a 8Sic@sdZddlmZddlZddlZddlZddlZddlZddlZddl Z ddl Z ddl Z ddl Z ddl m Z mZmZmZmZmZmZddlZddlmmZddlZddlZddlmZddlmZmZmZmZm Z ddl!m"Z"da#d d Z$iZ%ej&d d Z'ej&d dZ(ej&ddZ)ej&ddZ*dddddej+j,ddddddf ddZ-ddZ.ddZ/ddddddddZ0d d!Z1d"d#Z2d$dd%d&d'd(Z3d)d*Z4d+d,Z5d-d.d/d0Z6d1d2Z7dd3d4Z8d5d6Z9d7d8Z:d9d:Z;d;d<ZZ=d?d@Z>ej?ej@ejAejBejCejDejEejBiZFdAdBZGdCdDZHdEdFZIdddej+j,dddddf dGd.dHdIZJdddddej+j,ddddddddfdJdKZKejLjMdfdLdMZNdNdOZOdPdQZPdRdSZQdddddej+j,ddddddddddfdTdUZRdVdWZSdXdYZTdZd[ZUd\d]d^d_ZVd`dadbdcZWd`dddedfdgZXdhdhd%ddidjdkdlZYdmdnZZdd.dodpZ[d\dhdqdrdsZ\ej+j,fdd`d\dtdudvdwdxdyZ]dzd{Z^d|d}Z_d~dZ`ddZadS)zFunctions to export models into the ONNX IR format. These models can be loaded with the ONNX library and then converted to models which run on other deep learning frameworks.  annotationsN)AnyCallableDictListOptionalTupleUnion)_C) _constants _patch_torchsymbolic_caffe2symbolic_helpersymbolic_registry)GLOBALSFcCstSN)__IN_ONNX_EXPORTrrL/var/www/html/django/DPS/env/lib/python3.9/site-packages/torch/onnx/utils.pyis_in_onnx_export&src cst|tjjs|j}|dur2tjj}|r2t dd}|tjj ksR|tjj krr|rrt j dkrnt dt j d}t|||z"dVWt|tjjs||nt|tjjs||0dS)NaOYou are exporting the model to ONNX while in training mode with 'train' parameter not specified. The model will default to inference mode export. If you wish to export a training amenable ONNX model, specify training=TrainingMode.TRAINING or training=TrainingMode.PRESERVE (to preserve the original model state) in torch.onnx.export().F zYou are exporting the model in training mode with onnx opset version {}. Opset versions lower than opset 12 will not be able to export nodes such as Dropout and BatchNorm correctly.T) isinstancetorchjitScriptFunctiontraining_C_onnx TrainingModeEVALwarningswarnTRAININGPRESERVErexport_onnx_opset_versionformatr_set_training_modetrain)modelmodeZis_originally_trainingZis_export_trainingrrrselect_model_mode_for_export0s:      r*c cst|tjjsi}|D]h}|jD]2\}}t|jdkr(||vrNi||<||||<q(||vr|| D]}|j |qpqzFdVWt|tjjs|D]$\}}|D]\}}||j|<qqn@t|tjjs |D]$\}}|D]\}}||j|<qq0dS)NZO2StateDictHook) rrrrmodules_state_dict_hooksitemstype__name__keyspop)r(Ztmp_mapmodulekvZm_maprrrdisable_apex_o2_state_dict_hook^s* r5ccsJtj}|s|rtjzdVW|sFtjn|sDtj0dSr)ronnxis_onnx_log_enabled enable_log disable_log)verboseZis_originally_enabledrrrsetup_onnx_loggingys   r;c cst||l}t|D}t|}|||fVWdn1s@0YWdn1s^0YWdn1s|0YdSr)r*r5r;)r(r)r:Zmode_ctxZapex_ctxZlog_ctxrrrexporter_contextsr<TcCs*t|||||||||| | | | | |ddS)N)operator_export_type opset_versiondo_constant_folding dynamic_axeskeep_initializers_as_inputs custom_opsetsexport_modules_as_functions)_export)r(argsf export_paramsr:r input_names output_namesr=r>r?r@rArBrCrrrexports"rJcCsN|dkrdS|}|tjr0dS|ttjrJdSdS)Nzprim::ConstantFT) kindoutputr. isSubtypeOfr ListType ofTensors OptionalTypeofTensor)node output_typerrr_is_constant_tensor_lists  