Ng-ddlZddlZddlmZmZmZddlZddlm Z de dee dee dee dee d ee f d Z de d ee d ee fd Z d!dee dee dee dee deee dfded eee ee ffdZde dee dee dee d ee f dZdejde d efdZ d"dejdee dee dee dee de deee dfdeee dfde de d ejfdZGdd e ZdS)#N)SequenceTupleUnion)OpRunauto_padinput_spatial_shapekernel_spatial_shapestrides_spatialoutput_spatial_shapereturnct|}dg|z}|pdg|z}|dvrz)get_pad_with_auto_pad..(s"???a ! !???rc8g|]}||dzz Srr r!s rr#z)get_pad_with_auto_pad..(s>C C C 12IaL9Q<1, ,C C C rrc8g|]}||dzz Srr r!s rr#z)get_pad_with_auto_pad..,s+NNNQ ! y|q00NNNrc&g|] }|dzSrr r!s rr#z)get_pad_with_auto_pad..,s5R R R "#IaLA R R R rrrr)rrrpadss ` rget_pad_with_auto_padr(%sy>>L<????5+>+>???C C C C 6;L6I6IC C C   \ ! !NNNN% :M:MNNNR R R R ',\':':R R R  s\!A% KrFr' dilations ceil_modecbdgt|z}|pdgt|zdz}|pdgt|z}t|}|!tj|gtj}t |D]}||||z|||zz||||dz zz dz ||z dz} |r%t tj| ||<jt tj| ||<|dd} t |D]}||dz ||zdz} ||dz ||z| z} | ||z ||z |||zz } | dkr,| |xx| dzz cc<| ||zxx| | dzz z cc<|| fS)aCompute output shape according to: https://pytorch.org/docs/stable/generated/torch.nn.MaxPool1d.html?highlight=max+pool#torch.nn.MaxPool1d Pads are used to calculate output shape. Use output shape in turn to calculate the actual pads that are used to pad the input tensor so that computation in pool() will not cause out of bound error. Here is the detail. Thinking kernel as a sliding window, its size: sw = dilation * (kernel - 1) + 1 l_out = (l_in + pad[0] + pad[1] - sw) / stride + 1 # (ceiled if ceil_mode is True) l_in_required = (l_out - 1) * stride + sw l_in_required is used to for computation in pool() which may be larger than padded l_in, because of ceiling. as an example, l_in = 3, kernel = 2, stride = 2, dilation = 1, pad = [0, 0], then sw = dilation * (kernel - 1) + 1 = 1 * (2 - 1) + 1 = 2 l_out = ceil((l_in + pad[0] + pad[1] - sw) / stride + 1) = ceil((3 + 0 + 0 - 1 * (2 - 1) - 1) / 2 + 1) = 2 l_in_required = (l_out - 1) * stride + sw = (2 - 1) * 2 + 2 = 4 l_in_required (= 4) is not equal to l_in (= 3), so we need to pad the input tensor to l_in_required to make sure that the sliding window does not go out-of-bound w.r.t. input tensor. Otherwise pool() will fail. rrrNdtype)rnponesint64rintceilfloor)r'rr r r)r*r dimsdimdim_sizepads_spatial_shape_newsliding_window_sizeactual_padded_input_size extra_pads r!get_output_shape_explicit_paddingr;4sL23%8!9!99  5A30111A5D%G!s3F/G/G)GO " # #DGTF"(333 T{{ @ @  $3i 4#: n 4S 9A => ?   C ! $% %  @(+BGH,=,=(>(>  % %(+BHX,>,>(?(?  % %!!!!WT{{ M M3C81< #NQRR$8$=$A_ F $ $   %!