JThh^SrSSKrSSKrSSKJr SSKrSSKJr S'SjrS'Sjr S'Sjr Sr S r S'S jr S(S \S \S \S\\RS\4 SjjrS(S \S\S\S\\RS\4 SjjrS)S \S\S\S \S \S\\RS\4SjjrS \S\S\4SjrS \S\4SjrS \S\4SjrSrS*SjrSrS+S \S\S\\RS\4SjjrS+S \S\S\\RS\4SjjrSrS,S \S \S \S!\S\\R4 S"jjrS,S \S \S \S!\S\\R4 S#jjrS-S\\R4S$jjrS.S\\R4S%jjr S&r!\!"\5r"\!"\5r#\!"\5r$\!"\5r%\!"\5r&\!"\5r'\!"\5r(\!"\5r)\!"\5r*\!"\5r+\!"\ 5r,g)/zHThis file contains utilities for initializing neural network parameters.N)Optional)Tensorc[R"5 URXUS9sSSS5 $!,(df  g=fN generator)torchno_graduniform_tensorabrs E/var/www/auris/envauris/lib/python3.13/site-packages/torch/nn/init.py_no_grad_uniform_rs% qy9 0 >c[R"5 URXUS9sSSS5 $!,(df  g=fr)r r normal_r meanstdrs r_no_grad_normal_rs% ~~d9~= rcSnXSU-- :d XSU--:a[R"SSS9 [R"5 U"X1- U- 5nU"XA- U- 5nUR SU-S- SU-S- US9 UR 5 UR U[R"S5-5 URU5 URX4S9 UsSSS5 $!,(df  g=f) NchS[R"U[R"S5- 5-S- $)N?@)matherfsqrt)xs rnorm_cdf(_no_grad_trunc_normal_..norm_cdfs(dhhq499S>122c99zjmean is more than 2 std from [a, b] in nn.init.trunc_normal_. The distribution of values may be incorrect. stacklevelrr)minmax) warningswarnr r r erfinv_mul_rradd_clamp_) r rrrrrr!lus r_no_grad_trunc_normal_r2s: 1s7{1s7{ 2  ;  ah#% & ah#% & A 1q519 B   C$))C.() D  ! #+ s BC C-c[R"5 URU5sSSS5 $!,(df  g=fN)r r fill_r vals r_no_grad_fill_r8>s! ||C  s1 ?c[R"5 UR5sSSS5 $!,(df  g=fr4)r r zero_r s r_no_grad_zero_r<Cs ||~ rc/SQnX;dUS:XagUS:XagUS:Xa[R"S5$US:XavUcS nOQ[U[5(d[U[5(d[U[ 5(aUnO[ S US 35e[R"SSUS --- 5$US :Xag[ SU35e)avReturn the recommended gain value for the given nonlinearity function. The values are as follows: ================= ==================================================== nonlinearity gain ================= ==================================================== Linear / Identity :math:`1` Conv{1,2,3}D :math:`1` Sigmoid :math:`1` Tanh :math:`\frac{5}{3}` ReLU :math:`\sqrt{2}` Leaky Relu :math:`\sqrt{\frac{2}{1 + \text{negative\_slope}^2}}` SELU :math:`\frac{3}{4}` ================= ==================================================== .. warning:: In order to implement `Self-Normalizing Neural Networks`_ , you should use ``nonlinearity='linear'`` instead of ``nonlinearity='selu'``. This gives the initial weights a variance of ``1 / N``, which is necessary to induce a stable fixed point in the forward pass. In contrast, the default gain for ``SELU`` sacrifices the normalization effect for more stable gradient flow in rectangular layers. Args: nonlinearity: the non-linear function (`nn.functional` name) param: optional parameter for the non-linear function Examples: >>> gain = nn.init.calculate_gain('leaky_relu', 0.2) # leaky_relu with negative_slope=0.2 .. _Self-Normalizing Neural Networks: https://papers.nips.cc/paper/2017/hash/5d44ee6f2c3f71b73125876103c8f6c4-Abstract.html )linearconv1dconv2dconv3dconv_transpose1dconv_transpose2dconv_transpose3dsigmoidr'tanhg?relur leaky_relu{Gz?znegative_slope z not a valid numberr$selug?zUnsupported nonlinearity )rr isinstanceboolintfloat ValueError) nonlinearityparam linear_fnsnegative_slopes rcalculate_gainrTHsDJ!