o Ei/@sddlZddlmZddlZddlmZddlmZm Z ddl mZ e Z ejddddEd d ZejdFd d Zejjejddd   dGdedejde eefde edfde ejedfdede jdejfddZee_ejj   dGdedejde edfde ejedfdede jdejfddZddZejj   dGdedejde edfde ejedfdede jdejfddZdd Zejj   dGdedejde edfde ejedfdede jdejfd!d"Zd#d$Zejj   dGdedejde edfde ejedfdede jdejfd%d&Z d'd(Z!ejj   dGdedejde edfde ejedfdede jdejfd)d*Z"d+d,Z#ejj   dGdedejde edfde ejedfdede jdejfd-d.Z$d/d0Z%ejj   dGdedejde edfde ejedfdede jdejfd1d2Z d3d4Z&ejj   dGdedejde edfde ejedfdede jdejfd5d6Z'd7d8Z(ejj  dHdedejde edfde ejedfde jdejf d9d:Z)d;d<Z*ejj  dHdedejde edfde ejedfde jdejf d=d>Z+d?d@Z,ejj  dHdedejde edfde ejedfde jdejf dAdBZ-dCdDZ.dS)IN)util)partial)CallableUnioncdfdd S)Ncs||||dS)N)op)expr_in tensor_inexpr_outrbackendctr K/home/ubuntu/veenaModal/venv/lib/python3.10/site-packages/einx/op/reduce.py s ..Nr rrr rrr r)tracec Csb||fD]}|D]}t|tjjjrtdq qtj|g|g|g||d\}}|d|dfS)NzConcatenation not allowed)r r)all isinstanceeinxexprstage3 Concatenation ValueErrorvmap_with_axis_stage3) r r r rr rootr tensors_out exprs_outr r r reduce_stage3 s   r#Tc  stjj||\}}tjj|}t|dkrVtjjtj |dd|gdd| D|ddd}t dd| DsDt d fd d }tjj||}||fSdur^t d t|ddkrqt d t|dt|ddkrt dt|dtjjtj |dd|gtj |ddgdd| D|dddd\}}t dd| Dsdd| Dtjj|fdd}||fS)NrrcS4g|]\}}tjj|t|dtjfdddqS.N)depth1depth2rrEquationnpasarraynewaxis.0kvr r r &"zparse..T)csecse_in_markerscss |] }t|tjjjVqdSr)rrrrMarkerr.rr r r .szparse..z No axes are marked for reductioncs.t|tjjjrrtjjddgSgSdS)Nr)rrrrr5Axisr)keepdimsr rreplace2s zparse..replacez(keepdims cannot be given when using '->'z%Expected 1 input expression, but got z&Expected 1 output expression, but got cSr$r%r(r-r r rr1Gr2css|] }tjj|VqdSr)rrr is_markedr6r r rr7QscSs"h|] }t|tjjjr|jqSr rrrrr8name)r.axisr r r Rs zparse..cst|tjjjo |jvSrr>r9)axes_names_outr rrWs zparse..)rrr!_clean_description_and_parametersrstage1parse_oplensolver)items_anyrrrr;mark) description tensor_shaper:r3 parametersrr r;r r )rBr:rparsesb   "   rNcr)Ncs||fi|Srr )rKtensorr kwargsrr rr`s rrr rr rrr`rrKrOrr:r r3rMreturnc Ks>t|tj|f||d|\}}t|||||d\}}|S)a Applies a reduction operation on the given tensors. The operation reduces all marked axes in the input to a single scalar. It supports the following shorthand notation: * When no brackets are found, brackets are placed implicitly around all axes that do not appear in the output. Example: ``a b c -> a c`` resolves to ``a [b] c -> a c``. * When no output is given, it is determined implicitly by removing marked subexpressions from the input. Example: ``a [b] c`` resolves to ``a [b] c -> a c``. Args: description: Description string for the operation in einx notation. tensor: Input tensor or tensor factory matching the description string. op: Backend reduction operation. Is called with ``op(tensor, axis=...)