NgLM ,dZddlmZmZddlmZddlmZddlm Z m Z m Z m Z m Z mZddlZddlmZddlmZddlmZmZgd Z d#d ed e eee efd ede e eeffdZdeee edffde edffdZeee edffZGddZGddZd$dZdefdZ dZ!Gddej"Z#Gdde#Z$Gdd ej"Z%Gd!d"ej"Z&dS)%a PyTorch Feature Extraction Helpers A collection of classes, functions, modules to help extract features from models and provide a common interface for describing them. The return_layers, module re-writing idea inspired by torchvision IntermediateLayerGetter https://github.com/pytorch/vision/blob/d88d8961ae51507d0cb680329d985b1488b1b76b/torchvision/models/_utils.py Hacked together by / Copyright 2020 Ross Wightman ) OrderedDict defaultdict)deepcopy)partial)DictListOptionalSequenceTupleUnionN) checkpoint)Format_assert) FeatureInfo FeatureHooksFeatureDictNetFeatureListNetFeatureHookNetFeatureGetterNetfeature_take_indicesF num_featuresindicesas_setreturnc ttrBtdcxkokncdddfdtD}nQg}D]L}|dkr|zn|}td|cxkokncd|dd z d||Mt js |rt|t|fS|t|fS) aB Determine the absolute feature indices to 'take' from. Note: This function can be called in forwar() so must be torchscript compatible, which requires some incomplete typing and workaround hacks. Args: num_features: total number of features to select from indices: indices to select, None -> select all int -> select last n list/tuple of int -> return specified (-ve indices specify from end) as_set: return as a set Returns: List (or set) of absolute (from beginning) indices, Maximum index Nrzlast-n (z) is out of range (1 to )c g|] }z |z Sr).0irrs Q/var/www/html/ai-engine/env/lib/python3.11/site-packages/timm/models/_features.py z(feature_take_indices..8s#KKKq w.2KKKzfeature index z is out of range (0 to ) isinstanceintrrangeappendtorchjit is_scriptingsetmax)rrr take_indicesr idxs`` r!rrs[*'3 %G++++|++++-h-h-hYe-h-h-hiiiKKKKKE'NNKKK "$  % %A&'!ee,""C A++++|++++-mc-m-mZfijZj-m-m-m n n n    $ $ $ $ 9 ! ! # #44<  #l"3"333 \** **r#x.ct|trtt| dSt|S)Nr)r%r&tupler')r0s r!_out_indices_as_tupler3Fs7!S#UA2q\\""" 88Or#c peZdZdeedefdZdefdZddede e e ee ffdZ dd e eede e e ee ffd Z dde e e ee ffd Zdde e e ee ffd Zdde e e ee ffd ZdZdZdS)r feature_info out_indicesct|}d}t|D]M\}}d|vr |ddksJd|vr |d|ksJ|d}d|vsJ|d|N||_||_dS)Nr$num_chsr reductionmoduleindex)r3 enumerate setdefaultr6info)selfr5r6prev_reductionr fis r!__init__zFeatureInfo.__init__Rs ,K88 |,, & &EAr??r)}q'8'8'8'8"$$KN)J)J)J)J _Nr>>>> MM'1 % % % %&  r#cdt|}tt|j|SN)r3rrr>)r?r6s r! from_otherzFeatureInfo.from_othercs)+K88 8DI.. <<rr rFr?s r!r"z#FeatureInfo.get..ns$@@@!DIaL%@@@r#c6g|]}j|SrrIrJs r!r"z#FeatureInfo.get..ps$333!DIaL%333r#r6r%r2listr>)r?rFr/s`` r!getzFeatureInfo.getgso ;@@@@@t/?@@@ @ cE4= ) ) '33333s333 39S>#& &r#keysc)fdjDSfdjDStttfrfdDS jnfdDS)zm return info dicts for specified keys (or all if None) at specified indices (or out_indices if None) Nc*g|]}j|SrrI)rr r?s r!r"z)FeatureInfo.get_dicts..ys??? ! ???r#c0g|]fdDS)c8i|]}|j|SrrIrkr r?s r! z4FeatureInfo.get_dicts...{s%:::DIaLO:::r#rrr rOr?s @r!r"z)FeatureInfo.get_dicts..{s2UUUq:::::T:::UUUr#cNg|]  jnfdD!S)Nc8i|]}|j|SrrIrTs r!rVz4FeatureInfo.get_dicts...