Z j[ , S SK r S SKJr S SKrS SKJr S SKJs Jr S SK J
s J s Jr S SK
Jr SSKJr SSKJr SSKJr SSKJr SS KJr SS
KJr SS K J r J r SS K J r J!r!J"r" SS
K#J$r$ SSK%J&r& SSK'J(r(J)r) SSK*J+r+J,r,J-r-J.r.J/r/ SSK0J1r1J2r2 SSK3J4r4 SSK5J6r6 SSK7J8r8 SSK9J:r:J;r; \/Rx " \=5 r>\," SS9\ " S S\85 5 5 r? " S S \:5 r@\," SS9\ " S S \5 5 5 rA " S S \R 5 rC " S! S"\R 5 rD " S# S$\&5 rE " S% S&\;5 rF " S' S(\E5 rG " S) S*\E5 rH\,\ " S+ S,\$5 5 5 rI\," S-S.9 " S/ S0\E5 5 rJ\,\4" S1S29 " S3 S4\5 5 5 rK/ S5QrLg)6 N) dataclass)strict )initialization)ACT2CLS)filter_output_hidden_states)PreTrainedConfig)TorchvisionBackend) BatchFeature)group_images_by_shapereorder_images)IMAGENET_DEFAULT_MEANIMAGENET_DEFAULT_STDSizeDict)BaseModelOutput)PreTrainedModel) ImagesKwargsUnpack)TransformersKwargsauto_docstringcan_return_tupleis_torchdynamo_compilinglogging)
TensorTypemerge_with_config_defaults)requires)capture_outputs )GotOcr2VisionConfig)GotOcr2VisionAttentionGotOcr2VisionEncoderz&PaddlePaddle/SLANeXt_wired_safetensors)
checkpointc $ \ rS rSr% Sr\\S' Srg)SLANeXtVisionConfig2
image_size N)__name__
__module__ __qualname____firstlineno__r' int__annotations____static_attributes__r( |/root/GenerationalWealth/GenerationalWealth/venv/lib/python3.13/site-packages/transformers/models/slanext/modular_slanext.pyr$ r$ 2 s Jr0 r$ c \ rS rSrSrg)SLANeXtVisionAttention8 r( Nr) r* r+ r, r/ r( r0 r1 r3 r3 8 r0 r3 c ^ \ rS rSr% SrSrS\0rSr\ \- S- \
S' Sr \ \
S' Sr
\ \
S ' S
r\ \
S ' Sr\ \
S ' S
r\ \
S' U 4S jrSrU =r$ )
SLANeXtConfig< a
vision_config (`dict` or [`SLANeXtVisionConfig`], *optional*):
Configuration for the vision encoder. If `None`, a default [`SLANeXtVisionConfig`] is used.
post_conv_in_channels (`int`, *optional*, defaults to 256):
Number of input channels for the post-encoder convolution layer.
post_conv_out_channels (`int`, *optional*, defaults to 512):
Number of output channels for the post-encoder convolution layer.
out_channels (`int`, *optional*, defaults to 50):
Vocabulary size for the table structure token prediction head, i.e., the number of distinct structure
tokens the model can predict.
hidden_size (`int`, *optional*, defaults to 512):
Dimensionality of the hidden states in the attention GRU cell and the structure/location prediction heads.
max_text_length (`int`, *optional*, defaults to 500):
Maximum number of autoregressive decoding steps (tokens) for the structure and location decoder.
