. Note: this option is not supported when mode="max".

• mode (string, optional) – "sum", "mean" or "max". Specifies the way to reduce the bag. Default: "mean"

• sparse (bool, optional) – if True, gradient w.r.t. weight will be a sparse tensor. See Notes under torch.nn.Embedding for more details regarding sparse gradients. Note: this option is not supported when mode="max".

• per_sample_weights (Tensor, optional) – a tensor of float / double weights, or None to indicate all weights should be taken to be 1. If specified, per_sample_weights must have exactly the same shape as input and is treated as having the same offsets, if those are not None.

• include_last_offset (bool, optional) – if True, the size of offsets is equal to the number of bags + 1.

• last element is the size of the input, or the ending index position of the last bag (The) –

Shape:

• input (LongTensor) and offsets (LongTensor, optional)

  • If input is 2D of shape (B, N),

    it will be treated as B bags (sequences) each of fixed length N, and this will return B values aggregated in a way depending on the mode. offsets is ignored and required to be None in this case.

  • If input is 1D of shape (N),

    it will be treated as a concatenation of multiple bags (sequences). offsets is required to be a 1D tensor containing the starting index positions of each bag in input. Therefore, for offsets of shape (B), input will be viewed as having B bags. Empty bags (i.e., having 0-length) will have returned vectors filled by zeros.

• weight (Tensor): the learnable weights of the module of shape (num_embeddings, embedding_dim)

• per_sample_weights (Tensor, optional). Has the same shape as input.

• output: aggregated embedding values of shape (B, embedding_dim)

Examples:

>>> # an Embedding module containing 10 tensors of size 3
>>> embedding_matrix = torch.rand(10, 3)
>>> # a batch of 2 samples of 4 indices each
>>> input = torch.tensor([1,2,4,5,4,3,2,9])
>>> offsets = torch.tensor([0,4])
>>> F.embedding_bag(embedding_matrix, input, offsets)
tensor([[ 0.3397,  0.3552,  0.5545],
        [ 0.5893,  0.4386,  0.5882]])

one_hot¶

torch.nn.functional.one_hot(tensor, num_classes=-1) → LongTensor¶

Takes LongTensor with index values of shape (*) and returns a tensor of shape (*, num_classes) that have zeros everywhere except where the index of last dimension matches the corresponding value of the input tensor, in which case it will be 1.

See also One-hot on Wikipedia .

Parameters

• tensor (LongTensor) – class values of any shape.

• num_classes (int) – Total number of classes. If set to -1, the number of classes will be inferred as one greater than the largest class value in the input tensor.

Returns

LongTensor that has one more dimension with 1 values at the index of last dimension indicated by the input, and 0 everywhere else.

Examples

>>> F.one_hot(torch.arange(0, 5) % 3)
tensor([[1, 0, 0],
        [0, 1, 0],
        [0, 0, 1],
        [1, 0, 0],
        [0, 1, 0]])
>>> F.one_hot(torch.arange(0, 5) % 3, num_classes=5)
tensor([[1, 0, 0, 0, 0],
        [0, 1, 0, 0, 0],
        [0, 0, 1, 0, 0],
        [1, 0, 0, 0, 0],
        [0, 1, 0, 0, 0]])
>>> F.one_hot(torch.arange(0, 6).view(3,2) % 3)
tensor([[[1, 0, 0],
         [0, 1, 0]],
        [[0, 0, 1],
         [1, 0, 0]],
        [[0, 1, 0],
         [0, 0, 1]]])

pairwise_distance¶

torch.nn.functional.pairwise_distance(x1, x2, p=2.0, eps=1e-06, keepdim=False)[source]¶

See torch.nn.PairwiseDistance for details

cosine_similarity¶

torch.nn.functional.cosine_similarity(x1, x2, dim=1, eps=1e-8) → Tensor¶

Returns cosine similarity between x1 and x2, computed along dim.

$$\text{similarity} = \dfrac{x_1 \cdot x_2}{\max(\Vert x_1 \Vert _2 \cdot \Vert x_2 \Vert _2, \epsilon)}$$

Parameters

• x1 (Tensor) – First input.

• x2 (Tensor) – Second input (of size matching x1).

• dim (int, optional) – Dimension of vectors. Default: 1

• eps (float, optional) – Small value to avoid division by zero. Default: 1e-8

Shape:

• Input: $(\ast_1, D, \ast_2)$ where D is at position dim.

