RЦiUPDSrSSKJr SSKJr SSKJrJrJrJ r J r J r J r SSK r SSKJr SSKJrJr SSKJrJrJrJrJrJr SS KJr SS KJr SS KJr SS K J!r!J"r" SS K#J$r$J%r% S/r&\"\\RNS9r("SS\RR5r*S;S\+S\+S\,S\,S\+S\,S\\ \,\+\,\+44Sjjr-SS$\1S\04S1jj5r5\%S>S$\1S\04S2jj5r6\%S>S$\1S\04S3jj5r7\%S>S$\1S\04S4jj5r8\%S>S$\1S\04S5jj5r9\%S>S$\1S\04S6jj5r:\%S>S$\1S\04S7jj5r;\%S>S$\1S\04S8jj5r<\%S>S$\1S\04S9jj5r=\%S>S$\1S\04S:jj5r>g)?aReXNet A PyTorch impl of `ReXNet: Diminishing Representational Bottleneck on Convolutional Neural Network` - https://arxiv.org/abs/2007.00992 Adapted from original impl at https://github.com/clovaai/rexnet Copyright (c) 2020-present NAVER Corp. MIT license Changes for timm, feature extraction, and rounded channel variant hacked together by Ross Wightman Copyright 2020 Ross Wightman )partial)ceil)AnyDictListOptionalTupleTypeUnionNIMAGENET_DEFAULT_MEANIMAGENET_DEFAULT_STD)ClassifierHeadcreate_act_layer ConvNormActDropPathmake_divisibleSEModule)build_model_with_cfg)efficientnet_init_weights)feature_take_indices) checkpointcheckpoint_seq)generate_default_cfgsregister_modelRexNet) norm_layerc^\rSrSrSrSS\S\S\S\\\4S\S\S \S \S \S \ \ R4U4S jjjr SS\ S\4SjjrS\R S\R 4SjrSrU=r$)LinearBottleneck"zLinear bottleneck block for ReXNet. A mobile inverted residual bottleneck block as used in MobileNetV2 and subsequent models. in_chsout_chsstridedilation exp_ratiose_ratioch_div act_layer dw_act_layer drop_pathc >XS.n [TU]5 US:H=(a USUS:H=(a X:*UlXlX lUS:wa)[ [ X-5US9n[X4SU0U D6UlO UnSUl[UU4SUUSUS S .U D6Ul US:a)[U4S [ [X-5U50U D6Ul OSUl [U 5Ul[XS4S S 0U D6UlXlg) aInitialize LinearBottleneck. Args: in_chs: Number of input channels. out_chs: Number of output channels. stride: Stride for depthwise conv. dilation: Dilation rates. exp_ratio: Expansion ratio. se_ratio: Squeeze-excitation ratio. ch_div: Channel divisor. act_layer: Activation layer for expansion. dw_act_layer: Activation layer for depthwise. drop_path: Drop path module. devicedtyperr?divisorr)NF) kernel_sizer$r%groups apply_act rd_channelsr6)super__init__ use_shortcut in_channels out_channelsrroundrconv_expconv_dw SEWithNormintseract_dwconv_pwlr+)selfr"r#r$r%r&r'r(r)r*r+r.r/dddw_chs __class__s Q/var/www/html/ai-image-ml/venv/lib/python3.13/site-packages/timm/models/rexnet.pyr9LinearBottleneck.__init__(s: / "aK\HQK8A;,F\6K\!# ?#E&*<$=vNF'R)RrRDMF DM"    a[      a< j^CHYDZ\b5cjgijDGDG&|4 #FQN%N2N "expreturncTU(aURR$UR$)zwGet feature channel count. Args: exp: Return expanded channels if True. Returns: Number of feature channels. )r?r<)rErLs rI feat_channelsLinearBottleneck.feat_channelses!