h SSKrSSKrSSKrSSKrSSKrSSKJr SSKJr SSK J r SSK r SSK r SSK JrJrJrJr SSKrSSKJrJrJr S\ R.S \ R.S \ R.4S jrS-S\ R.S \ R.S \S-S \S \ R.4 Sjjr"SS5r"SS5r"SS\5r"SS\R>S9r "SS5r!"SS\!5r""SS\!5r#"SS\#5r$"S S!\#5r%"S"S#\#5r&"S$S%\R>S9r'"S&S'\'5r(SS(\RRS)S04S*\*\ -S+\\*S-4S,jjr+g).N)Sequence)Enum)Path) ModelProto TensorProtohelper numpy_helper) apply_plotload_model_with_shape_infersmooth_distributionpkqkreturnc[R"URURS9nUSS[R"USSUSS- 5-USS&US:HUS:-nSX#'US:US:-n[R X$)'U$)z See https://docs.scipy.org/doc/scipy/reference/generated/scipy.special.rel_entr.html#scipy.special.rel_entr. Python implementation. dtypeNr)npemptyshaperloginf)rrresc2c1s \/var/www/fran/franai/venv/lib/python3.13/site-packages/onnxruntime/quantization/calibrate.pyrel_entrrs ((288288 ,C URVVBqEBqEM* *CF 'bAg BCG q&R!V BvvCH JbaseaxiscJUb US:dS5eUcS5e[R"U5R[R5nSU-[R"XSS9- n[R"U5R[R5n[R "X5upSU-[R"XSS9- n[ X5n[R"XCS9nUbU[R"U5-nURUR5$)z Simplifeied version of entropy. Source: https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.entropy.html. This avoids taking a dependency on scipy just for this function. rz0base={base} must be a positive number or `None`.z qk is None?T)r keepdimsr ) rasarrayastypefloat32sumbroadcast_arraysrrr)rrrr vecss rentropyr,'s <4!8W%WW # >'<'> B  rzz *B rBFF248 8B B  rzz *B   (FB rBFF248 8B 2 C sA  RVVD\ 88BHH rch\rSrSr\"/SQ5r\"/SQ5rSr\S5r \S5r Sr Sr g ) TensorDataC)avgstdlowesthighesthist hist_edgesbins)r0r1r2r3r5c [UR55UlUR5Hup#U[R ;a![ SU<S[R S35eU[R;av[US5(d[ S[U5SU<35eUR[R[R4;a[ SURSU<35e[XU5 M g)NzUnexpected value z not in .rUnexpected type z for k=zUnexpected dtype )listkeys_attrsitemsr._allowed ValueError_floatshasattrtyperrfloat16r'setattr)selfkwargskvs r__init__TensorData.__init__Gs6;;=) LLNDA +++ #4QE*BUBUAVVW!XYYJ&&&q'**$'7Qyu%MNN772::rzz"::$'8 %NOO DQ #rc[US5(a[US5(d[S[U5S35eURUR4$)Nr2r3z0Attributes 'lowest' and/or 'highest' missing in r8)rAAttributeErrordirr2r3rEs r range_valueTensorData.range_valueSsLtX&&gdI.F.F #STWX\T]S^^_!`a a T\\**rc[US5(a[US5(d[S[U5S35eURUR4$)Nr0r1z)Attributes 'avg' and/or 'std' missing in r8)rArLrMr0r1rNs ravg_stdTensorData.avg_stdYsItU##74+?+? #LSQUYKWX!YZ Z$((##rcURVs0sHo[X5_M nnURRUS'U$s snf)NCLS)r<getattr __class____name__)rErGdatas rto_dictTensorData.to_dict_s?-1[[9[74##[9nn--U  :sA)r<N) rX __module__ __qualname____firstlineno__ frozensetr>r@rIpropertyrOrRrZ__static_attributes__rrr.r.CsJZ[HIJG  ++ $$ rr.cb\rSrSrS\\\\-44SjrSr Sr Sr Sr Sr S rS rS rS rg ) TensorsDatafrYc |Xl0UlUR5GHup4[U[5(d[ S[ U5S35e[U[5(aU[R:Xa.