• Output: $(\ast_1, \ast_2)$ where 1 is at position dim.

Example:

>>> input1 = torch.randn(100, 128)
>>> input2 = torch.randn(100, 128)
>>> output = F.cosine_similarity(input1, input2)
>>> print(output)

pdist¶

torch.nn.functional.pdist(input, p=2) → Tensor¶

Computes the p-norm distance between every pair of row vectors in the input. This is identical to the upper triangular portion, excluding the diagonal, of torch.norm(input[:, None] - input, dim=2, p=p). This function will be faster if the rows are contiguous.

If input has shape $N \times M$ then the output will have shape $\frac{1}{2} N (N - 1)$ .

This function is equivalent to scipy.spatial.distance.pdist(input, ‘minkowski’, p=p) if $p \in (0, \infty)$ . When $p = 0$ it is equivalent to scipy.spatial.distance.pdist(input, ‘hamming’) * M. When $p = \infty$ , the closest scipy function is scipy.spatial.distance.pdist(xn, lambda x, y: np.abs(x - y).max()).

Parameters

• input – input tensor of shape $N \times M$ .

• p – p value for the p-norm distance to calculate between each vector pair $\in [0, \infty]$ .

binary_cross_entropy¶

torch.nn.functional.binary_cross_entropy(input, target, weight=None, size_average=None, reduce=None, reduction='mean')[source]¶

Function that measures the Binary Cross Entropy between the target and the output.

See BCELoss for details.

Parameters

• input – Tensor of arbitrary shape

• target – Tensor of the same shape as input

• weight (Tensor, optional) – a manual rescaling weight if provided it’s repeated to match input tensor shape

• size_average (bool, optional) – Deprecated (see reduction). By default, the losses are averaged over each loss element in the batch. Note that for some losses, there multiple elements per sample. If the field size_average is set to False, the losses are instead summed for each minibatch. Ignored when reduce is False. Default: True

• reduce (bool, optional) – Deprecated (see reduction). By default, the losses are averaged or summed over observations for each minibatch depending on size_average. When reduce is False, returns a loss per batch element instead and ignores size_average. Default: True

• reduction (string, optional) – Specifies the reduction to apply to the output: 'none' | 'mean' | 'sum'. 'none': no reduction will be applied, 'mean': the sum of the output will be divided by the number of elements in the output, 'sum': the output will be summed. Note: size_average and reduce are in the process of being deprecated, and in the meantime, specifying either of those two args will override reduction. Default: 'mean'

Examples:

>>> input = torch.randn((3, 2), requires_grad=True)
>>> target = torch.rand((3, 2), requires_grad=False)
>>> loss = F.binary_cross_entropy(F.sigmoid(input), target)
>>> loss.backward()

binary_cross_entropy_with_logits¶

torch.nn.functional.binary_cross_entropy_with_logits(input, target, weight=None, size_average=None, reduce=None, reduction='mean', pos_weight=None)[source]¶

Function that measures Binary Cross Entropy between target and output logits.

See BCEWithLogitsLoss for details.

Parameters

• input – Tensor of arbitrary shape

• target – Tensor of the same shape as input

• weight (Tensor, optional) – a manual rescaling weight if provided it’s repeated to match input tensor shape

• size_average (bool, optional) – Deprecated (see reduction). By default, the losses are averaged over each loss element in the batch. Note that for some losses, there multiple elements per sample. If the field size_average is set to False, the losses are instead summed for each minibatch. Ignored when reduce is False. Default: True

• reduce (bool, optional) – Deprecated (see reduction). By default, the losses are averaged or summed over observations for each minibatch depending on size_average. When reduce is False, returns a loss per batch element instead and ignores size_average. Default: True

• reduction (string, optional) – Specifies the reduction to apply to the output: 'none' | 'mean' | 'sum'. 'none': no reduction will be applied, 'mean': the sum of the output will be divided by the number of elements in the output, 'sum': the output will be summed. Note: size_average and reduce are in the process of being deprecated, and in the meantime, specifying either of those two args will override reduction. Default: 'mean'

• pos_weight (Tensor, optional) – a weight of positive examples. Must be a vector with length equal to the number of classes.

