uki26tddlmZddlmZddlZddlmZddlmZddl m Z ddl m Z ddl m Z dd l mZdd lmZmZddlZeZ d dd Z d dd Zd ZddZeddZej4dZeddZej4dZy)) annotations)SequenceN)Any) custom_jvp)dtypes)lax) convolution)util)Array ArrayLikec Vtj|} t|}tj| j d|z|}|\} } } t j|} | j | d}tj| j| | f}tj||| | fd}tj||| zdg|}tj|d| d| df}tj| |||||||dn|tjj f|| }|S) a Extract patches subject to the receptive field of `conv_general_dilated`. Runs the input through a convolution with given parameters. The kernel of the convolution is constructed such that the output channel dimension `"C"` contains flattened image patches, so instead a single `"C"` dimension represents, for example, three dimensions `"chw"` collapsed. The order of these dimensions is `"c" + ''.join(c for c in rhs_spec if c not in 'OI')`, where `rhs_spec == dimension_numbers[1]`, and the size of this `"C"` dimension is therefore the size of each patch, i.e. `np.prod(filter_shape) * lhs.shape[lhs_spec.index('C')]`, where `lhs_spec == dimension_numbers[0]`. Docstring below adapted from `jax.lax.conv_general_dilated`. See Also: https://www.openxla.org/xla/operation_semantics#conv_convolution Args: lhs: a rank `n+2` dimensional input array. filter_shape: a sequence of `n` integers, representing the receptive window spatial shape in the order as specified in `rhs_spec = dimension_numbers[1]`. window_strides: a sequence of `n` integers, representing the inter-window strides. padding: either the string `'SAME'`, the string `'VALID'`, or a sequence of `n` `(low, high)` integer pairs that give the padding to apply before and after each spatial dimension. lhs_dilation: `None`, or a sequence of `n` integers, giving the dilation factor to apply in each spatial dimension of `lhs`. LHS dilation is also known as transposed convolution. rhs_dilation: `None`, or a sequence of `n` integers, giving the dilation factor to apply in each spatial dimension of `rhs`. RHS dilation is also known as atrous convolution. dimension_numbers: either `None`, or a 3-tuple `(lhs_spec, rhs_spec, out_spec)`, where each element is a string of length `n+2`. `None` defaults to `("NCHWD..., OIHWD..., NCHWD...")`. precision: Optional. Either ``None``, which means the default precision for the backend, or a :class:`~jax.lax.Precision` enum value (``Precision.DEFAULT``, ``Precision.HIGH`` or ``Precision.HIGHEST``). preferred_element_type: Optional. Either ``None``, which means the default accumulation type for the input types, or a datatype, indicating to accumulate results to and return a result with that datatype. Returns: A rank `n+2` array containing the flattened image patches in the output channel (`"C"`) dimension. For example if `dimension_numbers = ("NcHW", "OIwh", "CNHW")`, the output has dimension numbers `"CNHW" = "{cwh}NHW"`, with the size of dimension `"C"` equal to the size of each patch (`np.prod(filter_shape) * lhs.shape[lhs_spec.index('C')]`). rr)shape)r)rrrN) lhsrhswindow_stridespadding lhs_dilation rhs_dilationdimension_numbers precisionfeature_group_countpreferred_element_type)rasarraytupler conv_dimension_numbersrmathprod_eyedtypebroadcast_in_dimreshaper moveaxisconv_general_dilated PrecisionDEFAULT)r filter_shaperrrrrrr lhs_arraylhs_specrhs_specout_spec spatial_size n_channelsrouts M/home/cdr/jupyterlab/.venv/lib/python3.12/site-packages/jax/_src/lax/other.pyconv_general_dilated_patchesr2s1~kk#)|$,!88oov ,.?A"3(Hh<(,x{+* |(DE# S:|\"JFS# C*|3QFFG# c6HQK!