\Y%>    yy~  % =!N5$''5#&&%''#Nug5HIJ JyyNA$5 5677    4\NCDDr#r rrrreturnc [RRU5(a%[RR[U4XX#S9$[ XX#5$)aFill the input Tensor with values drawn from the uniform distribution. :math:`\mathcal{U}(a, b)`. Args: tensor: an n-dimensional `torch.Tensor` a: the lower bound of the uniform distribution b: the upper bound of the uniform distribution generator: the torch Generator to sample from (default: None) Examples: >>> w = torch.empty(3, 5) >>> nn.init.uniform_(w) r )r overrideshas_torch_function_variadichandle_torch_functionr rr s rr r sO( 226::44 viq5   V 55r#rrc [RRU5(a%[RR[U4XX#S9$[ XX#5$)aFill the input Tensor with values drawn from the normal distribution. :math:`\mathcal{N}(\text{mean}, \text{std}^2)`. Args: tensor: an n-dimensional `torch.Tensor` mean: the mean of the normal distribution std: the standard deviation of the normal distribution generator: the torch Generator to sample from (default: None) Examples: >>> w = torch.empty(3, 5) >>> nn.init.normal_(w) r)r rWrXrYrrrs rrrsO( 226::44 fYvc5   F# 99r#c [XX#XES9$)aFill the input Tensor with values drawn from a truncated normal distribution. The values are effectively drawn from the normal distribution :math:`\mathcal{N}(\text{mean}, \text{std}^2)` with values outside :math:`[a, b]` redrawn until they are within the bounds. The method used for generating the random values works best when :math:`a \leq \text{mean} \leq b`. Args: tensor: an n-dimensional `torch.Tensor` mean: the mean of the normal distribution std: the standard deviation of the normal distribution a: the minimum cutoff value b: the maximum cutoff value generator: the torch Generator to sample from (default: None) Examples: >>> w = torch.empty(3, 5) >>> nn.init.trunc_normal_(w) r)r2)r rrrrrs r trunc_normal_r\s8 "& OOr#r7c[RRU5(a$[RR[U4XS9$[ X5$)zFill the input Tensor with the value :math:`\text{val}`. Args: tensor: an n-dimensional `torch.Tensor` val: the value to fill the tensor with Examples: >>> w = torch.empty(3, 5) >>> nn.init.constant_(w, 0.3) r6)r rWrXrY constant_r8r6s rr^r^sK 226::44 y5   & &&r#c[US5$)zFill the input Tensor with the scalar value `1`. Args: tensor: an n-dimensional `torch.Tensor` Examples: >>> w = torch.empty(3, 5) >>> nn.init.ones_(w) r)r8r;s rones_r`s &# &&r#c[U5$)zFill the input Tensor with the scalar value `0`. Args: tensor: an n-dimensional `torch.Tensor` Examples: >>> w = torch.empty(3, 5) >>> nn.init.zeros_(w) )r<r;s rzeros_rbs & !!r#cUR5S:wa [S5e[R"5 [R"UR XR S.6 SSS5 U$!,(df  U$=f)aFill the 2-dimensional input `Tensor` with the identity matrix. Preserves the identity of the inputs in `Linear` layers, where as many inputs are preserved as possible. Args: tensor: a 2-dimensional `torch.Tensor` Examples: >>> w = torch.empty(3, 5) >>> nn.init.eye_(w) r$,Only tensors with 2 dimensions are supported)out requires_gradN) ndimensionrOr r eyeshaperfr;s reye_rjsYaGHH  6<1, each group of channels preserves identity Args: tensor: a {3, 4, 5}-dimensional `torch.Tensor` groups (int, optional): number of groups in the conv layer (default: 1) Examples: >>> w = torch.empty(3, 16, 5, 5) >>> nn.init.dirac_(w) >>> w = torch.empty(3, 24, 5, 5) >>> nn.init.dirac_(w, 3) )z5Only tensors with 3, 4, or 5 dimensions are supportedrz!dim 0 must be divisible by groupsr'rlr$rmN)rgrOsizer(r r r:range)r groups dimensionssizesout_chans_per_grpmin_dimgds rdirac_rx$s| ""$J"PQQ KKME Qx&A<==aF*#1X.G  vA7^?PQF0014aQ19LLM1_  -1 A!+ A!+- -1 A!+ A!+ A!