From 3f9e30c9b368a747a69cb9d78e90fa84c0b829ec Mon Sep 17 00:00:00 2001 From: TianYuan Date: Thu, 28 Oct 2021 06:52:21 +0000 Subject: [PATCH 1/2] refactor docs --- .readthedocs.yml | 2 +- docs/images/paddle.png | Bin 0 -> 5043 bytes docs/images/tuning_error_surface.png | Bin 0 -> 110461 bytes docs/requirements.txt | 6 + ...architecture.md => models_introduction.md} | 84 ++++---- .../{getting_started.md => quick_start.md} | 31 +-- docs/source/asr/released_model.md | 28 --- docs/source/conf.py | 11 +- docs/source/index.rst | 39 ++-- docs/source/{asr => }/install.md | 10 +- docs/source/introduction.md | 33 +++ docs/source/released_model.md | 55 +++++ docs/source/tts/advanced_usage.md | 53 +++-- docs/source/tts/basic_usage.md | 115 ----------- docs/source/tts/demo.rst | 4 +- docs/source/tts/demo_2.rst | 7 + docs/source/tts/gan_vocoder.md | 9 + docs/source/tts/index.rst | 45 ---- docs/source/tts/install.md | 47 ----- docs/source/tts/introduction.md | 27 --- ...eased_models.md => models_introduction.md} | 20 +- docs/source/tts/quick_start.md | 193 ++++++++++++++++++ ...n_text_frontend.md => zh_text_frontend.md} | 2 +- examples/other/text_frontend/README.md | 68 ++++++ 24 files changed, 502 insertions(+), 387 deletions(-) create mode 100644 docs/images/paddle.png create mode 100644 docs/images/tuning_error_surface.png create mode 100644 docs/requirements.txt rename docs/source/asr/{deepspeech_architecture.md => models_introduction.md} (81%) rename docs/source/asr/{getting_started.md => quick_start.md} (83%) delete mode 100644 docs/source/asr/released_model.md rename docs/source/{asr => }/install.md (92%) create mode 100644 docs/source/introduction.md create mode 100644 docs/source/released_model.md delete mode 100644 docs/source/tts/basic_usage.md create mode 100644 docs/source/tts/demo_2.rst create mode 100644 docs/source/tts/gan_vocoder.md delete mode 100644 docs/source/tts/index.rst delete mode 100644 docs/source/tts/install.md delete mode 100644 docs/source/tts/introduction.md rename docs/source/tts/{released_models.md => models_introduction.md} (90%) create mode 100644 docs/source/tts/quick_start.md rename docs/source/tts/{cn_text_frontend.md => zh_text_frontend.md} (96%) diff --git a/.readthedocs.yml b/.readthedocs.yml index 702ae6da..ffefb00e 100644 --- a/.readthedocs.yml +++ b/.readthedocs.yml @@ -7,7 +7,7 @@ version: 2 # Build documentation in the docs/ directory with Sphinx sphinx: - configuration: docs/src/conf.py + configuration: docs/source/conf.py # Build documentation with MkDocs #mkdocs: diff --git a/docs/images/paddle.png b/docs/images/paddle.png new file mode 100644 index 0000000000000000000000000000000000000000..bc1135abfab7aa48f29392da4bca614f688314af GIT binary patch literal 5043 zcmV;k6HM%hP)Px|ZAnByRCodHoe8iVMY+cbn;1d107fAM5|&a3LJ*P7Ld`R*M!^N#5=2ErP!zR< zA_!7S^daytPsEDif&wLr7zGqC2&iDt(hzpCh!7y4kbpu=uD}1`&Y@4AUS`hpOwPT| zSM|^Ie0%py&-8Tnxf!v>Fr=Zw{S0Rd&xAY^R)hjanbHz;NnVFVESY<8$iOrM7j>K$_5K zXo$4HaOzl3MCEPJ8B=PM^~U?QW)Un|Tr_|2ZIQ zL`@4+*8+U_enJoG)_deg@tDHFZ@+|05huw)o#i_Y{lQIz@k73U>l~eN&sZ4{+J!;pwTdT8IiM z_F949&r`7hpB($0k!@nh|Hb)Yuq0$_+2lSYrfvj1J?$GJ^P#7lcgi38m!h~wraX95 z9km000k+RrvQ($KX(pv|X8BQlPq28gQ|Y_f@CUXbZSq&zZ8rRaE!k=CFG2p0#ovtc zF&0nXtUj87K<#7GZ}D%(v~|pLUB~(4lXqRxv`tec;myE5L5TDb?<9;~0sa**eiqqm z@F9?QM}{+&$;8*GAViNMblos5H^_7|gh-o&&jtIKam$wRcT6Z~N!f)(@*~!0@Vq1D zAAEk$-3_>Yv2Zo~jjf?FyXbtF`c~36mbYUe(_2dGg8v3p&Y_pWPjw063=MP602Ah_$2-)13St~P*iE3jt?KMrp%|Y2;+*zAzZriCg*OmmVaA9Zi8}?I zD*}4D;kU=*ymfy(23(!tCh~b>-XQqe2wh!!O)i-9 z3@2OT>j{(TE8|D^M%Qehp|8o5*4xYAeFy(X4c2Fs!Ox0!QJ9yVvPzrRiR>sfe`6cy0tGVvMC|s0y4e1LNk&bDWDtsh3};u2G$AoH-Z$Tw{>kq!l55w z9#|w1e0|T;0ADNON4og=qQ{qyJ!{ZTLqU`Dji_OyW@`%(HkxR%Db(Q$`7fkQ7Mj8g z(ZwN$7>)vaSa?0>`+>v3*T91y4}CHeE0f`9&m7KzlML9u2H1IP=rRd>1nBQt#AZoR zX?>fz-X2-HNsq+4NkJX-7kIfxJLrDJ;X^FV?2hgwY3z;%uGvM{U1WDGbtp|<>}tm5 zd@R_&P97t(V`)}$Thk}!F9Fkm z9&YMz-nj>4Yl3;eCf!dOBGyX;)l)e{*Qd;gKzh6Rp14^a9}LArNz<~fo${Hzj)#PN z6qVsm!>`cWOEYNpH$>4eQGH@u3S6`pragJk_J>T%3cCW$WZfRDAn zQQ-H$N3D)!yFgh(!y|#6=MUl8-%l-b_{Rbl)$+jWr~_GS4`NT^$sI-sPKdb`&|AP{03F! z@Ef{G|KW=ZUyZJ{sSq7qMFSG>Wk8GMKGGC?Wkk^b^8|hperveH@oJAwjI9T3uAyX+ z3gUA-xX%;Z*|Fpw{eyw;p?