slanext
vision_configN post_conv_in_channelsr& post_conv_out_channelsr% out_channels hidden_sizei max_text_lengthc > U R c [ 5 U l O9[ U R [ 5 ( a [ S0 U R D6U l [ TU ] " S0 UD6 g Nr( )r; r$
isinstancedictsuper
__post_init__selfkwargs __class__s r1 rG SLANeXtConfig.__post_init__Y sS %!4!6D
**D
1
1!4!Jt7I7I!JD
''r0 )r; )r) r* r+ r, __doc__
model_typer$ sub_configsr; rE r. r= r- r> r? r@ rA rG r/
__classcell__rK s @r1 r8 r8 < sm J"$78K7;M4--4;!$3$"%C%L#K OS ( (r0 r8 c ^ \ rS rSrU 4S jrS\R S\R S\R S\\ 4S jr Sr
U =r $ ) SLANeXtAttentionGRUCella c > [ TU ] 5 [ R " XSS9U l [ R " X"5 U l [ R " USSS9U l [ R " X- U5 U l g )NF)bias ) rF __init__nnLinearinput_to_hiddenhidden_to_hiddenscoreGRUCellrnn)rI
input_sizer@ num_embeddingsrK s r1 rX SLANeXtAttentionGRUCell.__init__b s[
!yyu M " + C YY{AE:
::j 9;Gr0 prev_hidden batch_hidden char_onehotsrJ c U R U5 nU R U5 R S5 nXV- n[ R " U5 nU R U5 n[ R " US[ R S9R UR 5 nUR SS5 n[ R " X5 R S5 n [ R " X/S5 n
U R X5 n X4$ )NrW dimdtyper )r[ r\ unsqueezetorchtanhr] Fsoftmaxfloat32tori transposematmulsqueezecatr_ ) rI rc rd re rJ batch_hidden_projprev_hidden_projattention_scores attn_weightscontextconcat_context
hidden_statess r1 forward SLANeXtAttentionGRUCell.forwardk s ! 0 0 > 00=GGJ,? ::&67 ::&67yy!1q
NQQRbRhRhi #--a3 ,,|:BB1E G#:A> =
**r0 )r\ r[ r_ r] ) r) r* r+ r, rX rk FloatTensorr r r| r/ rP rQ s @r1 rS rS a sQ H+&&+ ''+ '' +
+, + +r0 rS c 2 ^ \ rS rSrSU 4S jjrS rSrU =r$ )
SLANeXtMLP c > [ TU ] 5 [ R " X5 U l [ R " X5 U l Uc [ R " 5 U l g [ U " 5 U l g N) rF rX rY rZ fc1fc2Identityr act_fn)rI r@ r?
activationrK s r1 rX SLANeXtMLP.__init__ sN
99[699[7'1'9bkkm wz?R?T r0 c l U R U5 nU R U5 nU R U5 nU$ r )r r r )rI r{ s r1 r| SLANeXtMLP.forward s2 /
/
M2
r0 )r r r r )r) r* r+ r, rX r| r/ rP rQ s @r1 r r s U r0 r c t ^ \ rS rSr% \\S' SrSrSrSr SS/r
\ R " 5 U 4S j5 r
S
rU =r$ ) SLANeXtPreTrainedModel configbackbone pixel_values)imageTstructure_attention_cellstructure_generatorc < > [ TU ] U5 [ U[ 5 ( a. UR b! [
R " UR S5 [ U[ 5 ( aS UR ( aB [
R " UR S5 [
R " UR S5 [ U[ R 5 ( a UR S: a# S[ R " UR 5 - OSn[
R " UR" U* U5 [
R " UR$ U* U5 UR& b# [
R " UR& U* U5 UR( b# [
R " UR( U* U5 [ U[* 5 ( a S[ R " U R, R S- 5 - nUR. 4 H nUR1 5 Hy n[ U[ R2 5 ( d M$ [
R " UR4 U* U5 UR6 c MV [
R " UR6 U* U5 M{ M gg)zInitialize the weightsNg r g ?) rF
_init_weightsrD SLANeXtVisionEncoder pos_embedinit constant_r3 use_rel_pos rel_pos_h rel_pos_wrY r^ r@ mathsqrtuniform_ weight_ih weight_hhbias_ihbias_hhSLANeXtSLAHeadr r childrenrZ weightrV )rI modulestd generatorlayerrK s r1 r $SLANeXtPreTrainedModel._init_weights s f% f 2 3 3 +v / /5 f 4 5 5!!v / /5v / /5 fbjj ) )9?9K9Ka9O# &"4"455UVC MM&**SD# 6 MM&**SD# 6~~)
fnnsdC8~~)
fnnsdC8 fn - - $++"9"9C"?@@C $ 8 8 : & / / 1E!%33