Examples:

>>> input = torch.randn(3, requires_grad=True)
>>> target = torch.empty(3).random_(2)
>>> loss = F.binary_cross_entropy_with_logits(input, target)
>>> loss.backward()

poisson_nll_loss¶

torch.nn.functional.poisson_nll_loss(input, target, log_input=True, full=False, size_average=None, eps=1e-08, reduce=None, reduction='mean')[source]¶

Poisson negative log likelihood loss.

See PoissonNLLLoss for details.

Parameters

• input – expectation of underlying Poisson distribution.

• target – random sample $target \sim \text{Poisson}(input)$ .

• log_input – if True the loss is computed as $\exp(\text{input}) - \text{target} * \text{input}$ , if False then loss is $\text{input} - \text{target} * \log(\text{input}+\text{eps})$ . Default: True

• full – whether to compute full loss, i. e. to add the Stirling approximation term. Default: False $\text{target} * \log(\text{target}) - \text{target} + 0.5 * \log(2 * \pi * \text{target})$ .

• size_average (bool, optional) – Deprecated (see reduction). By default, the losses are averaged over each loss element in the batch. Note that for some losses, there multiple elements per sample. If the field size_average is set to False, the losses are instead summed for each minibatch. Ignored when reduce is False. Default: True

• eps (float, optional) – Small value to avoid evaluation of $\log(0)$ when log_input`=``False`. Default: 1e-8

• reduce (bool, optional) – Deprecated (see reduction). By default, the losses are averaged or summed over observations for each minibatch depending on size_average. When reduce is False, returns a loss per batch element instead and ignores size_average. Default: True

• reduction (string, optional) – Specifies the reduction to apply to the output: 'none' | 'mean' | 'sum'. 'none': no reduction will be applied, 'mean': the sum of the output will be divided by the number of elements in the output, 'sum': the output will be summed. Note: size_average and reduce are in the process of being deprecated, and in the meantime, specifying either of those two args will override reduction. Default: 'mean'

cosine_embedding_loss¶

torch.nn.functional.cosine_embedding_loss(input1, input2, target, margin=0, size_average=None, reduce=None, reduction='mean') → Tensor[source]¶

See CosineEmbeddingLoss for details.

cross_entropy¶

torch.nn.functional.cross_entropy(input, target, weight=None, size_average=None, ignore_index=-100, reduce=None, reduction='mean')[source]¶

This criterion combines log_softmax and nll_loss in a single function.

See CrossEntropyLoss for details.

Parameters

• input (Tensor) – $(N, C)$ where C = number of classes or $(N, C, H, W)$ in case of 2D Loss, or $(N, C, d_1, d_2, ..., d_K)$ where $K \geq 1$ in the case of K-dimensional loss.

• target (Tensor) – $(N)$ where each value is $0 \leq \text{targets}[i] \leq C-1$ , or $(N, d_1, d_2, ..., d_K)$ where $K \geq 1$ for K-dimensional loss.

• weight (Tensor, optional) – a manual rescaling weight given to each class. If given, has to be a Tensor of size C

• size_average (bool, optional) – Deprecated (see reduction). By default, the losses are averaged over each loss element in the batch. Note that for some losses, there multiple elements per sample. If the field size_average is set to False, the losses are instead summed for each minibatch. Ignored when reduce is False. Default: True

• ignore_index (int, optional) – Specifies a target value that is ignored and does not contribute to the input gradient. When size_average is True, the loss is averaged over non-ignored targets. Default: -100

• reduce (bool, optional) – Deprecated (see reduction). By default, the losses are averaged or summed over observations for each minibatch depending on size_average. When reduce is False, returns a loss per batch element instead and ignores size_average. Default: True

• reduction (string, optional) – Specifies the reduction to apply to the output: 'none' | 'mean' | 'sum'. 'none': no reduction will be applied, 'mean': the sum of the output will be divided by the number of elements in the output, 'sum': the output will be summed. Note: size_average and reduce are in the process of being deprecated, and in the meantime, specifying either of those two args will override reduction. Default: 'mean'

Examples:

>>> input = torch.randn(3, 5, requires_grad=True)
>>> target = torch.randint(5, (3,), dtype=torch.int64)
>>> loss = F.cross_entropy(input, target)
>>> loss.backward()

ctc_loss¶

torch.nn.functional.ctc_loss(log_probs, targets, input_lengths, target_lengths, blank=0, reduction='mean', zero_infinity=False)[source]¶

The Connectionist Temporal Classification loss.