@o~?FRKU?DX9aUoDbr5?zwn9E658hzF)73FdHu?}-W! zpz#hfHTbCk8;O>H;G(9>Z(bh0zpo%3nw{5AQ9a^q0+xH?o-TjLw8f|wT87bkisls4 zmrtNh?;~%=8mv#~TI=4`weW4By8~I}L-VdSAZX}~fq39=RVIe-4HP6m6YZqY)pvrv zWkTfLzGVCB<($AL2ZG<+i_bhZfw&q{x7MIIUJ=O6Cz^v`$2>XY)xn=0+8VOODA=(S z4+k@1P8@Q*!3u5Z%3HUf9zjQfrH+J5mUsJ7(0KtN!8GLak*i~s+P+jEHLohkPW2@W zzmkrHU(Jp^;)(kWC5vnV@C}`8)2Crv4)*p|v^Er!cZ+&#L?9Ad{4X|r;b;|A2}TjLZgl>Z4KYA%}$JC>y5>^ z>UaXs=gEcOI7p?gq%O_BVQ>Jt30NVb}y^lo2DkqrJd6jej`=ngv$ilw$M+5UPt zCx=+u1=rttjCqi+k>Dp;xbGz|*%1G{KyI06Hia#K>)X`qSebs&qo(n-*xfM&@2gXl zL&1}LK?nJl@+w^v(J3qbIu5oEIQg|HqtP_XiG2?g-rHuzh?U7f!;#SL{sH83Vt#s% zf5v0;E3Q-P0Yq-cq4O4p-v6U>t%F|&OT*7oi9b`p$T4-y=WzI4FoPHmrgmP)AC=n& z#Hg~TAc4n#Ya$6~NY}B;p7UC^FZmji_(y=pL5RxklJ5#nu-yXz+Z_p=$sWB0@LTI! zUCt9Q@a=kIX;SP9pkwz2=c(RUx8duPV2fU=(0*g#U<3}{2DH7uEKQ3zxFLHAN{`qY zvrdTE#Kg*96D8UZI|alp^yTF60D3JVhw&6-7ozst3Vls-915cQ4MPr*X)SCFI1uFd z$w3ucXMArf{g0BrBCi%v*28KpCld!**T#j zT2!|=KtGB38^`iQgS`vXMU%6P38*!r*iYV5=k*j0?YGs;NRW>0OXyCE=vN`WIBU05 zN0J;1^!9eqyX`nA?dMV)0nzgGU zr>=c=1eqQQv;=v3`x4~?Bl@YuTD09MYqwNKeCkG9=1KNgUmWx|28%s*Vr%xy*(rTK z>)?~q0ZPt6N2HFg?-IQtFX!;~)Xp9L9)`CK<5&HBT5c3$yOWBK}!X)jOTOyW;1qHh94V>YCqT}hWv@LZm{9T04*J1k;u$9Aa9@bRiUgT25_fO#;;*ure=|N^b z^m}#tXX>DbbsD1Lv8nnWriql!3pwW*itdpvV*hxe?$-kseXmTGI2;AO30AnAM|j@F zW$a!GHc5c@dvZ9kOuZ|yCDDC7$(rolDhXb~DaozY?Mo$6FJ(%-cQr4JQPw0xk|}-s z>;UvTtRm#?Sfwg#7?aFoPeRG3+npcvMlN}ZvDNmD?MrzaR;@$hjwi;HNSyZerHI~8 z6Z@{$FBrzmOHmp8+=HKY(9Sz@)MtXQ3#gQcH<|!?>#l1B(l^|}7O(3R`zv7D6E|9> zIgUEr*=0}_gGBUX3Qe~n9%yf0sw$77e#vR35(`kr_NAi!`&FO%jsd=5+8e9Po4DJ7 zW5L-#v!3(7vmm3=Zw(yjrz0+!rMe*qzOj&h(Pa#8E?B3^_UZ6VFs-wMC^jqC;jfMp zlGS&Bc|l$IA6-W-b;qEhGhCI0r&+q%iMZN%qwgo7)I=Kg`Z1)7^v7VvsB#=*9ZP>7 zcjmp(Oa0zmTE4sm}^sg`FXzwl+a$67ho=92hCwCbR0|u3w zkX6ueJZPv^k-B^s_;%;tj!bK;EutLlSW-k!LscJ5f5BS?{}OQ|SnYp(sc6;6IktUC zinRPfCJCGvNurLwz7$EJGV!Ydo~P43W0Btq7K2rwZqnlF z5?@gA(>K^SpoM5H+P@9>zRR@9`yKdO$gwbC_ayPOApKjgp~Jx>c$3h%k)sy1ZwGf{ z*L%C+Y%TBr2+`RD>MJ7IDWQasX*jV>UnQL|8R|AIQvcO0KFL>I5<3(4{>HkUw>JL1 z)#VF*t~XX@i@r8o$Np3TDlO7p*jbsap5jI6?HFtuFw(Xddw*R(y*82e!VPAipXx8D zEt~oS{i*(t$s%e)@8Bt6!*#qihi+5_KXXyq59GNo)<;io(-!s8v0^u<{`!)J z6MaKGN!~y*dqooZYD7My#at6jsoyfBD-llqYQ#HH&!lz4-(C3K1$-wQzEfJLV{wD- zi+ODSTtDtu@a;CpT0?(4MC0V)PWm=^aF|u{4(g-tsY8MOifwYj?=o-{j^7M^ohLw{ zj)lY6+v<7}(37!mh=)O*?MuV@ZIxXNuFb=gcLvjep2S0ywEO;BK*PR?;EFmbLL6<+ z+n-n({hSN#^TqYes&wgdgYE!R7b9eI0D-URGG2`g@}5cfj|EcDZ>wHS-U9)hVJV_* zHlt#r!N#E18RArJqo1`u>T*V&+UKWP=!d{U^d4~O#d!GH2c0V&dLP8+YzIHe!si&2 zwgp>)mH7El)W}r8L9>0yY>=khv_RPwpi^xDn(a$v8?v!)TEJ`pI#siM$!w6O+_XU1 z7NApU`_dC2MD2}^W#gr?C)r~{2_qY}+m}+cVdJO57T^XvE^>oDjeNrjQz-13eDv3s z?v3PqUSY?LLDK>$T7Vm~*}jybAsZh}3z#iH$9iYvsjpw7YPzwnh@5VT{=kUcA*{7q51te%-b*Sr&gguX`O9pAn#SkP6G`!El^nt{0HuA9HmP~`H}zt002ov JPDHLkV1n`16BTMM8xZ5C}QsgSd#Ad-_4USG>-qN9DD* z7qsQ9U7PV-eUXvU`S}NAc+uDw?$q3_5pnArugN|td_34puZ|1#iA~ljd*18%{)0FF zE8}32^jBy&=~sj);iNd$1BBmS&XRW-II)oje7~V~DTU6o9O&Mr36;1Nx3nDT{@wf7 zKK7}-q}6h!@v>s?=Av@x(#x-M@{<+32qyUc7S_$Zj6{z7<>|YOfdYc?-^=u9SpNUL zgp3!2Ed0NBqKLs~|My1Gz&CJ8PZz`I^k6=t{qI^O!w?AWf43}R5r?4scdZ)X|9`vx z56<1bH<>><@}`JY_K3T-|M+S`-X%eP=T)zHDht-^^}$ur>pbReml z25%!7(4stTTaH7>xs0R7jPoM^i_L@KJ=9F8Pk8^ZYzMX7Rc>(0VV~Vs2C*of{$!u7ibxh}8VrR3XcNFDriQ9~)lk zEIOerwFA5MKbsIAaKMjT51rgFmQ&y`li@2o(gvYQOA1D07`Y;`%D;M9n{^*iTG0|# zQc@BjTKs3C;ZOBxD0Kl&Xr{e9e&wKr6;yDF-ANK#f)d0rZq(DKZMuJpc;v2L1;ubn2EZ%E?Le3wd#zyN~284K)(QcOiizAVghSJ|%Bx~XF5*DvI1 z%?JY)X%@oXZ5Sg_9_DYlA7PQyo^;zf2n=)d8C0pX$fEgAo9ZK0v4vi=^BSdU43aQ6 zNkM04*29CZ+*EH314Rt-ggLIiF!^`b0Q*S9fP06g!ZX>L?+T=1D4OshMeJbP_0W_5 zZ5@+^iJDf&d3H{x4i0|s^zwT3`nBo(&H2GndsnpMl>R`(ct9QkYXnJvUFoF0)d|J)G!OcYt+a4@S54w&^+%c2 z%iELN`#*pGGJ0Bn{`@VDMzO~4(U)Z4!#ivmYHI9Gsl16^_r8MkF`7YWupuGO)S5~c z1#-_P5s>4`F!kb}CUJGi=aW_cz`*YQKG+3Y+uL)kZp{1B94utUXXn;hQyc4i&UH%`sYYWA~f%Dnk9$$|zVw5X&cHz()Kx_dF* zwmr8TEv}KH<9i4Ff_SGAbDbeNv+r>DBA6c$LAq8#RB68jQDWC|O-;;Y%kb-qiLY;F zM~5IHA|kkp(NuwaU1OthDz({YI^73n#mGoQfn&(g)=015alpZRqKNQkSst1=x=nsT z!A#7^2Aj&=c(!zXJ`y?utZ%=cCMG__ddVlSY-s8e8z-lvogH&tUS3^8Lyg_CAQ3Tf zlDKc;TgcxP@(w#c;(aM0xPk7)rkGi~d;AT|qU|BXg75K8Eo%vX;%mpYhv=|p9Dl!Z zl)_Pc%xcrF+{_ZkL0IjP03X5-rv~lV;Ls^1*t1?6bRa=QLV8+&2Ah_amL?D@Y#$Hs z-$DNJL04L#f@N6`xUrdbXlMqn1!N8cpo4b{>b?