+ -$ , M- , Ms 4CE E'cUR5nUS:a [S5eURS5nURS5nSnUR5S:aURSSHnXE-nM X$-nX4-nXg4$)Nr$zNFan in and fan out can not be computed for tensor with fewer than 2 dimensionsr'r)dimrOrori)r rrnum_input_fmapsnum_output_fmapsreceptive_field_sizesfan_infan_outs r_calculate_fan_in_and_fan_outrYsJA~ \  kk!nO{{1~ zz|aab!A % "  3F5G ?r#gainc[U5up4U[R"S[X4-5- 5-n[R"S5U-n[ X*Xb5$)a[Fill the input `Tensor` with values using a Xavier uniform distribution. The method is described in `Understanding the difficulty of training deep feedforward neural networks` - Glorot, X. & Bengio, Y. (2010). The resulting tensor will have values sampled from :math:`\mathcal{U}(-a, a)` where .. math:: a = \text{gain} \times \sqrt{\frac{6}{\text{fan\_in} + \text{fan\_out}}} Also known as Glorot initialization. Args: tensor: an n-dimensional `torch.Tensor` gain: an optional scaling factor generator: the torch Generator to sample from (default: None) Examples: >>> w = torch.empty(3, 5) >>> nn.init.xavier_uniform_(w, gain=nn.init.calculate_gain('relu')) Note: Be aware that ``fan_in`` and ``fan_out`` are calculated assuming that the weight matrix is used in a transposed manner, (i.e., ``x @ w.T`` in ``Linear`` layers, where ``w.shape = [fan_out, fan_in]``). This is important for correct initialization. If you plan to use ``x @ w``, where ``w.shape = [fan_in, fan_out]``, pass in a transposed weight matrix, i.e. ``nn.init.xavier_uniform_(w.T, ...)``. r@)rrrrNr)r rrrrrrs rxavier_uniform_rnsRD4F;OF 3v'7!889 9C #A VR 66r#c[U5up4U[R"S[X4-5- 5-n[ USXR5$)aFFill the input `Tensor` with values using a Xavier normal distribution. The method is described in `Understanding the difficulty of training deep feedforward neural networks` - Glorot, X. & Bengio, Y. (2010). The resulting tensor will have values sampled from :math:`\mathcal{N}(0, \text{std}^2)` where .. math:: \text{std} = \text{gain} \times \sqrt{\frac{2}{\text{fan\_in} + \text{fan\_out}}} Also known as Glorot initialization. Args: tensor: an n-dimensional `torch.Tensor` gain: an optional scaling factor generator: the torch Generator to sample from (default: None) Examples: >>> w = torch.empty(3, 5) >>> nn.init.xavier_normal_(w) Note: Be aware that ``fan_in`` and ``fan_out`` are calculated assuming that the weight matrix is used in a transposed manner, (i.e., ``x @ w.T`` in ``Linear`` layers, where ``w.shape = [fan_out, fan_in]``). This is important for correct initialization. If you plan to use ``x @ w``, where ``w.shape = [fan_in, fan_out]``, pass in a transposed weight matrix, i.e. ``nn.init.xavier_normal_(w.T, ...)``. r)rrrrNr)r rrrrrs rxavier_normal_rs@B4F;OF 3v'7!889 9C FC 88r#cUR5nSS/nX;a[SUSU35e[U5up4US:XaU$U$)NrrzMode z" not supported, please use one of )lowerrOr)r mode valid_modesrrs r_calculate_correct_fanrsT ::>> w = torch.empty(3, 5) >>> nn.init.kaiming_uniform_(w, mode='fan_in', nonlinearity='relu') Note: Be aware that ``fan_in`` and ``fan_out`` are calculated assuming that the weight matrix is used in a transposed manner, (i.e., ``x @ w.T`` in ``Linear`` layers, where ``w.shape = [fan_out, fan_in]``). This is important for correct initialization. If you plan to use ``x @ w``, where ``w.shape = [fan_in, fan_out]``, pass in a transposed weight matrix, i.e. ``nn.init.kaiming_uniform_(w.T, ...)