rTcCs|D]}|D]}t||qt|rg}|D]4}||}||| || |q<| d| | tj}| |||qdS)Nzprim::ListConstruct)nodesblocks_split_tensor_list_constantsrTrLtoIValueinsertConstantrR moveBefore copyMetadataappendcreate insertBeforesetTyper rNrOreplaceAllUsesWith)gblockrRZsubblockinputsvalinputlcrrrrWs$      rWz_C.Graphz_C_onnx.OperatorExportTypesbool)graphr=_disable_torch_constant_propfixed_batch_sizecCst|t|t|t||dur:t|t||t|t|t|t|t |dt |t|t |dt|t ||t |t|t |t|t|tjt||ttr,t||ttdtd|t|tjrp|durLgn|}|dur^in|}t|||t|t||}t|t|t|dtj t|t!|tj |t|t"|t|t#|}t|tjrt$||tj |S)NFTz graph(%Pi): %Pq = quantized::nhwc2nchw(%Pi) %Pr = quantized::nchw2nhwc(%Pq) return (%Pr)z< graph(%Ri): return (%Ri))%r _jit_pass_inline_jit_pass_inline_fork_wait_jit_pass_lint_jit_pass_lower_all_tuples_jit_pass_constant_propagationrW _jit_pass_dce&_jit_pass_canonicalize_graph_fuser_ops_jit_pass_peephole_jit_pass_fuse_addmm*_jit_pass_onnx_remove_inplace_ops_for_onnx_jit_pass_onnx_preprocess#_jit_pass_prepare_division_for_onnx_jit_pass_onnx_remove_print _jit_pass_onnx_preprocess_caffe2r_quantized_opsclear'_jit_pass_onnx_unpack_quantized_weightsis_caffe2_aten_fallback+_jit_pass_onnx_quantization_insert_permutes,_jit_pass_custom_pattern_based_rewrite_graphtextwrapdedent_jit_pass_erase_number_typesronnx_shape_inference&_jit_pass_onnx_set_dynamic_input_shape_jit_pass_onnx_lint_jit_pass_onnx#_jit_pass_onnx_scalar_type_analysisr$_jit_pass_onnx_peephole4_jit_pass_dce_allow_deleting_nodes_with_side_effects_jit_pass_canonicalize)_jit_pass_onnx_graph_shape_type_inference)rhr=rirj params_dictr@rHr2rrr_optimize_graphs~                                    rcCsrt|d|dD]Z\}}t|trLt|t|krld}t|qt|tr||krd}t|qdS)aYWarns that changes to input dictionaries and strings won't take effect in the traced ONNX graph. We accept dictionaries and strings as ONNX inputs, but they should be only for configuration use. we detect here if these inputs are modified, and if so we warn the user that the changes won't take effect in the traced ONNX graph. raQWe detected that you are modifying a dictionary that is an input to your model. Note that dictionaries are allowed as inputs in ONNX but they should be handled with care. Usages of dictionaries is not recommended, and should not be used except for configuration use. Also note that the order and values of the keys must remain the same. zvThe model seems to have string inputs/outputs. Note that strings will not appear as inputs/outputs of the ONNX graph. N)ziprdictlistr0r r!str)Z input_statesreZ traced_inputwarningrrrwarn_on_static_input_changeRs   rcCs2||jur.tjr.