_-3Or^MJe%olOsuNEc?i+4BPDrU z_HbQXTx7-9EaiE`RwsJ@z$I2!vjznR&ow(0;#MK|AUqX1+E}?IpA@;Y_U#|<`S@h@ z^~t_|{rYEmy2b-$`fq+7CWA*%_ZnJ1bL_LPojfLxtbxZ7JEM+^Pou0O8v1?i>n|5%ndp482#^Qq z%!X>r?d1y!7R|k{n^HU~;8^mXVw!zQ?gbiXGoH{@S1lwq5fB$)Q zYinA&*X8L(oT6!a6G4~?NMW>LIDX3qmFaar^S^~JKq}m^9L6s z&~zTBp#FYI*Q13Gva&1W|p)1X$VQ8qu- zv<<3hof#8z^AM0NKvv$_*;!fX{%Ub3AwvN%Gh;<3>u-DX<(%ad&r|XrbbfQy0V7Q( zwuA5Vo47CKp`hFdR-wn5D=9A@Y0?D?Vr*<|{jZJ=8t>HL5*kRDTgjgCE%%Z);Ax=v zrI%B*$R_0a&<-6VM(oh>GkOK1^b`Pw;OQk%I*?)^78Vp8bLf}rlrZ&?S_j}XsM14=1#~q zRv@Z`8+@EZmqO9}=J*y)scD6-zg3a0{?}0-v6-c1*wT_7sD|dYwx%-`x@G#Ef_Cqc zzaV6<8{DlRC0{P3H8lO4nGd(M6qT39`Bm|WWVpL5iCME_F{d+F1f&NJ1=B-pK z6s8?%8hW<1>#Id43`+xYm}}JY{b`XD`0R=K@RK`O6A}iTqjE934Alipp2zDW_gqQ( z+Zbd_#Sv3kE)YUG=)@$^MiRLMABI16tx4GCKM()@63f2_L5i;JHeLl3@00a`SA>LJ z+pvknMR_wbvk8KPHVXewd6^!Fuzh{04DJ{oe(1#pChA}cSGL(tzgqK*Pe~$vIXjSs4 zvzFyuxjH;!Po`!NXAI}&UdUuJFB6d{V4;Aq;DiZX9O3p?>#rZps{fNQHI%RV2jAtW zG75kT$gkhbok?hn|I6>;eN{HZGcFo732gC|&jBtpwzZ9N*`NNvegY>BpIue(p}`(( zLDm?&y+s0mPke6J!S?$0)6mk=(%iygHSApl?Oe4fw&LEELqzLQkNZ)xC=&nWelPq{ zgAL(8^!B|oDO#=()XMF{S8;s!Y$cvD^=kdFc-UO6S+z~tjJ_W! z`Gik?I^DLTiX1yDcIff%ca`=d95d3-h7f-3-P_@PL;K-**?vB)4fcP7^_l6x=uQ6lGx_mMk?`>WS|x|GEg`kTJ^rWe2f27-BbJp#&o~yC zD>ZXM(Lm_X<$bVHR*Qw!z&m8C7h}{V4n(3LD3U^1`=s|`I0Dgx_bS`(&tBj_U zWQT*`+TTtu_otd~5yuK0k^)P(_u=9QR${J76X@NWy3UP;jy^dvBgIaXUs@WT!RJw1iXPbFtYF6BSYwNAZkHuxLN4;Ah1PcF=MB3WiYX2u>#hjknWHz%koUQlN@~{Mb z1N4gz@Ks-$#^3rk|GK9p9nEr3|F1X{YN0{H%R+`;bSHN=;M9}|p}@&d;bqw|<-gVWYj4F8M5#K*mXgL31pR)p zAOo!awL4INNc@ueY^DDz|TK!qHoux>MTiQuRL^PhpHQ@8$sH>xr z=QV{>q)BKRj2t9N81zLr6;)!LR}<5pO2Vqe*8}%{HCip6rD&thjjDZY@W-fFY3K$( znjPz}2kEvyS97znRu?6){ZQnQdLp||m{^!x=^`HH_(At4DSGe}(+SoZKx3oC`KL#! zGFXw4*XNafBfC?D#igaj0PX{Xi+_Juzk*w}Za~8G#bMc}$Zct`DFqQZPzl;WOb{R@ zhKb@q6B}}b#qx$fJT*fHdr_NRuRx51I8rFtomAEdOI3LOL_;~xEGNI5(io*QNd~{8 zR12~-NKnP^RC5N8L1!>S(tCCg58mR`}?_wUR3DNI^;uQ>BLrmUEbmWtsp6 z1~LUNAKzT3pU}&fFNXyC+f=)poCqH;>9_Xu7vdqnNAxX}q-|FV`tN*K3#> zopOlPbZ+JkWlJrN%Z1#L6D{VZrFXPo1gu(!>3X(jFLzS4tE)shATbVPbT+e5`>o+u zOr3X**7FJgJTh&jTj{RJsPo9ev+#k%#C!9%mHB$!B?m`@oyf$%z#pIn2&fH3#l?n~ zM~hT^OX+kgfj#K!tHFl89YsDnx_n9(lz&o-aLmH_CmF6#WWBcgzTh9<-yx6sHDM8@ z%5-9YDBgvuQW+x~#S5U#Kfb-cx_>C6+jd^m04c2Y+(}R2_nZR>$-p4deVH@ZekN2r zBg9~-uE+dFbHswd1qg~|Q&Ur*$~uCF2Xthv+8BwEkr53IjV1Q&%}3f4m9D+R#wRS} zz82auUqOfPDX*1NR29O(Mbhclp!sw6pfZbB0J#9~5{ZGs6?Yj&sO#YuS-h{c;a_sB zuHX3nf%pNMiGe0sfByXWyg;q{n}-A)tR4WF=4%~b-abzx1Mbv$+^0j~p8~K|d?3gW zOGtdnAO5=N0-}MF!2>7+kNfdTu3gUN@qu!7g{cMy;+!k!?|jE%pJ0zvwe$@PsKKL&neJ z>=K*`;$Z+cPMF67p2&c&9qafo5$*fUX1+|jF$%y)z+^4V&40}nK*=6Y$gWS5I)}%{ zF_Uy%U%4%L(A-{{HMDr{^t)W&G2Emn)WCCnM zbVRC5f&%O2_I4Kv9biENY<;fLj!ZDiZm@ys{6JfnE-*@x^|cSX-mJq~uOni3q1V9e z@kpd7c^{HefUSjChG%0)w^O>?i#)%7r&#kXuS}6x&Iv6Q4AT#P+RUrf_#l^t%P_&6 zY!%B8_;0V@CB-?uCMDhctD*k{Y{9l^uv?B&32SKJZ03fa93%&G>X7<^Y_1+O9TRUF;>3#K+ zY%8hrj;hu55bnS0X@ss{EZM+5JS>v{CKx30BlGD?FR8~d<>YtjLy0g0r zVEaognj9dBOwG(_@Yas)T5NUR>ZW@^^P9)n?NC|@QHi{ZnqP>*-k?oMTYbpmy2w_@ zWH3Ll=Z=`>z2%b3A9tzWX`Wk4^QV38Hs438TrsINxoXNelrR9qGakJBY zh_R936|A4w?lZt^WB_=F-lCImcP+JhHSbrnLT2l&2)LHq-)24DZR5-z%r{0aKO8P^ z*@0|_Gv={;yEL$TH?kZLqxu!PVf!ujL_DjpYUZ{fxjdI^u9n#&9XvtQYk}X)z9*<) z!pC#ObX@+b*VWZ+Y(CeFg|Y%D<#C&Ia$os~`ghp}*@*`+7fg{)Qtt4pKMUfcBY$V5 z@fDNM#s;OvEh24g8O;}*#t__w@D~;Z2&XB3M>2S#b@88AIvAr`1<6GEtDoLTMJMdQ zt1yY?C>fO-C@NwEqm!-)Jq$elG84Mh0uv$0PPCbyq_+lS0!4nX2`V}tJi!yDx()-j z%=GW|Q`|k~9%7L4_P&kg?^#(P|77KRGnM6dcje%_-{_0#Uu!mYdIuYunw;zbX%8^# z&6fR&^=+X?k5_tu2hu2-v>}_GmsFKAsKGx&*(=Ak^8j5P9UYCZ;&>mUptdE?5Oi}r z#`1XHT9oAHVmsW-b8z#i$~Z7zIYAV<6p9j^ORdQK(PHkC})$ zo#c%T4X4xEcKJS|F=jwlloFs=BX?I(857d>N54_rqwv2n;;XsM{PVir?X64}M8;qm*Kl2wuK)Qzhi6=sdSW$#$>zl{-58i&i;X?>(T!WukI3T)4Cl zm{?>wRYS)sCwe!^NgJ!+22V6ExC5pBL-uXU?;|*sYdR-FNm~Yi>Jn7p*|7Qt8;;d; zV!iNmW}F>9Xj|NNrdO2IYmaZ4nXfYe*cQCsRh*ib=sus(&qDrqH$irfm37ri18D*T zoX2qk`EtpNS^s7->EnUs%og9fGQ#2sLE6z)X8dpmP`Rw8cv&8g0{Usby1adt#?6MW0Xhs}m*cX>oH)O6b>5D)l=- z;(qm7j|4*a`gK7$qfb&^LqkK^e6Z*n+tx$Q4zTkbe!k>_RGE#v*7LnKvF*HfeF|St z(w;SK#LD7F);evEerr6Gspd3p71cV)^~uzTvzeEWx&z87kLv-UqTn?pP-}a(Vb$Os zt_QQq@4t?wV+jw4kb1!FPd)NYKVAeA3f&kgoD?d(u*p|w``YR>Fd2#V+-EL-oJc42 zNM6YNB3LS@3p&=JjB<@A7#Xu#_epu?