[U5S:Xa[USUSS9URU'M[U5S:Xa'[USUSUSUSS 9URU'M[ S US S [U5S US35e[U[5(d[ S[ U5S35eX@RU'GM g)NzKeys must be strings not r8rr r2r3)r2r3r4r6zUnexpected tuple for rz , it has z elements: zValues must be TensorData not ) calibration_methodrYr= isinstancestr TypeErrorrBtupleCalibrationMethodMinMaxlenr.)rErlrYrGrHs rrITensorsData.__init__gs""4 JJLDAa%%";DG9A FGG!U##%):)A)AAc!fPQk#-QqT1Q4#HDIIaLq6Q;#-QqT1Q4aPQdYZ[\Y]#^DIIaL"7!uIc!fX[YZX[[\ ]^^a,,"@a  KLLIIaL!rc#8# URShvN gN7fNrYrNs r__iter__TensorsData.__iter__ys99s cXR;$rvrwrEkeys r __contains__TensorsData.__contains__|siirc URU$rvrwr{s r __getitem__TensorsData.__getitem__syy~rc^XR;a[SU<S35eX RU'g)Nz)Only an existing tensor can be modified, z is not.)rY RuntimeError)rEr|values r __setitem__TensorsData.__setitem__s- ii !J3'QYZ[ [ #rc6URR5$rv)rYr;rNs rr;TensorsData.keyssyy~~rc6URR5$rv)rYvaluesrNs rrTensorsData.valuessyy!!rc6URR5$rv)rYr=rNs rr=TensorsData.itemssyy  rcbURRURURS.nU$)N)rUrYrl)rWrXrYrl)rErYs rrZTensorsData.to_dicts/>>**II"&"9"9   r)rlrYN)rXr\r]r^dictrnr.rprIrxr}rrr;rr=rZrarbrrrdrdfsEc:;M6M1N$   "!rrdc$\rSrSrSrSrSrSrSrg)rqrr rgrjrbN) rXr\r]r^rrEntropy Percentile Distributionrarbrrrqrqs FGJLrrqct\rSrSr\S5r\RS\4Sj5r Sr Sr Sr S\ S \ 4S jrS rg ) CalibrationDataReadercl[US5=(a [UR5=(d [$)Nget_next)rAcallablerNotImplemented)clssubclasss r__subclasshook__&CalibrationDataReader.__subclasshook__s%*-M(8;L;L2M`R``rrc[e)z9generate the input data dict for ONNXinferenceSession runNotImplementedErrorrNs rrCalibrationDataReader.get_nexts "!rcU$rvrbrNs rrxCalibrationDataReader.__iter__s rc8UR5nUc[eU$rv)r StopIteration)rEresults r__next__CalibrationDataReader.__next__s >  rc[ervrrNs r__len__CalibrationDataReader.__len__!!r start_index end_indexc[ervr)rErrs r set_rangeCalibrationDataReader.set_rangerrrbN)rXr\r]r^ classmethodrabcabstractmethodrrrxrrintrrarbrrrrsYaa "$"" ""S"S"rr) metaclassc\rSrSrSS\\-S\\S-4SjjrS/4SjrSr S \ 4S jr S r S r S \4SjrS\4SjrSrg)CalibraterBaseN model_pathop_types_to_calibratec,[U[5(a[[U55UlO1[U[5(a[U5UlO [ S5eX lX0lX@lXPl X`l SUl SUl S/Ul g)a :param model_path: ONNX model to calibrate. It should be a model file path :param op_types_to_calibrate: operator types to calibrate. By default, calibrate all the float32/float16 tensors. :param augmented_model_path: save augmented model to this path. :param symmetric: make range of tensor symmetric (central point is 0). :param use_external_data_format: use external data format to store model which size is >= 2Gb. :param per_channel: whether to compute ranges per each channel. z model_path should be model path.NCPUExecutionProvider)rmrnr rmodelr?raugmented_model_path symmetricuse_external_data_format per_channel augment_model infer_sessionexecution_providers)rErrrrrrs rrICalibraterBase.__init__s" j# & &4T*5EFDJ  D ) )4Z@DJ?