ellSD#> ::1 MM%**sdC @ 2 ; .r0 r( )r) r* r+ r, r8 r. base_model_prefixmain_input_nameinput_modalities supports_gradient_checkpointing _keep_in_fp32_modules_strictrk no_gradr r/ rP rQ s @r1 r r sF "$O!&*#$>@U#V
]]_"A "Ar0 r c \ rS rSrSrg)r r( Nr5 r( r0 r1 r r r6 r0 r c h ^ \ rS rSr S S\S- 4U 4S jjjrS\R S\\ 4S jr
Sr U =r $ )
SLANeXtBackbone Nr c > [ TU ] U5 [ UR 5 U l [
R " UR UR SSSSS9U l U R 5 g )Nr r rW F) kernel_sizestridepaddingrV ) rF rX r r; vision_towerrY Conv2dr= r> post_conv post_initrI r rJ rK s r1 rX SLANeXtBackbone.__init__ s^
01E1E F ( (&*G*GUV_`jkrw
r0 r{ rJ c U R " U40 UD6nU R UR 5 nUR S5 R SS5 n[ UUR UR S9$ )Nr rW )last_hidden_stater{
attentions)r r r flattenrq r r{ r )rI r{ rJ
vision_outputs r1 r| SLANeXtBackbone.forward sj ))-B6B
}'F'FG
%--a0::1a@
+'55$//
r0 )r r r )
r) r* r+ r, rE rX rk Tensorr r r| r/ rP rQ s @r1 r r s@ #
t
U\\
VDV=W
r0 r c ^ \ rS rSrS\0r S S\S- 4U 4S jjjr\\ \
S S\ R S\ R S- S\\ 4S jj5 5 5 rS
rU =r$ ) r r Nr c > [ TU ] U5 [ UR UR UR
5 U l [ UR UR
5 U l U R 5 g r )
rF rX rS r> r@ r? r r r r r s r1 rX SLANeXtSLAHead.__init__ s_
(? ) )6+=+=v?R?R)
% $.f.@.@&BUBU#V r0 r{ targetsrJ c [ R " UR S U R R 4[ R
UR S9n[ R " UR S /[ R UR S9n/ n/ n[ U R R S- 5 H n[ R " XPR R 5 R 5 n U R XAR 5 U 5 u pHU R U5 n
U
R! SS9nUR# U
5 UR# U5 [ R$ " USS9R' U R R S-
5 R) S5 R+ 5 ( d M O [ R, " [ R$ " USS9S[ R
S9R/ UR0 5 n [3 XS9$ ) Nr ri device)sizeri r rW rh rg )r r{ )rk zerosshaper r@ ro r longrangerA rm one_hotr? floatr r argmaxappendstackeqanyallrn rp ri r ) rI r{ r rJ featurespredicted_charsstructure_preds_liststructure_ids_list_embedding_featurestructure_stepstructure_predss r1 r| SLANeXtSLAHead.forward s ;;
#T[[%<%< =U]][h[o[o
++M,?,?,B+C5::^k^r^rs ! t{{22Q67A ! /;;;S;S T Z Z \ 77BUBUBWYjkKH ! 5 5h ?N , 3 3 3 :O ' ' 7 % %o 6{{-15889Q9QTU9UVZZ[]^bbdd 8 ))EKK0D!$LRT\a\i\ijmm
eer0 )r r r )r) r* r+ r, rS _can_record_outputsrE rX r r r rk r~ r r r r| r/ rP rQ s @r1 r r s - # t (,f ((f $f +, f ! fr0 r c j \ rS rSr% SrSr\R S- \S' Sr \R S- \S' Sr
g) SLANeXtForTableRecognitionOutputi aY
head_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Hidden-states of the SLANeXtSLAHead at each prediction step, varies up to max `self.config.max_text_length` states (depending on early exits).
head_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Attentions of the SLANeXtSLAHead at each prediction step, varies up to max `self.config.max_text_length` attentions (depending on early exits).
Nhead_hidden_stateshead_attentionsr( ) r) r* r+ r, rM r rk r~ r. r r/ r( r0 r1 r r s4 48))D0704OU&&-4r0 r z
SLANeXt Table Recognition model for table recognition tasks. Wraps the core SLANeXtPreTrainedModel
and returns outputs compatible with the Transformers table recognition API.
