ij Z4ddlZddlZddlZddlZddlZdZdZdZdZdZ dZ ddZ d Z d Z d Zd Zd ZdZdZdj'eZdedZdj'eZdj'eZdZddZdZdZdZddddZd dZdZy)!Nc:|jdvrtd||dz |zdz}|jdk(r$|t|||z n|}|dz }|dz |z}||fS| ||zdz }n||dzzdz |z}t|dz|dzz|dz }||z }||fS)aConvert the padding arguments from Keras to the ones used by JAX. JAX starts with an shape of size `(input-1) * stride - kernel_size + 2`, then adds `left_pad` on the left, and `right_pad` on the right. In Keras, the `padding` argument determines a base shape, to which `output_padding` is added on the right. If `output_padding` is None, it will be given a default value. >samevalidIThe `padding` argument must be one of 'valid', 'same'. Received: padding=r)lower ValueErrormaxmin) kernel_sizestride dilation_ratepaddingoutput_paddingleft_pad right_padpad_lens {/var/www/html/emotional.easysim.app/public_html/venv/lib/python3.12/site-packages/keras/src/backend/common/backend_utils.py6_convert_conv_transpose_padding_args_from_keras_to_jaxrs}}//!!(  +  ?m3a7K}}'! %  V ${ 2  ?!On4  Y   !{*Q.G"K!O3a7.HGw!|gk1;?Ch& Y c |jdvrtd||}|dz |zdz}|jdk(r|t|||z n|}d}|}n6|||dzz n|}t|dz|z |zdz d}d|z|dzz|z |z}|dkDr*|dkDr%tjd|d|d |d |d |d ||k\r,tjd |d|dz d|d t |dz }||fS)aConvert the padding arguments from Keras to the ones used by Torch. Torch starts with an output shape of `(input-1) * stride + kernel_size`, then removes `torch_padding` from both sides, and adds `torch_output_padding` on the right. Because in Torch the output_padding can only be added to the right, consistency with Tensorflow is not always possible. In particular this is the case when both the Torch padding and output_padding values are strictly positive. >rrrrrrrzbYou might experience inconsistencies across backends when calling conv transpose with kernel_size=z , stride=z, dilation_rate=z , padding=z, output_padding=.zHTorch backend requires output_padding < stride. Clamping output_padding z -> z for stride )r r r warningswarn UserWarning)r rrrroriginal_kernel_size torch_paddingtorch_output_paddings r8_convert_conv_transpose_padding_args_from_keras_to_torchr 5s}}//!!(  +  '?m3a7K}}'! %  V ${ 2   -% [1_ $  Q,~=!C Da    a /+ = N q1A5 77K6LMX-m_=i00@ C v% '';&|} || fS)Nrr"r#rr) r%r&r'r(r r)ziplisttupler )r*r+r,rrrr-r.torch_paddingstorch_output_paddingsr0r1r2r3rrcorrected_output_paddingssop max_alloweds r-compute_conv_transpose_padding_args_for_torchrAsk;'!+',N # $;%NC)H "  *'37GWQZ -- q!  E,Q/)+     m,$$%9:/;4!# ge} - Y) )  12 !QUm  % , ,[ 9 % , ,R 0 16 0 00rclt|dz }|dd}g}t|D]} |t|tr|n|| } t|tr|n|| } t|tr|n|| } t || | | || \} }|| dz | zdz}|dz | z }|dz |z }|j ||f|S)aZPer-spatial-dim asymmetric (left, right) crop amounts for torch. Torch's `conv_transpose*d` only supports symmetric `padding` plus a right-side `output_padding`, which cannot express the asymmetric crops that Keras "same" semantics requires when stride > 1 with an odd kernel. We work around this by calling torch with `padding=0, output_padding=0` (which produces the largest "natural" output of size `(input - 1) * stride + effective_kernel`), then asymmetrically slicing the spatial dims to the same window JAX would compute. The slice on each side is `(effective_kernel - 1) - left_pad` from the left and `(effective_kernel - 1) - right_pad` from the right, where `left_pad` / `right_pad` come from the existing JAX helper. A negative value means we need to extend with zero-padding instead of cropping. rNr"r#rr$)r*r+r,rrrr-r.cropsr0r1r2r3rreffective_kernel crop_left crop_rights r-compute_conv_transpose_output_crops_for_torchrGs.;'!