@ @%:"$8!"(@%&!!$:#; rrc0XlUR5 g)zj reset the execution providers to execute the collect_data. It triggers to re-creating inference session. N)rcreate_inference_session)rErs rset_execution_providers&CalibraterBase.set_execution_providerss$7  %%'rc[R"5n[RRUl[R "UR UURS9Ulg)z) create an OnnxRuntime InferenceSession. ) sess_options providersN) onnxruntimeSessionOptionsGraphOptimizationLevelORT_DISABLE_ALLgraph_optimization_levelInferenceSessionrrr)rErs rr'CalibraterBase.create_inference_sessionsN#113 0;0R0R0b0b -(99  % %%.. rrcURRVs0sHo"RU_M nnURURRVs0sHoDRU_M sn5 URURR Vs0sHoURU_M sn5 URR Vs1sHofRiM nn[5n[R[R1n URRHn UR(aU RUR;dM0[R"U R U R5Hnn X;dM X;nUR R#S5(dM0UR R$R&U ;dMVX;dM]UR)U 5 Mp M X4$s snfs snfs snfs snf)z select input/output tensors of candidate nodes to calibrate. returns: tensors (set): set of tensor name. value_infos (dict): tensor name to value info. tensor_type)graph value_infonameupdateoutputinput initializersetrFLOATFLOAT16noderop_type itertoolschainrBHasFieldr elem_typeadd) rErvi value_infosotitinitrtensors_to_calibratetensor_type_to_calibrater tensor_names rselect_tensors_to_calibrate*CalibraterBase.select_tensors_to_calibrates.3[[-C-CD-Crww{-C D%++2D2DE2DBGGRK2DEF%++2C2CD2CBGGRK2CDE-2[[-D-DE-DTyy-D E"u$/$5$5{7J7J#K KK$$D--A[A[1[#,??4::t{{#KK"1(5GG,,];;!#!4!4!>!>BZ!Z!,!?044[A$L%$00)EEDEsG4G9G> HcUR$)z@ return: augmented onnx model. Call after calling augment_graph )rrNs rget_augment_model CalibraterBase.get_augment_modelszzrc[e)z abstract method: augment the input model to prepare for collecting data. It will: 1. augment the model to be able to collect desired statistics data 2. save augmented model to augmented_model_paths rrNs r augment_graphCalibraterBase.augment_graphs "!r data_readerc[e)zp abstract method: collect the tensors that will be used for range computation. It can be called multiple times. r)rErs r collect_dataCalibraterBase.collect_data "!rrc[e)zU abstract method: compute data based on the calibration method stored in TensorsData rrNs r compute_dataCalibraterBase.compute_data"rr) rrrrrrrrr)Naugmented_model.onnxFFF)rXr\r]r^rnrrrIrrrrrrrrrdrrarbrrrrsz7;3!& <$J < (}t3 [T U]UUUUUU S9 /UlSUl[ UR R R5UlUR R RV s1sHoRiM sn Ul X`l U(aUS:dUS:a [S5eXpl Xlgs sn f)a' :param model_path: ONNX model to calibrate. It is a model path :param op_types_to_calibrate: operator types to calibrate. By default, calibrate all the float32/float16 tensors. :param augmented_model_path: save augmented model to this path. :param symmetric: make range of tensor symmetric (central point is 0). :param use_external_data_format: use external data