See CTCLoss for details.

Note

In some circumstances when using the CUDA backend with CuDNN, this operator may select a nondeterministic algorithm to increase performance. If this is undesirable, you can try to make the operation deterministic (potentially at a performance cost) by setting torch.backends.cudnn.deterministic = True. Please see the notes on Reproducibility for background.

Note

When using the CUDA backend, this operation may induce nondeterministic behaviour in its backward pass that is not easily switched off. Please see the notes on Reproducibility for background.

Parameters

• log_probs – $(T, N, C)$ where C = number of characters in alphabet including blank, T = input length, and N = batch size. The logarithmized probabilities of the outputs (e.g. obtained with torch.nn.functional.log_softmax()).

• targets – $(N, S)$ or (sum(target_lengths)). Targets cannot be blank. In the second form, the targets are assumed to be concatenated.

• input_lengths – $(N)$ . Lengths of the inputs (must each be $\leq T$ )

• target_lengths – $(N)$ . Lengths of the targets

• blank (int, optional) – Blank label. Default $0$ .

• reduction (string, optional) – Specifies the reduction to apply to the output: 'none' | 'mean' | 'sum'. 'none': no reduction will be applied, 'mean': the output losses will be divided by the target lengths and then the mean over the batch is taken, 'sum': the output will be summed. Default: 'mean'

• zero_infinity (bool, optional) – Whether to zero infinite losses and the associated gradients. Default: False Infinite losses mainly occur when the inputs are too short to be aligned to the targets.

Example:

>>> log_probs = torch.randn(50, 16, 20).log_softmax(2).detach().requires_grad_()
>>> targets = torch.randint(1, 20, (16, 30), dtype=torch.long)
>>> input_lengths = torch.full((16,), 50, dtype=torch.long)
>>> target_lengths = torch.randint(10,30,(16,), dtype=torch.long)
>>> loss = F.ctc_loss(log_probs, targets, input_lengths, target_lengths)
>>> loss.backward()

hinge_embedding_loss¶

torch.nn.functional.hinge_embedding_loss(input, target, margin=1.0, size_average=None, reduce=None, reduction='mean') → Tensor[source]¶

See HingeEmbeddingLoss for details.

kl_div¶

torch.nn.functional.kl_div(input, target, size_average=None, reduce=None, reduction='mean', log_target=False)[source]¶

The Kullback-Leibler divergence Loss

See KLDivLoss for details.

Parameters

• input – Tensor of arbitrary shape

• target – Tensor of the same shape as input

• size_average (bool, optional) – Deprecated (see reduction). By default, the losses are averaged over each loss element in the batch. Note that for some losses, there multiple elements per sample. If the field size_average is set to False, the losses are instead summed for each minibatch. Ignored when reduce is False. Default: True

• reduce (bool, optional) – Deprecated (see reduction). By default, the losses are averaged or summed over observations for each minibatch depending on size_average. When reduce is False, returns a loss per batch element instead and ignores size_average. Default: True

• reduction (string, optional) – Specifies the reduction to apply to the output: 'none' | 'batchmean' | 'sum' | 'mean'. 'none': no reduction will be applied 'batchmean': the sum of the output will be divided by the batchsize 'sum': the output will be summed 'mean': the output will be divided by the number of elements in the output Default: 'mean'

• log_target (bool) – A flag indicating whether target is passed in the log space. It is recommended to pass certain distributions (like softmax) in the log space to avoid numerical issues caused by explicit log. Default: False

Note

size_average and reduce are in the process of being deprecated, and in the meantime, specifying either of those two args will override reduction.

Note

:attr:reduction = 'mean' doesn’t return the true kl divergence value, please use :attr:reduction = 'batchmean' which aligns with KL math definition. In the next major release, 'mean' will be changed to be the same as ‘batchmean’.

l1_loss¶

torch.nn.functional.l1_loss(input, target, size_average=None, reduce=None, reduction='mean') → Tensor[source]¶

Function that takes the mean element-wise absolute value difference.

See L1Loss for details.

mse_loss¶

torch.nn.functional.mse_loss(input, target, size_average=None, reduce=None, reduction='mean') → Tensor[source]¶

Measures the element-wise mean squared error.

