2Vh`cddlZddlZddlmZddlmZddlmZ ddZdZ dZ dZ d Z d Z d Z dd Z dd ZdZdZy)N)tree)convert_to_tensor) get_devicec P  ./01234567xsjdd} |stj| tj} | djd} | }|t|jt j k7r|jt j }t|jdk(rt j|d}|s| |}gdd.|r| s tdt}g}g}fd}tjrtj|6n |f66fd }|t j|}r!t j||j }t| D] }||}||3|t|tz\}}.3|}|st j |}n|d}t j"|||}tj|}tj|}t.3fd |D}td t%|||D}tj&||} r$|j)||j)||g}|g}|d}|d}t j*|} rt j".|d||t j |}t j".||d |t j |}n^t| D]V}||}|t|tz\}} r#|j)||j)|Q|g}|g}X|d}|d}t j*|}nt}tfd| D2tj&| Dcgc]}|d c}} | ttz\}!}" r|nd}#g}$tj|!D]M}%t-|%}&t|%|#kr!|&j/gg|#t|%z z|$j)|&Ot j0dt j2}'|}(n/t5dr!t7t j8}(n}(|Brt j|dg}t-t j|44fd5.fd0n]t;t j<rAr6t j8d})t j>|)dz 77fd5nfd5d0nd55ntdtj|!D10125 f d}*d}+|$|1}}},|'|krp|+|(krk|*|'|,|g|}-|-dd\}'},}|-dd}|+dz }+|'|krH|+|(kr)nA2 fd}*d}+|$},|}|'|kr,|+|(kr'|*|'|,g|}-|-dd\}'},|-dd}|+dz }+|'|kr|+|(kr'd/-d}$t/fd|$D}td|D}tj&|!|}tj&|!|}|stj| |}|||fScc}w)Ncttt|j}d\|d<|d<t j ||S)Nrrrr)listrangelenshapetorchpermute)input_taxess K/home/dcms/DCMS/lib/python3.12/site-packages/keras/src/backend/torch/rnn.pyswap_batch_timestepz rnn..swap_batch_timesteps=E#gmm,-.Qa}}Wd++rctj|rtd|tj|rtd|t|jt|jz }t |D]}t j|d}dg|zt|j|dz}t j||S)Nz:mask_t is expected to be tensor, but got z;input_t is expected to be tensor, but got rr) r is_nested ValueErrorr r r r unsqueezer tile)mask_tr fixed_dim rank_diff_ multipless r _expand_maskzrnn.._expand_mask0s >>& !!($  >>' "")%  &V\\):: y! 1A__VR0F 1C)Od7==+D&EE zz&),,rz/Unrolling requires a fixed number of timesteps.cFtj|}r|ddd}|S)Nr)runbind)r go_backwardss r_process_single_input_tz$rnn.._process_single_input_tMs&ll7+G!$B$-Nrc^Dcgc]}|| }}tj|Scc}wN)rpack_sequence_as)timet_inpinputsprocessed_inputs r_get_input_tensorzrnn.._get_input_tensorZs2&562d86C6((5 57s *dimsc30K|] }|ywr').0sr!rs r zrnn..ts%01L+%sc3RK|]\}}}tj|||!ywr'rwhere)r3mr4pss rr5zrnn..ws+* 1bKK1b)*%'rc 3K|]W}sttj|n2ttjtj|dgYyw)rN)r rr#flip)r3input_r$s rr5zrnn..sR $U\\&)*%,,uzz&1#'>?@A sAA dtype__len__c|Sr'r2)r)mask_tas r masking_fnzrnn..masking_fns t}$rcjtfd|D}tdt|||DS)Nc3ZK|]"}|tj$yw)r<N)r r )r3or!rs rr5z5rnn..compute_masked_output..s.%!c&,,6GHH%s(+c3RK|]\}}}tj|||!ywr'r7)r3r9rHfms rr5z5rnn..compute_masked_output..s+ 1bKK1b)r;tuplezip)rflat_out flat_mask tiled_mask_tr!s` rcompute_masked_outputz"rnn..compute_masked_outputs?$%%%  $' h $Jrdimc0tj|Sr')rless)r)rev_input_lengths rrEzrnn..masking_fns ::&6==rc0tj|Sr')rgreater)r) input_lengths rrEzrnn..masking_fns ==t<.compute_masked_output..s*BKK2.s$'rK)rrNrOs` rrQz"rnn..compute_masked_outputs$#&x#;rc3FK|]}tj|ywr')r zeros_liker3rHs rr5zrnn..s %()  #%s!c( tfdD}tj|}}|t|tz\}}tj|}rntj|} ||| } tj|} tj|} || | } tj|| }rnd}t || D] \}}|||< dz|t| ft|zS)asRNN step function. Args: time: Current timestep value. output_ta_t: TensorArray. prev_output: tuple of outputs from time - 1. *states: List of states. Returns: Tuple: `(time + 1, output_ta_t, output) + tuple(new_states)` c3(K|] }| ywr'r2r3tar)s rr5z%rnn.._step.. %B2bh%BrrrLrr(flattenrM)r) output_ta_t prev_outputstates current_inputroutput new_states flat_outputflat_mask_outputflat_new_output flat_stateflat_new_stateflat_final_stateta_index_to_writercoutrQ constantsflat_zero_outputinput_tar,rEreturn_all_outputs step_functionzero_output_for_masks` r_stepzrnn.._steps0!