ijddlZddlZddlZddlZddlmZddlmZddlm Z ddlm Z ddlm Z ddl m Z ddl mZejd d Zd Zd Zd ZdZdZdZdZdZdPdZdZdZdQdZdZdRdZdZdZ dSdZ!dZ"dTdZ#dSdZ$dUdZ%d Z&dPd!Z'd"Z(dUd#Z)dUd$Z*dUd%Z+d&Z,d'Z- dVd(Z. dVd)Z/d*Z0dWd+Z1dWd,Z2dWd-Z3dWd.Z4dWd/Z5dWd0Z6dXd1Z7dXd2Z8d3Z9 dYd4Z:d5Z;d6Z< dXd7Z= dYd8Z> dYd9Z? dZd:Z@d[d;ZAd[d<ZBd=ZCd\d>ZDd\d?ZEd]d@ZFd^dAZGdBZHdCZI d_dDZJd`dEZK dadFZLdGZMdHZNdIZOdJZP dbdKZQdcdLZRdcdMZSdddNZTdddOZUy)eN)backend)check_conv_input_channels)#check_conv_transpose_input_channels)%compute_adaptive_pooling_window_sizes)#compute_conv_transpose_output_shape)cast)convert_to_tensor)maxsizec^tdtjjDS)zWhether the TF runtime sees only CPU devices. Cached because the device list does not change during a run and `tf.config.list_logical_devices()` is comparatively expensive to call on every conv invocation in eager mode. c3:K|]}|jdk(yw)CPUN) device_type.0ds t/var/www/html/emotional.easysim.app/public_html/venv/lib/python3.12/site-packages/keras/src/backend/tensorflow/nn.py z_cpu_only..sP!q}}%Ps)alltfconfiglist_logical_devicesr _cpu_onlyrs" Pryy/M/M/OP PPrc@tjj|SN)rnnreluxs rrr! 55::a=rc@tjj|Sr)rrrelu6rs rr#r#%s 55;;q>rcV|}tjj|}||_|Sr)rrsigmoid _keras_logits)r logitsoutputs rr%r%)s& F UU]]1 F!F Mrc t|}tj|dktjd|jtj|dk\tjd|jd|dzzS)Ndtype??)r rwhereconstantr-rs rsparse_sigmoidr30sa!A 88 R Cqww' aS8#Q-H rc@tjj|Sr)rrtanhrs rr5r59r!rcF|tjj|z Sr)rmathr5rs r tanh_shrinkr8=s rww||A rc@tjj|Sr)rr7softplusrs rr:r:As 77  A rc@tjj|Sr)rrsoftsignrs rr<r<Es 55>>! rc tj||kD||z tj|| k||ztj|Sr)rr1 zeros_liker thresholds r soft_shrinkrAIsD 88 I  I  iZY a0@A rc tj|dktj|tj|dkdtj|dzdz|S)Nr*r.g?)rr1r>powrs r sparse_plusrEQsN 88 R a Q"&&Q"22A6 rc@tjj|Sr)rrsilurs rrGrGYr!rct|}t||j}|tjtj||zz}|dz S)Nr,rC)r r-rsqrtsquare)r bys r squareplusrM]sE!A!177+A BGGBIIaL1$ %%A q5Lrc@tjj|Sr)rr7 log_sigmoidrs rrOrOds 77  q !!rcDtjj||S)N)alpha)rr leaky_relu)r negative_slopes rrRrRhs 55  A^  44rct|}t|tjd|jztjd|jz S)Ng@g@)r r#rr2r-rs r hard_sigmoidrUls?!A R[[agg.. /"++c1772K KKrc|t|zSr)rUrs r hard_silurWqs |A rctjj|}|dk(r|Stj|dkD|||zS)Nr.r)rrelur1)r rQress rrYrYus: %%))A,C z xxAsECK00rc@tjj|Sr)rrselurs rr\r\}r!rcZt|}tjj||S)N) approximate)r rrgelu)r r^s rr_r_s#!A 55::a[: 11rctj|d|tjjtj|d|z zzS)Nr+)rmaximumr7expm1minimum)r rQs rcelurdsC ::a   1cU")! rc|j|dzdk7rtd|jd|tj|d|\}}|tj|zS)NrCrz4axis size must be divisible by 2. Received: x.shape=z with axis=)num_or_size_splitsaxis)shape ValueErrorrsplitr%)r rgx1x2s rglurmshwwt}qA !!" TF <  XXaAD 9FB  2 rc2tj|ddS)Ngr/)clip_value_minclip_value_max)r clip_by_valuers r hard_tanhrrs  Ad3 GGrctjtj||kD|tj|Sr)rr1absr>r?s r hard_shrinkrus+ 88BFF1I )1bmmA.> ??rc6tj||kD||Sr)rr1)r r@ default_values rr@r@s 88A M1m 44rc"|}|btj|dg}tjj|d}tj|tj|}n!tjj||}||_|SNr*rg)rreshapersoftmaxrhr&)r rgr'r(s rr|r|sq F |At$vB/FBHHQK0qt,!F Mrc |atj|dg}tjj|d}tj|tj|Stjj||Sry)rr{r log_softmaxrh)r rgr(s rr~r~sg |At$""6"3zz&"((1+.. 