iWSSKrSSKrSSKrSSKrSSKrSSKrSSKJrJrJ r J r J r J r SSK Jr SSKJr SSKrSSKrSSKrSSKrSSKrSSKrSSKrSSK J!r" SSK#J$r$ SSK#J%r% SSK#J&r& \\'r(\\)\(4r*\\ \)\(4r+S\RXS\\)4Sjr-S\RXS\ \)S\RX4Sjr.S\RXS\RX4Sjr/S\RXS\ \+S\ \*4Sjr0Sr1SSSSSSSSSSSSS\1S4SS.S\ \)\RX4S\'S\2S \2S!\ \\)S"\2S#\ \)S$\2S\ \ \)S%\)S&\2S\ \RX\244S'jjjr3"S(S)\"Rh5r5S*r6g!\aO \R6SSSS/r\"\"S S R;\5S 3S S 95 SSKr\R>"\5 SSKrGNlf=f)+N)ListDictUnionOptionalTupleSequence)Text)printz-mpipinstall onnxruntimezInstalling onnxruntime by ` z`, please wait for a moment.. bold magentastyle) model_info)model_checking)versionmodelreturncnURRVs/sHoRPM nnU$s snfN)graphoutputname)ropt output_namess V/var/www/html/ai-image-ml/venv/lib/python3.13/site-packages/onnxsim/onnx_simplifier.pyget_output_namesr 's/(- (:(:;(:HH(:L; URRVs/sHoRPM nn[R"URR 5H:nURU;dMURR R U5 M< U$s snfr)r initializerrr%r&inputr')rxinitializer_names graph_inputs rremove_initializer_from_inputr3<sv).)@)@A)@A)@A}}U[[%6%67   0 0 KK   $ $[ 18 L BsB input_shapescn^UcgS[RS[[R4SjmS[RS[[4U4SjjnU"U5nSU;a([ U5S:XaUSXS'US O [ S5eUH"nXC;dM [ SRU55e U$) NrrcURRVs/sHoRPM nnURRVs/sHo3RU;dMUPM sn$s snfs snfr)rr.rr/)rr0r1ipts r get_inputs1check_and_update_input_shapes..get_inputsHsY-2[[-D-DE-DVV-DE${{00V0HHDU4U0VVFVsA(A-A-cT>T"U5Vs/sHoRPM nnU$s snfr)r)rr7 input_namesr8s rget_input_names6check_and_update_input_shapes..get_input_namesLs++5e+<=+s%rrzhThe model has more than 1 inputs, please use the format "input_name:dim0,dim1,...,dimN" in --input-shapez'The model doesn't have input named "{}")onnx ModelProtorValueInfoProtostrlenr$format)rr4r<r;r0r8s @rcheck_and_update_input_shapesrDDsW$//Wd43F3F.GWt49"%(K | { q +7+=LQ (T"z| |  :AA!DF F r,z1.5GBTF)r4check_nperform_optimization skip_fuse_bnskipped_optimizersdynamic_input_shape custom_libinclude_subgraphtensor_size_thresholdmutable_initializerc rU (a[[SSS95 Ub[[SSS95 UnUnU(dSnOUc/nU(aUbURS5 [U[5(a[ R "U5nUc0n[X5n[X5nUR5H|unnURRH\nURU:XdM[URRRR 5HunnUUUlM M^ M~ U b [%X 5nU(dUR&S:a [)U5nS[S [*4S jnU"U5nUS :a [-S 5eUR/5n[0R2"UUU(+U(+U5n[5U5S :Xa [-S5e[ R6"U5n[8R:"UXXYU 5nUU4$![,[ R<R>R@4Gav [S5 [BRD"5n[ RF"[HRJ"U5[LRNRQUS5SS9 [0RR"[LRNRQUS5[LRNRQUS5UU(+U(+U5n[8R:"[LRNRQUS5[LRNRQUS5XX5n[ R "[LRNRQUS55nSSS5 UU4$!,(df  WW4$=ff=f)aT :param model: onnx ModelProto object or file path :param check_n: The simplified model will be checked for `check_n` times by random inputs :param perform_optimization: Whether to run onnx optimizer on the model :param skip_fuse_bn: Skip fuse_bn_into_conv onnx optimizer :param overwrite_input_shapes: If the model has dynamic input shape, user must pass a fixed input shape for generating random inputs and checking equality. :param test_input_shapes: If the model has dynamic input shape, user must pass a fixed input shape for generating random inputs and checking equality. :param skipped_optimizers: Skip some specific onnx optimizers :param skip_constant_folding: Skip constant folding :param skip_shape_inference: Skip shape inference (sometimes shape inference will crash) :param input_data: Feed custom input data for checking if needed :param dynamic_input_shape: Deprecated. Not needed anymore. :param custom_lib: onnxruntime custom ops's shared library :param include_subgraph: Simplify subgraph (e.g. true graph and false graph of "If" operator) instead of only the main graph :param unused_output: name of unused outputs that will be eliminated from the model :param input_shapes: Deprecated. Please use `overwrite_input_shapes` and/or `test_input_shapes` instead. :return: A tuple (simplified model, success(True) or failed(False)) zWARNING: The argument `dynamic_input_shape=True` is not needed any more, onnxsim can now support dynamic input shapes natively, please refer to the latest documentation. An error will be raised in the future.bold redrNzWARNING: The argument `input_shapes` is deprecated. Please use `overwrite_input_shapes` and/or `test_input_shapes` instead. An error will be raised in the future.fuse_bn_into_convsizerc6SSSSSS.nUR5n[R"SU5(d[R"SS U5nUR 5Vs/sHo"R 5PM snup4[ [U5X-5$s snf) Nriii@l)BKBMBGBTBrz ([KMGT]?B)z \1)upperrematchsubsplitstripintfloat)rRunitsstringnumberunits r parse_sizesimplify..parse_sizes|uEeLzz|xxd##66-6D59ZZ\B\6 \B 5=,--CsBiz-tensor_size_threshold should be less than 2GBrz Simplified model larger than 2GBza[bold magenta]Simplified model larger than 2GB. Trying to save as external data...[/bold magenta]z model.onnxT)save_as_external_datazopt.onnx)*r r append isinstancerAr>loadrDitemsrr/r enumeratetype tensor_typeshapedim dim_valuer+ ir_versionr3r_ ValueErrorSerializeToStringCsimplifyrBload_from_stringrcompareonnx_cpp2py_exportcheckerValidationErrortempfileTemporaryDirectorysaver%r&ospathjoin simplify_path)rrErFrGoverwrite_input_shapestest_input_shapesrHskip_constant_foldingskip_shape_inference input_datarIrJrKr!rLrMr4r input_shaper7irpre model_bytesmodel_opt_bytes model_optcheck_ok tmpdirnames rrvrvdsRP c    u   ".( !  #*6!!"56% % %!#: '5 "499;k;;$$Cxx4'(<(<(B(B(F(FGFAs$/NCMH%<  $U: 5#3#3q#8-e4...''<=|+HII%H--/ **  % %$ $ !    1 $?@ @))/: !)) u'8j 4 h / //77GG HH qr  ( ( *j II e$ Z6&*   Z6 Z4"))((% H&-- Z4 Z6JH  "'',,z:"FGI'+ *( h )+ *( h /Hs-A=HAN6 D3N N6 N2 *N62N6c,\rSrSrS\S\\4SjrSrg)PyModelExecutor model_str inputs_strc4[R"5nURU5 Sn[XB5n[[RR U5nUR RVs/sHowRPM nn[[X55n [R"5n [R"S5U lSU l[R "UR#5U S/S9n U R%5Vs/sHowRPM n n[R&"5n SU lU R)XU S9nUVs/sH0n[RR+U5R#5PM2 sn$s snfs snfs snf)NcR[R"5nURU5 U$r)r> TensorProtoParseFromString)tp_strtps rdeserialize_tp+PyModelExecutor.Run..deserialize_tps#!!#B   v &Ir,rCPUExecutionProvider) sess_options providers) run_options)r>r?rmap numpy_helperto_arrayrr/rdictziprtSessionOptionsGraphOptimizationLevelgraph_optimization_levellog_severity_levelInferenceSessionrt get_outputs RunOptionsrun from_array)selfrrrr input_tps input_arrsr0r;inputsrsessrr output_arrss rRunPyModelExecutor.RunsQ! i(  3 **33Y? ',{{'8'89'8!vv'8 9c+23((* 020I0I!0L -*+ '""  # # %%-.  )-(8(8(:;(:1(: ;mmo )* &hh|hM IT ITAD   ( ( + = = ?  :< s1F  F7FN)__name__ __module__ __qualname____firstlineno__rArr__static_attributes__rr,rrrs S d3i r,rc[R"5nURSSS9 URSSS9 URSSS[S S 9 URS S S S9 URSSS S9 URSS[SS9 URSSS S9 URSS[SS9 URSS[SS9 URSS[SS9 URSS[SS9 URS S!S S9 URS"S#S S9 URS$S%S S9 URS&S'[SS9 URS(S%[S)9 URS*S+S S9 URS,S-[SS9 URS.S/[S0[ SS1S29 URS3S4S S9 URS5S6S S9 URS7S8S9S:[ R -S;9 UR5nUR(a[[S95 UR(a[[S?S=S>95 URbURbeUR(a$[[S@S=S>95 URUl UR(a[[SAS=S>95 URbUR beUR(a$[[SBS=S>95 URUlUR c/UlO [#UR 5S :XaSUlUR$(a(UR bUR R'SC5 UR cSDOSEnSFnU"UR(5nU"UR5nUR*(a[,R."5n[0R2"5n[0R4R6UlUR:Ul[0R>"UR@USG/SH9n[BRD"UR:5n O [BRD"UR@5n URF[ :XaBU RHRJH(n U RLSI;dM[[SJSKS>95 O URN(d[QU RHRRV s/sHoR:PM sn 5n [QU RHRTV s/sHoR:PM sn 