-0t||((FT5F5FFrKxcUnURbURU5nURU5nURbURU5nURU5nUR U5nUR (a^UR bUR U5n[R"USS2SUR24U-USS2URS24/SS9nU$)zGForward pass. Args: x: Input tensor. Returns: Output tensor. Nrr)dim) r>r?rBrCrDr:r+torchcatr;)rErQshortcuts rIforwardLinearBottleneck.forwardps == $ a A LLO 77  A KKN MM!    ~~)NN1% 1Q$"2"2 223h>!TEUEUEVBV@WX^_`ArK) rCr?r>rDr+r;r<rBr:) )rrr0rswishrelu6NNNF)__name__ __module__ __qualname____firstlineno____doc__rAr floatstrrnnModuler9boolrOrTTensorrW__static_attributes__ __classcell__rHs@rIr r "s)/" $ '-1;#;#;# ;# CHo ;#  ;#;#;#;#;#  *;#;#z G G# G%,,rKr r0 width_mult depth_mult initial_chs final_chsr'r(rMc @/SQn/SQnUVs/sHn[X-5PM nn[[U5V Vs/sHupU/S/XiS- --PM snn /5nS/US-S/[USS5--n [USS5S-n US:aX - OUn /n [U S-5H4nU R [ [ X-5US 95 XU S-S-- - n M6 S /USUS--U/[US S5--n[[XX55$s snfs snn f) aKGenerate ReXNet block configuration. Args: width_mult: Width multiplier. depth_mult: Depth multiplier. initial_chs: Initial channel count. final_chs: Final channel count. se_ratio: Squeeze-excitation ratio. ch_div: Channel divisor. Returns: List of tuples (out_channels, exp_ratio, stride, se_ratio). )rrpr3r3)rrprprprrprrNr3r0r1rYrp) rsum enumeraterangeappendrr=listzip)rkrlrmrnr'r(layersstrideselementidx exp_ratiosdepthbase_chs out_chs_listi se_ratioss rI _block_cfgrsF* F G8> ?Wd7'(F ?9U\K]^K]<3G9sfkAo66K]^`bcGvayA3VABZ#88J q NQ E+5+;{'HL 5A: N51F+GQWXY!c!122q F1I-.(c&*o1MMI LgA BB@^s DD block_cfgprev_chs output_strider)r*drop_path_ratec <XS.n U/n /n Sn Sn/n[U5n[U5HununnnnUnUS:a.US:XaSOSUS- 3nU [U SU US9/- n X:aUU-nSnUU-US- - nUS :a [U5OS nUR [ SUUUUUU4UUUUUS . U D65 U U-n UnUnXSR 5/- n M [S U-US 9nU [U SU S[U5S- 3S9/- n UR [UU4SU0U D65 X4$)aBuild ReXNet blocks from configuration. Args: block_cfg: Block configuration list. prev_chs: Previous channel count. width_mult: Width multiplier. ch_div: Channel divisor. output_stride: Target output stride. act_layer: Activation layer name. dw_act_layer: Depthwise activation layer name. drop_path_rate: Drop path rate. Returns: Tuple of (features list, feature_info list). r-rprrstemz features.)num_chs reductionmodulerYN) r"r#r&r$r%r'r(r)r*r+ir1r)) lenrtdictrrvr rOrr)rrrkr(rr)r*rr.r/rFfeat_chs feature_info curr_strider%features num_blocks block_idxchsr&r$r' next_dilationfname block_dprr+pen_chss rI _build_blocksrs6 +BzHLKHHYJ9B99M5 5CFH A:'1nFIi!m_2ME T(2,+V[\] ]L+ (6 1 "Y.*q.A +4r>HY't (   .