khdZddlmZmZmZmZmZmZmZddl Z ddl m Z ddl m Z m Z  ddlmZddlmZd Zdd lmZmZmZmZmZdd lmZmZmZdd lmZddl m!Z!m"Z"ddl#m$Z$m%Z%m&Z&m'Z'm(Z(m)Z)m*Z*gdZ+eee!eee"ee$e%e&e'e(e)e*hZ, ddlm-Z-e,j]e-dee j^fdZ0dee j^fdZ1dZ2eeehZ3defdZ4defdZ5dZ6de j^deee7ee7ffdZde j^deee7e7fee7ffdZGdd e j^Z8Gd!d"e j^Z9y#e$rd ZYwxYw#e$rYwxYw)#zm PyTorch FX Based Feature Extraction Helpers Using https://pytorch.org/vision/stable/feature_extraction.html )CallableDictListOptionalUnionTupleTypeN)nn)_get_feature_info_get_return_layers)create_feature_extractor)get_graph_node_namesTF) Conv2dSameScaledStdConv2dSame CondConv2d StdConv2dSameFormat)resample_abs_pos_embedresample_abs_pos_embed_nhwcmaybe_add_mask)BilinearAttnTransform) MaxPool2dSame AvgPool2dSame)BatchNormAct2dSyncBatchNormActFrozenBatchNormAct2d GroupNormAct GroupNorm1Act LayerNormActLayerNormAct2d) register_notrace_moduleis_notrace_moduleget_notrace_modulesregister_notrace_functionis_notrace_functionget_notrace_functionsrrFeatureGraphNetGraphExtractNet) InplaceAbnmodulec0tj||S)zq Any module not under timm.models.layers should get this decorator if we don't want to trace through it. ) _leaf_modulesaddr+s T/var/www/teggl/fontify/venv/lib/python3.12/site-packages/timm/models/_features_fx.pyr"r">sf Mc|tvSN)r-r/s r0r#r#Fs ] ""r1c ttSr3)listr-r1r0r$r$Js  r1funcc0tj||S)zF Decorator for functions which ought not to be traced through )_autowrap_functionsr.r7s r0r%r%VsD! Kr1c|tvSr3)r9r:s r0r&r&^s & &&r1c ttSr3)r5r9r6r1r0r'r'bs # $$r1modelreturncVt|ttttdS)N leaf_modulesautowrap_functions tracer_kwargs)_get_graph_node_namesr5r-r9)r=s r0rrfs&  '+M':RVWjRkl r1 return_nodescrtsJdt||ttttdS)NKPlease update to PyTorch 1.10+, torchvision 0.11+ for FX feature extractionr@rC)has_fx_feature_extraction_create_feature_extractorr5r-r9)r=rFs r0rrms7 $s&ss $ $ |'+M':RVWjRkl r1c eZdZUdZej j eed< d de jde e dfde ededef fd Zd ZxZS) r(zX A FX Graph based feature extractor that works with the model feature_info metadata return_dictr= out_indices.out_map output_fmtct|tsJdt|||_|t |t |k(sJt ||_t|j|}t|||_ ||_ y)NrH) super__init__rIr feature_infolenrrOr r graph_modulerL)selfr=rMrNrOrLrF __class__s r0rRzFeatureGraphNet.__init__zs (w*ww(-e[A  w<3{#33 33 ,)$*;*;WE 4ULI&r1cr|j|}|jr|St|jSr3)rUrLr5valuesrVxouts r0forwardzFeatureGraphNet.forwards1"   JCJJL!!r1)NNCHWF)__name__ __module__ __qualname____doc__torchjitFinalbool__annotations__r ModulerintrrstrrRr] __classcell__rWs@r0r(r(usp&& '+$ % '99'sCx'd^ '  '  '$"r1r(c eZdZUdZej j eed< d de jde e e e fee fdedeffd Zde eej ej ffdZxZS) r)a A standalone feature extraction wrapper that maps dict -> list or single tensor NOTE: * one can use feature_extractor directly if dictionary output is desired * unlike FeatureGraphNet, this is intended to be used standalone and not with model feature_info metadata for builtin feature extraction mode * create_feature_extractor can be used directly if dictionary output is acceptable Args: model: model to extract features from return_nodes: node names to return features from (dict or list) squeeze_out: if only one output, and output in list format, flatten to single tensor return_dict: return as dictionary from extractor with node names as keys, ignores squeeze_out arg rLr=rF squeeze_outc`t|||_t|||_||_yr3)rQrRrnrrUrL)rVr=rFrnrLrWs r0rRzGraphExtractNet.__init__s/ &4ULI&r1r>c|j|}|jr|St|j}|jrt |dk(r|dS|S)Nr r)rUrLr5rYrnrTrZs r0r]zGraphExtractNet.forwardsQ"   J3::< ))c#h!ms1vDDr1)TF)r_r`rarbrcrdrerfrgr rhrrrjrrRTensorr]rkrls@r0r)r)s && !% % '99 ' S#XS 9: ' '  'EE$u||"4ell"BCEr1r)):rbtypingrrrrrrr rcr _featuresr r %torchvision.models.feature_extractionrrJrrErI ImportError timm.layersrrrrrrrrtimm.layers.non_local_attnrtimm.layers.pool2d_samerrtimm.layers.norm_actrrrrrr r!__all__r-r*r.rhr"r#r$r9r%r&r'rjr(r)r6r1r0r{sFEE <&kc $ [Z[[<@ e 2M=   &j! DO#d299o# H'h'% eDItCy4H.IBIIU4S>SWX[S\C\=]"bii"E >E E  EE