Ph+?dZddlmZddlZddlmZmZmZmZm Z m Z m Z ddl Z ddl m Z ddlZddlZddlmZddlmZmZddlmZdd lmZmZmZdd lmZej> dd Z ej>ejBejDjFe  dd Z$ej>ddZ%ej>ddZ&ej> ddZ'ej> ddZ(ej> d ddZ)ej>ddZ*y)z9Module for handling op-level validation during exporting.) annotationsN)AnyCallableDictListSequenceTupleUnion) evaluator)symbolic_shapes) _constants _type_utils) _beartype) diagnosticsfx_onnx_interpreter type_utils)_pytreec d|jd|jd|jd|jd|jd S)Nz FX Node: z::z[name=z]. ONNX Node: z[opset=z].)optargetnameopset)fnselfnode symbolic_fnargskwargss pC:\Users\daisl\Desktop\realtime-object-detection\venv\Lib\site-packages\torch/onnx/_internal/fx/op_validation.py!_op_level_debug_message_formatterr sN DGG9Bt{{m6$))=!&&'w{/@/@.A E)diagnostic_message_formatterc  t|||\}}tjtj5 |j|i|} t!|j#||d\} }t%j&|} || i|}t+j,| \}}t+j,|\}}|sJt/|t/|k(sJt1||D]"\}}t3|t4j6rBt9j:|j<r#t5j>|jA} t3|tBjDj6r |jF}t4jHjKt5jD|jMt3|t4j6r|jMn"t5jD|jMdd % dddy#t$r}|j} | jtj d5| j tj |dddn #1swYnxYwtjj | _ Yd}~yd}~wwxYw#t$r} |j} | jtj d5| j tj | dddn #1swYnxYwtjj | _ Yd} ~ dddyd} ~ wt$r} |j} | jtj d5| j tj | dddn #1swYnxYwtjj | _ Yd} ~ dddyd} ~ wwxYw#t($r}|j} | jtj d5| j tj |dddn #1swYnxYwtjj | _ Yd}~dddyd}~wwxYw#t$r} |j} | jtj d5| j tj | dddn #1swYnxYwtjj | _ Yd} ~ dddyd} ~ wwxYw#tN$r}|j} | jtj d 5| j tj |dddn #1swYnxYwtjj | _ Yd}~d}~wwxYw#1swYyxYw) aValidate the op between ONNX Runtime and PyTorch. The function will run the op in ONNX Runtime and PyTorch and compare the results. It doesn't break the exporting process, but saves each op validated result into SARIF, under the section of `fx_onnx_interpreter`. There are three signs can be found: 1. Blue: Pass 2. Yellow: Bypass Args: node (torch.fx.Node): The validated fx.node symbolic_fn (Union[onnxscript.OnnxFunction, onnxscript.TracedOnnxFunction]): The corresponded ONNX node torch_args (list): torch argument inputs torch_kwargs (dict): torch keyword argument inputs fx_graph_module (torch.fx.GraphModule): The fx.GraphModule that contains the nodes z3Op level debug fails due to unsupported input typesNzOp level debug is bypassedzOp level debug fails on PyTorchT)allow_extra_kwargsz#Op level debug fails on ONNXRUNTIMEg-C6?gMbP?)rtolatolzValidation failed)(_wrap_fx_args_as_torch_args ValueErrorinflight_diagnostic log_sectionloggingWARNINGlog_source_exceptionrlevelslevelr default_as ort_evaluatorr IndexError RuntimeError(_convert_torch_args_to_onnxfunction_args param_schemasr,filter_incompatible_and_dtype_convert_kwargs TypeErrorr tree_flattenlenzip isinstancetorchTensor fx_type_utilsis_torch_complex_dtypedtype view_as_real resolve_conj onnxscripttensorvaluetesting assert_closecpuAssertionError)diagnostic_contextrrfx_args fx_kwargsfx_graph_module torch_args torch_kwargs value_error diagnosticexpected_outputs index_error runtime_errorfunction_eager_inputsfunction_eager_attributes type_error ort_outputsflattened_torch_outputs_flattened_function_outputs torch_outputfunction_outputes rvalidate_op_between_ort_torchr_)sB #> Y$  L   i55 6 *t{{JG,G & 9))+#'  %)$PP- & %&*CK&-%9%9:J%K"(/(<(<[(I%"A&&&*+s3M/NNNN-0 #%?. )L/ell66|7I7IJ$11,2K2K2MN  >oz/@/@/G/GH&5&;&;O **LL1557!, =!$$&l3779 +. A 7 6 ';;=  # # OOR   + +GOO[ I   '--55  +??AJ''9UV//MWVV*1199J   7 6 +??AJ''!B//O +1199J  ' 7 6 4 +??AJ''9UV//LWVV*1199J  W 7 6L  +??AJ''!F//O +1199J  q 7 6b F" >/CCE ++GOO=PQ33GOOQGRQQ#.#5#5#=#=    >i 7 6saHWJ, 5O-RB&>s&CD A#G##(')GI  ! !# & ::n %%r!ct|}|tjk(r5tjtj tj ||S|tjk(r5tjtjtj||S|tjk(r5tjtjtj||S|tjk(r5tjtjtj||S|tj k(r5tjtj"tj$||S|tj&k(rUtj(|}tj*|dkDtj,dtj,dSt/j0|rBtj2tj4g|dt/j6|Stj4||S)N)lowhighsizer@g?TFr@)rir<uint8randintr UINT8_MIN UINT8_MAXint8INT8_MININT8_MAXint16 INT16_MIN INT16_MAXint32 INT32_MIN INT32_MAXint64 INT64_MIN INT64_MAXboolrandwhererDr>r?view_as_complexrandnfrom_complex_to_float)rfr@random_numberss rgenerate_random_tensorsrs +E 2E  }}$$:+?