%     v  b\//1225:N6TJ.?GT(2,+PYZ]^fZgjkZkYlNmnooL OOK'MYM"MN  !!rKcD^\rSrSrSrS)S\S\S\S\S\S \S \S \S \S \S\S\S\S\4U4Sjjjr\ RRS*S\ S\ \\44Sjj5r\ RRS+S\ SS4Sjj5r\ RRS\R$4Sj5rS,S\S\\SS4SjjrS-S\ R,S\\\\\4S\ S\ S\S\ S\\\ R,\\ R,\\ R,444S jjrS.S\\\\4S!\ S"\ S\\4S#jjrS\ R,S\ R,4S$jrS*S\ R,S%\ S\ R,4S&jjrS\ R,S\ R,4S'jrS(rU=r $)/rzReXNet model architecture. Based on `ReXNet: Diminishing Representational Bottleneck on Convolutional Neural Network` - https://arxiv.org/abs/2007.00992 Nin_chans num_classes global_poolrrmrnrkrlr'r(r)r* drop_raterc >[TU]5 UUS.nX lXlXlSUlUS;deUS:aSU- OSn[ [UU-5U S9n[UUS4SU S .UD6Ul [XxXVX5n[UUUU UU U U40UD6unUl US R=UlUl[ R""U6Ul['URX#U 40UD6Ul[+U5 g ) a0Initialize ReXNet. Args: in_chans: Number of input channels. num_classes: Number of classes for classification. global_pool: Global pooling type. output_stride: Output stride. initial_chs: Initial channel count. final_chs: Final channel count. width_mult: Width multiplier. depth_mult: Depth multiplier. se_ratio: Squeeze-excitation ratio. ch_div: Channel divisor. act_layer: Activation layer name. dw_act_layer: Depthwise activation layer name. drop_rate: Dropout rate. drop_path_rate: Drop path rate. r-F) r0rr1r3rp)r$r)rN)r8r9rrrgrad_checkpointingrr=rrrrrr< num_featureshead_hidden_sizerd Sequentialrrheadr)rErrrrrmrnrkrlr'r(r)r*rrr.r/rF stem_base_chsstem_chsrrrHs rIr9RexNet.__init__sJ  /& ""' ++++5+;Z !% (B"CVT(A[a9[XZ[ z{x` &3         '  ' #$#5=RL4M4MMD1 x0 "4#4#4kPY`]_` !$'rKcoarserMc[SSS9nU$)zGroup matcher for parameter groups. Args: coarse: Whether to use coarse grouping. Returns: Dictionary of grouped parameters. z^stemz^features\.(\d+))rblocks)r)rErmatchers rI group_matcherRexNet.group_matcher>s& rKenablecXlg)zgEnable or disable gradient checkpointing. Args: enable: Whether to enable gradient checkpointing. N)r)rErs rIset_grad_checkpointingRexNet.set_grad_checkpointingNs #)rKc.URR$)z DIIDYDYDcDcI&t'8'8+RVR`R`gdfgDI IIOOK 5rKrQindicesnorm stop_early output_fmtintermediates_onlycUS;dS5e/nURVs/sH#n[USRS5S5PM% n n[[ U 5U5upU V s/sHoU PM n n Xn UR U5n[ RR5(dU(d URn OURSU S-n [U 5HeupUR(a/[ RR5(d [X5nOU"U5nX;dMTURU5 Mg U(aU$X4$s snfs sn f)aForward features that returns intermediates. Args: x: Input image tensor indices: Take last n blocks if int, all if None, select matching indices if sequence norm: Apply norm layer to compatible intermediates stop_early: Stop iterating over blocks when last desired intermediate hit output_fmt: Shape of intermediate feature outputs intermediates_only: Only return intermediate features Returns: )NCHWzOutput shape must be NCHW.r.rNr)rrAsplitrrrrTjit is_scriptingrrtrrrv)rErQrrrrr