See MSELoss for details.

margin_ranking_loss¶

torch.nn.functional.margin_ranking_loss(input1, input2, target, margin=0, size_average=None, reduce=None, reduction='mean') → Tensor[source]¶

See MarginRankingLoss for details.

multilabel_margin_loss¶

torch.nn.functional.multilabel_margin_loss(input, target, size_average=None, reduce=None, reduction='mean') → Tensor[source]¶

See MultiLabelMarginLoss for details.

multilabel_soft_margin_loss¶

torch.nn.functional.multilabel_soft_margin_loss(input, target, weight=None, size_average=None) → Tensor[source]¶

See MultiLabelSoftMarginLoss for details.

multi_margin_loss¶

torch.nn.functional.multi_margin_loss(input, target, p=1, margin=1.0, weight=None, size_average=None, reduce=None, reduction='mean')[source]¶

multi_margin_loss(input, target, p=1, margin=1, weight=None, size_average=None,

reduce=None, reduction=’mean’) -> Tensor

See MultiMarginLoss for details.

nll_loss¶

torch.nn.functional.nll_loss(input, target, weight=None, size_average=None, ignore_index=-100, reduce=None, reduction='mean')[source]¶

The negative log likelihood loss.

See NLLLoss for details.

Parameters

• input – $(N, C)$ where C = number of classes or $(N, C, H, W)$ in case of 2D Loss, or $(N, C, d_1, d_2, ..., d_K)$ where $K \geq 1$ in the case of K-dimensional loss.

• target – $(N)$ where each value is $0 \leq \text{targets}[i] \leq C-1$ , or $(N, d_1, d_2, ..., d_K)$ where $K \geq 1$ for K-dimensional loss.

• weight (Tensor, optional) – a manual rescaling weight given to each class. If given, has to be a Tensor of size C

• size_average (bool, optional) – Deprecated (see reduction). By default, the losses are averaged over each loss element in the batch. Note that for some losses, there multiple elements per sample. If the field size_average is set to False, the losses are instead summed for each minibatch. Ignored when reduce is False. Default: True

• ignore_index (int, optional) – Specifies a target value that is ignored and does not contribute to the input gradient. When size_average is True, the loss is averaged over non-ignored targets. Default: -100

• reduce (bool, optional) – Deprecated (see reduction). By default, the losses are averaged or summed over observations for each minibatch depending on size_average. When reduce is False, returns a loss per batch element instead and ignores size_average. Default: True

• reduction (string, optional) – Specifies the reduction to apply to the output: 'none' | 'mean' | 'sum'. 'none': no reduction will be applied, 'mean': the sum of the output will be divided by the number of elements in the output, 'sum': the output will be summed. Note: size_average and reduce are in the process of being deprecated, and in the meantime, specifying either of those two args will override reduction. Default: 'mean'

Example:

>>> # input is of size N x C = 3 x 5
>>> input = torch.randn(3, 5, requires_grad=True)
>>> # each element in target has to have 0 <= value < C
>>> target = torch.tensor([1, 0, 4])
>>> output = F.nll_loss(F.log_softmax(input), target)
>>> output.backward()

smooth_l1_loss¶

torch.nn.functional.smooth_l1_loss(input, target, size_average=None, reduce=None, reduction='mean', beta=1.0)[source]¶

Function that uses a squared term if the absolute element-wise error falls below beta and an L1 term otherwise.

See SmoothL1Loss for details.

soft_margin_loss¶

torch.nn.functional.soft_margin_loss(input, target, size_average=None, reduce=None, reduction='mean') → Tensor[source]¶

See SoftMarginLoss for details.

triplet_margin_loss¶

torch.nn.functional.triplet_margin_loss(anchor, positive, negative, margin=1.0, p=2, eps=1e-06, swap=False, size_average=None, reduce=None, reduction='mean')[source]¶

See TripletMarginLoss for details

triplet_margin_with_distance_loss¶

torch.nn.functional.triplet_margin_with_distance_loss(anchor, positive, negative, *, distance_function=None, margin=1.0, swap=False, reduction='mean')[source]¶

See TripletMarginWithDistanceLoss for details.

pixel_shuffle¶

torch.nn.functional.pixel_shuffle()¶

Rearranges elements in a tensor of shape $(*, C \times r^2, H, W)$ to a tensor of shape $(*, C, H \times r, W \times r)$ .