55  QT  **rc|t|}tj|d|}tj||}tjdtj ||dz|j }dgt|j z}d||<tj||}||dz |z z dkD}tj||d}tjtj||j |d } tj||d dz | z } tj|| z d} | S) N DESCENDING) directionrgrzr.r,r*rr+Trgkeepdims) r rsortcumsumrangerhr-lenr{r1 reduce_sumrra) r rgr' logits_sorted logits_cumsumrr_shapesupportlogits_cumsum_safektaur(s r sparsemaxrs q !FGGFlFMIIm$7M BHHV$T*Q.fllCAcC %%GGDM 1gA}q0A559G'=#> bgggv||44$OA ==+$ F Ja OC ZZ c *F Mrct|jdz }|dk(rtj|d}|S|dk(rtj|d}|S|dk(rtj|d}|St d|jd) NrCr.rrCr.rrCr.rrrCrr.zePooling inputs's shape must be 3, 4 or 5, corresponding to 1D, 2D and 3D inputs. But received shape: .)rrhr transposeri)inputsnum_spatial_dimss r_transpose_spatial_inputsrs6<<(1,1fi0 M Q fl3 M Q fo6 M  228,,q B  rct|jdz }|dk(rtj|d}|S|dk(rtj|d}|S|dk(rtj|d}|S)NrCr.rrrr.rCrrrr.rCr)rrhrr)outputsrs r_transpose_spatial_outputsrsu7==)A-1,,w 2 N Q ,,w 5 N Q ,,w8 Nrc tj|}||n|}|j}tdt |j }|dk(r t |}tjj|||||}|dk(r t|}|SN channels_lastchannels_first) rstandardize_data_formatupper_convert_data_formatrrhrrrmax_poolrr pool_sizestridespadding data_formattf_data_formatrs rrr11+>K"?iGmmoG)/3v||;LMN&&+62eenn G&&,W5 Nrc tj|}||n|}|j}tdt |j }|dk(r t |}tjj|||||}|dk(r t|}|Sr) rrrrrrhrrravg_poolrrs r average_poolrrrc 2tjtjtjtj|tjtj|tjztj|tjz tj }tjtj jtjtjd|dztjtj|tjztj|tjz tj }tj||}tj||dz }||z }tj||}td||z dz}|} ||z} tj|| | } tj| tj S)z>Compute gather indices for Two-Pool Gather method (corrected).r.) rrfloorrfloat32int32r7ceilrcequalmaxr1) input_dim output_size small_window big_window window_starts window_ends window_sizes is_big_windowsmall_pool_len small_indices big_indicesgather_indicess r_compute_static_gather_indicesr"stGG  GGBHH[)2:: 6ggi, -ggk2::. /  M''  GGBHHQ a0"** =ggi, -ggk2::. /  K**[)4KJJ}i!m288 ,,rc.t|tr|f}|dk(rtj|d}|jj }|d}|d}| t dt||\}}t||||}tjj||fdddd } tjj||fdddd } tj| | gd } tj| |d } |dk(rtj| d} | S) Nrrr.r:Input length must be statically known for adaptive poolingAVGr.VALIDNWC window_shape pooling_typerrrrz isinstanceintrrrhas_listrirrrpoolconcatgather rrr static_shapel_staticout_lsmall_lbig_lgather_l small_pool_l big_pool_l combined_lpooled_ls r_adaptive_average_pool1drF'+s#"n &&fi0<<'')LAH NE H  ;8UKNGU-hwNH55::Z LX JL*5A>JyyXA6H&&<<)4 Orc.t|tr|f}|dk(rtj|d}|jj }|d}|d}| t dt||\}}t||||}tjj||fdddd } tjj||fdddd } tj| | gd } tj| |d } |dk(rtj| d} | S) Nrrr.rrMAXrrrrrzrrs r_adaptive_max_pool1drqrrczt|tr||f}|dk(rtj|d}|jj }|d}|d}|\}}|| t dt||\}} t||\} } t|||| } t||| | } tjj||dfdddd }tjj|| dfdddd }tj||gd }tj|| d }tjj|d| fdddd }tjj|d| fdddd }tj||gd }tj|| d }|dk(rtj|d }|S) Nrrr.rCFInput spatial dimensions must be statically known for adaptive poolingrr.r.rNHWCrrzrrrrrrh_staticw_staticout_hout_wsmall_hbig_hsmall_wbig_wgather_hgather_w small_pool_h big_pool_h combined_hpooled_h small_pool_w big_pool_w combined_wpooled_ws r_adaptive_average_pool2drs+s#"K0 &&fl3<<'')LAHAHLE58+ 4  ;8UKNGU:8UKNGU-hwNH-hwNH55::q\ LQZ JL*5A>JyyXA6H55::\ LZ JL*5A>JyyXA6H&&<<,7 Orczt|tr||f}|dk(rtj|d}|jj }|d}|d}|\}}|| t dt||\}} t||\} } t|||| } t||| | } tjj||dfdddd }tjj|| dfdddd }tj||gd }tj|| d }tjj|d| fdddd }tjj|d| fdddd }tj||gd }tj|| d }|dk(rtj|d }|S) z5Adaptive Max Pooling 2D using Two-Pool Gather method.rrr.rCrrrrrrrzrrrs r_adaptive_max_pool2drs+s#"K0 &&fl3<<'')LAHAHLE58+ 4  ;8UKNGU:8UKNGU-hwNH-hwNH55::q\ LQZ JL*5A>JyyXA6H55::\ LZ JL*5A>JyyXA6H&&<<,7 Orct|tr|||f}|dk(rtj|d}|jj }|d}|d}|d}|\}}} ||| t dt||\} } t||\} } t|| \}}t||| | }t||| | }t|| ||}tjj|| ddfddd d }tjj|| ddfddd d }tj||gd }tj||d }tjj|d| dfddd d }tjj|d| dfddd d }tj||gd }tj||d }tjj|dd|fddd d }tjj|dd|fddd d }tj||gd }tj||d }|dk(rtj|d }|S)Nrrr.rCrrrr.r.r.rNDHWCrrzrrrrrrd_staticrrout_drrsmall_dbig_drrrrgather_drr small_pool_d big_pool_d combined_dpooled_drrrrrrrrs r_adaptive_average_pool3dr)s+s#"K= &&fo6<<'')LAHAHAH%E5%8+x/? 4  ;8UKNGU:8UKNGU:8UKNGU-hwNH-hwNH-hwNH55::q!_ LQ] JL*5A>JyyXA6H55::!_ L] JL*5A>JyyXA6H55::G_ LE] JL*5A>JyyXA6H&&<</: Orct|tr|||f}|dk(rtj|d}|jj }|d}|d}|d}|\}}} ||| t dt||\} } t||\} } t|| \}}t||| | }t||| | }t|| ||}tjj|| ddfddd d }tjj|| ddfddd d }tj||gd }tj||d }tjj|d| dfddd d }tjj|d| dfddd d }tj||gd }tj||d }tjj|dd|fddd d }tjj|dd|fddd d }tj||gd }tj||d }|dk(rtj|d }|S)z5Adaptive Max Pooling 3D using Two-Pool Gather method.rrr.rCrrrrrrrrzrrrs r_adaptive_max_pool3drs+s#"K= &&fo6<<'')LAHAHAH%E5%8+x/? 4  ;8UKNGU:8UKNGU:8UKNGU-hwNH-hwNH-hwNH55::q!_ LQ] JL*5A>JyyXA6H55::!_ L] JL*5A>JyyXA6H55::G_ LE] JL*5A>JyyXA6H&&<</: Orctj|}t|jdz }|dk(r t |||S|dk(r t |||S|dk(r t |||Std)NrCr.rz9adaptive_average_pool supports 1D, 2D, or 3D inputs only.)rrrrhrrrrirrrndimss radaptive_average_poolrsw11+>K   !E z' [II !' [II !' [II G  rctj|}t|jdz }|dk(r t |||S|dk(r t |||S|dk(r t |||Std)NrCr.rz5adaptive_max_pool supports 1D, 2D, or 3D inputs only.)rrrrhrrrrirs radaptive_max_poolrsw11+>K   !E z#FKEE !#FKEE !#FKEE C  rc|dk(r!|dk(ry|dk(ry|dk(rytd|d |d k(r!|dk(ry |dk(ry |dk(ry td|d td|d)NrrrrrrzInput rank not supported: z. Expected values are [3, 4, 5]rNCWNCHWNCDHWzInvalid data_format: z9. Expected values are ["channels_first", "channels_last"])ri)rndims rrrso% 19 QY QY,TF300  ( ( 19 QY QY,TF300  #K=1F F  rc\ fd tjd fd}dk(xrtjdk(}t j dk(rjd}njd }|xs|jd k7}|r|S S) Ncttj}tjj j |}|j}|jretj|jdk(r?tjjjdk7rtd|d|S)Nr dilationsrzEThe convolution operation resulted in an empty output. Output shape: z. This can happen if the input is too small for the given kernel size, strides, dilation rate, and padding mode. Please check the input shape and convolution parameters.) rrrhrr convolutionris_fully_definedr7prodrri) rresult result_shaper dilation_raterkernelrrs r_convzconv.._conv*s-k3v||;LM""    MMO&# # ||  ) ) + ,..01Q6 &,,..01Q6 >"  rT) jit_compilecSrr)r)sr _conv_xlazconv.._conv_xlaFs wrrrrr*r.)rfunctionrrhrr) rr(rrrr'r, needs_xlachannelsr)s `````` @rconvr1"s8[[T"#//JC 4E4JI11+>Ko%<<#<<?9Xb)99I{wrcT|ytd|Dxrtd|DS)NFc3&K|] }|dkD ywr.Nr)rss rrzA_needs_depthwise_stride_dilation_decomposition.._s&q1u&c3&K|] }|dkD ywr4rrs rrzA_needs_depthwise_stride_dilation_decomposition.._s.Lq1u.Lr6)any)rr's r._needs_depthwise_stride_dilation_decompositionr9Ys. &g& & L3.Lm.L+LLrcttj|tjdd}tjddgtjg}t |D]|\}}||}||} ||} |dz | zdz} || zdz | z} tj | dz | z| z|z d} | dz}| |z }|j tj||g~|j tjddgtjtj|tj|S)N)out_typer.r*rr,rC) rrhrr2 enumerateraappendstackpad)r kernel_sizerr' spatial_shapepaddingsirnr5reffective_kernel_sizeout_size total_pad pad_before pad_afters r_pad_same_spatial_channels_lastrJbs$ HHVbhh7"=M QF"((34H+& ;1 !  AJ ! !"Q! aEAI!#JJ \Q !6 6 :A !^  * *i!89: ; OOBKKAbhh78 66&"((8, --rct|jdz }|j}|dk(r t|}|dk(rVt j |d}t j |d}|t j |d}d|df}d|df} n|}|} t d|jddD} |dk(rt|| || }n|dk7rtd |tjj||d dd | } | dddd|ddd|dddf} |$tjj| |d dd } |dk(rt j| dg} |dk(r t| } | S)NrCrr.rzrc32K|]}t|ywr)r)rrCs rrz@_depthwise_conv_stride_dilation_decomposition..sC1ACsSAMErz7`padding` must be 'valid' or 'same'. Received: padding=)r.r.r.r.r)rrrr!)rrr)rrhrrr expand_dimstuplerJrirdepthwise_conv2dconv2dsqueezer) rdepthwise_kernelrrrr'pointwise_kernelrspatial_stridesspatial_dilation_rater@rs r-_depthwise_conv_stride_dilation_decompositionrW{s6<<(1,mmoG&&*621Q/>>*:C  '!~~.>QG gaj/!"M!$4 5! -C(8(>(>r(BCCK&0 K2G  G EgY O  ee$$' %Ga.OA..0E?13E0EqHIG#%%,,     1**Wqc*&&,W5 NrcVtj|}t|}t|}t|jdz }|dkDrt d|j dt|||t|d}|j}t|tr|f|z}t|tr|f|z}t||rt||||||S|dk(r|dk(xr t}|r t|}|s|dk(rd|dzzdz}d} n d |dzz}d} t!j"|| }t!j"|d }|dnd|z}|s|dk(r tdd} n tdd} t j$j'||||| | } t!j(| | g} |r t+| } | S|dk(xr t}|r t|}|s|dk(rd|zdz}tdd} nd |z}|} t j$j'||||| | } |r t+| } | S) NrCz<`inputs` rank must be 3 (1D conv) or 4 (2D conv). Received: rrr.rrrrrrzr )rrr rrhrirrrrrrr9rWrrrrNrrPrRr) rr(rrrr'rrneed_transposespatial_start_dimconv_data_formatrs rdepthwise_convr\s11+>K v &F v &F6<<(1,! J{{m1   ffk:*+q9NmmoG'3*//-%&(+;; 5g}M< FGWk=  1 %(88HY[ .v6F [O;Wq[(4/G ! w{*G ! (9:Q/ - 54-;O [O;3OQG 34DaH %%((    (# ) **W'8&9: 09G !$44DN*627.4'/C7")ee$$$ %G,W5 Nrc ~tj|}t|}t|}t|}t|jdz }|dkDrt d|dt |||t|d}|j}t|tr|f|z}t|tr|f|z}t||rt|||||||S|dk(r|dk(xr t} | r t|}| s|dk(rd |dzzd z}d} tdd} nd |dzz}d} tdd} tj || }tj |d }tj |d }|dnd |z}tj"j%|||||| | } tj&| | g} | r t)| } | S|dk(xr t} | r t|}| s|dk(rd |zd z}tdd} nd |z}|} tj"j%|||||| | } | r t)| } | S)NrCz>`num_spatial_dims` must be 1 or 2. Received: num_spatial_dims=rr)rTr.rrrrrrzr )rrr rrhrirrrrrr9rWrrrrNrseparable_conv2drRr) rrSrTrrrr'rrrYrZr[rs rseparable_convr_s11+>K v &F()9:()9:6<<(1,! 01 4  f&6 D*+q9NmmoG'3*//-%&(+;; 5g}M<      -  1%(88HY[ .v6F [O;Wq[(4/G ! 3OQG w{*G ! 34DaH (9:>>*:C>>*:C - 54-;O %%((     (#) **W'8&9: 09G !$44DN*627.4'/C7")ee$$$%G,W5 Nrc tj|}t|}t|}t|||t |t |j }|j dd}|j d} t|j } tj |} t| D]\} } |  | | | | <t| || |||||}tjj|||||j||S)Nr-)rrr!)rrr rrrrhlistrr<rrconv_transposer)rr(rroutput_paddingrr'rr@filters input_shapesymbolic_shaperCe output_shapes rrbrb}s 11+>K v &F v &F' D)+s6<<7HIN,,s#Kll2Gv||$KXXf%N+&/1 9+A.KN/7 L 55   "  rct|d}|d}ntj|}|r|dkr|t|jzdz}t j |j}tj||f}tjtj|d|}|jDcgc]}tj|}}tj|ddi}|j|tj|d|D cgc]} tj| |df}} |D cgc]&} tj| tj(}} tj |d}t#|j} | j||tj$||| S|d k(rd nd \} } tj&||| | || Scc}wcc} wcc} w) Nint64r,rrr.indexingijrzbool)TF)NN)on_value off_valuergr-)r rstandardize_dtyperrhr7r$rr{r greater_equalrmeshgridinsertrarjrra SparseTensorone_hot) r num_classesrgr-sparse values_countvaluesdimindicesarhrnros rrurus!7+A }))%0  !8#agg,&*Dyy) A /))&!4EB,-GG4S288C=44++w66tRZZ1-.=DE2::a,!23EE189A2771bhh'99))G!,QWW  T;'w66+0F?- Hi ::    5F9s:G%G*<+G/c<t|jdkDrdnd}tj|dk(r|rgt |||dd}t j j||d}|j}t j||}|j||St ||||}t j|| S|r2t ||||d}t j j||dSt ||||}t j|| S) Nr.rrmint8T)rgr-rw)rgoutput_is_sparse)rgr-rz) rrhrrprurrw reduce_maxr set_shape reduce_any)r rvrgr-rwreduction_axisr outputs_shapes r multi_hotrsagg,*QN  '61 ;TGii**nt+G$MMMgggu-G   m ,Na4uEG==~> > ;TtG99''nt( a4uEG==~> >rc|}|}t|d}|r|j}d}t|dxrPt|tjj tj f xr|jj|k(xr| }|rLt|jjdk7rtd|d|jjd}d}|r"|s|rtjd|d |d d ||fS) zCRetrieves logits tensor from maybe-softmax or maybe-sigmoid tensor.r&Topr.zExpected 1 input for rrz"`zK` received `from_logits=True`, but the `output` argument was produced by a zB activation and thus does not represent logits. Was this intended?rC) stacklevel)hasattrr&rr __internal__ EagerTensorVariablertyperrriwarningswarn)r( from_logitsop_typefn_nameoutput_ from_logits_has_keras_logitsfrom_expected_op_types r _get_logitsrsGLv7&&   &6BOO$?$?#MN N & IINNg %    vyy A %4WIQ?@ @))""1% (,A  77>i@! !   L  rctj|}tj|}t|jdkr%t d|jd|jt|jt|jk7r%t d|jd|jt |j|jD]5\}}| | ||k7st d|jd|jt ||dd\}}|r"tjj|||S|tj||d z }tj|tjd tjz }tj|tjj|z| S) a Categorical crossentropy between an output tensor and a target tensor. Args: target: A tensor of the same shape as `output`. output: A tensor resulting from a softmax (unless `from_logits` is `True`, in which case `output` is expected to be the logits). from_logits: Boolean, whether `output` is the result of a softmax, or is a tensor of logits. axis: Int specifying the channels axis. `axis=-1` corresponds to data format `channels_last`, and `axis=1` corresponds to data format `channels_first`. Returns: Output tensor. Example: >>> a = tf.constant([1., 0., 0., 0., 1., 0., 0., 0., 1.], shape=[3,3]) >>> print(a) tf.Tensor( [[1. 0. 0.] [0. 1. 0.] [0. 0. 1.]], shape=(3, 3), dtype=float32) >>> b = tf.constant([.9, .05, .05, .05, .89, .06, .05, .01, .94], ... shape=[3, 3]) >>> print(b) tf.Tensor( [[0.9 0.05 0.05] [0.05 0.89 0.06] [0.05 0.01 0.94]], shape=(3, 3), dtype=float32) >>> loss = categorical_crossentropy(a, b) >>> print(np.around(loss, 5)) [0.10536 0.11653 0.06188] >>> loss = categorical_crossentropy(a, a) >>> print(np.around(loss, 5)) [0. 0. 0.] r.zPArguments `target` and `output` must be at least rank 1. Received: target.shape=, output.shape=WArguments `target` and `output` must have the same rank (ndim). Received: target.shape=QArguments `target` and `output` must have the same shape. Received: target.shape=Softmaxcategorical_crossentropy)labelsr'rgTrr/)rr rrhriziprr!softmax_cross_entropy_with_logitsrrqrepsilonr7log)targetr(rrge1e2s rrrsN ! !& )F  ! !& )F 6<<1 "LL> H   6<<C -- "LL> H  fllFLL1B >bnr & ~_V\\NL & Y(BFKuu66&t7  bmmFD4@ @F  !3):#:F MM&277;;v#66 = ==rcj|dk7r)|t|jdz k7rtd|t||dd\}}t j |}t j |d}t j |}t|jt|jk(r)|jddk(rt j|d}t|jdkrtd |jt|jt|jd dk7r%td |jd |jt|j|jd dD]5\}}| | ||k7std |jd |j|s]t j|tjdtjz }tjj|}tjj||}|S)aNCategorical crossentropy with integer targets. Args: target: An integer tensor. output: A tensor resulting from a softmax (unless `from_logits` is True, in which case `output` is expected to be the logits). from_logits: Boolean, whether `output` is the result of a softmax, or is a tensor of logits. axis: Int specifying the channels axis. `axis=-1` corresponds to data format `channels_last`, and `axis=1` corresponds to data format `channels_first`. Returns: Output tensor. r*r.z4Only axis=-1 is currently supported. Received: axis=rsparse_categorical_crossentropyrjr,rzzBArgument `output` must be at least rank 1. Received: output.shape=NzRArgument `output` must have rank (ndim) `target.ndim - 1`. Received: target.shape=rzcArguments `target` and `output` must have the same shape up until the last dimension: target.shape=rr')rrhrirrr rrRrrqrrr7rr(sparse_softmax_cross_entropy_with_logits)rr(rrgrrr%s rrres " rzdc&,,/!33B4& I  & Y(IFK ! !& )F WWV7 +F  ! !& )F 6<<C --&,,r2Ba2GF, 6<<1 "LL> +   6<<C Sb 122 "LL> H  fllFLL"$56B >bnr & ~_V\\NL  !! GOO%q7??+<'< V$ UU ; ;f<F Mrc*tj|}tj|}t|jt|jk7r%t d|jd|jt |j|jD]5\}}| | ||k7st d|jd|jt ||dd\}}|r!tjj||Stj|tjdtjz }|tjj|z}|d|z tjjd|z zz }| S) apBinary crossentropy between an output tensor and a target tensor. Args: target: A tensor with the same shape as `output`. output: A tensor. from_logits: Whether `output` is expected to be a logits tensor. By default, we consider that `output` encodes a probability distribution. Returns: A tensor. rrrSigmoidbinary_crossentropyrr/r.)rr rrhrirrr!sigmoid_cross_entropy_with_logitsrqrrr7r)rr(rrrbces rrrsx ! !& )F  ! !& )F 6<<C -- "LL> H  fllFLL1B >bnr & ~_V\\NL & Y(=FKuu66&7    !3):#:F 277;;v& &CAJ"''++a&j1 11C 4Krcd}tj|j}|dvrd}t|d}|rt |||\}}nt |||\}}|rt j|t jjt jj}t j|t jjt jj}t||}t||}||fS)NF)float16bfloat16Tr) rrpr-r_compute_moments_sync_compute_momentsrrqrminr)r axesr synchronized need_cast ori_dtypemeanvariances rmomentsrsI))!''2I++ I .q$Ah)!T8<hbjjnnbjjnnE##HbjjnnbjjnnMD)$), >rctjj}|s t|||Stj|d}tj ||d}tj tj ||d}tj ||d}|jtjjj|}|jtjjj|}|jtjjj|} tjj|| } tjj|| } tj| tj | z d} |s,tj| |} tj| |} | | fS)Ncount)nameTrr+)r distributeget_replica_contextr ones_likerrJ all_reduceReduceOpSUMr7 divide_no_nanrarR) r rr replica_ctx local_count local_sumlocal_squared_sumy_sum y_squared_sum count_sumry_squared_meanrs rrrsY--335K 422,,qw/K adT:I biilM-- $FK  " "2==#9#9#=#=y IE** ""$5M&&r}}'='='A'A;OI 77  2DWW**=)DNzz.299T?:C@H zz$%::h- >rcFtjj|||S)Nr)rrr)r rrs rrr s 55==D8= 44rcd|dk7rdgt|jz}|jd||<tj||}tj||}|tj||}|tj||}tjj ||||||S)Nr*r.r)r rroffsetscalevariance_epsilon)rrhrr{rbatch_normalization)r rrrgrrrrhs rrrs rzc!''l"jjmd zz$&::h.  ZZ.F  JJue,E 55 $ $  % rcPt|}t|}tj|d}tj|j d}|dk(rdn|}tj||}tj j|||||d}tj||S)Nrr,rfloat64F)rr' label_length logit_length blank_indexlogits_time_major)r rrr result_typer-rctc_loss)rr( target_length output_length mask_index result_dtype compute_dtypelosss rrr&s v &F v &F WWV7 +F &&v||Y?L!-!:I M WWV] +F 55>>""  D 774 &&rct|}tj|}|d|d} }tj|d}t j |j d} tj|| }t|d}|dk(r'tjj||||\} } nx|d k(rd|;|d d|f} |d ||dzf}|d |dzdf}tj| ||gd }tjj|||| \} } ntd|dg}| D]_}tj|j|j|| f}|j!tj"j%|d atj&|d }tj|d}|d k(r,|*|dkr||d z}tj(||k\|dz|}|| fS)Nrr.)r.rrCrrr,greedy)rsequence_lengthmerge_repeatedr beam_search.r*rz)rr beam_width top_pathszInvalid strategy z2. Supported values are 'greedy' and 'beam_search'.)sp_inputrw)r rrhrrrr-rrctc_greedy_decoderrctc_beam_search_decoderrirtr{ryr=rwto_denser>r1)rsequence_lengthsstrategyrrrrre num_samples num_stepsr-decodedscores inputs_before inputs_mask inputs_after decoded_densests r ctc_decoder=s'v &F((6"K(^[^K \\&) ,F    i 8E WWVU #F()9I8EE44,)" 5 & ] "  !"3  #34M j:>&A!ABK!#zA~'7"78LYY k:FEE99,! : &z** *  MP __RZZ[)4L MRYY//2/NOPHH]3MGGM73M= Z%; >#k"o5J Z '):M  &  rcBddlm}|j|jk7r&td|jd|jdt ||j }t jt j||z }d||zd||zz }|S) Nr)log10z Input shapes z and z" must match for PSNR calculation. r, ) "keras.src.backend.tensorflow.numpyrrhrir r-r reduce_meanrJ)rkrlmax_valrmsepsnrs rrr~s8 xx288BHH:U288*5+ +   rxx8G ..27+ ,C g eCj 0D Krcrtj|}|dk(rdnd}tj|dz|S)Nrg@g̓$Ggffffffr,)rrprr2)r-vals r_get_large_negativers5  % %e ,Ei''ZC ;;sTz //rc||s|Stj|d}|tj||}|rytj|}|d|d}}tjj tj ||fddd}|ddddddf}tj||}tj||t|j}|S)Nrmr,rCrr*r) rr logical_andrhlinalg band_partonesr1rr-)r'mask is_causal combined_mask logits_shapeTS padded_logitss r _apply_masksrs |I LLv6M }d; xx' A Q1yy""277Aq66#:BBD$1$%}d; HHv26<<@M rctj|jd}tjd||d}tj ||}tj |tj ||j}|4tj|tj ||j}t|||} tj| jd} tj tjjtj | | d|j} tjd| |dS)NrzBTNH,BSNH->BNTSoptimal)optimizer*rzzBNTS,BSNH->BTNH) rrr-reinsumrmultiplyaddrrr|) querykeyvaluebiasr r r logits_dtyper'r probs_dtypeprobss r_dot_product_attention_xlar s&&u{{I>L YY(%y IF WWV\ *F [[ != >F fll ;< y9M%%m&9&99EK GG  bggm[9 CSYY E 99&uy IIrc @|d}|r tdt|}t|}t|}t|jdk7r3td|jd|jd|jdt j |j |j |j } t|| }t|| }t|| }|jd} |jd} | A| ?| | kDr:| d kDr5| | z} tj|| d }tj|| d }| t|| }tj|d } |,d tjtj| dz n|}t|||||||S)NFz7Flash attention is not supported in tensorflow backend.rzG`dot_product_attention` only supports 4D inputs. Received: query.shape=z , key.shape=z, value.shape=rrCr.)repeatsrgr,r*r/r) rir rrhrrr-rrrepeatrIr )rrrrr rr flash_attentionattn_logits_soft_caprnum_query_heads num_kv_headsgroupsHs rdot_product_attentionr*s E   e $E C C e $E 5;;1 %%*[[Mcii[I ;;-q *  '' SYY LM  &E sM "C  &Ekk!nO99Q. s Q1q5 r6rr.rr)imagessizesrratesr)rrrrr.rC) rrrhr8rr?rimageextract_patchesr{)inputr@dilationrstriderrpr5NCr)Wr patchesnHnWDs runfoldr>s k3 ' k"  !+8S 98xA(#6'GA&vs3AJAq!Q ! u1v1v!ad|adAaD\JK UL)Ahh&&!A$!a AaD!A$"!A$!a  'G==LAr2qjj!RQqT1Q4+Gll#Gjj1a!A$h1orBw"?@G Nrc t|tr||fn|}t|tr||fn|}t|tr||fn|t|tr||fn|}t|tr||fn|tjd} jd} |\|\} } | zz| d|dzzddz zz dz dzdz| d|dzzddz zz dz dzdztj| g| d|dzz| d|dzzf d} tj | }|ddkDs|ddkDr#|dddd|d|dz |d|dz f}|S)aTensorFlow implementation of Fold (col2im). Combine an array of sliding local blocks into a large tensor. Args: x: 3-D tensor, shape (N, C*kH*kW, L) **required**. output_size: int or (oH, oW) kernel_size: int or (kH, kW) dilation: int or (dH, dW), default 1 padding: int or (pH, pW), default 0 stride: int or (sH, sW), default 1 Returns: 4-D tensor, shape (N, C, oH, oW) rr.rCc  tjgj}tD]<}tD]*}|dz}|dz}tjdz|z}tjdz|z}|dd||ddddf}tjtjz} tj tj|g} tj tj |gg} tj | | | gd} tj|dg} tj|| | }-?|S)Nr,rr.rzr*) rzerosr-rr#tiler>r{tensor_scatter_nd_add)x_singler(rCjh_startw_start h_indices w_indicespatchc_idxh_idxw_idxr{ryr8rkHkWr;r<oH_padoW_padr5r s r _fold_singlezfold.._fold_singleSsM1ff-QWW=r KA2Y Kad(ad(HHRL1Q4/'9 HHRL1Q4/'9  Aq!Q/ "((1+rBw7 )R 81#> B4 81#>((E5)EB4011&'6J K K" rN)rrrrhr{vectorized_map)r rr@r4rr5ror6r7CKKoHoWrRr(r8rrNrOr;r<rPrQr5s` @@@@@@@@@rfoldrX#s" k3 ' k"  k3 ' k"  !+8S 98xA(#6'GA&vs3A  AA ''!*C FB FB RA q1Q4x-!A$"q&/ )A -!A$ 6 :B q1Q4x-!A$"q&/ )A -!A$ 6 :B 1q!RR,-A!ad(]F !ad(]F,  |Q /F tax1Q4!81adVad]2AaD6AaD=4HHI Mrc|dk(r"tjj||dStj|gd}tjj||d}tj|gd}|S)a"TensorFlow implementation of depth_to_space. Rearranges data from depth into blocks of spatial data. Args: x: 4-D tensor with shape (N, H, W, C) for channels_last or (N, C, H, W) for channels_first. block_size: An integer specifying the block size. data_format: "channels_last" or "channels_first". Returns: A tensor with shape (N, H*block_size, W*block_size, C/block_size**2) for channels_last or (N, C/block_size**2, H*block_size, W*block_size) for channels_first. rrrrr)rrdepth_to_spacerr block_sizers rr[r[rj o%uu##Azv#FF LLL ) EE JF C LLL )rc|dk(r"tjj||dStj|gd}tjj||d}tj|gd}|S)aTensorFlow implementation of space_to_depth. Rearranges blocks of spatial data into depth. Args: x: 4-D tensor with shape (N, H, W, C) for channels_last or (N, C, H, W) for channels_first. block_size: An integer specifying the block size. data_format: "channels_last" or "channels_first". Returns: A tensor with shape (N, H/block_size, W/block_size, C*block_size**2) for channels_last or (N, C*block_size**2, H/block_size, W/block_size) for channels_first. rrrZrr)rrspace_to_depthrr\s rr`r`r^r)r0)r)g?)r/)T)r*)NvalidN)rr)r.raNr.)r.raNNr.)r*NF)Fr*)F)FF)NNgMbP?)r)rdr.Tr)NNNFNN)r.rr.)r)V functoolsr7r tensorflowr keras.srcr&keras.src.backend.common.backend_utilsrrrr!keras.src.backend.tensorflow.corerr lru_cacherrr#r%r3r5r8r:r<rArErGrMrOrRrUrWrYr\r_rdrmrrrur@r|r~rrrrrrrrrrrrrrrr1r9rJrWr\r_rbrurrrrrrrrrrrrrrr r*r>rXr[r`rrrrisR L3?T"Q#Q"5L 12  H@5 +$$    >   8!-H(V(VCLDNZz[|     F   4nM.@AN   Y@  _J  (V!H?D!!HP>f<~-`.@5 ?C.'4>!B 0 *J.  6r,^L^8r