``. If ``op`` is a string, retrieves the attribute of ``backend`` with the same name. keepdims: Whether to replace marked expressions with 1s instead of dropping them. Must be None when ``description`` already contains an output expression. Defaults to None. backend: Backend to use for all operations. If None, determines the backend from the input tensors. Defaults to None. cse: Whether to apply common subexpression elimination to the expressions. Defaults to True. graph: Whether to return the graph representation of the operation instead of computing the result. Defaults to False. **parameters: Additional parameters that specify values for single axes, e.g. ``a=4``. Returns: The result of the reduction operation if ``graph=False``, otherwise the graph representation of the operation. Examples: Compute mean along rows of a matrix: >>> x = np.random.uniform(size=(16, 20)) >>> einx.mean("a b -> b", x).shape (20,) >>> einx.mean("[a] b -> b", x).shape (20,) >>> einx.mean("[a] b", x).shape (20,) Compute sum along rows of a matrix and broadcast to the original shape: >>> x = np.random.uniform(size=(16, 20)) >>> einx.sum("[a] b -> a b", x).shape (16, 20,) Sum pooling with kernel size 2: >>> x = np.random.uniform(size=(4, 16, 16, 3)) >>> einx.sum("b (s [s2])... c", x, s2=2).shape (4, 8, 8, 3) Compute variance per channel over an image: >>> x = np.random.uniform(size=(256, 256, 3)) >>> einx.var("[...] c", x).shape (3,) )r:r3)rr )rNrtracer get_shaper#) rKrOrr:r r3rMr r r r rreduce^sN  rTcKt||fd|||d|S)z7Specialization of :func:`einx.reduce` with ``op="sum"``sumrr:r r3rTrKrOr:r r3rMr r rrV rVcOt|ddi|S)NrrVr#argsrPr r r sum_stage3r_cKrU)z8Specialization of :func:`einx.reduce` with ``op="mean"``meanrWrXrYr r rrarZracOr[)Nrrar\r]r r r mean_stage3r`rbcKrU)z7Specialization of :func:`einx.reduce` with ``op="var"``varrWrXrYr r rrcrZrccOr[)Nrrcr\r]r r r var_stage3r`rdcKrU)z7Specialization of :func:`einx.reduce` with ``op="std"``stdrWrXrYr r rrerZrecOr[)Nrrer\r]r r r std_stage3r`rfcKrU)z8Specialization of :func:`einx.reduce` with ``op="prod"``prodrWrXrYr r rrgrZrgcOr[)Nrrgr\r]r r r prod_stage3r`rhcKrU)zASpecialization of :func:`einx.reduce` with ``op="count_nonzero"`` count_nonzerorWrXrYr r rris ricOr[)Nrrir\r]r r rcount_nonzero_stage3*r`rjcKrU)z7Specialization of :func:`einx.reduce` with ``op="any"``anyrWrXrYr r rrk.rZrkcOr[)Nrrkr\r]r r r any_stage3=r`rlcKrU)z7Specialization of :func:`einx.reduce` with ``op="all"``rrWrXrYr r rrArZrcOr[)Nrrr\r]r r r all_stage3Pr`rmcKt||fd||d|S)z7Specialization of :func:`einx.reduce` with ``op="max"``maxrr:r rXrKrOr:r rMr r rroT rocOr[)Nrror\r]r r r max_stage3`r`rscKrn)z7Specialization of :func:`einx.reduce` with ``op="min"``minrprXrqr r rrtdrrrtcOr[)Nrrtr\r]r r r min_stage3pr`rucKrn)z=Specialization of :func:`einx.reduce` with ``op="logsumexp"`` logsumexprprXrqr r rrvts rvcOr[)Nrrvr\r]r r rlogsumexp_stage3r`rwr)NT)NNT)NN)/rrnumpyr* functoolsrtypingrr numpy.typingnptrkrIjitr# lru_cacherNtraceback_utilfilterstrTensorboolBackend ArrayLikerTrVr_rarbrcrdrerfrgrhrirjrlrrmrorsrtrurvrwr r r rs     B   O