o hG@sddlZddlmZddlZddlmmZddlmZddl m Z m Z m Z ddl mZddlmZddlmZgd ZGd d d eZGd d d eZGdddeZGdddeZGdddeZGdddeZGdddeZGdddeZGdddeZGdddeZGdddeZGd d!d!eZGd"d#d#eZ Gd$d%d%eZ!Gd&d'd'eZ"Gd(d)d)eZ#Gd*d+d+eZ$Gd,d-d-eZ%Gd.d/d/eZ&Gd0d1d1eZ'Gd2d3d3eZ(d4eejd5e)fd6d7Z*d4eejd5e)fd8d9Z+d:d;Z,Gdd?d?eZ.Gd@dAdAeZ/GdBdCdCeZ0GdDdEdEeZ1GdFdGdGeZ2GdHdIdIeZ3GdJdKdKeZ4dS)LN)Optional)Tensor) constant_xavier_normal_xavier_uniform_) Parameter)NonDynamicallyQuantizableLinear)Module) ThresholdReLURReLUHardtanhReLU6Sigmoid HardsigmoidTanhSiLUMish HardswishELUCELUSELUGLUGELU Hardshrink LeakyReLU LogSigmoidSoftplus SoftshrinkMultiheadAttentionPReLUSoftsign TanhshrinkSoftminSoftmax Softmax2d LogSoftmaxc speZdZUdZgdZeed<eed<eed<ddedededdffd d Zd e de fd d Z ddZ Z S)r aThresholds each element of the input Tensor. Threshold is defined as: .. math:: y = \begin{cases} x, &\text{ if } x > \text{threshold} \\ \text{value}, &\text{ otherwise } \end{cases} Args: threshold: The value to threshold at value: The value to replace with inplace: can optionally do the operation in-place. Default: ``False`` Shape: - Input: :math:`(*)`, where :math:`*` means any number of dimensions. - Output: :math:`(*)`, same shape as the input. Examples:: >>> m = nn.Threshold(0.1, 20) >>> input = torch.randn(2) >>> output = m(input) ) thresholdvalueinplacer(r)r*FreturnNc t||_||_||_dSN)super__init__r(r)r*)selfr(r)r* __class__o/var/www/html/construction_image-detection-poc/venv/lib/python3.10/site-packages/torch/nn/modules/activation.pyr/Rs  zThreshold.__init__inputcCst||j|j|jSr-)Fr(r)r*r0r5r3r3r4forwardYzThreshold.forwardcC&|jrdnd}d|jd|j|S)N, inplace=Truez threshold=z, value=)r*r(r)r0 inplace_strr3r3r4 extra_repr\zThreshold.extra_reprF __name__ __module__ __qualname____doc__ __constants__float__annotations__boolr/rr8r? __classcell__r3r3r1r4r 0s  r cXeZdZUdZdgZeed<d deffdd Zdedefdd Z de fd d Z Z S) r aApplies the rectified linear unit function element-wise. :math:`\text{ReLU}(x) = (x)^+ = \max(0, x)` Args: inplace: can optionally do the operation in-place. Default: ``False`` Shape: - Input: :math:`(*)`, where :math:`*` means any number of dimensions. - Output: :math:`(*)`, same shape as the input. .. image:: ../scripts/activation_images/ReLU.png Examples:: >>> m = nn.ReLU() >>> input = torch.randn(2) >>> output = m(input) An implementation of CReLU - https://arxiv.org/abs/1603.05201 >>> m = nn.ReLU() >>> input = torch.randn(2).unsqueeze(0) >>> output = torch.cat((m(input), m(-input))) r*Fct||_dSr-r.r/r*r0r*r1r3r4r/  z ReLU.__init__r5r+cCtj||jdSNr*)r6relur*r7r3r3r4r8z ReLU.forwardcC|jrd}|Sd}|SNz inplace=Truer<rSr=r3r3r4r? zReLU.extra_reprrA rCrDrErFrGrJrIr/rr8strr?rKr3r3r1r4r as r csneZdZUdZgdZeed<eed<eed< ddededeffd d Zd e d e fd dZ ddZ Z S)r aApplies the randomized leaky rectified linear unit function, element-wise. Method described in the paper: `Empirical Evaluation of Rectified Activations in Convolutional Network `_. The function is defined as: .. math:: \text{RReLU}(x) = \begin{cases} x & \text{if } x \geq 0 \\ ax & \text{ otherwise } \end{cases} where :math:`a` is randomly sampled from uniform distribution :math:`\mathcal{U}(\text{lower}, \text{upper})` during training while during evaluation :math:`a` is fixed with :math:`a = \frac{\text{lower} + \text{upper}}{2}`. Args: lower: lower bound of the uniform distribution. Default: :math:`\frac{1}{8}` upper: upper bound of the uniform distribution. Default: :math:`\frac{1}{3}` inplace: can optionally do the operation in-place. Default: ``False`` Shape: - Input: :math:`(*)`, where :math:`*` means any number of dimensions. - Output: :math:`(*)`, same shape as the input. .. image:: ../scripts/activation_images/RReLU.png Examples:: >>> m = nn.RReLU(0.1, 0.3) >>> input = torch.randn(2) >>> output = m(input) )lowerupperr*r[r\r*?UUUUUU?Fcr,r-)r.r/r[r\r*)r0r[r\r*r1r3r4r/s  zRReLU.__init__r5r+cCst||j|j|j|jSr-)r6rrelur[r\trainingr*r7r3r3r4r8sz RReLU.forwardcCr:)Nr;r<zlower=z, upper=)r*r[r\r=r3r3r4r?r@zRReLU.extra_repr)r]r^FrBr3r3r1r4r s %r c seZdZUdZgdZeed<eed<eed<    ddededed eed eed d f fd d Z de d e fddZ d e fddZ ZS)raApplies the HardTanh function element-wise. HardTanh is defined as: .. math:: \text{HardTanh}(x) = \begin{cases} \text{max\_val} & \text{ if } x > \text{ max\_val } \\ \text{min\_val} & \text{ if } x < \text{ min\_val } \\ x & \text{ otherwise } \\ \end{cases} Args: min_val: minimum value of the linear region range. Default: -1 max_val: maximum value of the linear region range. Default: 1 inplace: can optionally do the operation in-place. Default: ``False`` Keyword arguments :attr:`min_value` and :attr:`max_value` have been deprecated in favor of :attr:`min_val` and :attr:`max_val`. Shape: - Input: :math:`(*)`, where :math:`*` means any number of dimensions. - Output: :math:`(*)`, same shape as the input. .. image:: ../scripts/activation_images/Hardtanh.png Examples:: >>> m = nn.Hardtanh(-2, 2) >>> input = torch.randn(2) >>> output = m(input) )min_valmax_valr*rarbr*?FN min_value max_valuer+csht|durtjdtdd|}|dur!tjdtdd|}||_||_||_|j|jks2JdS)NzBkeyword argument `min_value` is deprecated and rename to `min_val`) stacklevelzBkeyword argument `max_value` is deprecated and rename to `max_val`)r.r/warningswarn FutureWarningrarbr*)r0rarbr*rerfr1r3r4r/s& zHardtanh.__init__r5cCst||j|j|jSr-)r6hardtanhrarbr*r7r3r3r4r8 r9zHardtanh.forwardcCr:)Nr;r<zmin_val=z , max_val=)r*rarbr=r3r3r4r?r@zHardtanh.extra_repr)rcrdFNN)rCrDrErFrGrHrIrJrr/rr8rZr?rKr3r3r1r4rs4  rcs6eZdZdZd deffdd ZdefddZZS) raApplies the ReLU6 function element-wise. .. math:: \text{ReLU6}(x) = \min(\max(0,x), 6) Args: inplace: can optionally do the operation in-place. Default: ``False`` Shape: - Input: :math:`(*)`, where :math:`*` means any number of dimensions. - Output: :math:`(*)`, same shape as the input. .. image:: ../scripts/activation_images/ReLU6.png Examples:: >>> m = nn.ReLU6() >>> input = torch.randn(2) >>> output = m(input) Fr*cstdd|dS)Ng@)r.r/rOr1r3r4r/*zReLU6.__init__r+cCrVrWrSr=r3r3r4r?-rXzReLU6.extra_reprrA) rCrDrErFrJr/rZr?rKr3r3r1r4rsrc@"eZdZdZdedefddZdS)raApplies the Sigmoid function element-wise. .. math:: \text{Sigmoid}(x) = \sigma(x) = \frac{1}{1 + \exp(-x)} Shape: - Input: :math:`(*)`, where :math:`*` means any number of dimensions. - Output: :math:`(*)`, same shape as the input. .. image:: ../scripts/activation_images/Sigmoid.png Examples:: >>> m = nn.Sigmoid() >>> input = torch.randn(2) >>> output = m(input) r5r+cC t|Sr-)torchsigmoidr7r3r3r4r8F zSigmoid.forwardNrCrDrErFrr8r3r3r3r4r2srcNeZdZUdZdgZeed<d deddffdd Zdedefd d Z Z S) raApplies the Hardsigmoid function element-wise. Hardsigmoid is defined as: .. math:: \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} Args: inplace: can optionally do the operation in-place. Default: ``False`` Shape: - Input: :math:`(*)`, where :math:`*` means any number of dimensions. - Output: :math:`(*)`, same shape as the input. .. image:: ../scripts/activation_images/Hardsigmoid.png Examples:: >>> m = nn.Hardsigmoid() >>> input = torch.randn(2) >>> output = m(input) r*Fr+NcrMr-rNrOr1r3r4r/jrPzHardsigmoid.__init__r5cCt||jSr-)r6 hardsigmoidr*r7r3r3r4r8nzHardsigmoid.forwardrA rCrDrErFrGrJrIr/rr8rKr3r3r1r4rJs rc@ro)raApplies the Hyperbolic Tangent (Tanh) function element-wise. Tanh is defined as: .. math:: \text{Tanh}(x) = \tanh(x) = \frac{\exp(x) - \exp(-x)} {\exp(x) + \exp(-x)} Shape: - Input: :math:`(*)`, where :math:`*` means any number of dimensions. - Output: :math:`(*)`, same shape as the input. .. image:: ../scripts/activation_images/Tanh.png Examples:: >>> m = nn.Tanh() >>> input = torch.randn(2) >>> output = m(input) r5r+cCrpr-)rqtanhr7r3r3r4r8rsz Tanh.forwardNrtr3r3r3r4rrsrcrL) raApplies the Sigmoid Linear Unit (SiLU) function, element-wise. The SiLU function is also known as the swish function. .. math:: \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. Shape: - Input: :math:`(*)`, where :math:`*` means any number of dimensions. - Output: :math:`(*)`, same shape as the input. .. image:: ../scripts/activation_images/SiLU.png Examples:: >>> m = nn.SiLU() >>> input = torch.randn(2) >>> output = m(input) r*FcrMr-rNrOr1r3r4r/rPz SiLU.__init__r5r+cCrQrR)r6silur*r7r3r3r4r8rUz SiLU.forwardcCrVrWrSr=r3r3r4r?rXzSiLU.extra_reprrArYr3r3r1r4rs rcrL) raxApplies the Mish function, element-wise. Mish: A Self Regularized Non-Monotonic Neural Activation Function. .. math:: \text{Mish}(x) = x * \text{Tanh}(\text{Softplus}(x)) .. note:: See `Mish: A Self Regularized Non-Monotonic Neural Activation Function `_ Shape: - Input: :math:`(*)`, where :math:`*` means any number of dimensions. - Output: :math:`(*)`, same shape as the input. .. image:: ../scripts/activation_images/Mish.png Examples:: >>> m = nn.Mish() >>> input = torch.randn(2) >>> output = m(input) r*FcrMr-rNrOr1r3r4r/rPz Mish.__init__r5r+cCrQrR)r6mishr*r7r3r3r4r8rUz Mish.forwardcCrVrWrSr=r3r3r4r?rXzMish.extra_reprrArYr3r3r1r4rs rcru) raApplies the Hardswish function, element-wise. Method described in the paper: `Searching for MobileNetV3 `_. Hardswish is defined as: .. math:: \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} Args: inplace: can optionally do the operation in-place. Default: ``False`` Shape: - Input: :math:`(*)`, where :math:`*` means any number of dimensions. - Output: :math:`(*)`, same shape as the input. .. image:: ../scripts/activation_images/Hardswish.png Examples:: >>> m = nn.Hardswish() >>> input = torch.randn(2) >>> output = m(input) r*Fr+NcrMr-rNrOr1r3r4r/rPzHardswish.