iqddlZddlmZddlmZmZmZddlm Z ddl Z ddl Z e je jde jeZe GddZe Gd d ZGd d ZGd dZGddZGddZedk(rej/deddj1ddZedddzj5eZeddj1ddZedddzj5eZej1ddddZej1ddddZ ej/dedd d!d"#Z!e!jEeed$d%d!&e!jGe eZ$ej/d'e$d(ej/d)e dd*Z%e!jLjOe%Z(ej/d+e%jRej/d,e(jRej/d-e(jUd.d/e(jWd.d0yy)1N)read)TupleOptionalCallable) dataclassz)%(asctime)s - %(levelname)s - %(message)s)levelformatceZdZUdZeeeefed<dZeed<dZeeefed<dZ eed<d Z e ed <d Z e ed <d Z e ed<dZeed<y)TransformConfigz6Configuration for the Adversarial Reference Transform. input_shape@reference_channelsrreference_grid_sizecontemplation_depthMbP? learning_rateg?momentumg-C6?regularizationprogress_log_intervalN)__name__ __module__ __qualname____doc__rint__annotations__rrrrfloatrrrref01.pyr r sh@sC}%%  +1sCx1   M5 He NE !$3$r"r cleZdZUdZej ed<ej ed<ej ed<y) ReferenceBankz6Stores learned reference points and their activations. positionsfeaturesattention_weightsN)rrrrnpndarrayrr!r"r#r%r%s#@zzjjzz!r"r%c eZdZdZ ddeeeefdedeeefdefdZdZdejd ejd ejfd Z dejd ejfd Z d dejdejdede ede ed ejf dZ dejd ejfdZdejdejfdZdejdejfdZ d!dejdede ede ed ejf dZd"dejdedefdZdejd ejfdZy )#AdversarialReferenceTransformaH Non-convolutional reference transform inspired by Conv2D. Instead of sliding kernels, this transform: 1. Maintains a bank of reference points (learned positions) 2. Uses adversarial contemplation to select optimal references 3. Transforms input through reference-based interpolation Key Insight: Conv2D's power comes from local receptive fields. We achieve similar behavior by: - Learning what positions matter (generator) - Judging if transform is "conv-like" (discriminator) - Iterative refinement through adversarial training r rrrcLt|||||_|jy)Nr rrr)r config_initialize_components)selfr rrrs r#__init__z&AdversarialReferenceTransform.__init__6s*&#1 3 3    ##%r"c |jj\}}}tjtjdd|jj dtjdd|jj dd\}}tj |j|jgd}|jjt|z }|dkDr7tjj|d}tj||g}t|jtjtjj!t||jjdztj"t|tj |_t'|d zd |tj(gd  |_t-|dz|_g|_y)z$Initialize all learnable components.rij)indexingaxis{Gz?dtype)r&r'r(rr:r:r:r: input_size hidden_size output_sizer)input_channelsN)r/r r)meshgridlinspacerstackravelrlenrandomrandvstackr%astypefloat32randnonesreference_bank MLPClassifierarray transform_mlp Discriminator discriminatorcontemplation_history) r1HWCgrid_ygrid_xinitial_positionsn_extraextra_positionss r#r0z4AdversarialReferenceTransform._initialize_componentsFs++))1a KK1dkk==a@ A KK1dkk==a@ A  HHfllnflln%EAN++0037H3II Q; iinnWa8O " +/?zFART epoch %d batch %s contemplation %d/%d: %d/%d positions, %.1f pos/sr)realfakergffffff?zContemplation z : D_loss=.4fz, D_acc=.3f)"rbrrfr/rtime perf_counterreshapeminr)rIchoicererMrRlinalgnormrPr&argsortmeanr}rSforwardrmaxloggerinfo _mimic_conv2drU