|dur*td||d}|S)zVResolves the arguments that are ignored when export_type != operator_export_type.ONNX.Tz`{}' can be set to True only when 'operator_export_type' is `ONNX`. Since 'operator_export_type' is not set to 'ONNX', `{}` argument will be ignored.F)ONNXr_CAFFE2_ATEN_FALLBACKr r!r%)arg_name arg_valuer=rrr_resolve_args_by_export_typeosrzOptional[bool]int)rAr=r>cCsJ|dkr|durtddS|dur*dn|}|durF|tjjurFd}|S)aeDecides whether the initializers in the graph should be listed as ONNX graph inputs. This method encapsulates the logic to decide whether the initializers in the graph should be listed as ONNX graph inputs (i.e., whether to choose ONNX IR v3 or v4). If keep_initializers_as_inputs is not specified (None), then we decide whether to keep initializers as graph inputs (val_keep_init_as_ip) based on export type. If export type is ONNX, then do not keep initializers as input (val_keep_init_as_ip=False). For all other export types keep initializers as input (val_keep_init_as_ip=True). If keep_initializers_as_inputs is specified, then respect it. Unless opset version <= 8, in which case it must be ignored because for opset version <= 8, all initializers MUST be part of graph input (only ONNX IR v3 is allowed), i.e. val_keep_init_as_ip=True. Special handling is needed for opset version 8 or lower, because irrespective of user input for keep_initializers_as_inputs, the graph must follow ONNX IR v3 semantics, i.e. all initializers must be listed as ONNX graph input. FaSetting 'keep_initializers_as_inputs=False' for opset version8 or lower would lead to an invalid ONNX graph. Therefore, 'keep_initializers_as_inputs=False' is ignored during export.Exported model will have initializers as graph inputs (compliant to ONNX IR v3).TN)r r!rOperatorExportTypesr)rAr=r>val_keep_init_as_iprrr_decide_keep_init_as_inputs rcCs td||S)Nadd_node_names)r)rr=rrr_decide_add_node_namessrcCs2td||}|r.|dur.|tjjur.td|S)Nr?aIt is recommended that constant folding be turned off ('do_constant_folding=False') when exporting the model in training-amenable mode, i.e. with 'training=TrainingMode.TRAIN' or 'training=TrainingMode.PRESERVE' (when model is in training mode). Otherwise, some learnable model parameters may not translate correctly in the exported ONNX model because constant folding mutates model parameters. Please consider turning off constant folding or setting the training=TrainingMode.EVAL.)rrrrr r!)r?r=rrrr_decide_constant_foldings rzinspect.Signature)returncCs*t|d|}t|rt|StddS)Nforwardz/model has no forward method and is not callable)getattrcallableinspect signature ValueError)r(Zshould_be_callablerrr _signatures  rc Csz t|}Wn6tyB}ztd||WYd}~Sd}~00zt|j}|ddkrl|dd}i}t|tr|}nt|trt|}n|g}t|dt r|d}|dd}t |}||dD]>}||vr| ||q|j|} | j | j kr| | j qt|tr |nt|}WnXtyHtdYn<ty}z"td|jdWYd}~n d}~00|S)Nz!%s, skipping _decide_input_formatrselfrz,No input args, skipping _decide_input_formatz!Skipping _decide_input_format {})rrr r!r parametersr0rtuplerlenr\defaultempty IndexError Exceptionr%rE) r(rEsigeZordered_list_keys args_dict args_listZ n_nonkeywordZ optional_argparamrrr_decide_input_formats>         ,rcCsTt|tjr|f}tjj||dddd\}}}t|t||id}|rP||fS|S)NFTstrict_force_outplace_return_inputs_states)r)rrTensorr_get_trace_graphrr)funcrEr=Z return_outs trace_graph