) custom_introc ^ \ rS rSrS\4U 4S jjr\\S\R S\
\ S\ \R \
- 4S j5 5 rSrU =r$ ) SLANeXtForTableRecognitioni r c | > [ TU ] U5 [ US9U l [ US9U l U R
5 g )N)r )rF rX r r r headr )rI r rK s r1 rX #SLANeXtForTableRecognition.__init__# s2
'v6
"&1 r0 r rJ returnc U R " U40 UD6nU R " UR 40 UD6n[ UR UR UR
UR UR
S9$ )N)r r{ r r r )r r r r r{ r )rI r rJ backbone_outputs head_outputss r1 r| "SLANeXtForTableRecognition.forward) sl
== @@yy!1!C!CNvN / * < <*88'22 + 9 9 ( 3 3
r0 )r r )r) r* r+ r, r8 rX r r rk r~ r r tupler r| r/ rP rQ s @r1 r r s` }
! - -
9?@R9S
u !$D D
r0 r )rk )backendsc ^ \ rS rSrSr\r\rSSS.r SSS.r
Sr Sr Sr
SrSrSSS\S S4S
jrS \S S \S\S
SS\S\S\S\S\S\\\ - S- S\\\ - S- S\S- S\S- S\S- S\\- S- S \4 S jrS \\ 4U 4S jjrS r S r S r U =r $ )!SLANeXtImageProcessori9 r r& )heightwidthTr z torch.Tensorr r c f
UR u p4pVUR X4- XV5 nUR n[ UR UR
5 [ XV5 - n[
XX- 5 n [
Xh- 5 n
[ R " U
[ R US9n U S- [ U5 [ U
5 - - S-
n U R 5 R [ R 5 n
XR 5 -
n[ R " U
S: [ R " U5 U5 n[ R " U
S: [ R " U
5 U
5 n
[ R " XS-
: [ R " U5 U5 n[ R " XS-
: [ R" " XS-
5 U
5 n
US- S- R 5 R [ R 5 nSU-
n[ R " U [ R US9nUS- [ U5 [ U 5 - - S-
nUR 5 R [ R 5 nUUR 5 -
n[ R " US: [ R " U5 U5 n[ R " US: [ R " U5 U5 n[ R " UUS-
: [ R " U5 U5 n[ R " UUS-
: [ R" " UUS-
5 U5 nUS- S- R 5 R [ R 5 nSU-
nUR% SS5 R [ R& 5 nUR [ R 5 nU
R) 5 nU
S- R) 5 nUR) 5 nUS- R) 5 n US S 2US S 2S 4 US S S 24 4 n US S 2US S 2S 4 US S S 24 4 n US S 2U S S 2S 4 US S S 24 4 n US S 2U S S 2S 4 US S S 24 4 n UR SU S5 n!UR SU S5 n"UR SSU
5 n#UR SSU
5 n$U"U$U - U#U - - - U!U$U - U#U - - - - n%U%S- S - n%U%R% SS5 R [ R& 5 n&U&R X4X5 R UR* S
9$ ) Nr g ?r rW r i i )ri )r viewr maxr r roundrk arangero r floorrp int32where
zeros_like ones_like full_likeclampuint8r ri )'rI r r
batch_sizechannelsr r r scale
target_height target_width
target_colsrc_col
src_col_floor src_col_frac weight_right weight_left
target_rowsrc_row
src_row_floor src_row_frac
weight_bottom
weight_top image_uint8 image_int32col_left col_rightrow_top
row_bottompixel_top_leftpixel_top_rightpixel_bottom_leftpixel_bottom_rightweight_bottom_3d
weight_top_3dweight_right_3dweight_left_3dinterpresults' r1 _resize SLANeXtImageProcessor._resizeG s
/4kk+
f
:0&@ DKK,s6/AA fn-
U]+ \\,emmFS
#e u\7J(JKcQ
**5;;7
!4!4!66 {{=1#4e6F6F|6TVbc MA$5u7G7G
7VXef
{{=AI#=u|?\^jk QY &
qy(QS`
%t+c188:==ekkJ \) \\-u}}VT
#f
m8L(LMPSS
**5;;7
!4!4!66 {{=1#4e6F6F|6TVbc MA$5u7G7G
7VXef
{{=FQJ#>P\@]_kl VaZ ' QR
)SUb
&,s299;>>u{{K
M)
kk!S),,U[[9 !nnU[[1 %%'"Q&,,. $$&#a'--/
$Q4(8(47:K%KL%aD)99T1W;M&MN ':ag+>q@Q(Q R (Jq$w,?4QR7AS)S T(--a B"=!<
&++Aq,?$))!Q = ^ +o.O O
1B B_WiEi i jk G$+ a%((5{{: MPPW\WbWbPccr0 images do_resizeresamplez"tvF.InterpolationMode | int | Nonedo_center_crop crop_size
do_rescalerescale_factor do_normalize
image_meanN image_stddo_padpad_sizedisable_groupingreturn_tensorsc Ub$ [ 5 ( d [ R S5 [ X S9u nn0 nUR 5 H" u nnU( a U R UUS9nUUU' M$ [
UU5 n[ UUS9u nn0 nUR 5 H8 u nnU( a U R UU5 nU R UXxXU 5 nUUU' M: [
UU5 nU ( a U R UXS9n[ SU0US9$ )Nz&Resampling is not supported in SLANeXt)r= )r r )r< r= r )data tensor_type) r logger warning_oncer itemsr/ r