#& '3i 4#:   q== "3 ' ' '9> 9 ' ' ' "4#: . . .)i1n2L L . . . !7 77rc|pdgt|z}dgt|z}tt|D]z}|dvr-tj||dz ||z dz||<3|dkr3tj||||z ||z dz||<lt d|S)ahttps://www.tensorflow.org/api_docs/python/tf/keras/layers/AveragePooling2D output_shape = math.floor((input_shape - 1) / strides) + 1 (SAME) output_shape = math.floor((input_shape - pool_size) / strides) + 1 (VALID) IMPORTANT: this function assumes ceil_mode is False. In tenforflow, ceil_mode is always False. However, ONNX spec allow ceil_mode to be True because ORT does handle the case. rrrrz=auto_pad can only be NOTSET, SAME_UPPER, SAME_LOWER, or VALID)rrmathr3 ValueError)rrr r out_shapers rget_output_shape_auto_padr@us&G!s3F/G/G)GOc-...I 3*++ , , 3 3 3 /2Q6/!:LLMMPQQ aLL  (+.B1.EE%a()  aLLO  rxpcnd}tj|D]}|t||zz }|d|z zS)Nrg?)r.nditerabs)rArByvs rlp_poolrHsB A Yq\\ SVVq[ q>rrpaddedx_shapekernelstridesr? pooling_typecount_include_padc F tdz tjddgt||j} !tjgtj%tjdzgtjntdkrzdzpdgz fd} tjtdtdgfdtDRD]|ddftj fdttjfd tDD} |d kr tj } n/|d kr tj } n|d kr| } ntd |d|dkr|dvr| | | <| | tjtj| | <| S)a]This function is used to calculate the pooling result of a padded tensor padded: the padded tensor x_shape: the shape of the original tensor in [N, C, *spatial_shape] kernel: the pooling kernel strides: the strides out_shape: the shape of the output tensor pooling_type: the pooling type, can be "AVG", "LPPOOL", or "MAX" pads: the padding in an order of head_pad_1, head_pad_2, ..., tail_pad_1, tail_pad_2, ... dilations: the dilation count_include_pad: whether to include the padding in the calculation of average and lp pooling p: the p value for lp pooling rrrr,Nc$t|S)N)rH)rArBs r lp_pool_pzpool..lp_pool_psq!}}rc g|]`}tt|dz|z|zzd|dz |zzz |z dzaSrr)rr1)r"rr)rKr' spatial_sizerLrJs rr#zpool..s    Aq'"q</01q A 1==? aj !       rc g|] }| Sr r )r"rwindows rr#zpool..s-   q    rc g|]U}t||dzz||dzzd|dz |zzz|VSrS)r)r"rr)rKshaperLs rr#zpool..s!" " ' U1q5\ 9 ' U1q5\ 9#$q A 1'E#E!G )! rAVGMAXLPPOOLz Pooling type z% does not support. Should be AVG, MAX>rYr[)rr.zeroslistr-r/r0 itertoolsproductrarrayaveragemaxNotImplementedErrorwhereisnan)rIrJrKrLr?rMr'r)rNrBrFrQ window_valsfrXrTrVs ``` `` ` @@@rpoolrhs0w<.+JJJqQa&j!12JJJr)rrconstant)modeconstant_valuesc6g|]}||zfSr r rms rr#z#CommonPool._run..5ror) r.rXnanr@rr(rrpadrhr;)selfrMrNrArr*r) kernel_shaper'rLrBrJ pading_valuer? pads_shapepads_nprIrFrns ` @r_runzCommonPool._runs (1++!-!6!6:Kq:P:PrvvVW < < <!Y!^^^9&4^^1!'!""+|WI''!""+|WiJ):>>DYY!^FJJJJJE&MMJJJGV*'** , F !  A4K?gabbk<)YOItYY!^FJJJJJE&MMJJJGV*'** , F !  A4Kr)NNNNNNN)__name__ __module__ __qualname__r{r rrrjrjsE   IIIIIIrrj)NF)NNrr)r^r=typingrrrnumpyr.onnx.reference.op_runrstrr1rr(boolr;r@r`floatrHndarrayrhrjr rrrs   ))))))))))''''''!##3-c]  #3-  c] . C HSM hsm    (-1 >8>8 3->8!#>8#3->8c] >8 Xc]D() >8  >8 8C=(3- '(>8>8>8>8B""!#"#3-"c] " c] """"JrxCE(,,0 Z Z JZ c]Z  SMZ c] Z } Z  Z   t# $Z Xc]D()Z Z  Z ZZ Z Z Z zJJJJJJJJJJr