torch_out inputs_statesrrr_traces  rcCsXtj|}tjj||dddd\}}}t||tj|krPtd||fS)NFTrzXstate_dict changed after running the tracer; something weird is happening in your model!)rr_unique_state_dictr0rr RuntimeError)r(rEZorig_state_dict_keysrrrrrr_trace_and_get_graph_from_models rcCs^g}t||D]F\}}dt|vrJtj|\}}|t|q||duq|S)NZ PackedParams) rrcrr.rr_flattenr\r) method_graph args_paramsparam_count_listinput_Z arg_params_in_vars_rrr_get_param_count_listsrcsLfddt|}t|t|}|dks4J|rHtd|ddS)zAtorch.jit._flatten removes None. Check if it did so in this case.c3sft|ttfr.|D]}|D] }|Vqqn4t|tr\|D]}|D] }|VqLq@n|VdSr)rrrrvalues)xinneryflattenrrr"s      z._check_flatten_did_not_remove..flattenrzargs contained z None's after flattening. When exporting a ScriptModule or ScriptFunction, no args may be None because that breaks type propagation.N)rrr)originalZ jit_flattenedZflattened_with_noneZnum_nonerrr_check_flatten_did_not_removes   rc Csd}t|tjjtjjfriz)_get_named_param_dict..)rcrrr)rhrZinput_and_param_namesr _params_dictrrr_get_named_param_dicthsrcCsht|}i}|r4t|dtr4|d}|dd}||i|}t|trT|g}nt|tsd|f}|S)Nr)copydeepcopyrrrr)r(rE input_argsZ input_kwargsexample_outputsrrr_get_example_outputsos    rcCsxt|tjrn|jtvrn|}tj|tjd}tj| tj d}|||}|j t|jd}|||fS|fSdS)N)dtype) rrrr_qtype_vtype_map dequantizetensorq_scaledouble q_zero_pointint64to)valueZq_value_dequantizerrZq_valuerrrunpack_quantized_tensors  rcCsDtddt|dddDs2tdd|Dr@tj||S|S)zReturns `torch.jit.trace(model, args)` if model is quantized. Otherwise do nothing and return original model. This is due to https://github.com/pytorch/pytorch/issues/75761. css|]}t|dVqdS)_packed_paramsN)hasattr)rmrrr sz)_pre_trace_quant_model..r+cSsgSrrrrrrrz(_pre_trace_quant_model..css|]}t|ddVqdS is_quantizedFNr)rargrrrrr)anyrrrtrace)r(rErrr_pre_trace_quant_modelsr cs$ddfdd|||S)aTakes in ONNX graph, and mapping from _C.Node to node name in exported ONNX ModelProto. Returns: graph (_C.Graph): A TorchScript IR Graph with ONNX nodes, where each _C.Node gets its name in exported ONNX ModelProto assigned as attribute ``onnx_name``. cSs4|D]}|||q||vr0|d||dS)NZ onnx_name)rVs_)nb_fn node_namesbrrrn_fns  z$_assign_onnx_node_name..n_fncs|D]}||qdSr)rU)rr r r rrrr s z$_assign_onnx_node_name..b_fnr)rhr rrr_assign_onnx_node_names rzpTuple[_C.Graph, Dict[str, torch.Tensor], Optional[Union[torch.Tensor, Tuple[torch.Tensor], List[torch.Tensor]]]]c  Cst|tjtttfr|f}t||}t||\} } } }t| | }zt | ||||| ||d} Wn4t y}ztj d| WYd}~n d}~00t|tj jtj jf}|rt||}d}|D]}|t|7}qtj |\}}t| ||tj|nVt| ttfs| g}n| }tt|\}}tdd|DsLt| ||tj|t| ||t| | }| dusz| tjjkrt| |}|rtj t!j"vrt#| |tj }t$| tjrt%| |tj t&| |}tj dkrt'| t(|}t)| t*| || || fS)aConverts model into an ONNX graph. Returns: graph: A TorchScript IR Graph with ONNX nodes. params_dict: Dict from input param name to param value. torch_out: The output tensors resulting from the trace of ``model``. If ``model`` is a :class:`torch.jit.ScriptModule` or :class:`torch.jit.ScriptFunction`, this will be None, since we are not doing any tracing. )rirjrr@rHr2zTorch IR graph at exception: Nrcss|]}t|ddVqdSrr)routrrrrrz"_model_to_graph..r)+rrrrfloatrgr rrrrr6logrrrrrrr "_jit_pass_onnx_assign_output_shaperrrr _jit_flattenr_set_input_and_output_namesrrr_jit_pass_onnx_eval_peepholer$r onnx_constant_folding_opsets_jit_pass_onnx_constant_foldrr%_jit_pass_onnx_eliminate_unused_items,_jit_pass_onnx_cast_all_constant_to_floating%_jit_pass_filter_non_tensor_arguments#_jit_decay_packed_param_input_types_apply_friendly_debug_names)r(rEr:rHrIr=r?rirjrr@rhrrr2rrZ is_scriptrZexample_outputs_finalexample_outputout_varsdescZoutput_wrappedoutput_tensorsZout_descrrr_model_to_graphs                r$c Cs| durtj} | duri} t| t|tdt|||tt| || }t| |}t |||}t ||}t |||||||||d \}}}| || d|| || |WdS1s0YdS)NT)rr@F) r onnx_default_opsetr_set_opset_version_set_operator_export_type_set_onnx_shape_inferencer<rrrrr$_pretty_print_onnx)r(rErGr:rrHrIr= export_typeZgoogle_printerr>rArBrr?r@rval_add_node_namesval_do_constant_foldingrhrrrrrexport_to_pretty_string.sN     r-c Cs|ptj}t|t||d2t||}t||tjj d\}}}Wdn1sX0Yt }d}| D](} | dd|vrt|| qt||fS)aa Converts the model with operator_export_type set to torch.onnx.OperatorExportTypes.ONNX_FALLTHROUGH once in order to get a list of all the ops that are not supported/implemented by the exporter. Args: model: Same as corresponding arg to torch.onnx.export. args: Same as corresponding arg to torch.onnx.export. training: Same as corresponding arg to torch.onnx.export. opset_version: Same as corresponding arg to torch.onnx.export. Returns: Tuple[torch._C.Graph, List[str]], where the list includes the names of the unconvertible ops. F)r=N)r6prim quantized:r)r r%rr&r<rr$rrONNX_FALLTHROUGHrrUrKsplitr\) r(rErr>rhrrZunsupported_opsZsupported_namespacesrRrrrunconvertible_opsls   * r3csfdd}fdd}t|tr@|r@|}dd|D}n@t|tr|t|dkr|dd |}fd d|D}|}nd}|r||S) Ncs ddD}|tjj_|S)NcSsi|]}|tt|qSr)rtypenamer.)r_mrrr rzM_setup_trace_module_map..__setup_trace_module_map..)r+rrr_trace_module_map)trace_module_mapr(rr__setup_trace_module_maps z9_setup_trace_module_map..__setup_trace_module_mapcsBdfdd}fdd}D]}||||q$dS)NZ _onnx_attrscst|t|dSr)setattr_get_module_attributes)r2re attr_namerr)_track_module_attributes_forward_pre_hookszm_setup_trace_module_map..__register_attribute_hook.._track_module_attributes_forward_pre_hookcsJt}|sdS|}i}t|r:t|}t|t||dSr)r _get_tracing_staterhrrdelattr%_jit_pass_onnx_track_scope_attributes)r2rerL tracing_staterhZ onnx_attrsr=rr%_track_module_attributes_forward_hooks   zi_setup_trace_module_map..__register_attribute_hook.._track_module_attributes_forward_hook)r+register_forward_hookregister_forward_pre_hook)r?rDrr9r=r__register_attribute_hooks    z:_setup_trace_module_map..