+?eSX    }}##**=*=EQV    }}$$:+?+?eSX    }}$$:+?+?eSX    }}$$:+?+?eSX    E*{{ S %,,t"4ell56I  ++E2$$ KK %  =+N+Nu+U V   ;;uE **r!c g}|D]}t|tjjr|jj d}|%|j dk(rt||j}t|tjr2t|j|j}|j|t|ttt fr|j|t#j$|r&|jt#j&|t)d|d|dt+|dt|t,r|jt/||jt|tttjfs||j|t|tj0r|jt3|t)d|d|S) z)Recursively convert fx args to torch argsvalget_attrz:Unexpected input argument type found inside fx.Node. arg: z; arg.meta['val']/get_attr: z"; type(arg.meta['val']/get_attr): .z@Unexpected input argument type is found in node arguments. arg: z; )r;r<fxNodemetagetrgetattrrr=rrfr@rcintfloatrr rbrdr(typer_fx_args_to_torch_argsdevicestr)rKrM wrapped_argsarg fake_tensor real_tensors rrrs 24L c588== )((,,u-K"svv';%oszzB +u||45%%{'8'8 ##K0K#ud);<##K0 ))+6##O$<$<[$IJ PQTPUV11< =_K(), X &    6sO L M c5%++6 73;    $ U\\ *   C )RSVRWWYZ =D r!c"t||}||fS)zvPrepare torch format args and kwargs for op-level validation by using fake tensor to create real tensor to feed in ops)r)rKrLrMrNs rr'r's!0F0J y  r!c|Dchc]}|j}}t|j|}|r|std|dg}i}t |D]\} }|j r|j d|| dDg}1| t|kr\|jst|| tjr|jt|| |j|| |j|vrO|jr&t||j||j<||j||j<|jstd|d||fScc}w)aConvert Python args and kwargs to OnnxFunction acceptable with matching ONNX ParamSchema. NOTE: This is different from the param_schema separating in dispatcher, since at this point we are already sure that the args and kwargs are in order and matched. Args: param_schemas: The parameter schemas of an Op or a OnnxFunction. args: The Python positional arguments supplied by the caller. kwargs: The Python keyword arguments supplied by the caller. allow_extra_kwargs: Whether to allow extra keyword arguments. When set to True, extra/unknown arguments will be ignored. Returns: A tuple of two elements: - A list of Python positional argument. - An ordered dictionary of Python keyword argument names and its values. Raises: TypeError: When allow_extra_kwargs is False and there are unknown kwargs. TypeError: When a required input is not provided. zUnexpected keyword arguments ''c3 K|]}|yw)N).0rs r z;_convert_torch_args_to_onnxfunction_args..Ws7hsshs NzRequired input/attribute 'z' was not provided)rset differencer7 enumerateis_variadic_inputextendr9is_inputr;r<r@rc_convert_tensor_to_numpyrequired) r5rrr$paramall_param_names extra_kwargs tagged_args tagged_kwargsis rr4r4-sU>0==}euzz}O=v;))/:L.8aHIIK$&Mm,5  " "   7d12h7 7D  s4y=~~DGU[[!A""#;DG#DE""47+ ZZ6 !~~,DVEJJEW,X ejj),25::,> ejj) ^^8?QRS S#-&  %%7>sE0c ddl}t|tj rdt j |r#t j|j}|jjjSt|tjr6ttjj|j!St|t"t$frt'|dk(r|j)d|j*St|dtj r|Dcgc] }t-|c}St|dt.r|j)||j0St|dtr|j)||j*St|dt2r|j)|S|S#t$r}ttd|d}~wwxYwcc}w)Nrz% needs numpy, but it's not installed.rro)numpy ImportError__name__r;r<r= is_complexrArBdetachrHr@rjit_type_utils JitScalarType from_dtype onnx_typetuplelistr9arrayr}rrbool_r)inputnpexcxs rrrjsW%&   E "&&u'9'9';//::5AKKMNN%%' u:?88Bbhh8/ / eAh -9>?A,Q/? ? eAh %88E82 2 eAh $88E82 2 eAh &88E? " L/ WXJ&KLMSVVW@sG6G5 G2G--G2)rrr torch.fx.Noder=Union[onnxscript.OnnxFunction, onnxscript.TracedOnnxFunction]returnr) rJzdiagnostics.DiagnosticContextrrrrrKList[fx_type_utils.Argument]rL!Dict[str, fx_type_utils.Argument]rMtorch.fx.GraphModule)rf torch.Sizerr)rfrr@z torch.dtype)rKrrMrrr)rKrrLrrMrrzFTuple[List[fx_type_utils.Argument], Dict[str, fx_type_utils.Argument]])F) r5z'Sequence[onnxscript.values.ParamSchema]rrrrr$rrz Tuple[List[Any], Dict[str, Any]])rzfx_type_utils.Argumentrr)+__doc__ __future__rr+typingrrrrrr r rCr r<torch.fxtorch.fx.experimentalr torch.onnxr rrtorch.onnx._internalrtorch.onnx._internal.fxrrrr> torch.utilsrbeartyper diagnose_callrulesop_level_debuggingr_rirrr'r4rrr!rrs?"DDD 1@*       O     ((!BA>5A> A>OA>* A> 1 A> * A>  A>H &&4 !+!+H ' )'