@2Z~bi;~42`+k^?W)orUBe%GWwc*<`6nZUq z_0wjrvP};EReNqL4EN6EF`zMUJ8np?3C_4YEd~uSghAyYFOxOwrnIp0mM* z*<(a*N)Dcm)8?NzNa3IqXxBY4>fPPl!AE#me{o14{NR#LW%~t@c7t^ekS14_+WsdN zA0CJ9n2V6t{h9IXAdX+~a>gy!M^ZU>lX8 zhfO8y1N6wH{`GAS3~(-AACmanuS$%x8x%C4&kYsG zr=eh5MXggZXaQ+jS63I545x`;do2XwSJIJ?60YTtLJz+i9Qy+bB;BTS{8l=pb{Ypa zYcch-qCqxHdeL87VYB%sJj3oGfk#Vq-FkTToVINzlHlvAhoGYiARMpDqs<8^GI?B2Kx7$tAH{;v9j=ZI1uvV&9Oi8r zt6la#zV^LZ{f0!|b$_`$R^GVu8ww~bo1mZqNa6Z-YXk(ZL>{lR9(jQ7Fy?zv8xbAt zG-r}@arFJ44J&kOc3@;5SB9b!{=a- zCQHY0SxV?eS@32e+LWXc5JEfebz+`NFHm!ToD=DM_}oyz(ng!L&K7t%*U5knhgvu@p)TyjLTr~d8x-w-~>)a3OQ%gBos zQbLq`vbA*LxtLpDw?$+FC&n<>TihBE9qD2zkw+xPjk~RuyafK zmTEKqKk^^E&JEIJwbe{Tn~=H0+0~ZsF=^{ z`WD5?z+&r5W8dyfD*W(98$JQ0D@~MN+keth*JHZ8rJIH48|2xEz~8$2(u$W_UJoI! zS`q}3m3H78bMCfJ_6BCVIxIRhI<%4@mFw2ED&4n}BTit3inGc)DdSFfZzrIBf>iuP zU3Dz_jHNcXeWd&Xt4l@_eg$C|a9aXEp~ik_?Y7;Y5`V zp7GL@dUMui&ZYXlL`|2^}2Pxb9Mfa%{szX_=1Z= zxpTOqeQ<|#{&)n#K-@I3VjvuUo`R_r%h{+j<_!yOQ*%RvacyAn~^X4dtO{kz+j>E zUVRw(Kb{9c zu8?O9*EuF(LMtR({A)ps>L1c3%0o^tqxRhCnBSHThll0XE;Si~EK;=)j6nBB+ki4Y zSO!ld7EaF;e%$3m;M68!>b}<%`tC4bQ-l&br=tUH=nE6MIHbZMw*59XoHq>1_lfA0WKZnykFrDA{l*rA(x(f432C^p&)T|ME}fXA1Ytn^wqs@2@&yX4Jr3D z<=+X+@AG_0`FHtgsp0tkb1Xp$Jqv1W=Q>!Li4rz8Hfw&oc+hj{NQ%*s5vs|N7-{%& z4#6zuau&Gi!ogqHcH(Go zb}34Z90NDKV=?aKntRsUY!vGqf)2y^|IjmI^A@rhQr03BuIv}`++k(&`VVfjU-p%a+lmlL(3A6&z~vZV82;H zr`Il-gI;QsK~9VQz$yV!R5phn_uLz}BD;}X$aof_d;PlVl&g`Mss@pTivltO0mWW! ziorb=P}SVV#-#s|)~9CeT$SO^Sk|xxtCLK>JI1ok8`NY%AvZJg*RAa@A5&H%#@^YQ zp*F_QdfrL^QEe3Jkr>$FZN|TuX1F2;CmR{vOVjX$H*$js&cT=|#=PyZW^wX^7<~|ImJQa-m_#HJ-Rf=FjBr6MNg8&OQi*J zuVU%+)KnwTk&vboWhatH9yMqG-3slH1tEkTY%40d$NbD``Yo?>pqPXD6T0P8TqxfaR6AD$kleAQuyzgDH<+5 zERz%Idg|44P$@oR{&#uYbJI>jz_j=WS^HB>NN3=RW`RRpxUC&?k)2%w1?GT5h-5A| zD^k{&6lhLL!g$S>{cT}6&a{tb4w*YGh_s&|ri5&oF9Qh83I6QKXT1s%vG1;+R#k7^ zZ+T&O_0OtOSyvbMb&QRSDBcjY_d|ClhXyGe%0F^y)F79f(HlUC;$m19mZMnh zEhp}6y$Fs9!iU_y&BY)aJVoYvl#`<&s>z;Jmiznrv&&0kE&RpBMbI-`8_HX@R)$lu zia^3JetkILn7*@?7!QNRpsm%f!ak@GjV`FtecZvLSs%RU<= zPl+wj$|TsC)lxzqpOf8!cp3XKBIV6^N=amro5TAa)xN@j^R?y026X2E-r+sW0+c+_ z^_{)*A>IyHq#>c2y1KHudZ0xrkkuKn!mBl{Kj4C%3PZ8=1-u#4EwSOb(%>4wsDxkb z3t`w_f|EyBM$z~}>_GbSnS{gHQxSLr+t2-4RJiMlNeVweOA4qeAV5JOo37L!J6dej z>h$#yiPP$a8*gBJywzaIymFWxDNC`S7}+mEA=uAP%9OZZuC3h-ly{VERMA9~+(Pw> z5^R)YB#NAGx_a*GE%?~nvtGgUCqI&~Uu}W!9nqkmK}^vv{-Ybth-y*Q(-XCZ3#18w z#j4TXV~GH#eTqoXVdG$ez`EJxOG-;?KN3@GaMFbL78e)yF6)^%HPEhI7g<;Q2O-cY zw;t;%1NPT$AMUSw@7m9_uuU*qZ|+dKlxQjGR!Bl1{Lu_Et;||K`ixkCT}lizUqR9Y zE9MCLO`q!*ez_f_%gG!1oG>d_w#UwFxvD--rz?lRkCsRB@kJc%iM*uDH+%l<%tTwc zSXZDXJN|d-hmSaKS72Y;+5cu-96QH4)LLMpeO!exJ|OSCyUgYk7w#MQb1q*rf!4>C z-^8h|A(k09cJlm-y${ax#p^N`qI{zts#E=U$kkgJRNP(lih>Cvy9aiF0utn>wg2AI zp{V|oP?|vhg}xnSHtjK~4Y8KZAv}qTZS1mImPw3@Jpp!HlUq-X9UAgd4qY6Qi6pc6 zgg&Cy?GI*IK?@!k9?sd#4e;yg1Ls&SR|UkIst)^id&9eJ{p$X&n8Xll_J&N;Tg4jd zgf(xY^T9hlhkd?OIaDli#8adj(LfEE6RiOG+n@T(iH>6~~h-8^FRJizBp4 znpwNZHR6m8VR@}y2JO)X&Lr>5zy2nPe7}Xdzojw=l<8~lEANi7gb~0Iu|k0;3dB$> z^k4!l1QS_~+xv&Mz)0KxwSud_GdhFyKYXb8%z0W_`y}pmJQ$43lnCj>SpM(6(VX$4 zCK+khc^F3mDsj5;p>BG3WTe^+syAJ(tx@yhT^4e-=XH{%-)$00QqG~~KN`Bb^}O+_ z*NMje(s>yyl^sI&4khenAtR9v%6N1V12OEk<-Zxmt>>wULb08U9a>{b=yNB2VV`X0 zdnZ&olgih8*R8Sp#rHhj0@p;!LHi%;XaSTrH*Gqv^^>xTGTJ5_7BhDW-dnYZS^W1g z)_!p>>?}k|q=AEuMuT%3JM;wEeKRauh8gZ;blK}=S(ZT1W4GOgv*5ToM=w*WB^5gB zyI8oq#P*ZFRF7RK;tC&pz3~VC}rPR1$yGeYq2L zvFUR7i@i$9=B4M@$B5O&Hy;luGo_$Q@h<+`5n*}${ZI_0!%9gyyy&6=w4%#Hk@|Gm zYo_jzf(999htop2=Rp!>g-G>pLRL|8)KZ;&q6j1<3c6o>rNTLu`EMYVH9Yn9CLk%E zwA##vR_9L8g5)NrsL1?5K>ybT7dhREI6@U^-{YZIRn!A}<#u0%x1zVjum^}hX?Pak zHyQf2y~qWGrW7`_*)|U+qxW>qlj&Jmg&>UpeGxdcp7h?s!$aUl)Yo4=z2PwJVu?80 zC7bDMyR?mf3M&0+3Uy-sCB&Q+Kw!(Lzc4?R+HWrvUTS|HG@Q;O&^Xdddv)Km3e0^A ze4{LKU%DZJLrGeB>Kb_%N5Tk`;`gxw7~9x`Ly8lx*z(M$zWnoCsBn#Fp)J^jGfaH@ zK60GacCp`tEK7+u_FWp=Dx((g)ZN|8bW&p?8_aWp2z@cJ!>!QsK&|Zp`&40Fv9Uwb zTWQ!mNB2el0X51a7u|~#Dp@#P;N^L;?VCpPTL$$PVL`Zcrzk19J!{Gn$z)n z|9%*ABUN@4O0iB!