5n [#U RWU 55S :a[[SLSKS>95 SnURXbf0nURXHTn U R[SM5nSMR]USSN5USNnnUR_U[`RD"U505 MV [SO5 [cU URdUSDUUUR URfURhUSDURjURURlURFURN5unnURn(d"[BRp"UURr5 O [uSP5eU(a#[SS5 [R"U U5 g[ST5 [SU5 [R"U U5 [R"SV5 gs sn fs sn f![ta [vRxR{URr5SQ-n[vRxR}U5(a[vR~"U5 [BRp"[R"U5URrSESEUSR9 GN!f=f)WN input_modelzInput ONNX model)help output_modelzOutput ONNX modelrEz8Check whether the output is correct with n random inputs?r)rnargsrmdefaultz--enable-fuse-bnzEThis option is deprecated. Fusing bn into conv is enabled by default. store_true)ractionz--skip-fuse-bnz Skip fusing batchnorm into conv.z--skip-optimizationzSkip all ONNX optimizers or some of them. To skip all optimizers, use `onnxsim a.onnx b.onnx --skip-optimization`. To skip some of optimizers, use something like `onnxsim a.onnx b.onnx --skip-optimization fuse_bn_into_conv fuse_pad_into_pool`.*)rrmrz--skip-constant-foldingzSkip constant foldingz --input-shapezMThis argument has been renamed to --overwrite-input-shape, please refer to it+z--overwrite-input-shapeaCOverwrite the input shape. The format is "input_name:dim0,dim1,...,dimN" or simply "dim0,dim1,...,dimN" when there is only one input, for example, "data:1,3,224,224" or "1,3,224,224". Note: you might want to use some visualization tools like netron to make sure what the input name and dimension ordering (NCHW or NHWC) is.z--test-input-shapeaThe input shape to generated random inputs for test, useful when the input shape is dynamic. The format is "input_name:dim0,dim1,...,dimN" or simply "dim0,dim1,...,dimN" when there is only one input, for example, "data:1,3,224,224" or "1,3,224,224". Note: you might want to use some visualization tools like netron to make sure what the input name and dimension ordering (NCHW or NHWC) is.z--skip-optimizerz(Deprecated. Refer to --skip-optimizationz--skip-shape-inferencezSkip shape inferencez!--enable-onnxruntime-optimizationz:Enable ONNX Runtime's ORT_ENABLE_BASIC level optimization.z--dynamic-input-shapez Deprecated. Not needed any more.z--input-data-pathz|input data, The value should be "input_name1:xxx1.bin" "input_name2:xxx2.bin ...", input data should be a binary data file.z --custom-lib)rrmz--include-subgraphzyExperimental feature. Simplify subgraph (e.g. true graph and false graph of "If" operator) instead of only the main graphz--unused-outputz=Name of unused outputs that will be eliminated from the modelz--no-large-tensora[Some ops like Tile and ConstantOfShape can produce large tensor and make the model size much larger. Specifying this flag to skip folding these ops, with loss of some optimization chances. It can be followed with a threshold, for example, --no-large-tensor 1M or --no-large-tensor 100KB. A simple '--no-large-tensor' means '--no-large-tensor 1KB'.1KBrL)rrmconstrrdestz--mutable-initializeraBy ONNX specification, initializers can also serve as inputs. This allows users to overwrite their values during runtime, but some useful optimizations like fuse-conv-and-bn will not be applicable anymore. In almost all cases, having an initializer that is also an input is unintended (usually caused by a out-dated PyTorch). So onnxsim treats all initializers immutable to enabling all optimizations. If it is not wanted, you can specify '--mutable-initializer' to disable this behavior.z--save-as-external-datazSave