o "jsX@sddlZddlZddlmZddlTddlmZmZddZedd d Z ed d d Z edddZ edddZ edddZ edddZedddZedddZed d!d"Zed#d$d%Zed&d'd(ZdWd)d*Zed+d,d-Zed.d/d0Zed1d2d3Zed4d5d6Zed7d8d9Zed:dXd;d:Zed<d=d>Zed?d@dAZedBdCdDZedEdFdGZedHdIdJZedKdLdMZ edNdOdPZ!edQdRdSZ"edTdUdVZ#dS)YN)core)*)REGISTER_COMPOSITElookup_compositecGst|j}||g|RSN)rtype)opargsZ _lowerruler i/var/www/html/Deteccion_Ine/venv/lib/python3.10/site-packages/paddle/incubate/autograd/composite_rules.py _composites r Zsoftmaxc Csd}ddlm}||j}|dvrd}t|d}|js*t||}|r(t|d}|St||dd}d|_t||}t||dd }t ||}|rMt||}|S) z#define composite rule of op softmaxFr convert_dtypefloat16Zuint16Tfloat32r)keepdimaxisr) paddle.base.data_feederrdtypecastshapeexpmaxZ stop_gradientsumdivide) xris_amprrresZmax_tempZ molecular denominatorr r r softmax_composite"s&        r"Z batch_normc sHd} ddlm} | |j} | dvr*d} t|d}|rt|dn|}|r(t|dn|}|dvr0dnt|jd|r<| p=| }tfd d tt|jDtfd d t|jD}t dgd |j}|st |}t ||}|||}t |||}|d kr|||}n |t ||t ||}||d||}||d||}n6t |j|j}t |j|j}t |||}|d kr||t |||}n|t ||t t ||||}|d kr|||}n t |||t ||}| rt|| }t|}t|}t|}t|}d}|s||||||fS|||dd|fS)z define composite rule of op batch_norm As the same with op kernel, the position of savedvariance indeed return inverse std. FrrrTr)ZNCNCLNCHWZNCHWDrc3s|] }|kr|VqdSrr ).0i) feature_axisr r _sz&composite_batchnorm..c3s$|] \}}|vr dn|VqdS)rNr )r%r&s) reduce_axesr r r(`s ZNHWCN)rrrrlenrtuplerange enumeratefullmeanpowreshapezerosassign)rZrun_meanZrun_varscalebiasis_testZmomentumepsilon data_layoutZuse_global_statsZtrainable_statisticsrrrZ use_run_statZ stats_shapehalfZ batch_meantempZ batch_varZinv_stdZx_hatyZ batch_mean_Zinv_std_Z run_mean_Zrun_var_Z reserve_spacer )r'r*r composite_batchnorm=sb        r>Z layer_normcCs\d}ddlm}||j}|dvr*d}t|d}|rt|dn|}|r(t|dn|}tt|t|j}t||dd} || } | | } t| |dd} | |} t | }| |}|durr|j|d|jkrnt ||j|d}||}|dur|j|d|jkrt ||j|d}||}t | |jd|} t | |jd|} |rt||}|| | fS) z} define composite rule of op layer_norm out = (x - mean(x)) / sqrt(var + epsilon)) var = mean((x-mean(x))^2) FrrrTrrN) rrrrr-r.r,rr1rsqrtr3)rr6r7r9Zbegin_norm_axisrrrrmean_ differencevar_tmp1variancevar_tmp3Z rsqrt_varoutr r r layernorm_composites:     rFZ instance_normcCsFd}ddlm}||j}|dvr*d}t|d}|rt|dn|}|r(t|dn|}|j\}}} } ttdt|j} t|| dd} || } | | }t|| dd}||}t |t d gd |jd }| |}|d urvt |d |d d g}||}|d urt |d |d d g}||}t | d g} d |}t |d g}|rt||}|| |fS)z define composite rule of op instance_norm out = (x - mean(x)) / sqrt(var + epsilon)) var = mean((x-mean(x))^2) FrrrTrrr?rN) rrrrrr-r.r,r1r2r0r3)rr6r7r9rrrnchwrr@rArBrCrDZsqrt_varrEZ scale_tileZ bias_tileZsaved_variancer r r instancenorm_composites:       rOZgeluc Csd}d}t|jdkr|jndg}t||j}t|d|j}|rHt||||j}t|d|j}t||||||} |||| } | S||t|t|j||j} || } | S)z define composite rule of op gelug;f?gmBP ?rrrHgHm?)r,rZonesrr0tanherf) rZ approximateZ M_SQRT1_2Z M_2_SQRTPI full_shapeoner;ZkAlphaZ GELU_CONSTANTZtanh_outrEZcdfr r r gelu_composites  rTZ reduce_meanc sd}ddlm}|j}|dvrd}td|dgfvr(ttdtj}t|t r0|fn|}t ||d}fd d |D}|gkrId } nt t j|} tg| |jd } t|| } |rdt| |} | S) z define composite rule of op meanFrrrTrNrcsg|]}j|qSr r)r%rrr r z"mean_composite..r)rvaluer)rrrrr-r.r,r isinstanceintr functoolsreduceoperatormul fill_constantr) rrrrrrZaxesZsum_xZ ele_nums_listZ value_to_fillZnormr r rVr mean_composites.      