khMBdZddlmZddlmZddlmZmZmZm Z m Z m Z ddl Z ddl mZddlmZmZddlmZmZmZmZmZmZdd lmZdd lmZdd lmZdd lm Z m!Z!dd l"m#Z#m$Z$dgZ%eeejLZ'GddejPZ) d;de*de*de+de+de*de+dee e+e*e+e*ffdZ, dd$e0de/fd1Z4e$d>d$e0de/fd2Z5e$d>d$e0de/fd3Z6e$d>d$e0de/fd4Z7e$d>d$e0de/fd5Z8e$d>d$e0de/fd6Z9e$d>d$e0de/fd7Z:e$d>d$e0de/fd8Z;e$d>d$e0de/fd9Zd$e0de/fd:Z=y)?a ReXNet 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)AnyDictListOptionalTupleUnionNIMAGENET_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_layerceZdZdZ ddedededeeefdededed ed ed ee jffd Z dd e defdZ dejdejfdZxZS)LinearBottleneckzLinear 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 tt| |dk(xr|d|dk(xr||k|_||_||_|dk7r-t t||z|} t|| ||_ n |} d|_ t| | d||d| d |_ |dkDr)t| t t| |z| |_ nd|_ t| |_t| |dd |_| |_y) 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. rr?divisorr'NF) kernel_sizer"r#groups apply_act) rd_channels)r2)superr__init__ use_shortcut in_channels out_channelsrroundrconv_expconv_dw SEWithNormintseract_dwconv_pwlr)) selfr r!r"r#r$r%r&r'r(r)dw_chs __class__s N/var/www/teggl/fontify/venv/lib/python3.12/site-packages/timm/models/rexnet.pyr5zLinearBottleneck.__init__(s6 .0"aK\HQK8A;,F\6U\K\!# ?#E&9*<$=vNF')LDMF DM"  a[   a< ^CQYHYDZ\b5cdDGDG&|4 #FGQ%H "expreturncJ|r|jjS|jS)zGet feature channel count. Args: exp: Return expanded channels if True. Returns: Number of feature channels. )r;r8)rArFs rD feat_channelszLinearBottleneck.feat_channelsas!-0t||((FT5F5FFrExc|}|j|j|}|j|}|j|j|}|j|}|j |}|j r_|j |j |}tj|ddd|jf|z|dd|jdfgd}|S)zoForward pass. Args: x: Input tensor. Returns: Output tensor. Nrr)dim) r:r;r>r?r@r6r)torchcatr7)rArJshortcuts rDforwardzLinearBottleneck.forwardls == $ a A LLO 77  A KKN MM!    ~~)NN1% 1Q$"2"2 223h>!TEUEUEVBV@WX^_`ArE))rrr+rswishrelu6NF)__name__ __module__ __qualname____doc__r=r floatstrrnnModuler5boolrIrMTensorrP __classcell__rCs@rDrr"s)/" $ '-17#7#7# 7# CHo 7#  7#7#7#7#7#  *7#r G G# G%,,rErr+ width_mult depth_mult initial_chs final_chsr%r&rGc Dgd}gd}|Dcgc]}t||z}}tt|D cgc]\} }|gdg|| dz zzc}} g}dg|dzdgt|ddzz} t|dddz} |dkr||z n|} g} t| dzD]8}| j t t | |z| | || dzdzz z } :d g|d|dzz|gt|d dzz}tt| | ||Scc}wcc}} w) asGenerate 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). )rrfr/r/)rrfrfrfrrfrrNr/r+r,rQrf) rsum enumeraterangeappendrr9listzip)rarbrcrdr%r&layersstrideselementidx exp_ratiosdepthbase_chs out_chs_listi se_ratioss rD _block_cfgrysT* F G8> ?Wd7Z'( ?F ?9U\K]^<3G9sfSkAo66^`bcGvayA3VABZ#88J q NQ E+5+;{Z'HL 5A: 3N5J1F+GQWXYI!c!1223q F1I-.