+)fJ7ZR7XfX<_+Wec$+P3d*FmHZd?1gg2brsF=xcc?p!R_bH^H ze*DtCQgS$_vx_CXzq#FX5aZYVs|JkG#Phch=P~4s2_bYA1GThCy)~UZ({*!RK%wE!4 zayK96Yc$ek`-8Yep19_qsr{VFmLF~jfklR4OIVw&zP`TsrRv16ZcSi|?%x#Ej@_c- zug-AaC*8%Q8f^z`D*Cw+q2gJ~t}4hi7lUf>Mq2oZCMM3WjysSG!vJ)Sm*$e9i1~Ie z-;g=dZE)O%Tpbn(2@<;^YaxKmR2qzpi6jO+D=rJ#A?rmIhG;!cevaioME)N*frX8a zjE;i-_yACl!E@+%pY;9LmcDB9Nq7gBd>r6;OE~>_O~CRoL|+;)mPXCVBd&Dm&r9Nc z!lE1s^eOa~Ym0w(-P>q8~qhFT8o?a?D^7U>Wa?O7UW*ycz^5&K;= zaxeEa=Y!uvaMRwZg3 zldQCph{V%ND8c`2P4J5AUEnfVR~Ex>YRTGy+VlQMdkY+D!(VUrTjS5LmQ3_ham8I5 zzOH7|8OWPv@_o#SQD4SZ_@%&OV*qwo?cDN@{W*Tn_xtrrT9gtl^p_!LjAlr7X%TWZ zDAXsH$M?5idwIQByp|)0#qjKJs!3ZXX?T>7~sZ!J`E5`#^Uk5 z+J-M?%SAT{!2gSbxwv{yW}ItDLexH`u*59Efb-!vp4&bAp4^r&bNI!@NM0BE%I&>= z$=UiBCeb+uClbo-WiV&s=eTbmaqTm5Q%`d4&s^8Y`RS+$FU#GxgX$gC6wh5<3!>f* zwQ@^<$SePJ>=_X_E|YWWT?w0$>B}wJVdwq+NI+!h;?No8mkFZs5o?dze5}hJ&l~elYcF-Pk1hP_fB3DYBgj#;`3(WUg+-fx zs@HU2v)#hvKh=^#m!G^fJynN^;F*lz6o=AAUU^VkGI7rG5`Bg^>f&3m) z4R16DQ8)eBk_|1#3xxt)%f#44dqYF6fl#c-OzmDhRqgN;D<6Fpi8c-O_16mlF`o|( zBydqBBuu-vpY!R)S8zKVaaR1aok{m%)inzERAz5+j~#i9WdtY+d+9yJ-nrIj!_rVG ziacJZUJ2;lGM8AR!@~iaG7Sm7iWg_)#SEqXu)bjU$me)b zF3udFv~#Ez?U7fv-)d|m!$kN&;{Xg7_jL8L!RMrpASE-iATRIhb0qZ7cbwX^RM8Oh z3*(2WnHLWSbic0&QWU&HV}=3%pD-n2@V*z}sp-A_v_s19%BqO5tFd81nwEzGg;27)?Q z&MG8YR$PVRZg_5=;|w=A^FXadp2_PHrpb^f%@O=_wKu6pC#F~3&JPED^>LI7d&#$@ z0Do&sneIa8AG<@^jF^wWANcV^(N~<3*64wfSkEZ9wbA=#URGfDtX3g+@oR2A=35br z3DFB4`n_VscaC&`5yQ%eJW6s*f2$-G6ldMN<4|Ev1 zCmBuk;&0?BOs}UBCVtQ?qf<)MFXMbfO^8Jdyh_>yG*jFJ2Bck3c3=GTA-5UZvgd9q znh(>MOD{EpUDtH7dMRNFW z&sZ*^|EI;~)kWnWm6LHbj6%f$M%@ORfR94IQ%2Nozgw&vU<}|BRJ-p7<%0TNT2_Lx z{Jx`p-(`Qiwb7ur=0zs^b|&-mB<*LFMq|f;&9zA{i@i${+gH=FvQ)?a3RReMcY7;= z<6wH)wA@;O277yZ8>B($yF+zOvv3MTxPVD~)+VNDoX4O?Bn(Pk?q-|Aop4?jGo)uY z$Ps^UiI!aXExcOqbMF5vf2Rig>8*;BDunZ@t8$FuXxW2t)@FX!YRa1RxA7 z@^Md@^>99Ttntbysi9OdltsR_s#+?DwbP4my)gP~+NMYS-+YHx3U64|J@y|MP?>h& z%n&(*l<%5&wty;}a)$6?cm$j%C}{`AeRTn~Mx@$k*U z@$maQv~}@D8`-p*^YLnrI=BRw5AzCPT~y<@Qk${%@+x{jwJqp=)bQw*%-!CmGyYYH zN%khbJC}t=Uye-8iY)YEUixB$&BhyKYu=BXuTPwjbfFP8+wfQQZWRk9@W)G?*PwcQ z0c#dTCImXFm;eDC{HqT?-=8r{XZUUM7Y;!jA|-|YZP^%eh)E4H9P>t}7L7;5g~?Jv zitj)(cv5wz|KpCf!bI=n%u=mGC^46q`n{*+-X2Q&A5~M4AyNGr(!V4>mBNQUt9?%N zrj;3J=4`2W#*y}ic$WBWG-mdE6M8G-wu1dVM{>+0US-!8vTWoojt}`#kg=pI`3h@W zb93od`GVi3%zv%NM#o{p(H4HU>LmFg&aP{A>?gqrd1!$-Xht6)eu=PQmxt;oo1K3J zCRujHU}k)uD&u?uIJlwXG{W&p`wf-qD6)aVGj7hX>-JmLD1H=$2RcC$sTcKi<1Abn zryrEwNmws4mEB>s=6TH)C6@RO2%@?ME7$A(9HTLC~T${=W6GdMc+Sx_< zxM^+#TEXmu$FYV8yXyVfd;Nibp3?aK*Re%6Bli!j0sj`@s=B6Fwxm$BN9oFl=!=ml z6aM-h&C#c+Ltnr}F9j35?%kaew-#dz*rBXwn5dxM>b0-$7o1o86Y8%*D3gQVX47t! zsB*GrJ_h4V_?!HIP#rKxfaNbSzObP4J#xq+6!Za&e4!!{Z-wX;jv<;R$#krg^(?^RC<9gGz8T@EAtK)KCt1dDe#R8y*@34XQE{KEKPNsYRVCw?O21a-Cf+R zSC%Cd1x=wD4Toe5**3Yzw`!i@n`BgkeHgo((@kp#Ai#QOTy@T3PriME7hVB&9N zk-Y`UvvaC>;S4XZOK}4V(+%L`9bG z{rM3}m}?299V@Bmzg`VfKQ0uUi_f1vo+>xUte8~j(i?YO zjv3TM141^!2?lkGj;42kGb8jaqiDbZaUQk}zt%i|k{Hgz_FnSy3VVz+rJ-M(jeyZ0 zAWJKAUC47-&zj)kOZAy4B?xF}Sjaq(*9V0B&0pMH&?|ky*+1ywb};fVMtHt+pOig3 zI6QA2xilWeEAV~hJ5eNQJweH)4%$y~I=1a#J^3hq3lm+WVy?RE#>REG0sDISH)L^g z6gCrAeU@+d)I+ujdyh+e^AI#|Ht<&hU$;1l(t4pED%xedDYKmpL#F5GQ+rm){xmK! z{*aTccRPt6tI!rQbdx7e$_7zvw2D#=5)lUBxOFc}U1q#P`Y#wmZQ2`B2KTW*Ul+JO zwN>WQ2QT{f=h_x#bXB@Qrng6+ZBl83l#P=_O1x+ZeFK6WKK7sCdS%(Eb$+*Qnz`7P9ctW+%#Ycn4V=C8hjy zjuXbO9&s*Pb1zTb_IvAdz>Yen*X3M8yE%g*?&$FFQyab7VyXZ*^&hXIAAbRZnF%oQ zHXYOqHNi%K?NbQmA#gS>=KTQ6+Wu3lYQ5N!@Z>LB|4Mdeuy!&xr)Ar`i4WS@hJYe9 z?c|1EO$ptrK7d$Z2h*e*2xyx~*ryW;>NBlx@q>mb^AEwhhpU#} z0$LfyNaz0gPpep%x>F&k^s1|3>Xx!i>^OBvH}V!zNp^cY_(k2n=pd+Gx)I-<;kCl0 z;q6_Jh1WMu+DfWk)LC zQPbDU&d2*Fn;JMJM0M+ZF=uw&zI-dtahA_~l7rzB@-r9%wu<4Gtp9MD1PUtZf7fT- zZP0k{YcD??o#O}Ayt~1yyBDOSq<2Tq2V0{uMc>o7yL0_Vmo@0#-emnX11=Evgn20Z zJ;ydzV0jc1r}C3}>w1+OI)(=j1U9=ceRaKj>)8`p_DFO z(C-c=N@&s>lAd-;GRZ@JnS_Ea6ReK+6#^gjN1CpB!#~_p$F*WN9f#-Eebe@GhyAw1 z5_vabiRs@TYJ3v1oK^{0X>bvC5|DI8Q@6{eM%BN_Bj(DcR@`3tex~|xeeu919|ugj zR6Q2A7l(JU|oQwSkj$1rGs%LLhXv ziqQ<5OU;{!+KC4+vWHc&t1cwLt`V+fw5}%?-Z5}P!0Ib~7_r*Do@V!IKxp(S;2?tM z#fm4uQJq1c=WZASDXXXe?xI4v{V?*U^c4P5t_tO{=H1F%p)8**HVU3T1}9tO(UIi& zrP)$b1ODaja8|XWBK=srGYX)Rexe>8&c(gDf95Ty?