__register_attribute_hookcSsh|] \}}|qSrr)rr3r4rrr rz*_setup_trace_module_map..rcSs*t|trt|Stdt|jdS)NznOnly type of the `nn.Module` should be passed in the set for argument `export_modules_as_functions`. Got `%s`.)rr.rr4rr/)r4rrr_find_typenames  z/_setup_trace_module_map.._find_typenamecsh|] }|qSrr)rr4)rIrrrHr)rrgr-setr)r(rCr:rGr8Zmodule_typenamesr)rIr(r_setup_trace_module_maps"    rKcCsdtjj_tdSr)rrrr7r "_jit_pass_onnx_clear_scope_recordsrrrr_reset_trace_module_maps rMcs@ttttjj}fdd|DfddDS)Ncsg|]}|dqSr)r1rr3rrrrrz*_get_module_attributes..csi|]}|t|qSrrrN)r2rrr6rz*_get_module_attributes..)typingget_type_hintsr.rnnModule)r2Zbase_m_annotationsr)rr2rr<sr<c(Cs| durtjjj} t|tjjr(tdtdus4Jdaz>t || durTt j } |rh| dkrhtdt ||}|stjrtjj}ntjj}t | t |t|||t| || }t||}t| ||}t|tr|}nt}t||}| duri} t| |||t|||||||||| d \}}}| tjjju}|durNi}t|i}|rxt||t |!}t"||t#|dd}|r|$|| | ||| ||||| \}}}}n(|$i| | d|| ||||| \}}}}|rtj%dt&||| tjjjkr`t'|d ks"Jtj()|d } | *|Wdn1sR0Yn| tjjj+tjjj,fvr| tjjj,krt-j.nt-j/}!t-j0|d |!d @}"|"1tjj2||3D]\}#}$|"1|#|$qWdn1s0Yn| tjjj4krt5j67|r,t5j68|s6Jn t59|t5j6:|tjj2}%tj()|%d } | *|Wdn1sx0Y|3D]R\}#}$t5j6:||#}&tj()|&d } | *|$Wdn1s0Yqnt;d |tjjurB|sBztj<|ddWn2t;y@}'ztj=|'WYd}'~'n d}'~'00Wdn1sX0YWtsnJdat>ntsJdat>0|S)Nztorch.nn.DataParallel is not supported by ONNX exporter, please use 'attribute' module to unwrap model from torch.nn.DataParallel. Try torch.onnx.export(model.module, ...)FTz`export_modules_as_functions` is not supported for `opset_version` < 15.This is because `opset_version` < 15 implies IR version < 8, which means no local function support. )rjrr@rzExported graph: rwbw) compressionzUnknown export type)Z full_check)?rr6 ExportTypes PROTOBUF_FILErrQ DataParallelrrrr(r r%rKrrrONNX_ATEN_FALLBACKrr&r'r<rrrrr_validate_dynamic_axesr$r r"_jit_pass_onnx_function_extractionrr0'_jit_pass_onnx_deduplicate_initializersr _export_onnxrrr serialization_open_file_likewrite ZIP_ARCHIVECOMPRESSED_ZIP_ARCHIVEzipfile ZIP_DEFLATED ZIP_STOREDZipFilewritestrONNX_ARCHIVE_MODEL_PROTO_NAMEr- DIRECTORYospathexistsisdirmakedirsjoinr_check_onnx_proto CheckerErrorrM)(r(rErFrGr:rrHrIr=r*r>r?r@rArjrBrrrCrr+r,Zmodel_file_locationrhrrZdefer_weight_exportZnode_attr_to_nameprotoZ export_mapZval_use_external_data_formatr  opened_filerVzr3r4Zmodel_proto_fileZweight_proto_filerrrrrDs>               . 2 *0D  rDcCsp|D]b}|D]T}|}|t|kr2q|d|}||||vr||||<qqdS)Nr)rUrcrruniquerKrr1)rhrr r4old_namenew_namerrrrs   rcs8fdd}|t|d|t|ddS)Ncs|dur dSt|t|kr8td|t||t|ft}tt||D]\}\}}|dkr||vrd}||||| |  || | }||||krL||qLdS)Nz.set_namesrerL)rrcoutputs)rhrHrIrrrrrs rc CsTz||g|RWStyN}z$d|jd|f|_WYd}~n d}~00dS)z\ This trampoline function gets invoked for every symbolic method call from C++. z!