format to store model which size is >= 2Gb :param moving_average: compute the moving average of the minimum and maximum values instead of the global minimum and maximum. :param averaging_constant: constant smoothing factor to use when computing the moving average. :param max_intermediate_outputs: maximum number of intermediate outputs before an intermediate range is computed. :param per_channel: whether to compute ranges per each channel. )rrrrrNrr z;Invalid averaging constant, which should not be < 0 or > 1.)superrIintermediate_outputscalibrate_tensors_rangersrrrnum_model_outputsrmodel_original_outputsmoving_averager?averaging_constantmax_intermediate_outputs) rErrrrrr r rrrrWs rrIMinMaxCalibrater.__init__*s.  "7!5%=#  %'!'+$!$TZZ%5%5%<%.get_op_version^sO % 2 2 99==*=*=>>'///!3!GyPRST Trc<>SnUS-U-nUS-n[RRX/U/X#S9n[RRSUT/U/US9nTRRR Vs0sHowR U_M nnURTRRRV s0sHoR U _M sn 5 URTRRRV s0sHoR U _M sn 5 X;a#XRRRn O[SU<S35eTR(Ga,[XRRR R"5n S /[%S U 5Qn T"UTR5S :a1UR&R)[R*"S U 55 O[-[.R0"55n[2R4"[6R8"U [6R:S 9U5nURR)U5 TRRR<R)U5 TRRR>RAXV/5 TRRRR)[RB"X;S/55 gs snfs sn fs sn f)Nr __Reshape)r#rReshape)inputsoutputsrz'Unable to guess tensor type for tensor zE, running shape inference before quantization may resolve this issue.rrgaxesr)"rr make_noderrrrrrrrBrrr?rrsrdimrange attributeappendmake_attributernuuiduuid4r from_arrayrarrayint64rrextendmake_tensor_value_info)rreduce_op_namer# reduce_outputintermediate_output reduce_node reshape_noderroi onnx_type tensor_rank reduced_axesreduce_axes_name reduce_axesrreshape_shape_namerEs radd_reduce_min_max:MinMaxCalibrater.augment_graph..add_reduce_min_maxdsH(#->M"/*"< ++// 0C/Dx0K ;;00+-?@&( 1L261A1A1L1LM1L277B;1LKM   4::3C3C3J3JK3Ja 3JK L   4::3C3C3I3IJ3Ia 3IJ K)'499EEOO  =k_MZZ !+":"?"?"K"K"Q"Q"U"UV !:E![$9: !.$**=B))001F1Fv|1\]'*4::<'8$"."9"9"((JJ$$0077 D JJ   ! ! ( (+)D E JJ   # # * *6+H+Hdhci+j k3NKJs7L<LL ReduceMin ReduceMaxsave_as_external_dataN)rrrnr)r*r r+rr,r-rrr'rsaverr)rEtensorsr reshape_shaper=tensorrr<s` @@rrMinMaxCalibrater.augment_graphSs 55djjA  .$//"RXX0NPbc  $$++M: U , l\F v{ 3 v{ 3 JJ  % %"&"?"? rc/UlgrvrrNs rclear_collected_data%MinMaxCalibrater.clear_collected_data $&!rrc UR5nU(dOwURRURR SU55 UR b3[ UR5UR :XaUR5 M[ UR5S:XaURc [S5eUR5n[U[5(d[S[U5S35eUR5 g)NrNo data is collected.z+compute_data must return a TensorsData not r8)rrr'rrunrrsrJr r?rrmrdrorB)rErrts rrMinMaxCalibrater.collect_datas ))+F  % % , ,T-?-?-C-CD&-Q R--9112d6S6SS))+ t(( )Q .43O3O3W45 5    ![))I$q'RSTU U !!#rcXU(dU$UR5GH up4[U[5(aURSnURSnOUupV[X#[5(a#X#RSnX#RSnOX#upxUR(a'XPR Xu- --n X`R X- --n O[ XW5n [Xh5n [U[5(d[X#[5(a[XS9X#'GMX4X#'GM U$)Nrr rh)r=rmr.rOr r minmax) rE old_range new_ranger|rold_minold_maxnew_minnew_max min_value max_values r merge_rangeMinMaxCalibrater.merge_ranges #//+JC%,,++A.