+', E # $.%NC)H "  *'37GWQZ -- q!  C,Q/)+    13a7?JQN%)X5 &*i7  i,-3.4 Lrc|y|jdvrtd||dz |zdz}|jdk(r!|t|||z n|}|dz |z|z|zS|||zS|dz |z|dzz|zS)N>rrrrrr)r r r  input_sizer r,rrrs r"_get_output_shape_given_tf_paddingrKs}}//!!(  +  ?m3a7K}}'!%  W % 3  Q')K7.HH  !' 'Ng- a?.P Prc t|dz }|} t|tr|ft| z}t|tr|f|z}t|tr|f|z}|dk(r|dd} n|dd} g} t|D]9} |dn|| } t | | | | || || || }| j |;|dk(r|dg| z|gz} | S|d|g| z} | S)Nr channels_lastrrIr)r%r'r(r&rKr))r*r filtersr,rr data_formatrr-r.input_spatial_shape output_shaper0current_output_paddingshape_is r#compute_conv_transpose_output_shaperU,s;;'!+&.#&(*S1E-FF'3*//-%&(+;; o%)!B/)!"oL # $ %"*Dq0A 5*1-,Q/AJ1'*   G$ %o%#A',6'B  $A0<? rctj|}| |cxkr|ksntd|d|d|dkr||z}|S)z?Canonicalize an axis in [-num_dims, num_dims) to [0, num_dims).zaxis z. is out of bounds for an array with dimension rr)operatorindexr axisnum_dimss rcanonicalize_axisr\\sY >>$ D 9 'x 'D6Gj    axh Krcrt|ttfrtfd|DSt|fS)zACanonicalize an axis or axes to a tuple of non-negative integers.c36K|]}t|ywN)r\).0ar[s r z$canonicalize_axes..lsB&q(3Bs)r'r:r9r\rYs `rcanonicalize_axesrcis2$ &BTBBB dH - //rc<t|tr t|S|S)zDStandardize an axis to a tuple if it is a list in the numpy backend.)r'r9r:rZs rstandardize_axis_for_numpyrfps$T405;:d:rc |dk(r|jdn|jd}|jd}t|tr,t|tr||k7rtd|d|d|dy y y ) u;Validate a conv input against its kernel shape. Used by `conv`, `depthwise_conv`, and `separable_conv` — all share the convention that the kernel's input-channel dimension is `kernel.shape[-2]`. Produces a clear error message before the backend op raises its own implementation-specific one. rMrNrr"^The number of input channels must match the kernel's input channels. Received: input channels=, kernel input channels=, data_format=''.Nshaper'r(r inputskernelrPinput_channelskernel_input_channelss rcheck_conv_input_channelsrsus(?: R Q#LL, >3' ,c 2 3 3 22@1AB345'= ,   4 3 (rc |dk(r|jdn|jd}|jd}t|tr,t|tr||k7rtd|d|d|dyyy) aValidate a conv_transpose input against its kernel shape. `conv_transpose` kernels use the layout `(spatial..., out_channels, in_channels)`, so the input-channel dimension is at `kernel.shape[-1]` (vs. `kernel.shape[-2]` for regular conv). rMrNrrhrirjrkNrlrns r#check_conv_transpose_input_channelsrus(?: R Q#LL,>3' ,c 2 3 3 22@1AB345'= ,   4 3 (rc||St|tttfst d|t|tr|fS|S)z9Convert the non-`None` value to either a tuple or a list.z;`value` must be an integer, tuple or list. Received: value=)r'r(r:r9r )values rto_tuple_or_listrxsQ } ec5$/ 0 $g '  %x Lrz\w+z(?:{0:}(?:,{0:})*)?z\(z\)z{0:}(?:,{0:})*z ^{0:}->{0:}$ctjt|std|d|j dD\}}||fS)Nznot a valid gufunc signature: c 3K|]M}tjt|Dcgc]%}ttjt|'c}Oycc}wwr_)refindall _ARGUMENTr:_DIMENSION_NAME)r`arg_listargs rrbz4_vectorize_parse_gufunc_signature..sM zz)X6  "**_c2 3  s!A*A Az->)r{match _SIGNATUREr split) signatureargsretvalss r!_vectorize_parse_gufunc_signaturersO 88J *9)EFF "- MD' =rc (t|}|rt||kr0td|d|t||k7rtd|d||r|| dnd}t||D].