__init__r5cCrvr-)r6 hardswishr*r7r3r3r4r8rxzHardswish.forwardrAryr3r3r1r4rs rcjeZdZUdZddgZeed<eed<ddededdffdd Zd e de fd d Z de fd dZ Z S)rasApplies the Exponential Linear Unit (ELU) function, element-wise. Method described in the paper: `Fast and Accurate Deep Network Learning by Exponential Linear Units (ELUs) `__. ELU is defined as: .. math:: \text{ELU}(x) = \begin{cases} x, & \text{ if } x > 0\\ \alpha * (\exp(x) - 1), & \text{ if } x \leq 0 \end{cases} Args: alpha: the :math:`\alpha` value for the ELU formulation. Default: 1.0 inplace: can optionally do the operation in-place. Default: ``False`` Shape: - Input: :math:`(*)`, where :math:`*` means any number of dimensions. - Output: :math:`(*)`, same shape as the input. .. image:: ../scripts/activation_images/ELU.png Examples:: >>> m = nn.ELU() >>> input = torch.randn(2) >>> output = m(input) alphar*rdFr+Nct||_||_dSr-r.r/rr*r0rr*r1r3r4r/+  z ELU.__init__r5cCt||j|jSr-)r6elurr*r7r3r3r4r80z ELU.forwardcC|jrdnd}d|j|SNr;r<zalpha=r*rr=r3r3r4r?3zELU.extra_reprrdFrCrDrErFrGrHrIrJr/rr8rZr?rKr3r3r1r4rs rcr~)ra3Applies the CELU function element-wise. .. math:: \text{CELU}(x) = \max(0,x) + \min(0, \alpha * (\exp(x/\alpha) - 1)) More details can be found in the paper `Continuously Differentiable Exponential Linear Units`_ . Args: alpha: the :math:`\alpha` value for the CELU formulation. Default: 1.0 inplace: can optionally do the operation in-place. Default: ``False`` Shape: - Input: :math:`(*)`, where :math:`*` means any number of dimensions. - Output: :math:`(*)`, same shape as the input. .. image:: ../scripts/activation_images/CELU.png Examples:: >>> m = nn.CELU() >>> input = torch.randn(2) >>> output = m(input) .. _`Continuously Differentiable Exponential Linear Units`: https://arxiv.org/abs/1704.07483 rr*rdFr+Ncrr-rrr1r3r4r/Xrz CELU.__init__r5cCrr-)r6celurr*r7r3r3r4r8]rz CELU.forwardcCrrrr=r3r3r4r?`rzCELU.extra_reprrrr3r3r1r4r8s rc\eZdZUdZdgZeed<d deddffdd Zdedefd d Z de fd d Z Z S)raApplies the SELU function element-wise. .. math:: \text{SELU}(x) = \text{scale} * (\max(0,x) + \min(0, \alpha * (\exp(x) - 1))) with :math:`\alpha = 1.6732632423543772848170429916717` and :math:`\text{scale} = 1.0507009873554804934193349852946`. .. warning:: When using ``kaiming_normal`` or ``kaiming_normal_`` for initialisation, ``nonlinearity='linear'`` should be used instead of ``nonlinearity='selu'`` in order to get `Self-Normalizing Neural Networks`_. See :func:`torch.nn.init.calculate_gain` for more information. More details can be found in the paper `Self-Normalizing Neural Networks`_ . Args: inplace (bool, optional): can optionally do the operation in-place. Default: ``False`` Shape: - Input: :math:`(*)`, where :math:`*` means any number of dimensions. - Output: :math:`(*)`, same shape as the input. .. image:: ../scripts/activation_images/SELU.png Examples:: >>> m = nn.SELU() >>> input = torch.randn(2) >>> output = m(input) .. _Self-Normalizing Neural Networks: https://arxiv.org/abs/1706.02515 r*Fr+NcrMr-rNrOr1r3r4r/rPz SELU.__init__r5cCrvr-)r6selur*r7r3r3r4r8rxz SELU.forwardcCrVrWrSr=r3r3r4r?rXzSELU.extra_reprrArYr3r3r1r4res "rcr)ra5Applies the gated linear unit function. :math:`{GLU}(a, b)= a \otimes \sigma(b)` where :math:`a` is the first half of the input matrices and :math:`b` is the second half. Args: dim (int): the dimension on which to split the input. Default: -1 Shape: - Input: :math:`(\ast_1, N, \ast_2)` where `*` means, any number of additional dimensions - Output: :math:`(\ast_1, M, \ast_2)` where :math:`M=N/2` Examples:: >>> m = nn.GLU() >>> input = torch.randn(4, 2) >>> output = m(input) dimr+NcrMr-r.r/rr0rr1r3r4r/rPz GLU.__init__r5cCrvr-)r6glurr7r3r3r4r8rxz GLU.forwardcC d|jSNzdim=rr0r3r3r4r? zGLU.extra_repr)r) rCrDrErFrGintrIr/rr8rZr?rKr3r3r1r4rs rcs\eZdZUdZdgZeed<d deddffdd Zdedefd d Z defd d Z Z S)ra/Applies the Gaussian Error Linear Units function. .. math:: \text{GELU}(x) = x * \Phi(x) where :math:`\Phi(x)` is the Cumulative Distribution Function for Gaussian Distribution. When the approximate argument is 'tanh', Gelu is estimated with: .. math:: \text{GELU}(x) = 0.5 * x * (1 + \text{Tanh}(\sqrt{2 / \pi} * (x + 0.044715 * x^3))) Args: approximate (str, optional): the gelu approximation algorithm to use: ``'none'`` | ``'tanh'``. Default: ``'none'`` Shape: - Input: :math:`(*)`, where :math:`*` means any number of dimensions. - Output: :math:`(*)`, same shape as the input. .. image:: ../scripts/activation_images/GELU.png Examples:: >>> m = nn.GELU() >>> input = torch.randn(2) >>> output = m(input) approximatenoner+NcrMr-)r.r/r)r0rr1r3r4r/rPz GELU.