train_step_improve_generator_update_reference_positionsdebug)&r1r_rrrrrgrWrXrYspatial_contextcontemplation_step step_startX_flatn_refref_idx sampled_refsoutputpos_idxpos_ypos_xpos_norm distances local_refspixel ref_featurescontextcombinedrq combined_b transformedelapsedpositions_donepositions_per_second batch_labelconv_like_featuresd_loss d_accuracys& r#_adversarial_contemplationz8AdversarialReferenceTransform._adversarial_contemplations,WW 1a77:"' (G(G"HS  **,JYYq!a%+F*003R8Eii&&'9'?'?'BESX&YG-a!mF Q<. 1 ! 88eai%;$>5,*HrRq5A!)!Q$J"&"4"4"<"G%,q[N+9C!>'#"?" FJC''0BC  , ,Q 7 LL !3 4IfS\R\]`Qab cS j~~aAq))r"c L|j\}}}}d}|dz}tj|d||f||fdfd}tj||||ftj} t |D],} t |D]} | |dd| | |z| | |zddfz } .| ||zz} | S)z Create a pseudo-conv2d output for discriminator comparison. Uses reference-based aggregation to mimic convolution behavior. rr9)rrreflect)moder;N)rbr)padrerMrf) r1r_rgrWrXrY kernel_sizerX_paddedrdydxs r#rz+AdversarialReferenceTransform._mimic_conv2d,s WW 1a Q66!fsCj3*fEIV1aA,bjj9 $ ;BK( ;(1bAgr"Q$w#9:: ; ; + ++ r"rc 8|j\}}}}|jdk(r|j||||}n'|jdk7rtd|jt j t |jj}t|jjD]e\}} t| d|dz zt| d|dz z} } t jt j|dd| | ddf} | ||<g||jz } | dz}|jxj|ddtjft jddggzz c_t j|jjdd|j_|jxj d|dzzzc_|jxj |jj j#zc_y) z Update reference positions based on reconstruction quality. References that help reconstruct the transformed output move closer to important positions; others move away (adversarial positioning). rr{z:Expected transformed to have 3 or 4 dimensions, got shape rr4Nr:皙?)rbndimr ValueErrorr)rerHrPr& enumeraterrabsnewaxisrRrdr(sum)r1r_rrgrWrXrYerrorsrrposrirj contributiongradientposition_updatess r#rz9AdversarialReferenceTransform._update_reference_positionsDsWW 1a   q %--aAq9K    "L[M^M^L_`  #d11;;<= 3 3 = => %FAsQ1q5)*CA!a%0@,AB77266+aRl*C#DEL$F1I  %FKKM)#d? %%)9!RZZ-)H288VWYZU[T\K])]]%(*0C0C0M0MqRS(T% --!fsl2BC- --1D1D1V1V1Z1Z1\\-r"rrc T||z }tj|tt|jdz d}tj |tj jdd}tj|dd}||jddz}td D]|}tjjd|jjjd jtj }|jj!||~y) zE Improve generator when discriminator is too strong. r4Frr;rzgư>NTrr)r)rtuplerfrasarrayrMrrdrrIrJrSW1rbrLupdate)r1rrdiff mean_difftarget_X_samples r#rz0AdversarialReferenceTransform._improve_generatorjs d{GGDuU499q=-A'BUS IRZZ8@@BGt,&**!d*33q 8Ayy~~a););)>)>)D)DQ)GHOOPRPZPZ[H    % %h 7 8r"c|j\}}}}|j||jj} |j || |||} |jj } | j dddt| j|dddddddd| jdf} | tj| z} | S)z Forward pass through the adversarial reference transform. Args: X: Input tensor (B, H, W, C) Returns: Transformed output (B, H, W, C) rrrr4rzr7N) rbrwrPr&rr(rrHrepeatr)r)r1r_rrrrgrWrXrYrr attentionattention_reshapedrs r#rz%AdversarialReferenceTransform.forwardzs WW 1a"==    ) )  55 #' 6 ''99 &.. q!S^ && Q1&<{'8'8'<&<<> rwwy11 r"X_trainepochs batch_sizecHtjd|d|jd}t||zdz |zd}t |D]V}t j j|}d}d} tj} t d||D]} | |z} || | |z} || }tj}|j||| |}t j||z dz}||z }| dz } tj|z }tjd|dz|| dz|||t|t|dz |t| dz }tjd |dz||tj| z |jj|Yy ) z Fit the transform using adversarial contemplation. Args: X_train: Training data (N, H, W, C) epochs: Number of training epochs batch_size: Batch size for contemplation z0Starting adversarial contemplation training for z epochsrr4rr9zHART epoch %d/%d batch %d/%d: loss=%.4f, batch_time=%.2fs, samples/s=%.2frz9ART epoch %d/%d complete: avg_loss=%.4f, epoch_time=%.2fsN)rrrbrrfr)rI permutationrrrrrHrVappend)r1rrrNrrindices epoch_loss batch_count epoch_start batch_startr batch_idxX_batchbatch_start_timer recon_loss batch_elapsedavg_losss r#fitz!AdversarialReferenceTransform.fits  FvhgVW MM! Q^a/J>B 6]. 8Eii++A.GJK++-K$Q:6  )Z7 #K j0HI !),#'#4#4#6  +"/ & WWfw&61%<= j( q $ 1 1 36F F  ^AI!O!!L3}d#;; ' <"C Q$77H KKK !!#k1    & & - -h 7]. 8r"c$|j|S)z,Transform new data using learned references.)r)r1r_s r# transformz'AdversarialReferenceTransform.transforms||Ar")r rr)NN)rNN) )rrrrrrr2r0r)r*rwrrrrrrrrrr!r"r#r,r,&s $#%/5#$ &3S=)& &#38_ & ! & &(P.RZZ.BJJ.SUS]S].`G"**GGF&*'+ n* ::n*JJn* n* c] n*  } n* n*`rzzbjj0$]RZZ$]bjj$]L8rzz88&%)'+ * ::**c] *  } *  *X<82::<8s<8S<8|2::"**r"r,c eZdZdZddedefdZdejdejdee e ffdZ dd ejd ejd e dee e ffd Z y )rTz Discriminator network that judges if a transform is "conv-like". The discriminator compares reference-based transforms against ground-truth convolution outputs, providing adversarial feedback. rCrAc2tjj|dz|dz|_tjd|f|_tjj|ddz|_tjd|_d|_d|_ y)Nr9r:r4)r4r4?) r)rIrNrreb1W2b2real_probabilityfake_probability)r1rCrAs r#r2zDiscriminator.__init__st))//.1"4kBTI((A{+,))//+q1D8((6"!$ #r" real_features fake_featuresr`cr|j|jdd}|j|jdd}tj||gd}|jd}tj|ddd|f}tj|dd|df}t |t |fS)z Classify features as real (conv-like) or fake (reference-based). Returns: Tuple of (discriminator score for real, score for fake) rrzr4r7N)rrbr)r}rr ) r1rr real_flat fake_flat combined_real real_width score_real score_fakes r#rzDiscriminator.forwards"))-*=*=a*@"E !))-*=*=a*@"E   (  __Q' WW]1kzk>:; WW]1jk>:; Z % "333r"rrrc j|j||\}}d}||z }|xj||zdzz c_|xj||zdzz c_tdt dd|dzz|_tdt dd|dzz|_t|}ddt|zz }||fS)z} Single training step for the discriminator. Returns: Tuple of (loss, accuracy) rr:rgGz?r?)rrrrrrrr) r1rrrr r accuracyerrorlosss r#rzDiscriminator.train_steps"&dD!9 JZ' =5(4// =5(500!$D#dC*s:J4J*K L #D#dC*s:J4J*K L5z#E *+X~r"N)r )r) rrrrrr2r)r*rr rrr!r"r#rTrTs$s$$4RZZ4 4uUZ\aUaOb46 % jjjj  ue|  r"rTc eZdZdZ ddedededej fdZdej d ej fd Zdej d ej fd Z dej d ej fd Z d ej d ej fdZ d ej dej dej fdZ d ej dej fdZ y)rQz) r)rIrNrrerrrrRlr)r1r@rArBrs r#r2zMLPClassifier.