parameters as external data. This will make the .onnx file much smaller, but the .onnx file will depend on the external data file (.data).z-vz --versionrzonnxsim )rrzWARNING: "--enable-fuse-bn" is not needed any more, because fuse bn is enabled by default. "--enable-fuse-bn" flag is ignored now and will raise an error in the future.rOrzWARNING: "--dynamic-input-shape" is not needed any more, onnxsim v0.4 now handles dynamic input shapes automatically. "--dynamic-input-shape" flag is ignored now and will raise an error in the future.zXWARNING: "--input-shape" is renamed to "--overwrite-input-shape". Please use it instead.z@WARNING: subgraph optimization is not supported in v0.4 for now.zWWARNING: "--skip-optimizer" is renamed to "--skip-optimization". Please use it instead.rPFTc J0nUbUHnSU;a-[[[URS555US'M6URS5nSR USS5[[[USRS555pTUR XE05 M U$)N:,)listrr_r]rupdate) shapes_argshapesr0piecesrros r parse_shapesmain..parse_shapess  !a<#'C(>#?F4LWWS\F"%((s #%&*3sF2J4D4DS4I+J&K MM4-0  r,r)r)TileConstantOfShapezYour model contains "Tile" ops or/and "ConstantOfShape" ops. Folding these ops can make the simplified model much larger. If it is not expected, please specify "--no-large-tensor" (which will lose some optimization chances)raYour model contains initializers that are also inputs. This is usually caused by an out-dated PyTorch. onnxsim treats all initializers immutable to enabling all optimizations. If it is not wanted, please specify "--mutable-initializer" to disable this behavior.rrzSimplifying...rgz.data)rgall_tensors_to_one_filelocationzFinish! Here is the difference:zCheck failed. Please be careful to use the simplified model, or try specifying "--skip-fuse-bn" or "--skip-optimization" (run "onnxsim -h" for details).z,Here is the difference after simplification:r)FargparseArgumentParser add_argumentr_rA!DEFAULT_TENSOR_SIZE_THRESHOLDHOLDr parse_argsenable_fuse_bnr r rIroverwrite_input_shaperKskip_optimizerskip_optimizationrBrGrhtest_input_shapeenable_onnxruntime_optimizationr|NamedTemporaryFilerrrORT_ENABLE_BASICrroptimized_model_filepathrrr>rjrLrnodeop_typerMsetr.r/ intersectioninput_data_pathr]rrnprvrErrrJr!rgr~rrsrrbasenameexistsr'r%r&rprint_simplifying_infosysexit)parserargsrFrrrtmp_filer_rrr0r1r; input_tensorsrrdatarrexternal_data_paths rmainr s  $ $ &F  ,>? -@A  G    T  A, C    18OXde  \    !S    U     7     '=l + I   /  L    ?c I   L    k 1 $ x  !_  k)ZRYRaRaEab    D  {     [     ,1K1K1WX X  j   &*%5%5"  R   ##/D4J4J4VW W  i   "&!4!4 %!# T # # $ )"& T33? %%&9:$($:$:$B5 %T%:%:;)$*D*DE ++..0((* 020I0I0Z0Z -08 -    0 0,KaJb c (--( $**+ !!%FFKK$$D||::z, %  # #1H1H I1HA1H IJ5;;+<+<=+  --k: ;a ? \( M ' %%AWWS\F&"+.r $D  $ !6 7&  "     "" !!     ""   !Ix& )) IIi!2!2 345 5 /0))%;  g  <=))%;  G!J=V    WW--d.?.?@7J 77>>, - - II( ) MM) $   "&$('   s%*\#\(2\ \B&__)7rr%rrrZr|typingrrrrrr rich.textr richr numpyrr> onnx.checker onnx.helperonnx.shape_inferenceonnx.numpy_helperr r ImportError executablecommandr subprocess check_callonnxsim.onnxsim_cpp2py_exportonnxsim_cpp2py_exportrurrrr_ TensorShaperA TensorShapesTensorShapesWithOptionalKeyr?r r+r3rDrboolrv ModelExecutorrrrr,rrs{ ?? *3i C$% "8C=+#=>DOOS   ?? +3C=  __  T__QlHmrz|HsI:%,! !%.2 % $"-1!B %!G$%G doo% &G GG G!c+GG GGHSM*GG !G& 4??D !'GT aoo >Ks~~tUI}EG $,SXXg->,??\]es tu'" sE**AF?>F?