^_;)l3cjh-l~*w&Z46Et;S#A zv$uGR$laEJb6AzE1+$Lx9@h(9;e6)~!klXN?N`Pwv*NfiR8l>ojmW(E>0=a_>ls|@ zJ<#%*wTX^$?QSz;@xnLQk{W-FaMMiePo3GvcU8MaFaFM4zfAy^T3`l&f)OS1<3~z* z?B*vmffF1!m#>1Jj9z0Cc~VnRZomd>+p=3AaoPTO58Yl3CR4cc*%^QH+G7Lj;TYrb zJLp?xdY^uH!fwmpnAnS*7h3U&Q^jDiJuf?0NgddE zy~|IoT+6HN>W5ib{d(#&$1qH1Ox<}ot=+?CHGYSao=wf*y8T? zJ0Gm73tJfe!yWIxxL)eEvxp=pKJ%P`aCj3j;H3Ykk4Ov@tRgm%_h%AbVfx{3Tum5b zjyup$jX!<)SN&QEA-t2jxkBS6Y!P+Wf;_d=M!VVyqVA3;N7kE2(>sYRJB4k#3Wz12 z{qv#aLf35=z)+aRzTc7o$`zw3^w&6B5Q8N%3x=Qfs1iOe9ADIJN;)$THzSYr4Glvf zn7~3_a~cqTK+r4RCY$&6=n=B#z(Ehz;4g9A&Tu;gM#qz1Z==VjrhZN-3XK#B-CF?i z)%0+)AK?CJ))=_w!4X|pTIY3II=90YM(u{rL@n0;C1_Ll#o2G|Cd4<9GcrP)bjAOG z8Z`Bl_r&&YNt|r?1BwEdC1|zvyXf9M;I3~SGqg0IbMUFoR&O~n&NUMMwP!ldIq*PM zGf6i{3a_x>*05GNt=^A?H@NFv1OQ7rDd|R_@hjm!{aS0(Y_*8%+TNxgKa5c}ff(n( zvgCmZOUW`yKa$`vG%pcx`J1~iQTCC^c0H&q&s`RofTgnF`gH$5GoyT2FU^rA8ehZGyOZ9F1{>q~i&!hI3aiH<=1w$WFE6hnoOjOwQhN4#4rO%7jbqy9 zGSS4k5}8uO9R1Qq!+=}%@YDNMrewc=6STisKnqOV22kGjiWLDW_u7_c2}wD7Y(JE1 zk`w#6XdPdY`^!8I%OS8Jyv%s-y0KRNo$B0b-qq-(r6cp$#ynMVCPPA3ACqys%u4+S zDk+GpI=m5y7wLZS0INQ4m@QPeaoR|-zl+X(vCc5(w95WNa79UweWIh4Wta#p?r%#frzv@^lapw_^rxdoz#il@K?Q81dgAz9`SaZCRBnb;Iedlw2@ zX)~(CBegZ9n6Xdew`$9$}cwaz}KPf!ZH2DFUw~BvpS~h>UOS5 znk((N#!+Y*pPAtA17lHTD@Xvl&3r79F7;@0^;?efE8a<;?=GMb68n8U<1A?GSPTOquMmUY%V6uMO{82^Yft7mF@h5$CSl7d7G+sv-JTD-SKA%G zzf_p2>7u>W(Lt!3UkGA#C^)7eLV4Twf6;W6QBiemn`Y<%5s;FQRJywx>F#bsx}{<0 zknZkIm2RZFL%O^BJ3QZdf6ZF_7}o4__FY%p=Bke{dwpibWx=$trs4|00bS5$NSx@! zj{0#KI~ig`m-m3K%AC!=4d5+(`0!yJMgI;4ke9x(r6oRkvnuXQ%*{1PK7HhN`md<@ zJ?0(tcIPUzn=a0}F{pTWkKGu(go^uu;mGBPzec3SBiMm#^ ztVKLP?UT`6OA>9A2~Q`d&E3#kkwzxXhIF^Zy!M4Fi+jAu)X8 zw@+K_h{!oy%3j3yLC_6iNY~$SpZpBxSe81$R`H=p+cvO!J>7d<0ANs_rHpw|9vcIU z<*m5KHwNVeprm2lbOSg*0Eu#K;Lc55eJ+lkWEX&Q#d6~Z9|;_|lsgJ7ng6&G0I@{8 znPX#3F$?Td&H4tE4VpdELXekz7E!G+GaXugen{3_)4fI+t34Z z3z_hJu_*MOx8d01auFq?k2Ch=h`wP&)@aM@Xi@fnNhJELk?hXNS~IA-( z6iR0&8Rgf=d8s}hE<>d~A(d_PR_+A*qIzVl2;fkY5GKJ2&9Re<9ZJiChGd`YWxbS| zJc~MOXHcAnkle@lh+r;m3jlz*2&bL!r~T5{1Z4@*^3c0FuN>3S*BfJulp&1rEq=9i z+J^h3_z@`^c|PB|xrb{ypV-xqo9Yl}K*1k`a1SO*HCe@YMy!iG*0gUN__fO`-v1X< zbYLy;5VUBF9?=bm?er*sJnJ;|lN=5_VOh5I6pZD{;BTChB_z@DlF3^!;soVNbd zQ|l&YWiX@exY*^ifzJ{DecqHHiAW3x~D+U?^zCCWsYuimKvPlWOrDS7D=Th;>AU4D-Qgb?#FbVj;Iyo_INBqw3&sq+UbRWEW#tR# zNfj@{@gDcJs-6AM9z2s^inpg24@s<_M^s&)OFY5_qH^h9-hotHFf&p<*wl`Zmd==r zq>4g48Vf_oL-p&zg23$;GkKVTG>TP!-fI^7?xp8JHj+DT7s9_tYyBMBX5zjnJVu)+ zoZqV+>4ch)rBSGANl7Q(*w-Qo&~zlq(4vq0&M@;WD&w+fzHSI^TUFZE@YCRd{Z9cX zQ&)!Elj~!eZnNhr_sqG(ySrM%E|Y?y`)JA0!tZxuwpXol zQ}z9Od6nY*D8uQej`ly#zk|vBZB6xzBK7=H7G!`g3;Gv>w?*l=3>4|>Uz?10= zllOw)e>o%3M1N6$sprOiKRuK0SR`-E@9mCScj`wiI5tYo{uslZlJ9L=i24|#@bhXd z^33zjDAEWP7IHvusXcuy;{Jkt=olt>WPp-p#3Uo0SX?f^bTPk@9lVF6lgZUgqp-2U*=!o0v2eMDaUHwNLF-Ta_e8L;7)@H0SmHM zVIAdbpBOOkUxF>D9yYoIp)#RRyEz+I%@;+`Rotijxg_%Wa&|Ca|#Cs$_69 z9L+ZvW^l)8^y-O)kZ{sQ+n+Uck}2sTvF=_f`%r$Q(Dhp{ZD&B%S>1MCwi?6t4(0xIB??YFSBJOB6 zHJfP=%n{n`N-?yh%MoopigNK33SJ*SWOUy!A%lqXd8yXt2mCQ%p5Y6V0U4B3G!ir~ zH0xx&J#Qw=Gv%~ux-`vI*@@YV!-U`!Ynw*7BQZszgc{E>^;d+ps z0O8;!4eMequyezQ%>;Se4=M`&8%`?7K#P2jF+$_>=th?F_gWlHAj^M{IOl)BI~~)y zg;3662!UL1MlbK;f|w_nwB=91hfpFBO-03)sau!s7x!NMw+TvjZEJ~Et7~g^PnGdq ztRw77#qqg1^qYU^T|!Pbf*`n_$n$*joli-RhVJ&PWz^k*D#3OrC1yCWJY^w{Lpu}Z zYB4v^2(3~D?2)mP?dQuQA}3GFIjaaKz276F!rp%HG&yp=CSq~CaO;Ij1oo=q!wvy> z2)}221GC?rJkl(c^0!VU<}(lSdeYB$QRkLcqa}so$+Km~iK;fFefHS)hhF}Osx`24 zM2Wo(NLtw7Fs)HFNoBnd^$Nf!Rlv<=6|3DM1;374N>eXN+VI5Ee5#| zhQTz80@;(-$9DI9S7LGe^#T4It8?sH?(}n(aSfLG&D_z<&hMWg{5=>@mLvh|#5V6P z7$9bLcvi2wSkJQARGdA5Q0SpTK-Fm6;IAkxR0C;S$LVSJF1rj)`SPI3AjTOIE`cWo z*CtQcKQ9rH<~gFCU>4Jj)e+y|+{GifivkHK>>XtsulkL&UN8P4u9dq<4`)Tl(Jr^O zKhR0#9q0W-V&b}wdO;zE+?a5%Nj?SREei)0I>L!Ttmx0q|I6a|`AARwVDX~znjd3PotT&(t9SBf-9s23y znVd%UFN@d1;I zO1TnMkTYA|t~*+dG?l}=PS%}{_G(MAQb!-h$G(S;uW0Vq$OnI9sEO9U@uK}6-G;Nw zsk_f;zNu&@)f1yzl6*a>Q8MRzg%(2=( zpC5JDzWOjLjnw&}p;xzqKc~_t6j!b0hQuggN7CrG&Ns|5D88xTYw8Ut!ZDN&7oie0 zFORRG=X>qY*`xX)AoObVMx?$oQ1K29!qv;W;@=d>f5}Zg zY^>PMx%HuEl7cSk2f{Tts+mACK!^TI&X{h9+XmBB$@^dSI0d&L-YiB`rV!R|{K$$c z1Z~kLP_|c~R{YS`9K?HTRNgzCgYR|U zEc60%_IQ(n8JtUMT2bYB7N~SdzLQASeX4OqhD0PkhTIbaDPt%vLZHVP!I51 zhboS~hLDJ}=Hg@rMT_4?