intermediatesinfo stage_ends take_indices max_indexrstagesfeat_idxstages rIforward_intermediatesRexNet.forward_intermediatesos4*Y&D(DD& EIEVEVWEVTc$x...s3B78EV W"6s:"P /;<|!1 | <)  IIaL 99 ! ! # #:]]F]]>IM2F(0OH&&uyy/E/E/G/Gu(!H'$$Q' 1  /X*  c 0%,,5<< Sell S S S  %,,  rKvariant pretrainedc <[SS9n[[UU4SU0UD6$)zCreate a ReXNet model. Args: variant: Model variant name. pretrained: Load pretrained weights. **kwargs: Additional model arguments. Returns: ReXNet model instance. T)flatten_sequential feature_cfg)rrr)rrkwargsrs rI_create_rexnetrs8$/K       rKurlc 2USSSSS[[SSSS . UE$) zCreate default configuration dictionary. Args: url: Model weight URL. **kwargs: Additional configuration options. Returns: Configuration dictionary. r)r3r)rg?bicubicz stem.convzhead.fcmit) rr input_size pool_sizecrop_pct interpolationmeanstd first_conv classifierlicenser )rrs rI_cfgrs54}SYI%.B!   #  rKztimm/) hf_hub_idgffffff?)r3 rz apache-2.0)rr test_crop_pcttest_input_sizeri-.)rrrrrr) zrexnet_100.nav_in1kzrexnet_130.nav_in1kzrexnet_150.nav_in1kzrexnet_200.nav_in1kzrexnet_300.nav_in1kzrexnetr_100.untrainedzrexnetr_130.untrainedzrexnetr_150.untrainedzrexnetr_200.sw_in12k_ft_in1kzrexnetr_300.sw_in12k_ft_in1kzrexnetr_200.sw_in12kzrexnetr_300.sw_in12kc [SU40UD6$)zReXNet V1 1.0x rexnet_100rrrs rIrr#s , =f ==rKc [SU4SS0UD6$)zReXNet V1 1.3x rexnet_130rk?rr s rIr r ) , Ms Mf MMrKc [SU4SS0UD6$)zReXNet V1 1.5x rexnet_150rk?rr s rIrr/r rKc [SU4SS0UD6$)zReXNet V1 2.0x rexnet_200rk@rr s rIrr5r rKc [SU4SS0UD6$)zReXNet V1 3.0x rexnet_300rk@rr s rIrr;r rKc [SU4SS0UD6$)z*ReXNet V1 1.0x w/ rounded (mod 8) channels rexnetr_100r(rrr s rIrrAs - HA H HHrKc "[SU4SSS.UD6$)z*ReXNet V1 1.3x w/ rounded (mod 8) channels rexnetr_130r rrkr(rr s rIrrG - XA XQW XXrKc "[SU4SSS.UD6$)z*ReXNet V1 1.5x w/ rounded (mod 8) channels rexnetr_150rrrrr s rIrrMrrKc "[SU4SSS.UD6$)z*ReXNet V1 2.0x w/ rounded (mod 8) channels rexnetr_200rrrrr s rIr r SrrKc "[SU4SSS.UD6$)z+ReXNet V1 3.0x w/ rounded (mod 16) channels rexnetr_300rrrrr s rIr"r"Ys - YB YRX YYrK)r0r0rrrYr)rrrZr[rYNN)rr\)?ra functoolsrmathrtypingrrrrr r r rTtorch.nnrd timm.datar r timm.layersrrrrrr_builderr_efficientnet_builderr _featuresr _manipulaterr _registryrr__all__ BatchNorm2dr@rer rbrArrcrrrfrr default_cfgsrr rrrrrrr r"rrKrIr1s @@@ Aii*<+3< *X".. 9 cryycN  %C%C%C%C %C  %C  %C %UC& '(%CX # "@"c5#u456@"@"@" @"  @"  @"@"@" 4 ?Dc3h0 01@"FkRYYk\CT*cT#s(^&%'2'2'2'2'2!V!V!V$(S-Q]%_%)S-Q]%_!S-Q]_!S-Q]_'& 4>4>f>> N4NfNN N4NfNN N4NfNN N4NfNN IDIvII YDYvYY YDYvYY YDYvYY ZDZvZZrK