``. )r rrrPrr,Initializing zero-element tensors is a no-oprrN)r rWrXrYkaiming_uniform_rir*r+rrTrrr r ) r rrrPrfanrrbounds rrrsV 226::44  I%5   FLL DE  .C , *D 3 C IIcNS E vuB s C-- C;c0SUR;a[R"S5 U$[X5n[ X15nU[ R "U5- n[R"5 URSXtS9sSSS5 $!,(df  g=f)a+Fill the input `Tensor` with values using a Kaiming normal distribution. The method is described in `Delving deep into rectifiers: Surpassing human-level performance on ImageNet classification` - He, K. et al. (2015). The resulting tensor will have values sampled from :math:`\mathcal{N}(0, \text{std}^2)` where .. math:: \text{std} = \frac{\text{gain}}{\sqrt{\text{fan\_mode}}} Also known as He initialization. Args: tensor: an n-dimensional `torch.Tensor` a: the negative slope of the rectifier used after this layer (only used with ``'leaky_relu'``) mode: either ``'fan_in'`` (default) or ``'fan_out'``. Choosing ``'fan_in'`` preserves the magnitude of the variance of the weights in the forward pass. Choosing ``'fan_out'`` preserves the magnitudes in the backwards pass. nonlinearity: the non-linear function (`nn.functional` name), recommended to use only with ``'relu'`` or ``'leaky_relu'`` (default). generator: the torch Generator to sample from (default: None) Examples: >>> w = torch.empty(3, 5) >>> nn.init.kaiming_normal_(w, mode='fan_out', nonlinearity='relu') Note: Be aware that ``fan_in`` and ``fan_out`` are calculated assuming that the weight matrix is used in a transposed manner, (i.e., ``x @ w.T`` in ``Linear`` layers, where ``w.shape = [fan_out, fan_in]``). This is important for correct initialization. If you plan to use ``x @ w``, where ``w.shape = [fan_in, fan_out]``, pass in a transposed weight matrix, i.e. ``nn.init.kaiming_normal_(w.T, ...)``. rrrN) rir*r+rrTrrr r r)r rrrPrrrrs rkaiming_normal_r slV FLL DE  .C , *D 3 C ~~a~: s -B BcUR5S:a [S5eUR5S:XaU$URS5nUR5U-nUR X445R SSUS9nX4:aUR 5 [RRU5upg[R"US5nUR5n Xi-nX4:aUR 5 [R"5 URU5RU5 URU5 SSS5 U$!,(df  U$=f)alFill the input `Tensor` with a (semi) orthogonal matrix. Described in `Exact solutions to the nonlinear dynamics of learning in deep linear neural networks` - Saxe, A. et al. (2013). The input tensor must have at least 2 dimensions, and for tensors with more than 2 dimensions the trailing dimensions are flattened. Args: tensor: an n-dimensional `torch.Tensor`, where :math:`n \geq 2` gain: optional scaling factor generator: the torch Generator to sample from (default: None) Examples: >>> # xdoctest: +REQUIRES(env:TORCH_DOCTEST_LAPACK) >>> w = torch.empty(3, 5) >>> nn.init.orthogonal_(w) r$z4Only tensors with 2 or more dimensions are supportedrr'rN)rgrOnumelro new_emptyrt_r linalgqrdiagsignr view_ascopy_r-) r rrrowscols flattenedqrrwphs r orthogonal_r>s,QOPP ||~ ;;q>D <<>T !D  $.66q!y6QI {  <>> w = torch.empty(3, 5) >>> nn.init.sparse_(w, sparsity=0.1) r$rdrrN) rgrOrirMrceilr r rrprandperm) r sparsityrrrr num_zeroscol_idx row_indices zero_indicess rsparse_rqs.aGHHJDDIIho./I q#3T{G...K&z 2L,-F( )#  M  Ms $AB22 Ccn^^^TRmTSSmUUU4SjnSTSTSTS3UlTUlU$)NcV>[R"STSTS3[SS9 T"U0UD6$)Nz `nn.init.z)` is now deprecated in favor of `nn.init.z`.r$r%)r*r+ FutureWarning)argskwargsmethnew_nameold_names rdeprecated_init(_make_deprecate..deprecated_inits; z!J8*TV W  T$V$$r#z z_(...) .. warning:: This method is now deprecated in favor of :func:`torch.nn.init.z"`. See :func:`~torch.nn.init.z` for details.)__name____doc__)rrrrs` @@r_make_deprecaters\}}H}H%$ JHIQzR'j :O (O r#r4)rrN)rrgrN)r')rN)rrrHN)r'N)rIN)-rrr*typingr _Optionalr rrrr2r8r<rTrN Generatorr rr\r^r`rbrjrxrrrrstrrrrrruniformnormalconstantrhdiracxavier_uniform xavier_normalkaiming_uniformkaiming_normal orthogonalsparser#rrsON ( : > "J!  CEP,0 6 6 6 6) 6  6:,0 : : : :) :  ::,0 P P P P P  P ) P P>'f'5'V'$ '& 'V ' "6 "f "*2j.,0&7 &7 &7)&7 &7V,0$9 $9 $9)$9 $9N3$,0 >C >C >C >C >C ) >CF$,0 2; 2; 2; 2; 2; ) 2;n ,00)0l ,0 #) #N. ( #  ! 9 %d 1/ !"23 1 [ )  !r#