__init__>s~))//*k:TA((A{+,))//+{;dB((A{+,#0#<-"((KcBdr"xr`ctj|tjtj|dzzS)z2SPDER: sin(x) * sqrt(|x|) - Periodic with damping.:0yE>)r)sinsqrtr)r1rs r#spderzMLPClassifier.spderFs+vvay277266!9t#3444r"c(tj|dz}tj|}tj|}tj|dk(d|}|tj |z|d|zz tj |zzS)zDerivative of SPDER.rrr4r9)r)rrsignwherecosr)r1rabs_x sqrt_abs_xsign_xs r#spder_derivativezMLPClassifier.spder_derivativeJsrq D WWU^ &A+q&1BFF1I%1z>)BbffQi(OOOr"ctj|tj|ddz }|tj|ddz S)Nr4Tr)r)exprr)r1rexp_xs r#softmaxzMLPClassifier.softmaxRs:q266!!d;;<rvve!d;;;r"r_cd|dz dz }tj||j|jz|_|j |j|_tj|j |j|jz|_ |j|jS)N_@r ) r)dotrrz1ra1rrz2r%)r1r_X_norms r#rzMLPClassifier.forwardVsxUS&&)DGG3**TWW%&&$''*TWW4||DGG$$r"y_truey_predc|jd}|dz dz }||z }tj|jj||z }tj |dd|z }tj||j j} | |j|jz} tj|j| |z } tj | dd|z } | | ||fS)Nrr'r Tr) rbr)r(r*Trrr!r)) r1r_r-r.mr,dz2dW2db2da1dz1dW1db1s r#backwardzMLPClassifier.backward]s LLOUSvoffTWWYY$q(ffSq4014ffS$''))$D))$''22ffVXXs#a'ffSq4014Cc!!r"c~|j|}|j|||\}}}}|xj|jd|zzc_|xj|jd|zzc_|xj |jd|zzc_|xj |jd|zzc_y)Nrr4r9r)rr9rrrrr)r1r_r-r.r7r8r3r4s r#rzMLPClassifier.updatelsa!]]1ff=S#s 4771:## 4771:## 4771:## 4771:##r")N)rrrrrr)r*r2rr!r%rr9rr!r"r#rQrQ;sF.2e3eSese " e5rzz5bjj5P"**PP<< <%% % ""** "bjj ""** "$ $BJJ$r"rQc 2eZdZdZ ddeeeefdededefdZdejdejfd Z dd ejd ejd ed edef dZ dejdejfdZ dejdejde fdZ y)ART_MLP_Classifierz Combined Adversarial Reference Transform + MLP Classifier. The ART layer preprocesses the input, then feeds to MLP for classification. r rrArBc t||dd|_ttj|||tj gd|_y)Nrrr.r>r?)r,artrQr)prodrRmlp)r1r rrArBs r#r2zART_MLP_Classifier.__init__sI1#1 & !  !ww{+##((#;<  r"r_r`c$t|jdk(r?ttj|jd}|j d||d}|j j|}|j |jddS)z!Apply ART preprocessing to input.r9r4rzr)rHrbrr)rrr>r)r1r_sizers r# preprocesszART_MLP_Classifier.preprocesssw qww<1 rwwqwwqz*+D "dD!,Ahh&&q) "";#4#4Q#7<rrCrfr)rIrandintrHr@reyescore) r1rrDrErFrX_preprocessediidxry_batchaccs r#rzART_MLP_Classifier.fits  JK  WZJ G 9:1z" BA))##As7|Z@C$S)GclG HHOOGRVVBZ%8 91uzjj9 j;s3i@A Br"c|j|}tj|jj |dS)Nr4r7)rCr)argmaxr@r)r1r_X_procs r#predictzART_MLP_Classifier.predicts0#yy))&1::r"r-c|jdk(r1tj|jj |d}n|j |}tj ||k(S)Nr9r4r7)rr)rQr@rrSr)r1r_r-r.s r#rJzART_MLP_Classifier.scoresM 66Q;YYtxx//2;F\\!_Fwwv'((r"N)rVr4rdr)rrWrW)rrrrrrr2r)r*rCrrSr rJr!r"r#r<r<ys-8"$  3S=)       . =BJJ =2:: =KNB2::B BB-0BDGB&;; ;)rzz)2::)%)r"r<__main__zLoading data...z../X_train.wavr4rziz../y_train.wav z ../X_test.wavz ../y_test.wavrVz$Initializing ART + MLP Classifier...rUrrWr)r rrArBrr9)rErFrzFinal Test Accuracy: rz)Demonstrating standalone ART transform...rz Input shape: zART output shape: zOutput range: [rz, ]),numpyr)scipy.io.wavfilertypingrrr dataclassesrloggingr basicConfigINFO getLoggerrrr r%r,rTrQr<rrrrLrrDX_testy_test X_train_img X_test_img classifierrrJr demo_inputr>r art_outputrbrrr!r"r#rjsb!,,! ',,/Z[   8 $  % %  % "" "@@NGG\7$7$|G)G)\ z KK!"#$Q'//C8G$%a(1,44S9G / "1 % - -b# 6F?#A&* 2 23 7F//"b"a0KBA.J KK67# JNN;A!PSNT F3H KK'~67 KK;<BQJ'' 3J KK- 0 0123 KK$Z%5%5$678 KK/*.."23!7r*..:J39OqQREr"