raZ expand_v2cCs|j}t|}t|}||kr|dksJg}t|D]8}||}||}|dkr-||nd} ||} | dkr@|dks=Jd} n| | dksHJt| | } || q||krxg} t||D]}| dq`| |t|| } t| |dSt||dS)z define composite rule of op expnad_v2, expand_v2->expand repeat_times = shape / x.shape out = tile(x, repeat_times = repeat_times) rrrJ repeat_timesrr,r.r[appendextendr3Ztile)rrshape_indim_outdim_inrcr&offsetdimsize_insize_outrepeatshape_in_expand x_reshaper r r expand_v2_composite"s0         rqZ expand_as_v2cCs|j}|dur |j}|dusJt|}t|}||kr |dks"Jg}t|D]8}||}||} | dkr:|| nd} ||} | dkrM| dksJJd} n| | dksUJt| | } || q(||krg} t||D]}| dqm| |t|| }t||dSt||dS)z define composite rule of op expnad_as_v2, expand_as_v2->expand_as repeat_times = target_shape / x.shape out = tile(x, repeat_times = repeat_times) NrrrJrbrd)rr=Z target_shapergrhrircr&rjrkrlrmrnrorpr r r expand_as_v2_compositeDs6          rrstackcsZ|dj}|dkr|t|d7}|d|d||dtfdd|D|}|S)zt define composite rule of op stack unsqueeze each dimension of the input (use reshape), and then concat rrN)rcsg|]}t|qSr )r3)r%itemZ out_shaper r rWsrXz#stack_composite..)rr,concat)rrx_shaperEr rur stack_compositeis rxZflatten_contiguous_rangec Cs|j}t|dkr |nd}t|dkr|nd}||ksJt|dkr,t|dgddfS||kr8t||ddfSd}t||dD]}|||9}qAg}t|D] }|||qP||t|dt|D] }|||qht||ddfS)a define composite rule of op flatten, flatten_contiguous_range -> flatten. xshape is the dim with 0 added to the front of x, keep the shape information of x to calculate the grad. CINN doesn't need xshape for backward pass, return none instead of xshape. shape_out is the parameter of reshape, get from start_axis and stop_axis. out = reshape(x, shape=shape_out), xshape rrrUN)rr,r3r.re) rZ start_axisZ stop_axisrgZ start_dimZend_dimZ slice_numelr&Z shape_outr r r "flatten_contiguous_range_compositews$     ryZdropoutc Cs|durdn|}|r |nd}|dk}t|j|j||d}|rM|sA|dkr1d|t|jtjjjfS||d|t|tjjjfSt |t|tjjjfS|s[||t|tjjjfS|d|t|tjjjfS)zdefine composite rule of op dropout. upscale_in_train: train: out = input * mask / ( 1.0 - p ) inference: out = input downscale_in_infer train: out = input * mask inference: out = input * (1.0 - p) NTrupscale_in_train)rrpseed?) bernoullirrr4rZVarDescZVarTypeZUINT8rr5) rZ seed_tensorr{r8moder|Zfix_seedrzmaskr r r dropout_composites  rc CsXddlm}||dvrdn|}ttt||dd|dtt|dkr#|ndg|||S) Nrrrrr~r})minrr|r)rrrZ greater_equaluniformr`r,)rrr{r|rZ new_dtyper r r rs rZ hard_swishcCsvd}d}d}t|jdkr|jndg}tt|t|||jdt|d|jdt|||jd|t|||jd}|S)zdefine composite rule of op hard_swish. offset=3, threshold=6, scale=6 out = minimum( maxmum(x + offset, 0), threshold ) * x / scale g@g@rrrIr~)r,rminimummaximumr0r)r thresholdr6rjrRr r r r hard_swish_composites" rZ index_selectcCs(|dkr t|j|}t|||d}|S)z)define composite rule of op index_select.r)r)r,rZgather)rindexrr r r r index_select_compositesrZsigmoidcCsXd}ddlm}||j}|dvrd}t|d}dt| }d|}|s'|St||S)zI define composite rule of op sigmoid res = 1 / (1 + exp(-x)) FrrrTrrrrrrrrrrrZsum_tempr r r r sigmoid_composite   rZsilucCsXd}ddlm}||j}|dvrd}t|d}dt| }||}|s'|St||S)zF define composite rule of op silu res = x / (1 + exp(-x)) FrrrTrrrrr r r silu_compositerrZmeshgridcCst|}dg|}t|D]}||}|dks|dksJ|dkr,||jd||<q g}t|D]}dg|}||||<|||||q3|S)a define composite rule of op meshgrid If the input has N tensors of size S_0, ... S_n-1, then the output will also have N tensors, where each tensor is of shape (S_0, ..., S_n-1). E.g. a1 is Tensor [1,2,3] b1 is Tensor [4,5] r1, r2 = paddle.meshgrid([a1, b1]) r1 is Tensor [[1,1], [2,2], [3,3]] r2 is Tensor [[4,5], [4,5], [4,5]] rr)r,r.rkrrer3Z broadcast_to)Zinputssizerr&rkZoutputsZ view_shaper r r meshgrid_composites      r fill_any_likecCs t|j||}|S)z&define composite rule of op full_like.)r0r)rZ fill_valuerZplacevalr r r r+sZsqueeze2cs t|jdkrt|dgSt|dkrtt}n fdd|D}g}t|jD]\}}|dkr9||vs>||q-t||}|dgS)z define composite rule of squeezerNcsh|]}|qSr r )r%axrankr r Dsz%squeeze2_composite..r)r,rr5setr.r/rer3)rrdimsZ new_shapedr)rEr rr squeeze2_composite4s     rsqrtcCpd}ddlm}||j}|dvrd}t|d}tt|jdkr#|jndgd|j}t||}|s3|St||S) z@ define composite rule of op sqrt res = pow(x, 0.5) FrrrTrrrHrrrrr0r,rr2rrrrr=r r r r sqrt_compositeMs   $ rr2cCs~d}ddlm}||j}|dvrd}t|d}t|ttfr1tt|j dkr*|j ndg||j}t ||}|r=t||}|S)z7 define composite rule of op pow res = x^y FrrrTrr) rrrrrZr[floatr0r,rr2)rr=rrrr r r r pow_composite`s   $  rZrelucCs:t|jdkrt|t|jd|jSt|tdgd|jS)z!define composite rule of op relu.rr~r)r,rrr0rrVr r r relu_compositevsrZ unsqueeze2cCsf t|j}t|}|D]}|dkr|t|d7}|d|dg||d}q t||}|dgS)z%define composite rule of op unsqueezerrN)listrr,r3)rrrwZ axis_listr&rEr r r unsqueeze_composites    rr?cCr) z"define composite rule of op rsqrt.FrrrTrrr+rrr r r rsqrt_composites   $ rZ group_normcCs4|dksJ|j\}}}} d} ddlm} | |j} | dvr/d} t|d}t|d}t|d}t|||df}t|d dd } t||d dd | | }t|t|}d t ||}|| |}t||||| f}|d uru|t|d }|d ur|t|d }t| ||f}t|||f}| rt|| }|||fS) z define composite rule of op group_norm. x = ((x - mean) / sqrt(var + epsilon)) * scale + bias mean and var are computed from groups r$FrrrTrrJrrN)rJrr) rrrrrr3r1rZ zeros_liker)rr6r7r9groupsr:NCHWrrrr@Zvar_Zvar_invrEZ ret_mean_Zret_var_r r r group_norm_composites4         rrcCsd}|D]}||7}q|S)Nrr )rZansxir r r sum_composites rZ leaky_relucCs$|dkr t|||St|||S)z'define composite rule of op leaky_relu.r})rr)rZnegative_sloper r r leaky_relu_compositesr)rr)$r\r^Z paddle.baserZ primitivesZprimregrrr r"r>rFrOrTrarqrrrxryrrrrrrrrrrrrrrrrrr r r r sv   U * +   ! $    !             (