++A.#( ).*55#.44Q7#.44Q7#,> ""#&=&=AR&SS #&=&=AR&SS 1 1 %,, 9>:0V0V!+9!P "+!7 3,6rrc [UR5S:Xa UR$[[URS55Vs/sH)oRR 5UR PM+ nnURVs/sHn[[X#SS95PM nn0nUH=nUR5H&upxURU/5RU5 M( M? X RSn [S[U 5S5Vs/sHoURS5SPM n nUVs0sHoUR;dMXU_M n n/n [S[U 5S5HnUR(a8[ R""XUSS9n [ R""XUS-SS9nO7[ R$"XUSS9n [ R&"XUS-SS9nUR((aU[ R&"[ R*"U 5[ R*"U5/SS9nU RU*U45 MU RX45 M [-[.R0[[XSS955nUR(a-UR3URU5UlUR$UUlUR$s snfs snfs snfs snf) zt Compute the min-max range of tensor :return: dictionary mapping: {added node names: (ReduceMin, ReduceMax) pairs } rFstrictNrgrr$r )rsrr r%r get_outputsrrzipr= setdefaultr'r  rpartitionr r rnanmeannanminnanmaxrabsrdrqrrr])rEr6 output_namesr2output_dicts_listmerged_output_dictdrGrHadded_output_namescalibrate_tensor_namesmerged_added_output_dictpairsmin_value_arraymax_value_arraymax_absolute_valuenew_calibrate_tensors_ranges rrMinMaxCalibrater.compute_datas t(( )Q .// /JOPSTXTmTmnoTpPqJrsJrQ**668;@@Jr s(,'@'@ '@# \uE F'@   "A "--a4;;A>"#**@*@*BC>CAsK]G^`a>b" >bq ! , ,S 1! 4>b " /A$ .@TMhMhDh $A!$ $.@ !$ q#0115A"""$**-EYZF[-\cd"e"$**-EYZ]^Y^F_-`gh"i"$)),DXYEZ,[bc"d"$)),DXY\]X]E^,_fg"h~~%'YY0GP_I`/ahi%j" 113EFG o?@6'2  $ $d3/EUZ+[&\' #  ' '+/+;+;D[T U]UUUUUS9 /UlSUl[ UR R R5UlUR R RV s1sHoRiM sn Ul SUl XPl Xpl XlXlSUlXlgs sn f)a :param model_path: ONNX model to calibrate. It is a model path. :param op_types_to_calibrate: operator types to calibrate. By default, calibrate all the float32/float16 tensors. :param augmented_model_path: save augmented model to this path. :param use_external_data_format: use external data format to store model which size is >= 2Gb :param method: A string. One of ['entropy', 'percentile']. :param symmetric: make range of tensor symmetric (central point is 0). :param num_bins: number of bins to create a new histogram for collecting tensor values. :param num_quantized_bins: number of quantized bins. Default 128. :param percentile: A float number between [0, 100]. Default 99.99. :param scenario: see :class:`DistributionCalibrater` )rrrrN)rrIrr rsrrrr rr  collectormethodnum_binsnum_quantized_bins percentilerscenario) rErrrrr~rrrrrrrWs rrIHistogramCalibrater.__init__ s2  "7!5%=  %'!'+$!$TZZ%5%5%<%9K9K9V9V9XY9XX==9XY6:6H6H6T6T6VW6V( 6V W ))+F((,,T6:GM(1'(:$ -@!((6):;!((0 );  % % , ,] ;" t(( )Q .45 5(,'@'@ '@# \uE F'@   "A &&q"-44Q7"#9Df 1qDLeLeGe.Q A. f~~/{{..#'#:#:?? DN 01 !!