__init__r5cCrQ)N)r)r6gelurr7r3r3r4r8rUz GELU.forwardcCsdt|jS)Nz approximate=)reprrrr3r3r4r?rUzGELU.extra_repr)r) rCrDrErFrGrZrIr/rr8r?rKr3r3r1r4rs rcr)raApplies the Hard Shrinkage (Hardshrink) function element-wise. Hardshrink is defined as: .. math:: \text{HardShrink}(x) = \begin{cases} x, & \text{ if } x > \lambda \\ x, & \text{ if } x < -\lambda \\ 0, & \text{ otherwise } \end{cases} Args: lambd: the :math:`\lambda` value for the Hardshrink formulation. Default: 0.5 Shape: - Input: :math:`(*)`, where :math:`*` means any number of dimensions. - Output: :math:`(*)`, same shape as the input. .. image:: ../scripts/activation_images/Hardshrink.png Examples:: >>> m = nn.Hardshrink() >>> input = torch.randn(2) >>> output = m(input) lambd?r+NcrMr-r.r/rr0rr1r3r4r/rPzHardshrink.__init__r5cCrvr-)r6 hardshrinkrr7r3r3r4r8rxzHardshrink.forwardcCs|jSr-)rrr3r3r4r? szHardshrink.extra_reprr rCrDrErFrGrHrIr/rr8rZr?rKr3r3r1r4rs rcsjeZdZUdZddgZeed<eed<ddededdffdd Zd e de fd d Z de fd dZ Z S)ra}Applies the LeakyReLU function element-wise. .. math:: \text{LeakyReLU}(x) = \max(0, x) + \text{negative\_slope} * \min(0, x) or .. math:: \text{LeakyReLU}(x) = \begin{cases} x, & \text{ if } x \geq 0 \\ \text{negative\_slope} \times x, & \text{ otherwise } \end{cases} Args: negative_slope: Controls the angle of the negative slope (which is used for negative input values). Default: 1e-2 inplace: can optionally do the operation in-place. Default: ``False`` Shape: - Input: :math:`(*)` where `*` means, any number of additional dimensions - Output: :math:`(*)`, same shape as the input .. image:: ../scripts/activation_images/LeakyReLU.png Examples:: >>> m = nn.LeakyReLU(0.1) >>> input = torch.randn(2) >>> output = m(input) r*negative_slope{Gz?Fr+Ncrr-)r.r/rr*)r0rr*r1r3r4r/6rzLeakyReLU.__init__r5cCrr-)r6 leaky_relurr*r7r3r3r4r8;rzLeakyReLU.forwardcCr)Nr;r<znegative_slope=)r*rr=r3r3r4r?>rzLeakyReLU.extra_repr)rF)rCrDrErFrGrJrIrHr/rr8rZr?rKr3r3r1r4rs "rc@ro)raApplies the Logsigmoid function element-wise. .. math:: \text{LogSigmoid}(x) = \log\left(\frac{ 1 }{ 1 + \exp(-x)}\right) Shape: - Input: :math:`(*)`, where :math:`*` means any number of dimensions. - Output: :math:`(*)`, same shape as the input. .. image:: ../scripts/activation_images/LogSigmoid.png Examples:: >>> m = nn.LogSigmoid() >>> input = torch.randn(2) >>> output = m(input) r5r+cCrpr-)r6 logsigmoidr7r3r3r4r8VrszLogSigmoid.forwardNrtr3r3r3r4rCrcsjeZdZUdZddgZeed<eed<ddededdffdd Zd edefd d Z de fd dZ Z S)rayApplies the Softplus function element-wise. .. math:: \text{Softplus}(x) = \frac{1}{\beta} * \log(1 + \exp(\beta * x)) SoftPlus is a smooth approximation to the ReLU function and can be used to constrain the output of a machine to always be positive. For numerical stability the implementation reverts to the linear function when :math:`input \times \beta > threshold`. Args: beta: the :math:`\beta` value for the Softplus formulation. Default: 1 threshold: values above this revert to a linear function. Default: 20 Shape: - Input: :math:`(*)`, where :math:`*` means any number of dimensions. - Output: :math:`(*)`, same shape as the input. .. image:: ../scripts/activation_images/Softplus.png Examples:: >>> m = nn.Softplus() >>> input = torch.randn(2) >>> output = m(input) betar(rd4@r+Ncrr-)r.r/rr()r0rr(r1r3r4r/{rzSoftplus.__init__r5cCrr-)r6softplusrr(r7r3r3r4r8rzSoftplus.forwardcCsd|jd|jS)Nzbeta=z , threshold=)rr(rr3r3r4r?rnzSoftplus.extra_repr)rdrrr3r3r1r4rZs rcr)raApplies the soft shrinkage function element-wise. .. math:: \text{SoftShrinkage}(x) = \begin{cases} x - \lambda, & \text{ if } x > \lambda \\ x + \lambda, & \text{ if } x < -\lambda \\ 0, & \text{ otherwise } \end{cases} Args: lambd: the :math:`\lambda` (must be no less than zero) value for the Softshrink formulation. Default: 0.5 Shape: - Input: :math:`(*)`, where :math:`*` means any number of dimensions. - Output: :math:`(*)`, same shape as the input. .. image:: ../scripts/activation_images/Softshrink.png Examples:: >>> m = nn.Softshrink() >>> input = torch.randn(2) >>> output = m(input) rrr+NcrMr-rrr1r3r4r/rPzSoftshrink.