See PixelShuffle for details.

Parameters

• input (Tensor) – the input tensor

• upscale_factor (int) – factor to increase spatial resolution by

Examples:

>>> input = torch.randn(1, 9, 4, 4)
>>> output = torch.nn.functional.pixel_shuffle(input, 3)
>>> print(output.size())
torch.Size([1, 1, 12, 12])

pad¶

torch.nn.functional.pad(input, pad, mode='constant', value=0)¶

Pads tensor.

Padding size:

The padding size by which to pad some dimensions of input are described starting from the last dimension and moving forward. $\left\lfloor\frac{\text{len(pad)}}{2}\right\rfloor$ dimensions of input will be padded. For example, to pad only the last dimension of the input tensor, then pad has the form $(\text{padding\_left}, \text{padding\_right})$ ; to pad the last 2 dimensions of the input tensor, then use $(\text{padding\_left}, \text{padding\_right},$ $\text{padding\_top}, \text{padding\_bottom})$ ; to pad the last 3 dimensions, use $(\text{padding\_left}, \text{padding\_right},$ $\text{padding\_top}, \text{padding\_bottom}$ $\text{padding\_front}, \text{padding\_back})$ .

Padding mode:

See torch.nn.ConstantPad2d, torch.nn.ReflectionPad2d, and torch.nn.ReplicationPad2d for concrete examples on how each of the padding modes works. Constant padding is implemented for arbitrary dimensions. Replicate padding is implemented for padding the last 3 dimensions of 5D input tensor, or the last 2 dimensions of 4D input tensor, or the last dimension of 3D input tensor. Reflect padding is only implemented for padding the last 2 dimensions of 4D input tensor, or the last dimension of 3D input tensor.

Note

When using the CUDA backend, this operation may induce nondeterministic behaviour in its backward pass that is not easily switched off. Please see the notes on Reproducibility for background.

Parameters

• input (Tensor) – N-dimensional tensor

• pad (tuple) – m-elements tuple, where $\frac{m}{2} \leq$ input dimensions and $m$ is even.

• mode – 'constant', 'reflect', 'replicate' or 'circular'. Default: 'constant'

• value – fill value for 'constant' padding. Default: 0

Examples:

>>> t4d = torch.empty(3, 3, 4, 2)
>>> p1d = (1, 1) # pad last dim by 1 on each side
>>> out = F.pad(t4d, p1d, "constant", 0)  # effectively zero padding
>>> print(out.size())
torch.Size([3, 3, 4, 4])
>>> p2d = (1, 1, 2, 2) # pad last dim by (1, 1) and 2nd to last by (2, 2)
>>> out = F.pad(t4d, p2d, "constant", 0)
>>> print(out.size())
torch.Size([3, 3, 8, 4])
>>> t4d = torch.empty(3, 3, 4, 2)
>>> p3d = (0, 1, 2, 1, 3, 3) # pad by (0, 1), (2, 1), and (3, 3)
>>> out = F.pad(t4d, p3d, "constant", 0)
>>> print(out.size())
torch.Size([3, 9, 7, 3])

interpolate¶

torch.nn.functional.interpolate(input, size=None, scale_factor=None, mode='nearest', align_corners=None, recompute_scale_factor=None)[source]¶

Down/up samples the input to either the given size or the given scale_factor

The algorithm used for interpolation is determined by mode.

Currently temporal, spatial and volumetric sampling are supported, i.e. expected inputs are 3-D, 4-D or 5-D in shape.

The input dimensions are interpreted in the form: mini-batch x channels x [optional depth] x [optional height] x width.

The modes available for resizing are: nearest, linear (3D-only), bilinear, bicubic (4D-only), trilinear (5D-only), area

Parameters

• input (Tensor) – the input tensor

• size (int or Tuple[int] or Tuple[int, int] or Tuple[int, int, int]) – output spatial size.

• scale_factor (float or Tuple[float]) – multiplier for spatial size. Has to match input size if it is a tuple.