#]ZW, gsIII#)I( I(rcUR(d [S5e[U[5(a[R nOd[U[ 5(a[RnO>[U[5(a[RnO[S[U5S35e[XRR55$)zh Compute the min-max range of tensor :return: dictionary mapping: {tensor name: (min value, max value)} z9No collector created and can't generate calibration data.zUnknown calibrater z". This method must be overwritten.)r}r?rmEntropyCalibraterrqrPercentileCalibraterrDistributionCalibraterrrorBrdcompute_collection_result)rEcals rr HistogramCalibrater.compute_data~s ~~XY Y d- . .#++C 2 3 3#..C 4 5 5#00C1$t*=_`a a3 H H JKKr) r r}rr~r rr rrrr) NrFrF-X@same)rXr\r]r^rnrrrIrrJrrrdrrarxrys@rr{r{ sv7;3!&*!$J*! (}t3*!*!X '2$(=2$hLkLLrr{cX^\rSrSrSS\\-S\\S-4U4SjjjrSrU=r $)riNrrc .>[T U]UUUUUUUUS9 g)a| :param model_path: ONNX model to calibrate. It is a model path :param op_types_to_calibrate: operator types to calibrate. By default, calibrate all the float32/float16 tensors. :param augmented_model_path: save augmented model to this path. :param use_external_data_format: use external data format to store model which size is >= 2Gb :param method: A string. One of ['entropy', 'percentile', 'distribution']. :param symmetric: make range of tensor symmetric (central point is 0). :param num_bins: number of bins to create a new histogram for collecting tensor values. :param num_quantized_bins: number of quantized bins. Default 128. )r~rrrNrrI) rErrrrr~rrrrWs rrIEntropyCalibrater.__init__s/*   ! $1  rrb)NrFr,Frr rXr\r]r^rnrrrIrarxrys@rrrsC7;3!& $J  (}t3  rrcX^\rSrSrSS\\-S\\S-4U4SjjjrSrU=r $)riNrrc .>[T U]UUUUUUUUS9 g)ag :param model_path: ONNX model to calibrate. It is a model path :param op_types_to_calibrate: operator types to calibrate. By default, calibrate all the float32/float16 tensors. :param augmented_model_path: save augmented model to this path. :param use_external_data_format: use external data format to store model which size is >= 2Gb :param method: A string. One of ['entropy', 'percentile', 'distribution']. :param symmetric: make range of tensor symmetric (central point is 0). :param num_quantized_bins: number of quantized bins. Default 128. :param percentile: A float number between [0, 100]. Default 99.99. )r~rrrNr) rErrrrr~rrrrWs rrIPercentileCalibrater.__init__s/*   ! $!  rrb)NrFrFrrrrys@rrrsC7;3!& $J  (}t3  rrcV^\rSrSrSS\\-S\\S-4U4SjjjrSrU=r $)riNrrc ,>[TU]UUUUUUUS9 g)a, :param model_path: ONNX model to calibrate. It is a model path :param op_types_to_calibrate: operator types to calibrate. By default, calibrate all the float32/float16 tensors. :param augmented_model_path: save augmented model to this path. :param use_external_data_format: use external data format to store model which size is >= 2Gb :param method: A string. One of ['entropy', 'percentile', 'distribution']. :param symmetric: make range of tensor symmetric (central point is 0). :param num_bins: number of bins to create a new histogram for collecting tensor values. :param scenario: for float 8 only, if `scenario="same"`, the algorithm weights and float 8 follow the same distribution, if `scenario="p3"`, it assumes the weights follow a gaussian law and float 8 ~ X^3 where X is a gaussian law )r~rrNr) rErrrrr~rrrWs rrIDistributionCalibrater.