__init__r5cCrvr-)r6 softshrinkrr7r3r3r4r8rxzSoftshrink.forwardcCs t|jSr-)rZrrr3r3r4r?rszSoftshrink.extra_reprrrr3r3r1r4rs rxr+cCs$|dur|jjddtjjjfvSdS)NcpucudaT)devicetyperqutilsbackend_registration_privateuse1_backend_namerr3r3r4_check_arg_devicesrcCs|dur|jSdS)NF) requires_gradrr3r3r4_arg_requires_gradsrcCs,tjstjj}tdd|DSdS)Ncss$|] }t|tjjjjkVqdSr-)rrqfx experimental proxy_tensorProxyTorchDispatchMode.0rr3r3r4 s  z&_is_make_fx_tracing..F)rqjit is_scriptingr_python_dispatch _get_current_dispatch_mode_stackany)torch_dispatch_mode_stackr3r3r4_is_make_fx_tracings  rcseZdZUdZdgZeejed<eejed<         d dfd d Z d d Z fddZ     ddedededeede deede de d e eeeffddZdeedeeded e eeeeffddZZS)r a Allows the model to jointly attend to information from different representation subspaces. .. note:: See `this tutorial `_ for an in depth discussion of the performant building blocks PyTorch offers for building your own transformer layers. Method described in the paper: `Attention Is All You Need `_. Multi-Head Attention is defined as: .. math:: \text{MultiHead}(Q, K, V) = \text{Concat}(\text{head}_1,\dots,\text{head}_h)W^O where :math:`\text{head}_i = \text{Attention}(QW_i^Q, KW_i^K, VW_i^V)`. ``nn.MultiheadAttention`` will use the optimized implementations of ``scaled_dot_product_attention()`` when possible. In addition to support for the new ``scaled_dot_product_attention()`` function, for speeding up Inference, MHA will use fastpath inference with support for Nested Tensors, iff: - self attention is being computed (i.e., ``query``, ``key``, and ``value`` are the same tensor). - inputs are batched (3D) with ``batch_first==True`` - Either autograd is disabled (using ``torch.inference_mode`` or ``torch.no_grad``) or no tensor argument ``requires_grad`` - training is disabled (using ``.eval()``) - ``add_bias_kv`` is ``False`` - ``add_zero_attn`` is ``False`` - ``kdim`` and ``vdim`` are equal to ``embed_dim`` - if a `NestedTensor `_ is passed, neither ``key_padding_mask`` nor ``attn_mask`` is passed - autocast is disabled If the optimized inference fastpath implementation is in use, a `NestedTensor `_ can be passed for ``query``/``key``/``value`` to represent padding more efficiently than using a padding mask. In this case, a `NestedTensor `_ will be returned, and an additional speedup proportional to the fraction of the input that is padding can be expected. Args: embed_dim: Total dimension of the model. num_heads: Number of parallel attention heads. Note that ``embed_dim`` will be split across ``num_heads`` (i.e. each head will have dimension ``embed_dim // num_heads``). dropout: Dropout probability on ``attn_output_weights``. Default: ``0.0`` (no dropout). bias: If specified, adds bias to input / output projection layers. Default: ``True``. add_bias_kv: If specified, adds bias to the key and value sequences at dim=0. Default: ``False``. add_zero_attn: If specified, adds a new batch of zeros to the key and value sequences at dim=1. Default: ``False``. kdim: Total number of features for keys. Default: ``None`` (uses ``kdim=embed_dim``). vdim: Total number of features for values. Default: ``None`` (uses ``vdim=embed_dim``). batch_first: If ``True``, then the input and output tensors are provided as (batch, seq, feature). Default: ``False`` (seq, batch, feature). Examples:: >>> # xdoctest: +SKIP >>> multihead_attn = nn.MultiheadAttention(embed_dim, num_heads) >>> attn_output, attn_output_weights = multihead_attn(query, key, value) .. _`FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness`: https://arxiv.org/abs/2205.14135 batch_firstbias_kbias_vrmTFNr+c s|dks|dkrtd|d|d| | d} t||_|dur&|n||_|dur/|n||_|j|ko;|j|k|_||_||_| |_ |||_ |j ||jksWJd|jst t j ||ffi| |_t t j ||jffi| |_t t j ||jffi| |_|ddn"t t j d||ffi| |_|d d|d d|d d|rt t j d|fi| |_n|d dt||fd |i| |_|rt t j dd|ffi| |_t t j dd|ffi| |_nd|_|_||_|dS)Nrz>embed_dim and num_heads must be greater than 0, got embed_dim=z and num_heads=z insteadrdtypez(embed_dim must be divisible by num_headsin_proj_weight q_proj_weight k_proj_weight v_proj_weight in_proj_biasbiasr) ValueErrorr.r/ embed_dimkdimvdim_qkv_same_embed_dim num_headsdropoutrhead_dimrrqemptyrrrregister_parameterrrr out_projrr add_zero_attn_reset_parameters) r0rrrr add_bias_kvrrrrrrfactory_kwargsr1r3r4r/sl          zMultiheadAttention.__init__cCs|jr t|jnt|jt|jt|j|jdur*t|jdt|jj d|j dur4t |j |j dur@t |j dSdS)Nrm) rrrrrrrrrrrrrrr3r3r4r\s         z$MultiheadAttention._reset_parameterscs d|vrd|d<t|dS)NrT)r. __setstate__r0stater1r3r4rlszMultiheadAttention.