(c&*o1MMI L*gyA BB@^s DD block_cfgprev_chs output_strider'r(drop_path_ratec<|g}g} d} d} g} t|} t|D]\}\}}}}| }|dkDr0|dk(rdnd|dz }| t|d| |gz } | |k\r| |z}d}||z| dz z }|dkDr t|nd }| j t ||||| |f||||| | |z} |} |}|| dj gz }td |z| }| t|d| dt| dz gz } | j t||| | | fS)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). rfrrstemz features.)num_chs reductionmodulerQN) r r!r$r"r#r%r&r'r(r)ir,r.) lenrjdictrrlrrIrr)rzr{rar&r|r'r(r}feat_chs feature_info curr_strider#features num_blocks block_idxchsr$r"r% next_dilationfname block_dprr)pen_chss rD _build_blocksrs2zHLKHHYJ9B99M35 5CFH A:'1nFIi!m_2ME T(2,+V[\] ]Lm+ (6 1 "Y.*q.A +4r>HY't ( .%   v  Xb\//122334TJ.?GT(2,+PYZ]^fZgjkZkYlNmnooL OOK'YGH \ !!rEceZdZdZ d'dededededededed ed ed ed ed ededeffd Zejjd(de de ee ffdZejjd)de ddfdZejjdej"fdZd*dedeeddfdZ d+dej*deeeeefde de dede deeej*eej*eej*fffdZ d,deeeefd e d!e deefd"Zdej*dej*fd#Zd(dej*d$e dej*fd%Zdej*dej*fd&ZxZS)-rzReXNet model architecture. Based on `ReXNet: Diminishing Representational Bottleneck on Convolutional Neural Network` - https://arxiv.org/abs/2007.00992 in_chans num_classes global_poolr|rcrdrarbr%r&r'r( drop_rater}c tt| ||_| |_d|_|dvsJ|dkrd|z nd}t t||z| }t||dd| |_ t||||| | }t|||| || | |\}|_ |d jx|_|_t!j"||_t'|j||| |_t+|y ) aInitialize ReXNet. Args: in_chans: Number of input channels. num_classes: Number of classes for classification. global_pool: Global pooling type. output_stride: Output stride. initial_chs: Initial channel count. final_chs: Final channel count. width_mult: Width multiplier. depth_mult: Depth multiplier. se_ratio: Squeeze-excitation ratio. ch_div: Channel divisor. act_layer: Activation layer name. dw_act_layer: Depthwise activation layer name. drop_rate: Dropout rate. drop_path_rate: Drop path rate. F) r+rr,r/rf)r"r'rN)r4rr5rrgrad_checkpointingrr9rrryrrr8 num_featureshead_hidden_sizer[ Sequentialrrheadr)rArrrr|rcrdrarbr%r&r'r(rr} stem_base_chsstem_chsrzrrCs rDr5zRexNet.__init__sF fd$&&""' ++++5+;Z !% (B"CVT(Aa9U z:{IxY_` &3         ' #$#5=RL4M4MMD1 x0 "4#4#4k;PYZ !$'rEcoarserGc tdd}|S)zGroup matcher for parameter groups. Args: coarse: Whether to use coarse grouping. Returns: Dictionary of grouped parameters. z^stemz^features\.(\d+))rblocks)r)rArmatchers rD group_matcherzRexNet.group_matcher1s& rEenableNc||_y)zEnable or disable gradient checkpointing. Args: enable: Whether to enable gradient checkpointing. N)r)rArs rDset_grad_checkpointingzRexNet.set_grad_checkpointingAs #)rEc.|jjS)zTGet the classifier module. Returns: Classifier module. )rfc)rAs rDget_classifierzRexNet.get_classifierJsyy||rEcJ||_|jj||y)zReset the classifier. Args: num_classes: Number of classes for new classifier. global_pool: Global pooling type. N)rrreset)rArrs rDreset_classifierzRexNet.reset_classifierSs'  [1rErJindicesnorm stop_early output_fmtintermediates_onlyc||dvsJdg}|jDcgc]"}t|djdd$} }tt | |\} } | D cgc]} | |  } } | | } |j |}t jjs|s |j} n|jd| dz} t| D]Z\}}|jr+t jjs t||}n||}|| vsJ|j|\|r|S||fScc}wcc} w)a Forward features that