__init__s,.   ! $  rrb)NrF distributionrrrrys@rrrs@7;3!& $J  (}t3  rrc`\rSrSrSr\R S5r\R S5rSr g)CalibrationDataCollectorizD Base class for collecting data for calibration-based quantization. c[e)zk Generate informative data based on given data. name_to_arr : dict tensor name to NDArray data rrE name_to_arrs rr CalibrationDataCollector.collects "!rc[e)z/ Get the optimal result among collection data. rrNs rr2CalibrationDataCollector.compute_collection_results "!rrbN) rXr\r]r^__doc__rrrrrarbrrrrs; "" ""rrcn\rSrSrSrSrSrSrSrSr Sr S r S r S r \SS j5rS rSrSrg)riaL Collecting histogram for each tensor. Percentile and Entropy method are supported. ref: https://github.com//apache/incubator-mxnet/blob/master/python/mxnet/contrib/quantization.py ref: https://docs.nvidia.com/deeplearning/tensorrt/pytorch-quantization-toolkit/docs/_modules/ pytorch_quantization/calib/histogram.html cZ0UlXlX lX0lX@lXPlX`lgrv)histogram_dictr~rrrrr)rEr~rrrrrs rrIHistogramCollector.__init__s)  " "4$ rcUR$rv)rrNs rget_histogram_dict%HistogramCollector.get_histogram_dict s"""rc[S5 URS;aURU5$URS:Xa3UR(aUR U5$URU5$[ S5e)Nz/Collecting tensor data and making histogram ...>r,rrDOnly 'entropy', 'percentile' or 'distribution' methods are supported)printr~ collect_valuercollect_absolute_valuer?rs rrHistogramCollector.collect#sl ?@ ;;5 5%%k2 2 [[L (~~22;??))+66cd drc UR5GHiup#[U[5(aUH9n[U[R5(aM$S[ U5SU<35e UVs1sHoUR iM nn[U5S:XdSUSU<35e[R"U5nO<[U[R5(d[S[ U5SU<35eUnUR5nURS:a-[R"U5n[R"U5n O>[R"SUR S9n[R"SUR S9n [R"U5nX R ;at[R""XpR$S9upU R'UR 5n UR [R(:wdS5eXX4UR U'GM UR Un U S n U S n[+U S 5(dS [ U 535e[+US 5(dS [ U535eU SnU Sn[R"U5nUUS :aDUSUS- n[R,"US U-UU-U5n[R."UU45n[R""UUS9upU R'UR 5n U S[U5===U- sss&UR [R(:wdS5eX[1X5[3X54UR U'GMl gs snf)z% Collect histogram on absolute value r9z for tensor=r z6The calibration expects only one element type but got rr)r6zMonly float32 or float16 is supported, every constant must be explicitly typedrgrjrz'old_min should be a numpy array but is rN)r=rmr:rndarrayrBrrsr%r?flattensizergrhr,absoluter histogramrr&float64rAarangehstackrSrT)rErrFdata_arrarradtypes data_arr_npr[r\r4r5 old_histogramrWrXold_histold_hist_edges temp_amaxwidth new_bin_edgess rr)HistogramCollector.collect_absolute_value2sE!, 1 