__setstate__querykeyr)key_padding_mask need_weights attn_maskaverage_attn_weights is_causalc Csd} |dur t|s|durt|rd} |dk} tj|dt|d|jd}tj|ddd|jdd }tjj } | sAd } n| sKd |} n||usS||urVd } n{|j duro|j|j jkrod |jd|j jd} nb|j durwd} nZ|j|j jkrd |jd|j jd} nF|j rd} n@|j ddkrd} n6|jsd} n0|jdurd} n(|jdurd} n |jrd} n|jsd} n|jr|dus|durd} ntrd} | sG||||j |j |jj|jjf} tj| rd} n*trd} n$tdd | Dsd!tjjj} ntrt d"d | Drd#} | sG|!|||\} }|j durG|j durGt"||||j#|j |j |j |jj|jj| ||| S|jpQ|jpQ|j}|r^Jd$d%| |jr| r||ur||urz|$d&d}}}nd'd ||fD\}}|}n d(d |||fD\}}}|jstj%||||j#|j |j |j |j|j|j|j&|jj|jjf |j |||d)|j'|j(|j)||d* \}}n'tj%||||j#|j |j |j |j|j|j|j&|jj|jj|j |||||d+\}}|jr| r|$d&d|fS||fS),aICompute attention outputs using query, key, and value embeddings. Supports optional parameters for padding, masks and attention weights. Args: query: Query embeddings of shape :math:`(L, E_q)` for unbatched input, :math:`(L, N, E_q)` when ``batch_first=False`` or :math:`(N, L, E_q)` when ``batch_first=True``, where :math:`L` is the target sequence length, :math:`N` is the batch size, and :math:`E_q` is the query embedding dimension ``embed_dim``. Queries are compared against key-value pairs to produce the output. See "Attention Is All You Need" for more details. key: Key embeddings of shape :math:`(S, E_k)` for unbatched input, :math:`(S, N, E_k)` when ``batch_first=False`` or :math:`(N, S, E_k)` when ``batch_first=True``, where :math:`S` is the source sequence length, :math:`N` is the batch size, and :math:`E_k` is the key embedding dimension ``kdim``. See "Attention Is All You Need" for more details. value: Value embeddings of shape :math:`(S, E_v)` for unbatched input, :math:`(S, N, E_v)` when ``batch_first=False`` or :math:`(N, S, E_v)` when ``batch_first=True``, where :math:`S` is the source sequence length, :math:`N` is the batch size, and :math:`E_v` is the value embedding dimension ``vdim``. See "Attention Is All You Need" for more details. key_padding_mask: If specified, a mask of shape :math:`(N, S)` indicating which elements within ``key`` to ignore for the purpose of attention (i.e. treat as "padding"). For unbatched `query`, shape should be :math:`(S)`. Binary and float masks are supported. For a binary mask, a ``True`` value indicates that the corresponding ``key`` value will be ignored for the purpose of attention. For a float mask, it will be directly added to the corresponding ``key`` value. need_weights: If specified, returns ``attn_output_weights`` in addition to ``attn_outputs``. Set ``need_weights=False`` to use the optimized ``scaled_dot_product_attention`` and achieve the best performance for MHA. Default: ``True``. attn_mask: If specified, a 2D or 3D mask preventing attention to certain positions. Must be of shape :math:`(L, S)` or :math:`(N\cdot\text{num\_heads}, L, S)`, where :math:`N` is the batch size, :math:`L` is the target sequence length, and :math:`S` is the source sequence length. A 2D mask will be broadcasted across the batch while a 3D mask allows for a different mask for each entry in the batch. Binary and float masks are supported. For a binary mask, a ``True`` value indicates that the corresponding position is not allowed to attend. For a float mask, the mask values will be added to the attention weight. If both attn_mask and key_padding_mask are supplied, their types should match. average_attn_weights: If true, indicates that the returned ``attn_weights`` should be averaged across heads. Otherwise, ``attn_weights`` are provided separately per head. Note that this flag only has an effect when ``need_weights=True``. Default: ``True`` (i.e. average weights across heads) is_causal: If specified, applies a causal mask as attention mask. Default: ``False``. Warning: ``is_causal`` provides a hint that ``attn_mask`` is the causal mask. Providing incorrect hints can result in incorrect execution, including forward and backward compatibility. Outputs: - **attn_output** - Attention outputs of shape :math:`(L, E)` when input is unbatched, :math:`(L, N, E)` when ``batch_first=False`` or :math:`(N, L, E)` when ``batch_first=True``, where :math:`L` is the target sequence length, :math:`N` is the batch size, and :math:`E` is the embedding dimension ``embed_dim``. - **attn_output_weights** - Only returned when ``need_weights=True``. If ``average_attn_weights=True``, returns attention weights averaged across heads of shape :math:`(L, S)` when input is unbatched or :math:`(N, L, S)`, where :math:`N` is the batch size, :math:`L` is the target sequence length, and :math:`S` is the source sequence length. If ``average_attn_weights=False``, returns attention weights per head of shape :math:`(\text{num\_heads}, L, S)` when input is unbatched or :math:`(N, \text{num\_heads}, L, S)`. .. note:: `batch_first` argument