{} (occurred when translating {})rN) TypeErrorr%rE)raop_name symbolic_fnrErrrr_run_symbolic_methods rz_C.NoderRcCs|Sr)addBlockrrrr _add_blocksrz_C.BlockrbcCs|Sr)addInputToBlockrrrr_add_input_to_blocksrz_C.Value)rbrcCs||}|Sr)registerOutput)rbr new_outputrrr_add_output_to_blocks rrzOptional[Callable])domainrr>r=rcCs,t|||s|tjjkrdSt|||S)aTLooks up for the symbolic function in the registry. Args: domain: The domain of the symbolic function. op_name: The name of the op. opset_version: Currect opset used. operator_export_type: An enum in _C_onnx.OperatorExportTypes. Returns: The symbolic function if found, None otherwise. N)ris_registered_oprrr1Zget_registered_op)rrr>r=rrr_find_symbolic_in_registrys rcCs4t|d|}|tjjk}|tjjk}|p2| o2|S)N)rrrr ONNX_ATENrZ)nsrr>r=Zis_exportable_aten_opZis_onnx_aten_exportZis_aten_fallback_exportrrr_should_aten_fallbacks  rcCsRtt|j}|sdS|dj}t|}||vr.ZPythonOpr6cs(i|] }|d|d|qSrrkindOfrNrrrr6rrcs(i|] }|d|d|qSrrrNrrrr6rrcs(i|] }|d|d|qSrrrNrrrr6rrz (Occurred when translating z).) rr$rr|rKendswithr2rZregister_versionrZregister_quantized_opsrrattributeNamesrrr6rrop outputsSizeratrZUnsupportedOperatorErrorrrrr1rZrrE)rarbr rcrr=r>Z ns_op_namerrrrattrsctxrrrrr_run_symbolic_functionEs       &  rcCs`ttd|s td|d|d\}}|dkrLtd|d|d|dkrXd }||fS) Nz+^[a-zA-Z0-9-_]*::[a-zA-Z-_]+[a-zA-Z0-9-_]*$zFailed to register operator z.The symbolic name must match the format Domain::Name, and should start with a letter and contain only alphanumerical charactersrr6z. z domain cannot be modified.rr)rgrematchrr2) symbolic_namerrrrrget_ns_op_name_from_custom_ops  rcCs@t|\}}ttjtjgD]}||krt||||qdS)avRegisters a symbolic function for a custom operator. When the user registers symbolic for custom/contrib ops, it is highly recommended to add shape inference for that operator via setType API, otherwise the exported graph may have incorrect shape inference in some extreme cases. An example of setType is `test_aten_embedding_2` in `test_operators.py`. N)r itertoolschainr onnx_stable_opsetsonnx_main_opsetrZ register_op)rrr>rrversionrrrregister_custom_op_symbolics   rcCs>t|\}}ttjtjgD]}||krt|||qdSr)rrrr rrrZ unregister_op)rr>rrrrrrunregister_custom_op_symbolics   rc Cs0t|dkrdSt|drj|dus.t|dkrBdd|jD}|dusVt|dkrjdd|jD}t|prg|pxg}|D]\}}||vrtd |t |t rtd |i}t |D]T\}} t | t std | |vrtd | |qt|d t|d || <q|||<qdS) z=Ensures dynamic axes argument is follows the expected format.rNrhcSsg|] }|qSrr)rrrrrrrz*_validate_dynamic_axes..cSsg|] }|qSrr)rrrrrrrzAProvided key {} for dynamic axes is not a valid input/output namezNo names were found for specified dynamic axes of provided input.Automatically generated names will be applied to each dynamic axes of input {}z3The type of axis index is expected to be an integerz:Duplicate dynamic axis index {} was provided for input {}.Z_dynamic_axes_r)rrrhrcrrJr-r r!r%rrrrrr) r@r(rHrI valid_nameskeyr value_dictrrrrrr[sF     r[)FFNNNN)F)b__doc__ __future__r contextlibrrrrkrrrOr rdrrrrrr r rZtorch._C._onnxr _onnxrtorch.jit._traceZtorch.serialization torch.onnxr r rrrtorch.onnx._globalsrrrrcontextmanagerr*r5r;r<rrrJrTrWrrrrrrrrrrrrrrrquint8uint8qint8int8qint32int32quint4x2rrr rr$r-rrr3rKrMr<rDrrrrrrrrrrrrrrr[rrrrs $   -    %x+'  /  { ? (;  [ _