Siz9SSKJr SSKrSSKJrJrJrJrJ r J r J r SSK J r JrJrJrJrJrJrJrJr SSKJr S SjrS SjrS SjrS SS jjrg)) annotationsN)AttributeProto FunctionProto GraphProto ModelProto NodeProtoSparseTensorProto TensorProto) make_attribute make_function make_graph make_model make_node make_tensormake_tensor_value_infoset_model_propstensor_dtype_to_np_dtype) from_arrayc URS:wa[SUR<S35eURGH*nURS:Xa[ SUR<S35eURS:XaUR nURS3nUR n[R"U5nXq::aU/s $[[R"[U5[RS 9US 9n[UR5n [S/U/US 9n [S U/UR [[R"U/U S 95S 9n X/s $[ S UR<35e U/$)zReplaces a Constant node with a large tensor (with more than threshold elements) by a sequence of nodes that produces a dummy constant of same shape as original tensor.Constantz!Node type must be 'Constant' not . sparse_valuezEThis feature is not yet implemented for a sparse constant (node name=).value__SHAPEdtypenamerConstantOfShapez'Replacement of constant with attribute )op_type TypeError attributerNotImplementedErrortdimsnpprodrarraylistint64r data_typeroutput) node thresholdvalue_constant_of_shapeattrnew_namer'sizeinitr node_shapenew_nodes [/var/www/html/ai-image-ml/venv/lib/python3.13/site-packages/onnx/tools/replace_constants.py_replace_constantr9sV ||z!;DLL;K1MNN~~ 88~ %%"ii]".  88w EEE**W-H::D774=D v bhhtDzBRD,U__=E"  J !!   +B*C5!QR H ) )!5chh\ B  9> 6Mc ^[U[5(a[UR5nOC[U[5(a[UR5nO[ S[ U5S35e[5mUH3nT[UR5-mT[UR5-mM5 U4Sjn[S/U"S5/SS9n[S/U"S5/S S9n0n[U5GHuuprURS :waMURSn[S U/U"US 35/5n [S URSU RSURS/U"US35/5n [UR5S :Xa$URSRR n O["R$n [SU RS/U"US35/U S9n [SU RS/U"US35/U S9n [SU RSU RS/U"US35/5n[SURSU/UR5nXXX/Xg'GMx ['UR)5SS9H unnUXUS -&M UR+SU5 UR+S U5 [U[5(a@[-UUR.URURUR0UR2S9$[U[5(a@[5UR6UR.URURUUR8S9$[ S[ U5S35e)zReplaces all *ConstantOfShape* by node *Range* to avoid constant tensors. The function is not recursive. The recursivity is done by *replace_initializer_by_constant_of_shape*. Not implemented for type rc>UT;aTRU5 U$SnUSU3nUT;aTRU5 U$US- nM')N_)add RuntimeError)prefixirexisting_namess r8 _find_name9_replace_constant_of_shape_with_range.._find_name[s^  '   v &M XQqc?D>)""4( FA r:rzeror) value_intoner@r! ReduceProd_NRange_RANGECast_RANGEf)to_NfDiv_FLATReshapeT)reverse initializersparse_initializer opset_imports) isinstancerr+r/rr#typesetinputr.r enumerater"lenr$r&r-r FLOATsorteditemsinsertr rrXrYr domain opset_import)onxnodesr/rFcst0cst1updateinodeshapenarQaccldreshuprEs @r8%_replace_constant_of_shape_with_rangervGs'#z""SXX C ' 'SXX3DI;a@AAUN#djj/)#dkk** N Zj&8%9Q GD Zj&7%8A FD F '  <<, ,  1  lUGjE7".F-G H   [[^QXXa[$++a. 9 5'( ) *  t~~ ! #"$$..B""B v }zUG7:K/L.MRT U v }zUG3-/H.Ib Q  BIIaL"))A,/*we_2M1N QXXa[%$8$++Frq/ +(.FLLND9 r#%eai : LLD LLD#z""  HH II JJ"55   #}%% JJ HH II JJ **    /S {!< ==r:c [U[5(a[UR5nOC[U[5(a[UR5nO[ S[ U5S35e[5nUH3nU[UR5-nU[UR5-nM5 0n[U5HupdURS:waMURSRn[URUR S/U/5n[#SURUR5n [%URSRUS9n U RR'U 5 XU'M UR)5H upkXU'M [U[5(a@[+UURURURUR,UR.S9$[U[5(a@[1UR2URURURUUR4S9$[ S[ U5S35e) z7Replaces all fill value of all nodes *ConstantOfShape*.r<rr!rr@r rWrZ)r\rr+r/rr#r]r^r_r.r`r"r$r&rrr-rr appendrdr rXrYr rfrg) rhr1rirEr/rlrmtensor new_tensorr7r2rus r8 _replace_constant_of_shape_valuer{s#z""SXX C ' 'SXX3DI;a@AAUN#djj/)#dkk**F '  <<, , "$$ KK))A31H0I . DKKHT^^A.33:F!!