is ignored for unbatched inputs. r<Nz5floating-point masks are not supported for fast path.rrr)mask mask_name other_type other_name target_typeF)rrrrr check_otherz6torch.backends.mha.get_fastpath_enabled() was not Truez5input not batched; expected query.dim() of 3 but got zKnon-self attention was used (query, key, and value are not the same Tensor)zdtypes of query (z) and self.in_proj_bias (z ) don't matchzin_proj_weight was Nonez) and self.in_proj_weight (ztraining is enabledrgrzself.num_heads is not evenzbatch_first was not Truezself.bias_k was not Nonezself.bias_v was not Nonezadd_zero_attn was enabledz _qkv_same_embed_dim was not Truezsupplying both src_key_padding_mask and src_mask at the same time is not supported with NestedTensor inputzautocast is enabledz'some Tensor argument has_torch_functionzwe are running make_fx tracingcs|]}t|VqdSr-)rrr3r3r4rsz-MultiheadAttention.forward..z=some Tensor argument's device is neither one of cpu, cuda or csrr-)rrr3r3r4rs zhgrad is enabled and at least one of query or the input/output projection weights or biases requires_gradzKMultiheadAttention does not support NestedTensor outside of its fast path. z"The fast path was not hit because rcs|] }|ddVqdSrrN transposerr3r3r4r=csrrrrr3r3r4r@rT) r`rrruse_separate_proj_weightrrrrr)r`rrrrr)*rqis_floating_pointrr6_canonical_mask_none_or_dtyperbackendsmhaget_fastpath_enabledrrr`rrrrrr is_nestedis_autocast_enabledrweightr overrideshas_torch_functionrallrrris_grad_enabledr merge_masks_native_multi_head_attentionrrmulti_head_attention_forwardrrrr)r0rrr)rrrrrwhy_not_fast_path is_batchedis_fastpath_enabled tensor_args merged_mask mask_type any_nested attn_outputattn_output_weightsr3r3r4r8ssDG           zMultiheadAttention.forwardc Csd}d}|dur d}|}|durO|j\}}}d}|dkr'||d||} n|dd||||jdd} | }|durO||dd|d|jdd} | | }||fS)aDetermine mask type and combine masks if necessary. If only one mask is provided, that mask and the corresponding mask type will be returned. If both masks are provided, they will be both expanded to shape ``(batch_size, num_heads, seq_len, seq_len)``, combined with logical ``or`` and mask type 2 will be returned Args: attn_mask: attention mask of shape ``(seq_len, seq_len)``, mask type 0 key_padding_mask: padding mask of shape ``(batch_size, seq_len)``, mask type 1 query: query embeddings of shape ``(batch_size, seq_len, embed_dim)`` Returns: merged_mask: merged mask mask_type: merged mask type (0, 1, or 2) Nrrgrr)shaperviewexpandr) r0rrrrr batch_sizeseq_len_attn_mask_expandedkey_padding_mask_expandedr3r3r4rws,   zMultiheadAttention.merge_masks) rmTFFNNFNN)r+N)NTNTF)rCrDrErFrGrrqrrIr/rrrJtupler8rrrKr3r3r1r4r sl C G    r csjeZdZUdZdgZeed< ddededdffdd Zd d Z d e de fd dZ de fddZ ZS)r!aApplies the element-wise PReLU function. .. math:: \text{PReLU}(x) = \max(0,x) + a * \min(0,x) or .. math:: \text{PReLU}(x) = \begin{cases} x, & \text{ if } x \ge 0 \\ ax, & \text{ otherwise } \end{cases} Here :math:`a` is a learnable parameter. When called without arguments, `nn.PReLU()` uses a single parameter :math:`a` across all input channels. If called with `nn.PReLU(nChannels)`, a separate :math:`a` is used for each input channel. .. note:: weight decay should not be used when learning :math:`a` for good performance. .. note:: Channel dim is the 2nd dim of input. When input has dims < 2, then there is no channel dim and the number of channels = 1. Args: num_parameters (int): number of :math:`a` to learn. Although it takes an int as input, there is only two values are legitimate: 1, or the number of channels at input. Default: 1 init (float): the initial value of :math:`a`. Default: 0.25 Shape: - Input: :math:`( *)` where `*` means, any number of additional dimensions. - Output: :math:`(*)`, same shape as the input. Attributes: weight (Tensor): the learnable weights of shape (:attr:`num_parameters`). .. image:: ../scripts/activation_images/PReLU.png Examples:: >>> m = nn.PReLU() >>> input = torch.randn(2) >>> output = m(input) num_parametersr?Ninitr+csD||d}||_t||_ttj|fi||_|dS)Nr) rr.r/rrrqrrreset_parameters)r0rrrrrr1r3r4r/s   zPReLU.