\}}||vr|||<|||k7std|d|d||y) Nzinput with shape z9 does not have enough dimensions for all core dimensions z output shape z does not match core dimensions z%inconsistent size for core dimension z: z vs )r%r r8) dim_sizesrm core_dimsis_input num_core_dims core_shapedimsizes r_vectorize_update_dim_sizesrs NM u: %#E7+'[*  u: &w'))2 5  ,5 ~'"JJ/ T i !IcN Ys^ #7uB&Ys^,.  rcddlm}t|t|k7r#tdt|dt|g}i}t ||D]T\}}t ||j |d|jt|z }|j|j d|V|d}|D]} |j|| }||fS)Nropsz/wrong number of positional arguments: expected , got Tr) keras.srcrr% TypeErrorr8rrmndimr)broadcast_shapes) rinput_core_dimsrshapesrrrrbroadcast_shaper>s r!_vectorize_parse_input_dimensionsrs 4yC(( O,-VCI; @  FIdO4(Y# syy)d xx#i.( cii&' ( QiO C..BC I %%rc*ddlmfd}|S)Nrrc |}t|ttfr|Dcgc]} j|}}n |jg} dgt |z}nh }t |dkDrt|tst d|t |t |k7r#t dt |dt |t }t||D]\}}t|||d|Scc}w)Nrrz@output must be a tuple when multiple outputs are expected, got: z+wrong number of output arguments: expected rFr) r'r9r:rmr%rdictr8r) routx out_shapesoutput_core_dimssizesrmrrexpected_output_core_dimsfuncrs rwrappedz-_vectorize_check_output_dims..wrapped sDk cD%= )0341#))A,4J4))J $ , "tc*o5 8 #$q(C1G**-0:#&6"77 #$4 56fS_.3sBaz!S/ABscft|}D]\}}|j|||i|Sr_)r9insert)rkwargsr0rr static_args static_kwargss rnew_funcz+_vectorize_apply_excluded..new_func8sBDz! FAs KK3  T5V5}55r) enumerateitemssortedr%) rexcludedrrr0r dynamic_argskeyvaldynamic_kwargsrrrs ` @@r_vectorize_apply_excludedr)s T6!!&/oKFAs(9JCKLK!'S#3h3FSN B8BB s4y= DG K /5llnP(#sxS#XPM6 \> 11#L Qs( B=B= C C)C  C(C)rrcrddlmttjfd}|S)z*Implementation adapted from JAX and NumPy.rrct||\}}}t\}ndgt|zd}t|Dchc] \}}| | }}}t |rYt fd|Drt d|dt|||i\}}}tDcgc] \}}||vr| c}}t tj|}t|\} } t|| |} g} g} t|D]\}}|jd|jt|z }t| t|z }|dz|z}| j|dddt dt|D}j||}| j|| }gtt| D]S\}}t d |D}t!d |Dr!jt| d z |z K||}U|| }s|St#|t rt fd |DSj%|Scc}}wcc}}w) Nrc3.K|] }|dk7yw)rNr)r`r0rs rrbz2vectorize_impl..wrapped..Xs??1%+?szCannot pass None at locations z with signature=)rrNc32K|]\}}|dk(s |yw)rNr)r`r0rs rrbz2vectorize_impl..wrapped..ws$aTQY$s rec3.K|] }|dk(rdndyw)rNrr)r`rs rrbz2vectorize_impl..wrapped..sLDAID14Lsc3$K|]}|du ywr_r)r`rZs rrbz2vectorize_impl..wrapped..s4D44<4src3DK|]}j|yw)reN) expand_dims)r`rdims_to_expandrs rrbz2vectorize_impl..wrapped..s#<=7s )rrr%ranyr r:mapconvert_to_tensorrrr8rmrr)squeezeallr'r) rr excluded_funcrr0r none_args_rrr checked_func squeezed_argsrev_filled_shapesr noncore_shapepad_ndim filled_shapesqueeze_indices squeezed_argvectorized_funcnegdim axis_sizesin_axesresultrrrrpyfuncrvmap_fns @@rrzvectorize_impl..wrappedHs&? HdF' # tV  1)< .O- "dSY.O# %.t_D61c QD D y>?Y?? 4YK@&&/[2&?y$& "M4 (8AsI%O S..56%F /& "4 9&6  !$8 /NCII&A3y>(ABM?+c-.@@H#d?]:L  $ $\$B$%7 8#$!*=!9$O;;s;AL   . /'"+C1B,C"D D FJLLLG4G44%%c/&:Q&>&GH")/7"C  D !-0M  &AG ??6?? ?qEs II(I)rrset functoolswraps)rrrrrrs```` @rvectorize_implrAs6uH__VE@E@N Nrc|dk\rtdg|zt|||gz}n&tt|||gtdgd|z zz}|t|S)z$Slice a Tensor along the given axis.rNrN)sliceEllipsisr:)rstartstopsteprZslicess rslice_along_axisrsi qy+%udD)A(BBE%t45t I9   U6] rcDtj||z }|dz}||fS)z8Compute small and big window sizes for adaptive pooling.r)mathceil) input_dim output_dimsmallbigs r%compute_adaptive_pooling_window_sizesrs( IIi*, -E !)C #:r)NrMr)T)rNrr) rrrWr{rrr r6rArGrKrUr\rcrfrsrurxr~format_CORE_DIMENSION_LISTr}_ARGUMENT_LISTrrrrrrrrrrrrrs  *Z?/D$N61r5pQF-` 0;  6 0 ",33OD&'r * !((3  " "> 2  6&2#L2015Od r