%%h`6W@($`C7Z6HFSXU%D+JW)74?KiL*0n9l1n#H2haW) zxhSkO)6>`QzibH(YgboKU`^y;C)^a28`C;_^$j7|8ssFPpQ#9aq4;BxTy` za!Wt{r)33|uyIgu8K3#$S)YWR=URSB0IOn;%moK6`xV+2@i{EqVoB)ehRT=x{zmM0 zqIW0%g3+gyI_%6lE1)*_6UDuk9}$)O_T(2iOI{+$IOCfM7bHw4#jhBFfp-;2goA3R zaPzr>jDqLxS}N1rT6#Rf0lP!ILE!)V6KstVxZ{;Nzkxs zW!}LKZ>5H=52O8GudtG)*G08HCqR5XJu?V_(S1%&Hpk9s@Y^a_3 zTXTdk_^-&I2nL^%?X~&pTJO_NH{O~X-e#Hbe} z_AGl#R$Ca^Nu>5H;x;RVms5}x^sv@=HeY90+X`p0>KfZXR#4)YLgpCOUuKp5Dg2WG zxX`nZmgMGl?c8L#QcA-O?VX&P2Bxi1bii5XPuRFCf5GATR@czb>k9aU0UF^4_MtSP z2bL^R8Ty*$$n(e$tjXKc*V)3oDiMYZVg6^g&D*#W!a0f2*+p}jJ#jc)BJwCPtW^eZF?CgLyAD~kPq|70O86Gc318~(J13AjU z`TaLQc5b``9iE-52YMQ4fN$l}DEG&qizZP$&vMPbM&7YCe5m_^V#|ee{+T8olRN`8 zaC_na+jE#SYRKYq2xw|i?277#b5GzrnS_QNqd=dEg_tcP9Ycg|3sN5<4TX}lnYlSn zTh=_@{au@UgO;r*`k~d%AC2u_4=mvok@)+bpv@YIKXGk+UUZL)54a)4hhxUrw?cj{ z4w6+|{;OJdq!7&bpC_Vn$NEwogV?F=tmxG5{P9XjJ^nxjv12?w2Jla`u&F`}>B`NT+|gx+U-CPOHPJxYZc-5PJo5YdpZGGZ;>Ky~n>>)j(CFO$=^ zOe!{7Ii&;n=UWK;>}390p9bI&VbV;Z97w!QPT$L~RbU_)FlPEN$(eU--909}s^JyX zhu64S>t}5KKi0y*xk4KV$73xuH4`M!0axZY1nNv}FQ&tGyb)d-6Mj}dW-crvihlUc zFF@gLbvcKeZcF2B&&rA{Dl37>K<^y({8Q7qA&nu9`L#OvYYS1dGcOA%ev|EkG+?uE zxG1R^&b^)~VXk#hRH@B;F5D~;by z_}XJxMDC*{3aR+gNXx-G(9g9jpt?H8lIqE6f~a2xV#izi7N>dzY<{Qj;GlyovzC4U`Cq!8cQcJ+`G3=2vbv(rLfX~awCx6J`yXioX z&KA127p*u2gt$?svKB5|@&z0lkge4R%JoQ3kL&KH$+6jI6_mhIp=rcI zMD(`>v^;#_#5ow96Lt)=PA_VpF&I!X%E%JC>;51CYj7p^O3NdWOo#akC-Vaq0cj|R zqQgnf*jzu=rl@iSyZZDI&>HZ~P5q(--Uq5cLo(kPtAwE>hRdrdd1Y$sUw|T9Lr2F( z`-mG!_UdTg#E*dDIDN;=4ucd9GdBosUk6jqrS$VXtl)JOtSni0fP7Jnj{?Gw-x7q`^Q*aeXfs-1ba4?M1qG$$K()%q%BuA6hUk4?a8RbEkPR>2Rl4#* z_Fm^l=d?VlKMMXIwS@4z4<5U`8EqjC*X;N~$snh1$o4%i*?L3MM+JBz^mXkDJxzx_ zRj-jsA)da(j%e8y_cSij$!suqJGj5KoT#-iwIs-KKc;qZeeyLH6{}(yDyWI(>D~a|L2spa}uMmR2U3TD!~D)~lIkod>z-&MM!2 z{`OXj()Xtv$v)Ft`~Xs^mUX3G>)`fK(!$cx=GFexGGOOXl4+73k2VyX$h12aEee%J zImd=iUD_<-Iyl}-^UJh^;ojOtKUv7;_M#B~*{r~p09w>O@RTY{kCfuxhMn}sWzj;G zD^dlNq{I`x@#l{e4%vOjC9VTUWSF^WkhSsO#pntT<6w{UPmQ@86I!G_4es6UIY#-Z zsuh@r)**1Gp+iXf%@@Ohnn=zMs19|mpV>%7$sgw2Z0%5Lnp3vX$agfq7qagk$TcQT3LP2Y)yTbk0ONc3Alco=@K-R5l!KZ zwf<*HBy60ul`8MELRGSsR3BAw&Gr?FiY&(It>5vFkE6C~r_Q6^STN7pzSth>u^;~J zxa-#YS&u2k+20V)P0#Z|aV#t76kT9fDdZlkaiNd4+^eeZOxFFHEc@)(pzQ*U`iTr9r9z z`W=)q9eLAi_Nst7NCDaV9DAb=#Yszuc}f}L5U7M-@pI2`1&#+*I9|;nLB~`$jGSgC zA6+pSz3Zc%cUf5YK(Xr*$2K;^QH-pg@~vU;8~4Q!PgFpkNZ6M=Xgb%FzCOmf^y>Wt zlMGbXoG=fiCTRwE;(330K2h!el;{7eEYBxK`Ez};K;{b_9oqHu%!rBBSQdd>x!j&7 zId!4|7)B#iD3jR`e1Vwlp-oQs-X=eU5+nk%M%6`boubhgz-T-BHR<3t7IWTylDHJC zh&rS%9d0&RMY?F=J1-czv3mT#?V!mfw$pfdOR1W^Qh}+m_OokA7;P~7xhal3tsYhe zKRe3-l85m>M@gFW*NgtVp8|rqE$F{fm;*3_1}apG)wJ}RP_Jw-pG(uYPYS=HXf7a- zGNM8^T*`h966b&f&SpDXhiO`i5mJfTM}N7`7!k*=5yFa|CU?6W^wMT?Y0ocD#xA0K~t1RQnEE>=I&Q3C}ksIsJwoXU;pC*$b+u&JD#7t z>Gj)&6F#daXJl+QE`2iw%DDVAXLT6P*7QLEI}_Aug}r^vUzX@)EhE{^>}0%9JU_2U zPg=wdYDh}@Py(RVFu*W2{iJr9)LbC~EF!=0Fja3D>Dbs|$RLnuHyou=5E$7iedGFqa=AnYv51ik#P!I68IWZ?B_&kkDUk5Ea`s$DFzO)2!7}b>0kS zmBx+|@2k(Cun)EbQ)8l)hUBSL7$rQsZXU5JNfhtDSLvf{d|y+Jh@zDI z4<$engZTL{{-I)1s|q^3-d`ruRjE?>RxN$ zcX6}cf+{b}m76@d6206X#>g-F6=Il?TNiuvx^!W9 zHUIf1$JW&vQ?nb^)XvGV6{U7(okO_qO0dc3KyHdqkY+o0ES;fkv~>hH21Xdh8I!QB zaFlsig0W-{@#3Idb6hZoTmWwb30miCod3zUOW#aKRjnD`Ct=hnlAnL4Td?65|2kB+ z1e&XppvSmi5~*5+u=l=UK*&DoQu>zXCER_3L%oyw6LK@x6}oinA1V z6K)H$!=|+LEe13m!WPS(7lQYu`@l^JJcWKIFeL|Vet*)aCPFhL`l+z(G{BIRi^8u_(H-LWhQP$$b6FN@0sVk@25RhRqSXj0!x3LON- z$(~!P+bOmC32UREw|hW;Gm><=n`k)oQK$rO4Fl8-HB~>bU5h9FJUtg!_xVUJATi`x zgmzH`Pt%C+gc}c9%TPXSnKCkL>GmWNl`Hlu_i(_3Jl|Qse+YABEFElVjV5ZcoNSx zEF+mLg za!C+<5S9~(uZQ7>^EhGc*vs+DP+AMBfHR(azvz0~aBDSsMQnEWJE9l1dP`+JF0q;8 zbfsB)_)pZari3stCQ(ox*&e*MN%OQ@)M-RJ zD6CiVYNt>(mh3qq*O7VBk924(D@ehjrH3yDA9sCTh z2)cuYN4&wAzx~|Kzx2svNg7s-4JLpQbt}Ce1on!%cx!7z8|!+4XjJ;dG%>Y2*t)M@ zFv@BaU@7q62;)S$JBEjih<~@G%?vx&;z?j4(vJBdX|JvI$LL#GH_a^vGem7#GsA%1 zy+0!X3IfcW1~X63G~n3k+S*1hE|{y_N35xF!s(VGCy^^ZdSD7lqhvH|M-UJNl6cs| zpU3GO@85#{T+M$CT4n_f@EZF&w%DDYJYnyAqZq%wh4ygnBuuj-`ruy0mqar0P0}MS zCr0kr(lsTOUq3ltZy6_NCkvHy45n_l3uULWOI@qQf)&(5F076~IV9h+MG*dixp+T! zCobr%xng657It`4p)<0T|0SYl2AY7-wu4v3p8LEtiONE-b-u5b77S)CFehvr$?vb! z7Y`)%=bP>#3|{r`Ey?$kaGyS`ne!<3oD|5Q!=xX{sesXX_{3HN8)rZ+QZ8Gke{egxMfY8B`;^Js%si4Uv!