#=>#((  #)!) 03 0E0E0G$3 # *c tj|} tj|} | d} n2t| trt | dk(r| d} nt d|t| ||||||| } tj| jdt |z|\} }}|dg|dg}}|dd|dd}ttt |fd Dcgc]}|| }}t||f|ff}tj| ||| }tj|d |d|df}|Scc}w) aTGeneral n-dimensional unshared convolution operator with optional dilation. Also known as locally connected layer, the operation is equivalent to convolution with a separate (unshared) `rhs` kernel used at each output spatial location. Docstring below adapted from `jax.lax.conv_general_dilated`. See Also: https://www.openxla.org/xla/operation_semantics#conv_convolution Args: lhs: a rank `n+2` dimensional input array. rhs: a rank `n+2` dimensional array of kernel weights. Unlike in regular CNNs, its spatial coordinates (`H`, `W`, ...) correspond to output spatial locations, while input spatial locations are fused with the input channel locations in the single `I` dimension, in the order of `"C" + ''.join(c for c in rhs_spec if c not in 'OI')`, where `rhs_spec = dimension_numbers[1]`. For example, if `rhs_spec == "WHIO", the unfolded kernel shape is `"[output W][output H]{I[receptive window W][receptive window H]}O"`. window_strides: a sequence of `n` integers, representing the inter-window strides. padding: either the string `'SAME'`, the string `'VALID'`, or a sequence of `n` `(low, high)` integer pairs that give the padding to apply before and after each spatial dimension. filter_shape: a sequence of `n` integers, representing the receptive window spatial shape in the order as specified in `rhs_spec = dimension_numbers[1]`. lhs_dilation: `None`, or a sequence of `n` integers, giving the dilation factor to apply in each spatial dimension of `lhs`. LHS dilation is also known as transposed convolution. rhs_dilation: `None`, or a sequence of `n` integers, giving the dilation factor to apply in each input spatial dimension of `rhs`. RHS dilation is also known as atrous convolution. dimension_numbers: either `None`, a `ConvDimensionNumbers` object, or a 3-tuple `(lhs_spec, rhs_spec, out_spec)`, where each element is a string of length `n+2`. precision: Optional. Either ``None``, which means the default precision for the backend, a ``lax.Precision`` enum value (``Precision.DEFAULT``, ``Precision.HIGH`` or ``Precision.HIGHEST``) or a tuple of two ``lax.Precision`` enums indicating precision of ``lhs``` and ``rhs``. Returns: An array containing the unshared convolution result. In the string case of `dimension_numbers`, each character identifies by position: - the batch dimensions in `lhs`, `rhs`, and the output with the character 'N', - the feature dimensions in `lhs` and the output with the character 'C', - the input and output feature dimensions in rhs with the characters 'I' and 'O' respectively, and - spatial dimension correspondences between `lhs`, `rhs`, and the output using any distinct characters. The examples below use 'W' and 'H'. For example, to indicate dimension numbers consistent with the `conv` function with two spatial dimensions, one could use `('NCHW', 'OIHW', 'NCHW')`. As another example, to indicate dimension numbers consistent with the TensorFlow Conv2D