}s%6X6X6Xaq$)A,q/6X6X6Xr#rIrWs @r!r"z)FeatureInfo.get_dicts..}sBfff]^DLDIaLL6X6X6X6X6XSW6X6X6Xfffr#c8i|]}|j|SrrI)rrUr/r?s r!rVz)FeatureInfo.get_dicts..s'7[7[7[QR49S>!;L7[7[7[r#rL)r?rOr/s```r! get_dictszFeatureInfo.get_dictsts ;|????d.>????UUUUUDDTUUUU cE4= ) ) \fffffbefff f%)\49S>>7[7[7[7[7[VZ7[7[7[ [r#c.|d|S)z# feature channels accessor r8rNr?r/s r!channelszFeatureInfo.channelssxx 3'''r#c.|d|S)z4 feature reduction (output stride) accessor r9r]r^s r!r9zFeatureInfo.reductionsxx S)))r#c.|d|S)z& feature module name accessor r:r]r^s r! module_namezFeatureInfo.module_namesxx#&&&r#c|j|SrDrI)r?items r! __getitem__zFeatureInfo.__getitem__syr#c*t|jSrD)lenr>)r?s r!__len__zFeatureInfo.__len__s49~~r#rD)NN)__name__ __module__ __qualname__rr OutIndicesTrBrEstrr r r&rNr[r_r9rbrerhrr#r!rrPs!t*!%!!!!"=k==== ' 's '%T#Y*?!@ ' ' ' ' \ \htCy1 \xcSWX[S\nH]?^ \ \ \ \((HU3S >%:;(((( **XeCcN&;<**** ''xc49n(=>'''' r#rc eZdZdZ d deeeefdedeee efdefdZ d Z d eee j ffd ZdS) rz Feature Hook Helper This module helps with the setup and extraction of hooks for extracting features from internal nodes in a model by node name. FIXME This works well in eager Python but needs redesign for torchscript. Nforwardhooks named_modulesout_mapdefault_hook_typec*tt|_g|_d|D}t |D]\}}t |t r|n|d}||} |r||n|} t|j| } |} t |tr| d|} | dkr| | } n&| dkr| | } n Jd|j | dS)Nci|]\}}|| Srr)rrUvs r!rVz)FeatureHooks.__init__..s222DAq1a222r#r: hook_type forward_preroFzUnsupported hook type)rr_feature_outputs_handlesr<r%rmr_collect_output_hookdictrNregister_forward_pre_hookregister_forward_hookr()r?rprqrrrsmodulesr h hook_namemhook_idhook_fnrwhandles r!rBzFeatureHooks.__init__s1!,K 8 8 22M222e$$ ) )DAq'3//@Qx[I "A$+:gajjGd7AAG)I!T"" BEE+/@AA M))44W==i''009955555 M  ( ( ( ( ) )r#cz|d}t|tr|d}||j|j|<dS)Nr)r%r2rydevice)r?rargsr0s r!r{z!FeatureHooks._collect_output_hooks@ H a   !A34ah'000r#rcL|j|}t|j|<|SrD)ryr)r?routputs r! get_outputzFeatureHooks.get_outputs%&v.(3 f% r#)Nro)rirjrk__doc__r r rmrr|r&rBr{r)tensorrrr#r!rrs26%. ))E#t),-) )eCHo. ) # ))))6555 Del):$;r#rc`g}|D]\}}|rwt|tjr]|D]G\}}||g}|d|d||fH~||||f|S)N_.)named_childrenr%nn Sequentialr(join)r:flatten_sequentialmlname child_name child_modulecombineds r! _module_listrs B--//,, f  ,*VR]"C"C ,,2,A,A,C,C R R( L *- 388H--sxx/A/A<PQQQQ R IItT6* + + + + Ir#r6ct|d}t|tr||St|tt frt|j|SJd)Nr5Fz"Provided feature_info is not valid)getattrr%rrErMr2r5)netr6r5s r!_get_feature_inforso3//L, ,,;&&{333 L4- 0 0;3+[999:::::r#c|}i}t|D]\}}|||n |j|||< |SrD)rbr<r6)r5rr module_names return_layersr rs r!_get_return_layersrs^++--LM\**aa4,3,?gajj\E]^_E` d r#c eZdZdZ ddejdedeee e fd e d e d e f fd Z dde fdZ dee ejffdZdee ejffdZxZS)ra9 Feature extractor with OrderedDict return Wrap a model and extract features as specified by the out indices, the network is partially re-built from contained modules. There is a strong assumption that the modules have been registered into the model in the same order as they are used. There should be no reuse of the same nn.Module more than once, including trivial modules like `self.relu = nn.ReLU`. Only submodules that are directly assigned to the model class (`model.feature1`) or at most one Sequential container deep (`model.features.1`, with flatten_sequent=True) can be captured. All Sequential containers that are directly assigned to the original model will have their modules assigned to this module with the name `model.features.1` being changed to `model.features_1` rr$NNCHWFmodelr6rr output_fmtfeature_concatrctt|t|||_t ||_||_d|_i|_ t|j|}t||}t| } t} |D]E\} } } | | | <| | vr2t|| |j | <| | | snF| s%t#|j t#|ksJd| d|| dS)a Args: model: Model from which to extract features. out_indices: Output indices of the model features to extract. out_map: Return id mapping for each output index, otherwise str(index) is used. feature_concat: Concatenate intermediate features that are lists or tuples instead of selecting first element e.g. `x[0]` flatten_sequential: Flatten first two-levels of sequential modules in model (re-writes model modules) FrReturn layers () are not present in modelN)superrrBrr5rrconcatgrad_checkpointingrrrr,rOrrmremovergupdate)r?rr6rrrrrrr remaininglayersnew_nameold_namer: __class__s r!rBzFeatureDictNet.__init__sf$ nd##,,...-e[AA ,,$ "'*4+.sB( {a/?/?Q{^^EVr#)prefixrrr)rr)rrBr)r*r+rr5rrrrrhasattrrextendr[rrqr(r|rrrp)r?rr6rrrrrrrsrrprrrrr:fnfmrs ` r!rBzFeatureHookNet.__init__js. 9))+++++-e[AA& ,,"'  //J  Z) ) ) )u011 *&&q)))"F6N LL*4466 7 7 7 7"5=OPPPG*4466I/6  *(F#)x $22(2CC**FBY Ty}%M%M%MNNN%bM E Y Y"YI"Y"Y"Y Y Y Y F!%)<)<)>)>PPP r#Trc||_dSrDrrs r!rz%FeatureHookNet.set_grad_checkpointingrr#ct|D]\}\}}|jrftjsH|dkp#|t t|dz dk}|r ||nt||}u||}|j |j }|j r|n t|Sr)r<rrr)r*r+r-rgr rprrrrMr)r?r0r rr:rrs r!rozFeatureHookNet.forwards!*4::<ssD,>,>>r#)rNFrNFror)rirjrkrrrrlr r r r&rmrrBrrorrs@r!rr]s  (7;? %$)-',%.7Q7Q97Q%7QhuS#X78 7Q  7Q  7Q!7Q!%7Q #7Q7Q7Q7Q7Q7Qr))T)))) ? ? ? ? ? ? ?r#rceZdZdZ ddejded eee e e fd e d e d e d e ffd Z dZxZS)rze FeatureGetterNet Wrap models with a feature getter method, like 'get_intermediate_layers' rNFrTrr6rrrrnormprunec4t|r(t|dr||| }t |||_||_||_||_||_ t||_ ||_ dS)an Args: model: Model to wrap. out_indices: Indices of features to extract. out_map: Remap feature names for dict output (WIP, not supported). return_dict: Return features as dictionary instead of list (WIP, not supported). norm: Apply final model norm to all output features (if possible). prune_intermediate_layers) prune_normN) rrBrrrr5rr6rrrrrr) r?rr6rrrrrrrs r!rBzFeatureGetterNet.__init__s&   WU$?@@ 99#8:K.e[AA & & ,, r#cb|j||j|j|jd}|S)NT)rrrintermediates_only)rforward_intermediatesr6rr)r?r0featuress r!rozFeatureGetterNet.forwards;:33 $# 4  r#)rNFrFT)rirjrkrrrrlr r r r&rmrrBrorrs@r!rrs();? %$  9 % huS#X78            Dr#r)NF)F)'r collectionsrrcopyr functoolsrtypingrrr r r r r)torch.nnrtorch.utils.checkpointr timm.layersrr__all__r&rrr3rlrrrrr ModuleDictrrrrrr#r!rs  10000000???????????????? ------''''''''   48&+&+&+%T#Y/0&+&+ 49c> &+&+&+&+RU3c3h#78U38_CsCx() DDDDDDDDN--------`    ; ;;;;R R R R R R]R R R j/////^///BT?T?T?T?T?R]T?T?T?n00000r}00000r#