returns intermediates. Args: x: Input image tensor indices: Take last n blocks if int, all if None, select matching indices if sequence norm: Apply norm layer to compatible intermediates stop_early: Stop iterating over blocks when last desired intermediate hit output_fmt: Shape of intermediate feature outputs intermediates_only: Only return intermediate features Returns: )NCHWzOutput shape must be NCHW.r.rNr)rr=splitrrrrMjit is_scriptingrrjrrrl)rArJrrrrr intermediatesinfo stage_ends take_indices max_indexrwstagesfeat_idxstages rDforward_intermediateszRexNet.forward_intermediates]s=*Y&D(DD& EIEVEVWTc$x...s3B78W W"6s:"P i/;IM2F(0 (OHe&&uyy/E/E/Gua(!H<'$$Q'  (  -/X*  c 0%,,5<< Sell S S S  %,, rEvariant pretrainedc Btd}tt||fd|i|S)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 rD_create_rexnetrs8$/K       rEurlc 2|dddddttdddd |S) zCreate default configuration dictionary. Args: url: Model weight URL. **kwargs: Additional configuration options. Returns: Configuration dictionary. r)r/r)rg?bicubicz stem.convzhead.fcmit) rr input_size pool_sizecrop_pct interpolationmeanstd first_conv classifierlicenser )rrs rD_cfgrs54}SYI%.B!   #  rEztimm/) hf_hub_idgffffff?)r/ rz apache-2.0)rr test_crop_pcttest_input_sizeri-.)rrrrrr) zrexnet_100.nav_in1kzrexnet_130.nav_in1kzrexnet_150.nav_in1kzrexnet_200.nav_in1kzrexnet_300.nav_in1kzrexnetr_100.untrainedzrexnetr_130.untrainedzrexnetr_150.untrainedzrexnetr_200.sw_in12k_ft_in1kzrexnetr_300.sw_in12k_ft_in1kzrexnetr_200.sw_in12kzrexnetr_300.sw_in12kc td|fi|S)zReXNet V1 1.0x rexnet_100rrrs rDrrs , =f ==rEc td|fddi|S)zReXNet V1 1.3x rexnet_130ra?rrs rDrr , Ms Mf MMrEc td|fddi|S)zReXNet V1 1.5x rexnet_150ra?rrs rDrrrrEc td|fddi|S)zReXNet V1 2.0x rexnet_200ra@rrs rDrr#rrEc td|fddi|S)zReXNet V1 3.0x rexnet_300ra@rrs rDrr)rrEc td|fddi|S)z*ReXNet V1 1.0x w/ rounded (mod 8) channels rexnetr_100r&rrrs rDrr/s - HA H HHrEc "td|fddd|S)z*ReXNet V1 1.3x w/ rounded (mod 8) channels rexnetr_130rrrar&rrs rDrr5 - XA XQW XXrEc "td|fddd|S)z*ReXNet V1 1.5x w/ rounded (mod 8) channels rexnetr_150rrrrrs rDrr;rrEc "td|fddd|S)z*ReXNet V1 2.0x w/ rounded (mod 8) channels rexnetr_200rrrrrs rDr r ArrEc "td|fddd|S)z+ReXNet V1 3.0x w/ rounded (mod 16) channels rexnetr_300rrrrrs rDr r Gs - YB YRX YYrE)r+r+rrrQr)rrrRrSrQ)rrT)>rX functoolsrmathrtypingrrrrr r rMtorch.nnr[ timm.datar r timm.layersrrrrrr_builderr_efficientnet_builderr _featuresr _manipulaterr _registryrr__all__ BatchNorm2dr<r\rrYr=ryrZrrr]rr default_cfgsrrrrrrrrr r rErDrs :: Aii*<+3< *X".. 9 _ryy_F  %C%C%C%C %C  %C  %C %UC& '(%CX # "<"c5#u456<"<"<" <"  <"  <"<"<" 4 ?Dc3h0 01<"~aRYYaHCT*cT#s(^&%'2'2'2'2'2!V!V!V$(S-Q]%_%)S-Q]%_!S-Q]_!S-Q]_'& 4>4>f>> N4NfNN N4NfNN N4NfNN N4NfNN IDIvII YDYvYY YDYvYY YDYvYY ZDZvZZrE