operation, one could use `('NHWC', 'HWIO', 'NHWC')`. When using the latter form of convolution dimension specification, window strides are associated with spatial dimension character labels according to the order in which the labels appear in the `rhs_spec` string, so that `window_strides[0]` is matched with the dimension corresponding to the first character appearing in rhs_spec that is not `'I'` or `'O'`. If `dimension_numbers` is `None`, the default is `('NCHW', 'OIHW', 'NCHW')` (for a 2D convolution). Nrrz6Unsupported precision for conv_general_dilated_local: )rr)rrrrrrrrc|SN)k lhs_b_dimss r1z,conv_general_dilated_local..s JqMr3)key)rr))rrcanonicalize_precision isinstancerlen ValueErrorr2r rrsortedrange dot_generalr r%)rrrrr)rrrrr* c_precision lhs_precisionpatchesr+r,r- lhs_c_dims rhs_c_dims rhs_b_dimsidnr0r9s @r1conv_general_dilated_localrM~sx^kk#)**95+M+u%#k*:a*?NM  @ L NN ) #) '"-!C!Coovl 335F"H(Hh%QK=8A;-j*|*|*'-eC O.D2I(KL! 1 L*Lj!* Z :z":;" iP# c8hqk8A;%?@# *Ls! D<ctj||}tj|d|z}tj|d}tjtj|||}tjtj ||tj|||}tj ||S)zWraps `x` between `[-a, a]`.rr)r_constremaddselectltsub)x_aatwo_azerorPs r1 _wrap_betweenrZs jjB! **QB % Aq $ 1 u%# 366#t$cggc5&93?# ar3cFtj|jtjrt j |n|}t j|td}t jt j||t j||S)Ninf) r issubdtyper"npcomplexfloatingrrealrOfloatrReq_zeros)rUre_xr\s r1 _replace_infresd))!''23E3EF!A$ 4u&# CFF4%szz!}a 88r3ctj|}tj|}|j|jk(sJtj||}t j |jdrtj ||}tjtj|tj||tj|tjtjtjtj|St j |jdrtj tj||tj|tj|d}tj|tjtj|}tj tj"|t%tj&|t(j*St-d|j)z1Compute log(exp(x1) + exp(x2)) avoiding overflow. real floatingcomplex floatingrz9logaddexp requires floating-point or complex inputs; got )rrr"maxrisdtyperTrR_isnanrQlog1pexpnegabsmulrOcomplexr`rZimagr^pirA)x1x2x1_arrx2_arramaxdeltar0s r1 logaddexprzsp ;;r?& ;;r?&  %% %  $ ^^FLL/2 GGFF #E ::cjj'ggff-ggdCIIcggcggcggen6M.N$OP RR~~fll$67 GGCGGBOSWWT3::dA3F%G HE ''$ #''%.1 2C ;;sxx}mCHHSM255&I JJ PQWQ]Q]P^_ ``r3c|\}}|\}}t||}tjtj|tjtj t |t |tj|tjtj t |t |}||fSr6)rzrrQrprmrTreprimalstangentsrtrut1t2 primal_out tangent_outs r1_logaddexp_jvprs &"b &"bR *CGGCGGL4DlS]F^,_$`aCGGCGGL4DlS]F^,_$`ac+ [  r3cjtj|}tj|}|j|jk(sJtj||}tj|dt j dz }tj|jdrtj||}tjtj|tj||tj|tj|tjtjtj tj"|Stj|jdr tjtj||tj|tj|d}tj|tj|tjtj|}tj$tj&|t)tj*|t j,t j dz St/d|j)z4Compute log2(exp2(x1) + exp2(x2)) avoiding overflow.rrrgrhz:logaddexp2 requires floating-point or complex inputs; got )rrr"rirOr^logrrjrTrRrkrQrprlexp2rnrorqr`rZrrrsrA)rtrurvrwrxinvln2ryr0s r1 logaddexp2r$s ;;r?& ;;r?&  %% %  $ ::dAbffQiK (& ^^FLL/2 GGFF #E ::cjj'ggff-ggdCGGFCIIchhswwswwW\~G^>_4`$ab dd~~fll$67 GGCGGFF+SWWT3::dA;N-O PE ''$ #((5/(BC DC ;;sxx}mCHHSM255266RS9CT&U VV QRXR^R^Q_` aar3c|\}}|\}}t||}tjtj|tjtj t |t |tj|tjtj t |t |}||fSr6)rrrQrprrTrer|s r1_logaddexp2_jvpr:s &"b &"b"b!*CHHSWW\"5E|T^G_-`$abCHHSWW\"5E|T^G_-`$abd+ [  r3)NNNNN)rr r) Sequence[int]rrrstr | Sequence[tuple[int, int]]rSequence[int] | Nonerrr5convolution.ConvGeneralDilatedDimensionNumbers | Nonerzlax.Precision | Nonerz DType | Nonereturnr )NNNN)rr rr rrrrr)rrrrrrrrzlax.PrecisionLikerr )rUr rr )rtr rur rr ) __future__rcollections.abcrrtypingrjax._src.custom_derivativesrjax._srcr jax._src.laxrr r jax._src.typingr r numpyr^DTyper2rMrZrerzdefjvprrrr7r3r1rs#$ 2$, *.)-OS&*+/\ \ \ "\ - \ ' \ ' \ M\ $\ )\  \ J*.)-OS#'t t t "t - t  t ' t 't Mt !t  t n9  a a( !! b b* !!r3