__init__cCstjj|j|jdSr-)rqnnrrrrr3r3r4r szPReLU.reset_parametersr5cCrvr-)r6prelurr7r3r3r4r8rxz PReLU.forwardcCr)Nznum_parameters=)rrr3r3r4r?rzPReLU.extra_repr)rrNN)rCrDrErFrGrrIrHr/r rr8rZr?rKr3r3r1r4r!s 1 r!c@ro)r"aApplies the element-wise Softsign function. .. math:: \text{SoftSign}(x) = \frac{x}{ 1 + |x|} Shape: - Input: :math:`(*)`, where :math:`*` means any number of dimensions. - Output: :math:`(*)`, same shape as the input. .. image:: ../scripts/activation_images/Softsign.png Examples:: >>> m = nn.Softsign() >>> input = torch.randn(2) >>> output = m(input) r5r+cCrpr-)r6softsignr7r3r3r4r8rszSoftsign.forwardNrtr3r3r3r4r"rr"c@ro)r#aApplies the element-wise Tanhshrink function. .. math:: \text{Tanhshrink}(x) = x - \tanh(x) Shape: - Input: :math:`(*)`, where :math:`*` means any number of dimensions. - Output: :math:`(*)`, same shape as the input. .. image:: ../scripts/activation_images/Tanhshrink.png Examples:: >>> m = nn.Tanhshrink() >>> input = torch.randn(2) >>> output = m(input) r5r+cCrpr-)r6 tanhshrinkr7r3r3r4r8rszTanhshrink.forwardNrtr3r3r3r4r# rr#cjeZdZUdZdgZeeed<ddeeddffdd ZfddZ d e de fd d Z d d Z Z S)r$a:Applies the Softmin function to an n-dimensional input Tensor. Rescales them so that the elements of the n-dimensional output Tensor lie in the range `[0, 1]` and sum to 1. Softmin is defined as: .. math:: \text{Softmin}(x_{i}) = \frac{\exp(-x_i)}{\sum_j \exp(-x_j)} Shape: - Input: :math:`(*)` where `*` means, any number of additional dimensions - Output: :math:`(*)`, same shape as the input Args: dim (int): A dimension along which Softmin will be computed (so every slice along dim will sum to 1). Returns: a Tensor of the same dimension and shape as the input, with values in the range [0, 1] Examples:: >>> m = nn.Softmin(dim=1) >>> input = torch.randn(2, 3) >>> output = m(input) rNr+crMr-rrr1r3r4r/CrPzSoftmin.__init__c$t|t|dsd|_dSdSNrr.rhasattrrrr1r3r4rG   zSoftmin.__setstate__r5cCtj||jddSN _stacklevel)r6softminrr7r3r3r4r8LrzSoftmin.forwardcCrrrrr3r3r4r?OrzSoftmin.extra_reprr-rCrDrErFrGrrrIr/rrr8r?rKr3r3r1r4r$!s   r$cspeZdZUdZdgZeeed<ddeeddffdd ZfddZ d e de fd d Z de fd d Z ZS)r%aApplies the Softmax function to an n-dimensional input Tensor. Rescales them so that the elements of the n-dimensional output Tensor lie in the range [0,1] and sum to 1. Softmax is defined as: .. math:: \text{Softmax}(x_{i}) = \frac{\exp(x_i)}{\sum_j \exp(x_j)} When the input Tensor is a sparse tensor then the unspecified values are treated as ``-inf``. Shape: - Input: :math:`(*)` where `*` means, any number of additional dimensions - Output: :math:`(*)`, same shape as the input Returns: a Tensor of the same dimension and shape as the input with values in the range [0, 1] Args: dim (int): A dimension along which Softmax will be computed (so every slice along dim will sum to 1). .. note:: This module doesn't work directly with NLLLoss, which expects the Log to be computed between the Softmax and itself. Use `LogSoftmax` instead (it's faster and has better numerical properties). Examples:: >>> m = nn.Softmax(dim=1) >>> input = torch.randn(2, 3) >>> output = m(input) rNr+crMr-rrr1r3r4r/~rPzSoftmax.__init__cr&r'r(rr1r3r4rr*zSoftmax.__setstate__r5cCr+r,)r6softmaxrr7r3r3r4r8rzSoftmax.forwardcCrrrrr3r3r4r?rzSoftmax.extra_reprr-)rCrDrErFrGrrrIr/rrr8rZr?rKr3r3r1r4r%Ss '  r%c@ro)r&a|Applies SoftMax over features to each spatial location. When given an image of ``Channels x Height x Width``, it will apply `Softmax` to each location :math:`(Channels, h_i, w_j)` Shape: - Input: :math:`(N, C, H, W)` or :math:`(C, H, W)`. - Output: :math:`(N, C, H, W)` or :math:`(C, H, W)` (same shape as input) Returns: a Tensor of the same dimension and shape as the input with values in the range [0, 1] Examples:: >>> m = nn.Softmax2d() >>> # you softmax over the 2nd dimension >>> input = torch.randn(2, 3, 12, 13) >>> output = m(input) r5r+cCs0|dvrtd|dtj|dddS)N)rz.Softmax2d: expected input to be 3D or 4D, got z D insteadr-r.)rrr6r2r7r3r3r4r8s zSoftmax2d.forwardNrtr3r3r3r4r&sr&cr%)r'aApplies the :math:`\log(\text{Softmax}(x))` function to an n-dimensional input Tensor. The LogSoftmax formulation can be simplified as: .. math:: \text{LogSoftmax}(x_{i}) = \log\left(\frac{\exp(x_i) }{ \sum_j \exp(x_j)} \right) Shape: - Input: :math:`(*)` where `*` means, any number of additional dimensions - Output: :math:`(*)`, same shape as the input Args: dim (int): A dimension along which LogSoftmax will be computed. Returns: a Tensor of the same dimension and shape as the input with values in the range [-inf, 0) Examples:: >>> m = nn.LogSoftmax(dim=1) >>> input = torch.randn(2, 3) >>> output = m(input) rNr+crMr-rrr1r3r4r/rPzLogSoftmax.__init__cr&r'r(rr1r3r4rr*zLogSoftmax.__setstate__r5cCr+r,)r6 log_softmaxrr7r3r3r4r8rzLogSoftmax.forwardcCrrrrr3r3r4r?rzLogSoftmax.extra_reprr-r1r3r3r1r4r's   r')5ritypingrrqtorch.nn.functionalr! functionalr6r torch.nn.initrrrtorch.nn.parameterrlinearr moduler __all__r r r rrrrrrrrrrrrrrrrrrrJrrrr r!r"r#r$r%r&r'r3r3r3r4sZ     !1+<L(,'*0-2#*+4-)  `I2;