&%B%B B $ 5 5fm L #D)%2!5=53C#C&" #ll62 ,%k2! #8K)9# "\\&1 !%j!9#8NJ$ "22:?OP ,>DA!";@0GB,/B()0q+u_/EFJrctfd D}tj|}|t|t z\}}tj|}tj|}rnd}t ||D] \} } | | |< tj |}dz|ft|zS)a)RNN step function. Args: time: Current timestep value. output_ta_t: TensorArray. *states: List of states. Returns: Tuple: `(time + 1,output_ta_t) + tuple(new_states)` c3(K|] }| ywr'r2rbs rr5z%rnn.._step..Jrdrerrrf)r)rhrjrkrlrmrrrnrtrcrurvinitial_statesrxr,ryrzs` rr|zrnn.._step?s!&%B%B B $ 5 5fm L %2!5=53C#C&" "&j!9"ll62 ,>DA!"; <0GB,/B()0"22"N q+.z1BBBrct|Dcgc]}|jc}}g}t|D]&\}}|j|k(s|j|(t j |Scc}wr')maxndim enumerateappendrstack) tensor_listt max_ndimsmax_listis r_stackzrnn.._stackdsf[9QVV9:IH!+. '166Y&OOA& ';;x( ( :sA/c3.K|] }|ywr'r2)r3rHrs rr5zrnn..ns5aq 5sc3&K|] }|d yw)rNr2r_s rr5zrnn..os3aAbE3s)r) r r map_structurergrArbooltyper rrrLrr#r>r r^r8rMr(rrr extendtensorint32hasattrrr isinstanceTensorsubtract)8rzr,rr$maskrvunrollrY time_majorr{ryrflattened_inputs time_steps time_steps_trjsuccessive_statessuccessive_outputsr%r. mask_listrr+rlrmrPri flat_statesflat_new_statesflat_final_states last_outputoutputsinput_time_zerooutput_time_zeroroutput_ta_size output_taruout_listr)max_iterationsmax_lenr|itrh final_outputsr!rrQrwrxrrDrEr-rVs8```` ` ` `` @@@@@@@@@@rrnnr s 26<<?L, ##$7@||F+!!$**1-JL  :: #99UZZ(D tzz?a ??4,D&t,D -"NO O~&  >>& !"00'O 7v>@O 6   T*I!JJyyG :&! 1'*"1%2vy)99&"  ,FF; )"'"2"26":K"4R"8K\6;G"ll62 "&,,z":$%5@%  %**$'$o{%*%! ..v7HI%&--f5%,,V4*0&)/%C! 1D-R0K*2.Jkk"45G##kk 2 <$$[1  ++ w!<$$W-:& 1'*!.vy)99"&&--f5%,,V4*0&)/% 1-R0K*2.Jkk"45G~&  +   // '78SV8 , U>2U95EE !*< << 01 'CCyH3x.(S(A BC   X &  ' ||AU[[1  )N|Y/0> !&>'A+|#L > = J  ! %%-1\\:J-K%  , , \B &1K %"~*= %+{!5?! 2?r1B.k;*12. a %"~*= C C8B#KJ%"~*= %dK E* E $1"1$5!k*12. a %"~*=  )"!$ 59553733 ''(8'B++,$++>EEAG -"9"9!"<< 99T.. //rcXtjtj|dS)NrrR)ranyr)rs r_has_fully_masked_sequencers 99UYYu!, --rcvt|}t|}||z}|js d}t|y)NaYou are passing a RNN mask that does not correspond to right-padded sequences, while using cuDNN, which is not supported. With cuDNN, RNN masks can only be used for right-padding, e.g. `[[True, True, False, False]]` would be a valid mask, but any mask that isn't just contiguous `True`'s on the left and contiguous `False`'s on the right would be invalid. You can pass `use_cudnn=False` to your RNN layer to stop using cuDNN (this may be slower).)rritemr)rno_fully_maskedis_right_paddedvalid error_messages r_assert_valid_maskrsI1$77O/5O o -E ::< B '' rcX|sdnd}tj|j|S)a]Calculate the sequence length tensor (1-D) based on the masking tensor. The masking tensor is a 2D boolean tensor with shape [batch, timestep]. For any timestep that should be masked, the corresponding field will be False. Consider the following example: a = [[True, True, False, False] [True, True, True, False]] It is a (2, 4) tensor, and the corresponding sequence length result should be 1D tensor with value [2, 3]. Note that the masking tensor must be right padded that could be checked by, e.g., `is_sequence_right_padded()`. Args: mask: Boolean tensor with shape [batch, timestep] or [timestep, batch] if time_major=True. time_major: Boolean, which indicates whether the mask is time major or batch major. Returns: sequence_length: 1D int32 tensor. rrrR)rrint)r batch_firsttimestep_indexs r"_compute_sequence_length_from_maskrs$**QqN 99TXXZ^ 44rc,|j|}|j}tj|dd\}}}} tj|||| gdj } tj|dd\} } } }tj| | | |gdj }|#t |d}tj|}n4tjd|z|}tjd|z|}t | dj}t |dj}t |dj}t |dj}|j|}|j|}|j|}|j|}tj5|jj||jj||jj||j j|ddd|j#|j%D]2}|j&|k7s|j(j||_4y#1swY_xYw)aVCopies kernel and recurrent kernel weights in the Pytorch format We split the kernel and recurrent kernel weights, create associated torch tensors adapted to be in line with the Cudnn optimization. After we have copied the weights, we ensure the paramters are on the same device and memory layout is optimized for Cudnn. r)axisNfloat32r@r)to hidden_sizenpsplit concatenateTrrr^zeros contiguousno_grad weight_ih_l0copy_ weight_hh_l0 bias_ih_l0 bias_hh_l0flatten_parameters parametersrdata)lstmkernelrecurrent_kernelbiasrri_kf_kc_ko_kweight_ih_datai_rf_rc_ro_rweight_hh_data bias_ih_data bias_hh_data weight_ih weight_hhbias_ihbias_hhparams rprepare_lstm_weightsrs! 776?D""K&!!4Cc3^^S#sC$8qACCN"2AA>Cc3^^S#sC$8qACCN (Y? '' 5 {{1{?6B {{1{?6B ". BMMOI!. BMMOI I>IIKG I>IIKG V$I V$Ijj Gjj G '  *  * g& g& ' "/ <<6 !v.EJ/''s ?A-J  Jctjjxr(tjjjSr')rcuda is_availablebackendscudnnr2rr_is_cuda_cudnn_availabler s- :: " " $ L)=)=)J)J)LLrcddlm}ddlm}||jtj|jfvxr>||j tj |j fvxr| xr|xr t S)Nr) activations)ops) keras.srcrrtanhrsigmoidr) activationrecurrent_activationruse_biasrrs rcudnn_okrsv & {''SXX>> '    < = 'J '   ' % & rc t||| |du} | stt}ddlm}t ||r |j }t ||r |j }t ||r |j }t|d}t|d}t|d}| t|d}| r8| rdnd}tj||g}|tj||g}|j|}|j|}|j|}||j|} t|||||||| | | | S#t$rtwxYw) N)rr)Variablerr@rrr/) rNotImplementedErrorrkeras.src.backend.torchr rvaluerrr>r _cudnn_lstm Exception)r,initial_state_hinitial_state_crrrrrrreturn_sequencesr$rrcudnn_supportedrr seq_dims rrr$spT! O !!\F0&(#"H-+11$!zzvY 7F'yIO'yIO  V4"!F'3  ::d'3DYYv F%((0O%((0O wwv"              "!!"s )D>>Ec |t|t||} |s|jddd}|sdnd\} } |jdk(r#|j d}|j d}nK|jdk(r8|j ddk(r&|jddd}|jddd}|j d}|j d}t jj||d|d}t||||| |t j d \}}||}|dd|f}|dd|f}t jjjj||j|}||||f\}\}}t jjjj||\}}n||||f\}\}}|j!j#j}|j!j#j}|j!j#j}|j%| }|j%| }||}n|r |ddd fn|d }| s$|r|j dn|j d}|r| rt j&|| g }||||gfS) Nrrr)rrr r}F) input_sizer num_layersr bidirectionalT) descendingrr/)rrrrSrr rnnLSTMrsortutilsrpack_padded_sequencecpupad_packed_sequencedetachclonesqueezer>)r,rrrrrrrr$rrsequence_lengthsseq_axis batch_axisrrrsorted_lengthssorted_indices sorted_inputssorted_initial_hsorted_initial_c packed_inputspacked_outputsh_nc_nrrrs rr r ps 4 =dKP 1a()46&Hj!)33A6)33A6    ! #(=(=a(@A(E)11!Q:)11!Q:aJ"((+K 88==  Dv'7vF ). * &~. *1n+<=*1n+<=**?? >--/ &* ,.>?& " c XX^^'';; K  #6O_+MN#snn$$&**,G **,    " " $C **,    " " $C ++j !C ++j !C  (3gaen   ! !! $&&q)  (**WH:6 #s ++rctr')r )argskwargss rgrur2s r)FNNFNFFT)T)FFFT)numpyrrrrkeras.src.backend.torch.corerrrrrrrrrrrr r2r2rrr5s :3  n,b 0F.((526/rM : I"Xb,Jr