#& u (\\^ e $#z""  HH II JJ"55   #}%% JJ HH II JJ **    /S {!< ==r:c [U[5(aSn/nURHTnURS:Xa0[ XqU5n[ U5S:XaSnUR U5 MCURU5 MV U(ae[URURURURUURS9n U(a [U 5$US:wa [X5$U $U(a [U5$US:wa [X5$U$[U[ 5(Ga[#UR$U=(d UR&UUUS9n UR(V s/sH#n [#U UU=(d UR&UUS9PM% n n [+U U UR,UR.U=(d UR&UR0URUR2S 9n [ UR45S :a;UR4Vs0sHoR6UR8_M nn[;X5 U RS S 2 URHnU RR=5nURS :Xa0UR>S :a U(a[ASUR>S35eURS :Xa)UR>S:a[ASUR>S35eURUl UR>UlM U $[U[B5(d[ES[GU5S35eS n/n[I5n/n/nURJGHXn[MURN5n[PRR"U5nUU::aURU5 MHUS- nURS3nUR[U[PRV"[YU5[PRZS9US95 []UR^5n[aSU/UR/[U[PRV"S/US95S9nURU5 UR=UR5 UcGMUS::dGM$UR[cU[dRf[ U5/55 GM[ /nURhHn[MURN5n[PRR"U5nUU::aURU5 MG[kSURlR<SURnR<S35e URGHsnURS:Xa3[ XqU5n[ U5S:XaUS- nUR U5 MGSn/nURpHnURF[rRt:Xao[wUS5(a^URxbQ[#URxUUUUS9n [{U 5[{URx5:waSn[}URU 5nURU5 M U(a_[aURURUR5n!U!RpR U5 URU!5 US- nGMbURU5 GMv US :ax[UURURV"s/sHn"U"RU;dMU"PM sn"U-URUUS9n#U(a [U#5$US:wa [U#U5$U#$U(a [U5$US:wa [X5$U$s sn fs snfs sn"f)aReplace initializers or constant node by nodes *ConstantOfShape* to reduce the size. This reduce the cost to write a unit test about a specific graph structure. Args: onx: ModelProto threshold: every initializer under this threshold is not impacted ir_version: initializer must be specified as input for `ir_version <= 3`, this must be specified if onx is :class:`FunctionProto` or :class:`GraphProto` use_range: if uses operator *Range* instead of *ConstantOfShape* to avoid constant tensors value_constant_of_shape: value to use as a value for all nodes *ConstantOfShape*, a high value may produce nan or inf predictions Returns: onx, modified ModelProto The function is designed so that the function can be reapplied on a modified model and either replace *ConstantOfShape* with *Range* operators, either replace the fill value for every *ConstantOfShape*. Frr>TrZr@) ir_versionr0 use_ranger1)r0r}r~r1) functions producer_nameproducer_versionr} doc_stringrf model_versionrN z.Range was introduced in opset 11 but opset is r z7ConstantOfShape was introduced in opset 9 but opset is z-onx should be a GraphProto at this stage not rrrr!?r zKThis feature is not yet implemented for a sparse initializer (indices.name=z, values.name=rgrW)@r\rr/r"r9raextendrxr rfrr_r.rgrvr{r(replace_initializer_by_constant_of_shapegraphr}rrrrrrmetadata_propskeyrrrAversionrBrr#r]r^rXtupler'r(r)rr*r+r,rr-rrr INT64rYr%indicesvaluesr$rGRAPHhasattrridr r )$rhr0r}r~r1modified new_nodesr/ cst_nodesnew_onx new_graphf new_functionsmodelproimpop_setn_modificationsremovedadditional_inputs new_initsr5r'r4r3rnew_sparse_initssp_init shape_nodesattsr2rr7rDrs$ r8rrs>#}%%%' HHD||z)-d?VW y>Q&#H  +   T " #   !.. G<!KKFM!\\FN% c: & &GS {RSTUUOIeG#%ITYYwwt} 9    T " 1ii[( rxxT "((;( K )8  J YYKRXXse59:    DII  !jAo  $ $&x1B1BSYKP + 213))W\\"wwt} 9   # #G , !$__1145">>..1 5  * <<: %+D=TUK;1$1$   [ ) >>CN000C%%EE%<EE')',C a5BsuuI%#H(15C KK !""  tzz4;;GH    % %d +   X & q O   T "AD  HH ; 1QVV7%:Q ;>O O JJ!/   8? ? "a '3E;RS S 4S99!#/MM JW  *BP rs#   )% % #%>C%%PS> #S>S>l2> #2>>C2>2>n!%( S 0SSS S # Sr: