o 2h$@spddlZddlZddlZddlmZddlmZddlm Z ddlm Z ddZ dd Z d d Z d d ZddZddZddZddZd|ddZddZddZd}ddZd d!Zd~d#d$Zd%d&Zd'd(Zdd*d+Zd,d-Zdd/d0Zdd1d2Zdd4d5Zd6d7Z d|d8d9Z!d:d;Z"ddd?Z$dd@dAZ%dBdCZ&dDdEZ'  F ddGdHZ(  F ddIdJZ)dKdLZ* M F  MddNdOZ+ M F  MddPdQZ, M F  MddRdSZ- M F   MddTdUZ.ddWdXZ/ddYdZZ0d[d\Z1dd]d^Z2dd_d`Z3ddadbZ4ddcddZ5dedfZ6dgdhZ7 iddjdkZ8ddldmZ9 n o M . ddpdqZ:drdsZ;dtduZ    V  ddzd{Z?dS)N)backend)#compute_conv_transpose_output_shape)cast)convert_to_tensorcC tj|SN)tfnnreluxr Z/var/www/html/chatgem/venv/lib/python3.10/site-packages/keras/src/backend/tensorflow/nn.pyr  r cCrr)rr relu6r r r rrrrcCs|}tj|}||_|Sr)rr sigmoid _keras_logits)r logitsoutputr r rrs rc CsHt|}t|dktjd|jdt|dktjd|jdd|dS)Ndtype??)rrwhereconstantrr r r rsparse_sigmoids $rcCrr)rr tanhr r r rr&rrcCs|tj|Sr)rmathrr r r r tanh_shrink*r!cCrr)rr softplusr r r rr#.rr#cCrr)rr softsignr r r rr$2rr$rc Cs0t||k||t|| k||t|Sr)rr zeros_liker thresholdr r r soft_shrink6s r(c Cs6t|dkt|t|dkdt|dd|S)Nrrg?)rrr%powr r r r sparse_plus>s  r+cCrr)rr silur r r rr,Frr,cCs6t|}t||jd}|tt||}|dS)Nrr))rrrsqrtsquare)r byr r r squareplusJsr2cCrr)rr log_sigmoidr r r rr3Qrr3皙?cCstjj||dS)N)alpha)rr leaky_relu)r negative_sloper r rr6Ur"r6cCs,t|}t|td|jtd|jS)Ng@g@)rrrrrr r r r hard_sigmoidYs$r8cCs |t|Sr)r8r r r r hard_silu^rr9rcCs.tj|}|dkr |St|dk|||S)Nrr)rr elur)r r5resr r rr:bs r:cCrr)rr selur r r rr<jrr<TcCst|}tjj||dS)N) approximate)rrr gelu)r r=r r rr>nsr>cCs(t|d|tjt|d|S)Nr)rmaximumr expm1minimum)r r5r r rcelussrBrcCsJ|j|ddkrtd|jd|tj|d|d\}}|t|S)Nr)rz4axis size must be divisible by 2. Received: x.shape=z with axis=)num_or_size_splitsaxis)shape ValueErrorrsplitr)r rDx1x2r r rgluysrJcCstj|dddS)Ngr)clip_value_minclip_value_max)r clip_by_valuer r r r hard_tanhr"rNcCstt||k|t|Sr)rrabsr%r&r r r hard_shrinksrPcCst||k||Sr)rr)r r' default_valuer r rr'r'cCsX|}|durt|dg}tjj|dd}t|t|}ntjj||d}||_|SNrrD)rreshaper softmaxrEr)r rDrrr r rrVsrVcCsH|durt|dg}tjj|dd}t|t|Stjj||dSrS)rrUr log_softmaxrE)r rDrr r rrWs rWc Cst|}tj|d|d}tj||d}tjdt||d|jd}dgt|j}d||<t||}||d|dk}t ||d}tj t ||j|d d } tj ||d d d| } t || d} | S) N DESCENDING) directionrDrTrrrrrTrDkeepdims) rrsortcumsumrangerErlenrUr reduce_sumrr?) r rDr logits_sorted logits_cumsumrr_shapesupportlogits_cumsum_safektaurr r r sparsemaxs  ricCsht|jd}|dkrt|d}|S|dkrt|d}|S|dkr+t|d}|Std|jd) Nr)rrr)r)rr)rrk)rr)rkr-rzePooling inputs's shape must be 3, 4 or 5, corresponding to 1D, 2D and 3D inputs. But received shape: .)r_rEr transposerF)inputsnum_spatial_dimsr r r_transpose_spatial_inputss    rpcCsVt|jd}|dkrt|d}|S|dkrt|d}|S|dkr)t|d}|S)Nr)rrj)rrkrr)rk)rr-rr)rk)r_rErrm)outputsror r r_transpose_spatial_outputss   rrvalidcCjt|}|dur |n|}|}tdt|j}|dkr!t|}tj |||||}|dkr3t |}|SN channels_lastchannels_first) rstandardize_data_formatupper_convert_data_formatr_rErprr max_poolrrrn pool_sizestridespadding data_formattf_data_formatrqr r rr{ r{cCrtru) rrxryrzr_rErprr avg_poolrrr|r r r average_poolrrcCs|dkr|dkr dS|dkrdS|dkrdStd|d |d kr<|dkr(d S|dkr.d S|dkr4d Std|d td|d)NrvrkNWCr-NHWCNDHWCzInput rank not supported: z. Expected values are [3, 4, 5]rwNCWNCHWNCDHWzInvalid data_format: z9. Expected values are ["channels_first", "channels_last"])rF)rndimr r rrzs.   rzrc sfddtjddfdd}dko!tjdk}td kr1jd }njd }|p>|jd k}|rD|SS) Ncs,ttj}tjj|dS)Nr dilations)rzr_rErr convolutionry)r)r dilation_raternkernelrr~r r_conv7szconv.._convT) jit_compilecsSrr r )rr r _conv_xlaDszconv.._conv_xlarwrrvrr)rfunctionr_rErrx) rnrr~rrrr needs_xlachannelsr )rrrrnrrr~rconv/s    rc Cs4t|}t|jd}|dkrtd|jdt|d}|}t|t r,|f|}t|t r6|f|}|dkr{|dkrId|dd}d}nd|d}d}t ||}t j |d d }|durddnd|}t j j ||||||d } t | |gS|dkrd|d}d}nd|}d}t j j ||||||d S) Nr)z<`inputs` rank must be 3 (1D conv) or 4 (2D conv). Received: rlr-rrvrrrrrTr)rrxr_rErFrrzry isinstanceintr expand_dimsr depthwise_conv2dsqueeze) rnrr~rrrrorspatial_start_dimrqr r rdepthwise_convWsZ         rc Cs<t|}t|jd}|dkrtd|dt|d}|}t|tr+|f|}t|tr5|f|}|dkr|dkrHd|dd}d} nd|d}d} t || }t j |d d }t j |d d }|durjdnd|}t j j |||||||d } t | | gS|dkrd|d}nd|}t j j |||||||d S) Nr)z>`num_spatial_dims` must be 1 or 2. Received: num_spatial_dims=rlr-rrvrrrrTr)rrxr_rErFrzryrrrrr separable_conv2dr) rndepthwise_kernelpointwise_kernelr~rrrrorrrqr r rseparable_convs\         rc Cst|}t|t|j}|jdd}|jd} t|j} t|} t| D]\} } | dur5| | | | <q't| || |||||}tj j ||||| ||dS)Nr)rrr) rrxrzr_rElistr enumeraterr conv_transposery)rnrr~routput_paddingrrr kernel_sizefilters input_shapesymbolic_shapeie output_shaper r rrs:      rFc s"t|dd}|dur d}nt|}|r||dkr!|t|jd}t|jt|f}tj t |d|d}dd|jD}tj |dd i}| |t |dfd d|D}d d|D}tj|dd }t|j}| ||t|||S|d krdnd\}} tj|||| ||dS)Nint64rfloat32rrcSsg|]}t|qSr )rr^).0dimr r r szone_hot..indexingijcsg|] }t|dfqSr)rrUra values_countr rrscSsg|] }t|tjqSr )rrrrr r rrsrTbool)TF)NN)on_value off_valuerDr)rrstandardize_dtyper_rEr prodrrUr greater_equalmeshgridinsertr?concatr SparseTensorone_hot) r num_classesrDrsparsevaluesindicesrErrr rrrs8     rcCst|jdkr dnd}t|dkrE|r6t|||ddd}tjj||dd}|j}t||}| ||St||||d}tj ||d S|rYt||||dd}tjj||ddSt||||d}tj||d S) Nrrrint8T)rDrr)rDoutput_is_sparse)rDrrT) r_rErrrrr reduce_maxr set_shape reduce_any)r rrDrrreduction_axisrq outputs_shaper r r multi_hot s0    rcCs|}|}t|d}|r|j}d}t|do%t|tjjtjf o%|jj|ko(| }|r=t |jj dks5J|jj d}d}|rQ|sC|rQt j d|d|dd d ||fS) zCRetrieves logits tensor from maybe-softmax or maybe-sigmoid tensor.rToprrz"`zK` received `from_logits=True`, but the `output` argument was produced by a zB activation and thus does not represent logits. Was this intended?r)) stacklevel) hasattrrrr __internal__ EagerTensorVariablertyper_rnwarningswarn)r from_logitsop_typefn_nameoutput_ from_logits_has_keras_logitsfrom_expected_op_typer r r _get_logitsBs2     rcCst|}t|}t|jdkrtd|jd|jt|jt|jkr3td|jd|jt|j|jD]\}}|durV|durV||krVtd|jd|jq:t||dd\}}|rktjj|||d S|tj ||d d }t |t d t }t |tj || 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.] rzPArguments `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=NQArguments `target` and `output` must have the same shape. Received: target.shape=Softmaxcategorical_crossentropy)labelsrrDTr[r)rrr_rErFziprr !softmax_cross_entropy_with_logitsr`rMrepsilonr log)targetrrrDe1e2r r rresP ' rcCsl|dkr|t|jdkrtd|t||dd\}}t|}tj|dd}t|}t|jt|jkrF|jddkrFtj|dd}t|jdkrUtd |jt|jt|jd dkrotd |jd |jt|j|jd dD]\}}|d ur|d ur||krtd |jd |jqz|st |t dt }tj |}tjj||d}|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. rrz4Only axis=-1 is currently supported. Received: axis=rsparse_categorical_crossentropyrrrTzBArgument `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)r_rErFrrrrrrrMrrr rr (sparse_softmax_cross_entropy_with_logits)rrrrDrrresultr r rrsZ  " rcCst|}t|}t|jt|jkr td|jd|jt|j|jD]\}}|durC|durC||krCtd|jd|jq't||dd\}}|rWtjj||dSt |t dt }|tj |}|d |tj d |7}| 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. rrNrSigmoidbinary_crossentropyrrr)rrr_rErFrrr !sigmoid_cross_entropy_with_logitsrMrrr r)rrrrrbcer r rrsB rcCsd}t|j}|dvrd}t|d}|rt|||\}}nt|||\}}|rHt|tjj tjj }t|tjj tjj }t||}t||}||fS)NF)float16bfloat16Tr) rrrr_compute_moments_sync_compute_momentsrrMrminmax)r axesr[ synchronized need_cast ori_dtypemeanvariancer r rmoments's    rc Cstj}|s t|||Stj|dd}tj||dd}tjt||dd}tj||dd}|tjjj |}|tjjj |}|tjjj |} tj || } tj || } t | t| d} |sqt | |} t | |} | | fS)Ncount)nameTrZr)r distributeget_replica_contextr ones_liker`r/ all_reduceReduceOpSUMr divide_no_nanr?r) r rr[ replica_ctx local_count local_sumlocal_squared_sumy_sum y_squared_sum count_sumry_squared_meanrr r rr>s&     rcCstjj|||dS)Nr)rr r)r rr[r r rr^rRrMbP?cCs~|dkr3dgt|j}|jd||<t||}t||}|dur)t||}|dur3t||}tjj||||||dS)Nrrr)r rroffsetscalevariance_epsilon)r_rErrUr batch_normalization)r rrrDrrrrEr r rrbs"    rcCslt|}t|}tj|dd}t|jd}|dkrdn|}t||}tjj|||||dd}t||S)Nint32rrfloat64F)rr label_length logit_length blank_indexlogits_time_major)rrrr result_typerr ctc_loss)rr target_length output_length mask_index result_dtype compute_dtypelossr r rr%ys  r%greedydcCst|}t|}|d|d}} t|d}t|jd} t|| }t|dd}|dkrBNTSoptimal)optimizerrTzBNTS,BSNH->BTNH) rr$rreinsumrmultiplyaddrZr rV) querykeyvaluebiasrSrTr logits_dtyperrY probs_dtypeprobsr r r_dot_product_attention_xlas  rgc Cs|durd}|r tdt|}t|}t|}t|jdkr0td|jd|jd|jdt|d} |durFd tt| d n|}t|||||||S) NFz7Flash attention is not supported in tensorflow backend.r-zG`dot_product_attention` only supports 4D inputs. Received: query.shape=z , key.shape=z, value.shape=rlrrr)rFrr_rErr.rrg) r`rarbrcrSrrTflash_attentionattn_logits_soft_capHr r rdot_product_attentions0 "rk)r)r-)r4)r)T)r)NrsN)rrsNr)rrsNNr)rNF)Fr)F)FF)NNr)r)r,r-rTr)NNNFNN)@r r tensorflowr keras.srcr&keras.src.backend.common.backend_utilsr!keras.src.backend.tensorflow.corerrr rrrrr!r#r$r(r+r,r2r3r6r8r9r:r<r>rBrJrNrPr'rVrWrirprrr{rrzrrrrrrrrrrrrrrr%rDrLrNrZrgrkr r r rs                  # + B B ( $" # S ? 0   A