center_croprescale_and_normalizepadr )rI r1 r2 r r3 r4 r5 r6 r7 r8 r9 r: r; r< r= r> rJ grouped_imagesgrouped_images_indexresized_images_groupedr stacked_imagesresized_imagesprocessed_images_groupedprocessed_imagess r1 _preprocess!SLANeXtImageProcessor._preprocess s/ & (@(B(B H I 0EV/o,,!# %3%9%9%; !E>!% N !N,: "5 ) &< ((>@TU 0E^fv/w,,#% %3%9%9%; !E> !%!1!1.)!L ! 7 7
LV_ N /= $U + &< **BDXY #xx(88x o .2B!CQ_``r0 rJ c F > [ TU ] " S0 UD6 U R 5 g rC )rF rX init_decoderrH s r1 rX SLANeXtImageProcessor.__init__ s
"6" r0 c / SQnU[ S5 Vs/ s H nSUS- S3PM sn-
nU[ S5 Vs/ s H nSUS- S3PM sn-
nSU; a UR S5 SU; a UR S5 S /U- S
/- n[ U5 VVs0 s H u p#X2_M snnU l Xl / S QU l U R S U l U R S
U l g s snf s snf s snnf )
a
Initialize the decoder vocabulary for table structure recognition.
Builds a character dictionary mapping HTML table structure tokens (e.g., ``, `
`, `
`, colspan/
rowspan attributes) to integer indices. The dictionary includes special `"sos"` (start-of-sequence) and
`"eos"` (end-of-sequence) tokens. Merged `
` tokens are used in place of standalone `
` tokens
when applicable.
)
zzz
zz
z
z
z
colspan="r "z
rowspan="
rU soseos)rU rV rZ N) r r remove enumeraterE charactertd_tokenbos_ideos_id)rI dict_characterichars r1 rR "SLANeXtImageProcessor.init_decoder s
%)D)QZAwa0)DD%)D)QZAwa0)DD n , ! !+ . ^ # ! !& ) >1UG;,5n,EF,ETW,EF '4
ii& ii& ED Gs C&C+C0c UR U l U R SS n[ U R 5 [ U R 5 /n[ U R 5 nUR SS9nUR
SS9R n/ nUR S n[ U5 H n/ n / n
[ UR S 5 H[ n [ XXU 4 5 n U S: a X:X a O@X; a M& U R U n
U R U
5 U
R X(U 4 5 M] UR U 5 [ R " U
5 R 5 R 5 nM / SQUS - / SQ- nUWS.$ )a
Post-process the raw model outputs to decode the predicted table structure into an HTML token sequence.
Converts the model's predicted probability distributions over the structure vocabulary into a sequence of
HTML tokens representing the table structure. The decoded tokens are wrapped with ``, ``, and
`
` tags to form a complete HTML table structure.
Args:
outputs ([`SLANeXtForTableRecognitionOutput`]):
Raw outputs from the SLANeXt model. The `last_hidden_state` field contains the predicted probability
distributions over the structure vocabulary at each decoding step, with shape
`(batch_size, max_text_length, num_classes)`.
Returns:
`dict`: A dictionary containing:
- **structure** (`list[str]`): The predicted HTML table structure as a list of tokens, wrapped with
``, ``, and `
` tags.
- **structure_score** (`float`): The mean confidence score across all predicted tokens.
r rW r r )zzz
)z
zz) structurestructure_score)r predr- ra rb r r valuesr r r_ r rk r meanitem)rI outputsstructure_probsignored_tokensend_idx
structure_idxstructure_str_listr batch_indexstructure_list
score_listpositionchar_idxtextri rh s r1 post_process_table_recognition4SLANeXtImageProcessor.post_process_table_recognition sn ( -- ))Aa. dkk*C ,<=dkk"'..1.5
)--!-4;; "((+
,K NJ !-"5"5a"8 9 }(-B CDar s ! 2 2 . & " 9 3 ; 2 E P P / - 4 l l C * 5 A
H % CD- E 3 CD ($ ( E (F +bii +B +A_ +A\ /
,
01f+ 1fh
5 5 5
!7
. : VL . VL VLr r0