1 3 F(D))#C%c2::66l:J4PS9+Uabhak8ll6$+348a''846{a'LVHT`ag`jk'!jj2 "**55 #3DN3C[R"SURS9n[R"SURS9n[R"[[U5[U55URS9nX R;a2URUnURXsXEU5URU'GM#[R"X0RU*U4S9upUU UUU4URU'GM^ g)z! Collect histogram on real value rrr%N)r=rr%rrrgrhr,rrTrirmerge_histogramrr) rErrFrr[r\ thresholdrr4r5s rr HistogramCollector.collect_valuels+!, 1 1 3 Fzz(+H'')H}}q IIh/ IIh/ HHQhnn= HHQhnn= S^S^!DHNN[I,,, $ 3 3F ; .2.B.B!Y9/##F+$&<<--QZPZ\eOf#g  /##F+)!4rcUupgpn XZ::a?[R"U[U5U *U 4S9upX-U[X5[ X5U 4$U S:Xa)[R"U[U5U*U4S9upX- n Ob[U5nSU -U- n[ XZ- U-S-5nUSU--nUU-U -n[R"UUU*U4S9upU UUU- ===U- sss&U U[X5[ X5U4$)Nrrrgr )rrrsrSrTr)rErrrYrZ new_thresholdrrrWrX old_thresholdnew_histrr4r5 old_num_bins old_stridehalf_increased_bins new_num_binss rr"HistogramCollector.merge_histograms:FSC7]  ),,xX ~WdFefKH#G%G%  !#%<<#h-Q^P^`mOn#o  "8} .= &)=+HZ*WZ[*[&\#+a2E.EE 3j @= P #%<<,P]~_lNm#n (<:M+MNRZZNG%G%  rcvUR(a[UR5S:Xa [S5e[SUR<S35 URS:XaUR 5$URS:XaUR 5$URS:XaUR5$[S5e) Nrz=Histogram has not been collected. Please run collect() first.z0Finding optimal threshold for each tensor using z algorithm ...r,rrr)rrsr?rr~compute_entropycompute_percentilecompute_distributionrNs rr,HistogramCollector.compute_collection_results""c$*=*=&>!&C\] ] @~^_ ;;) #'') ) [[L (**, , [[N *,,. .cd drcvURS:dURS:a [S5eURnURn0n[S[ U535 [SUR 35 [SSU- SUS 35 UR 5GHupEUSnUS nUR5n[R"Xh- 5n UR(a_[R"XS- 5n [R"XzURS 9*[R"XzURS 94X4'O|SU- S - n [R"U S U - 5n [R"X5n [R"X|URS 9[R"XzURS 94X4'USn USnX4SU :a XUS 4X4'X4S U:a X4SU4X4'/X4QUSSQ7X4'[RR!SS5S;dGM[#Xg5 GM U$)NrdzrcURnURn0n[S[U535 [SURS35 [SUR35 UR 5H^upEUR XR5nXcU'/UQUSSQ7X4'[RRSS5S;dML[US US 5 M` U$) Nrrz: (The number may increase depends on the data it collects)zNumber of quantized bins : rgrrrrr ) rrrrsrr=get_entropy_thresholdrrrr )rErrrrFroptimal_thresholds rr"HistogramCollector.compute_entropys,,!44 $S%8$9:; +DMM?:tuv +D,C,C+DEF!/!5!5!7 F $ : :9 Y &7F #&J(9&JIbqM&JO #zz~~2C8HD9Q<16"8rcUS::a[SUS35eUSSUSS-S-nUS:XaX-R5UR5- nXS--R5UR5- US-- S-n[R"XARS9[R"XQRS94$[ U5U:Xa[ U5S-S:XaXU--R5UR5- nXU-U- S--R5UR5- S-n[R"XARS9[R"XQRS94$[R "U5U- nSU[R"U5'SU[R"U5'[R "U5U-U-nX-R5UR5- nXS--R5UR5- US-- S-n[R"XARS9[R"XQRS94$) Nrzpower=z <= 0 is invalid.rr g?rgr) r?r(rr,rrriisnanisinf)r4r5powerrr0r1facts r_avg_stdHistogramCollector._avg_stds A:veW,=>? ?Sb/JqrN2c9 A:=%%'$((*4C19$))+dhhj836AcIC88C'7'78"((3N^N^:__ _ u: 3u:>Q#6%-',,.;CEMC/A55::>RZZ/ //>>RZZ/ //##rzz1 11&0%!~~'"( 'O #zz~~2C8HD4,3"86rc &USnUSnURnUS-nUS-nUSRn[R"Xg- S-5n [ U R5V s/sH-n [R "SUS9[R "SUS94PM/ n n [ XvS-S5GHn Xj- n [ Xj-S-U5n XLXM4XU- '[R"Xr4s  $ >>UU    "    $,   ZZ 8  F11h"ckk"4k"k"\^,~^,BDL.DLN + D . D  0  F"",E!1E!T 37/&--"NK :NK#C=4/NKr