• mode (str) – algorithm used for upsampling: 'nearest' | 'linear' | 'bilinear' | 'bicubic' | 'trilinear' | 'area'. Default: 'nearest'

• align_corners (bool, optional) – Geometrically, we consider the pixels of the input and output as squares rather than points. If set to True, the input and output tensors are aligned by the center points of their corner pixels, preserving the values at the corner pixels. If set to False, the input and output tensors are aligned by the corner points of their corner pixels, and the interpolation uses edge value padding for out-of-boundary values, making this operation independent of input size when scale_factor is kept the same. This only has an effect when mode is 'linear', 'bilinear', 'bicubic' or 'trilinear'. Default: False

• recompute_scale_factor (bool, optional) – recompute the scale_factor for use in the interpolation calculation. When scale_factor is passed as a parameter, it is used to compute the output_size. If recompute_scale_factor is `False or not specified, the passed-in scale_factor will be used in the interpolation computation. Otherwise, a new scale_factor will be computed based on the output and input sizes for use in the interpolation computation (i.e. the computation will be identical to if the computed output_size were passed-in explicitly). Note that when scale_factor is floating-point, the recomputed scale_factor may differ from the one passed in due to rounding and precision issues.

Note

With mode='bicubic', it’s possible to cause overshoot, in other words it can produce negative values or values greater than 255 for images. Explicitly call result.clamp(min=0, max=255) if you want to reduce the overshoot when displaying the image.

Warning

With align_corners = True, the linearly interpolating modes (linear, bilinear, and trilinear) don’t proportionally align the output and input pixels, and thus the output values can depend on the input size. This was the default behavior for these modes up to version 0.3.1. Since then, the default behavior is align_corners = False. See Upsample for concrete examples on how this affects the outputs.

Warning

When scale_factor is specified, if recompute_scale_factor=True, scale_factor is used to compute the output_size which will then be used to infer new scales for the interpolation. The default behavior for recompute_scale_factor changed to False in 1.6.0, and scale_factor is used in the interpolation calculation.

Note

When using the CUDA backend, this operation may induce nondeterministic behaviour in its backward pass that is not easily switched off. Please see the notes on Reproducibility for background.

upsample¶

torch.nn.functional.upsample(input, size=None, scale_factor=None, mode='nearest', align_corners=None)[source]¶

Upsamples the input to either the given size or the given scale_factor

Warning

This function is deprecated in favor of torch.nn.functional.interpolate(). This is equivalent with nn.functional.interpolate(...).

Note

When using the CUDA backend, this operation may induce nondeterministic behaviour in its backward pass that is not easily switched off. Please see the notes on Reproducibility for background.

The algorithm used for upsampling is determined by mode.

Currently temporal, spatial and volumetric upsampling are supported, i.e. expected inputs are 3-D, 4-D or 5-D in shape.

The input dimensions are interpreted in the form: mini-batch x channels x [optional depth] x [optional height] x width.

The modes available for upsampling are: nearest, linear (3D-only), bilinear, bicubic (4D-only), trilinear (5D-only)

Parameters

• input (Tensor) – the input tensor

• size (int or Tuple[int] or Tuple[int, int] or Tuple[int, int, int]) – output spatial size.

• scale_factor (float or Tuple[float]) – multiplier for spatial size. Has to match input size if it is a tuple.

• mode (string) – algorithm used for upsampling: 'nearest' | 'linear' | 'bilinear' | 'bicubic' | 'trilinear'. Default: 'nearest'

• align_corners (bool, optional) – Geometrically, we consider the pixels of the input and output as squares rather than points. If set to True, the input and output tensors are aligned by the center points of their corner pixels, preserving the values at the corner pixels. If set to False, the input and output tensors are aligned by the corner points of their corner pixels, and the interpolation uses edge value padding for out-of-boundary values, making this operation independent of input size when scale_factor is kept the same. This only has an effect when mode is 'linear', 'bilinear', 'bicubic' or 'trilinear'. Default: False

Note

With mode='bicubic', it’s possible to cause overshoot, in other words it can produce negative values or values greater than 255 for images. Explicitly call result.clamp(min=0, max=255) if you want to reduce the overshoot when displaying the image.

Warning

With align_corners = True, the linearly interpolating modes (linear, bilinear, and trilinear) don’t proportionally align the output and input pixels, and thus the output values can depend on the input size. This was the default behavior for these modes up to version 0.3.1. Since then, the default behavior is align_corners = False. See Upsample for concrete examples on how this affects the outputs.