=?PeSA}N70yjl7uv$k( zBw7Qv_=w}}3AxnKhs z6@jeWPCG=OCJyZkz~(y(EQ2JxI~hFHCeWWEBP6M)3+@io+Xpu`$SM!nB{u*U^8n6S zTO(j3Dym_g;fBu@Mz+D3ISo|^zz|Ws4!?&72a${9vnoT8{%mrR`#$&0qU;7!`dW&T zK;_m>74J`dHV5ObtRhb<&IbLG>%d~O#S6Y`i&nz#jxllk9krjvDbT%Y4T7)Y*B0}+>+maoy98k*0ze+4LAuEqNDS1Zl1D{6n+C4KfUYoNf%XAmb zmRX2oZhQ}JRX2zOD8g76VW3s+QySZT(^>F13AH%k=tf}1E8dWaCdhdf_7&T{C&zU1 zL2oGJ`VQwOnHxv(C|ud+GYOVMi-3wVV$-Zse1E^{-L|Lj^>(Vd;0J>#vK5Za~| z;(d7hH8q?deUTwc=DQDc#N_{glC&hxqn`-+h^;Y|$McE(k3EEbL@>wq*Qk`C09o)! z*ymi@!I5{zwnQFx3v%5x?U7GNXNjYAS+|0yjeE#&ZCt%-dU?(tH0#vPQ9q9PcF)34 z?vXV%evc0#3iu!f84wXo5GhNi`^s|u6gD)nt1sN|aeIzGcSi?(8P8HZ!tHww(>NjwUC zdmrkg>c0_!Vf=3{CAHZfp7wzZ2QED8KZK`()EHej^RZ+*Pbl zTUcHF)>7F3{JnB(+*wz4ijX=?;}pVH0xkPippo*Lv-wWPcZe9pe4~0e`mp81mTV#)O^G4kcj{@IuVjjR#4v)T(Y?Y*yMz5m@!jejd*qkc1Epm;I6nx; zBwj}o-4~iWx`2+*G?$it#vmkJSy3JrGW2;G{;&w$aBqB*gO)+*gtDmL=r_mR$OO-R*a@L|- zdPec_kl}=Ejm2y3EYG2)hY`&n0vs@h--N$)%*#?9xL0*T_T2RXNIWtnu*L9#ClUy` zqXITL=`OE$F-UuF)AcgN{pHQBQX`V7)Mvu zdI!fZ1O})H%5^fonkQoYQhjFOP|O$T#7GeIE!e=@NehupSo#(C`PEW+C)R%8Kr*s; zOhfjrp9tI-0J^9-*aw-2;Ab?ADvJu54)e7@g**7}lLgC6E2DbDkI)n{BBP}eG}czE zH2w1-^QT}{!=T$%+U7lKxA0Lz9G7mof6y1^BOT)+M_@pVIo+NZQ0_z(vsllSoK?uk zfT`A8Eh06cEij22rx2Z;n#iu`nDjBFa|pX+VAfd*^tZ(kF&Hl$m8c0V0vP!HlP(yg zJXOSQh!I)4rAZnB>Uj1ah`?l4-fQEX3hjGfP^zh+@y{{qkr1fVb^*sgB5R^IL*Yva zpDPJ~L4Hx2)EW`(ePsOR6RNxW$l-Iz@icn+gr3Bp|HfwD#dtl#aGy$BopyWu_s`e! zN2LDCw*EH)8S9a>)ho(3n%nulyiaOV`}1+Y%Y9$)+bcZ% zUb%n5r1=GI+PM=4hR+rK{Hh&|?MIbL1^p;kKUpMM^{kBgb2#ebPr+Db`11HJ46iPd z9@ei83pyjV?LR4`M{L(dzgO1=nqwpxO$aJ*_7*A!mD-f}P3KiS*-n2RM@%+>h}Wd- zW9G1#W|l<(EK2;W^1+Spb0lHb)MVNymoqVQp%AIi8Gd#K%NnaAg^0Ha`?x({Ng`7w zzMje~DuDVJAOG!M+rjP_=9jkBP9kY3r?A-D0!5O7e-g~tbY4qND zd#&xF*}1{4J)gMVcVhhT-F0tyW4DJ=WoGNpv+(ha0KQzmPum{fCWm)#1npG-Q+GIR zy}bYuRciE)E5Kqc(r)W>#7U6}_eMAk#J_#?mb1s;&1W!NF2QBr)aP=qq0StyYI;yNmE#nmcsrGKeR&^B||Vj-h}xBa|)`!t)l>9hkDz^ z`!?X1&yIoROTGZy@uoA#blX5uQ;T!nB>#91I`x=qz6;X7gNHh4z14fUv3pHyzYA~w zm+5(8`Is(n|5ISS-{-0S%~lz)r{ z@+)IQISYI)3hqQW)`U*qc!5kT{dP&>P%#+_>bnw6rQ-qwxs@^8%QMrT|W3%{9^SalmI< z6=1ze)c-W7-Zojw*Fms$*Oz#F`fC%N|8ZgL0_d}{8+P7Y$zvsm*MX}iVw?uRG-FT0 zV`p#d>VksZkfQxhz)-qa<}f+{Y8H1t6Yl%T(YP^(1kCithpRC7735cAj0|nT_^KP? zruC{gk}xu_H7YgarA8VP=4bM9W_JO4x@wkPaNm1S$kpyUO3wm!^kHGfrN<;BfGO!G{rPtZ1clfrqw zt&SYHiRl~IIt?tE5fZxGuC&Rc%N)fbuFc~}tC|Bw54j*@#4z~QkwKaDf-oM6!lN5 z#9%Zy{7zg>=jZ1EfF1`IS6feCM`!CzYt-)zv<^Jq+W&ZHggn$ldq=poKdzw3aMDot zyzKgnM!)Puvp3$4tUY}_0SIF< z_i#w6UDCjH5H-!lof;gH1#J7P`Mc<173;62N?XLBP?E-tF{`={o-`|B73tj*?BKp5 zaVbTU4wsXlgd#AfhH+z+HpHi>oX?JqY4_lvCZ%|<_*FMx0WIUE{yQiroNBmc_5*Qm zY7Gm}h55?S7~BtIfDk5n0{QVeV^O{=+ui}EBi=`W?&q2+?qQY=t zCQ2sTAZZD=NVDF|b#n}&upK;~5|p*zR9Ugy*`h1oJCsP3r@zaBC|lzNmxDi6wiBIJmf}Nn(pf{R*dnW5zvSad(j3L)(b`rHmQ`f|K>0Iw3e`WHKB7%e4SZgkrc{ISq1TVOA zaZzjN(AwHscdAxKMuxAcO0CT6AMD3}Q&}&pZ#L!7cR3#)NZam5>@MEo=rW+@TK0K; z22AY*3;>+SG(8KD4{ibCmaR_2kAznW{r*6WA;giNdzQ!yM5SB*Roz(sWUNfJH~v540I#W+|=X{_cOCtqeErW`erE zW<#7PJ%DZt_B`Rpe-5V|=~N*mjEtu_JdK)$qIE^snCPLZ$v~Xp7RTC8>vCj1;z8Qa zOGE-|s9d4HU&`6g6CLVax{gmJ8#*suuiD10-gW2WgR))PNvQIocyn7OxcMwi({|U( zaFKc2Qnblm<$Y{@59kvH?4J7USoI#?Y@Mz3AGJvS8qM%zvF3461Mcy+XEN9YSrRcD@_p3WK$o#MpwU7NxW@K(qz;0{w~+ zYRK<}|7piNeqir@Zrsa>havqmL!ZdF*_FDS0`4&GQ^bVKFA1}2Hlo5dRQpp~m~grg zEoa{_&gEp91WLC0O?2hH%d__Ptu}UiJZ`!KG!5|%zn2x@U!TTa->~nE zw}K3+{pGNI^ic1X`@fV97=Nb#aUuHcDLgG?J%4}#I_kOB6XzD-3(0@}{eH5{^`DCN zul$d=Xs<1DK`gN&Y(rkpn=iQZI(WNT1|)B2`^GPF-%T=KtNSk=3vh~lB)Sa$*n9YM zSRxnQgR|wtYZrn|zv78`|4n5-(&`zLxQ0;nPM{~Ugh1vdXXfUo%C}9e8on2ihA;=7 zFcAyp55gAt1yR?imd1%~dLH*w-Hv9gB?YFP6vp zPKCaiSt7M4pG1`j(D&Arek~3}oIIP$=Y1g<%P?MxvWjiH@#srVZXIStjDmNfWKL{M zR=@0)FI>~9QiNx`iiU!DnwEdYK0G{p6AaS>jEiNLC0nl($1H0QA5fqDP26k;aEt${ z-?yC>mwp&tzK2!qnQy%c8it_n+qthjRIF{hIkqn>E^Yvw(3|k-$O7)!tLrUoWgdb^ zzM<}rgZu*ZsWI0?LvKH}ZF$y6Q--(!OCNNnC z_viI(ekuOa6m*%c8ia&N$Py(E6;ko4j55TyDNGu@>}B7vg2@U{M@7{eE`UV7nr zaXi%@=s8mouc=ih^K@NM4Pj7Pi#>9o9rneGaizx!Y=cf@-hX##O5Ki3ib5+^*}uwI zA-PAOf9qd?-{Op=gaMts#SQDhCS|@=p4D^o)$=G4tI`XWp0et@8hEvC^p;?>#1C_O-;yAsJxSUMGc z#Iql&Z>)cuS7DjmEX*|er8<$Q4cC^XOYy10DYyNY6-_4Xu4IR!7b92KpIW#>u&0U% zt%Gbz910#LoUA$6K%@~6qBiZplZswc8+W9q8of#`0@V6Hvu%u|LH5`bG&|^U6Ka%u zsOq}V%%XeUSJSGR>zAYNWIu;2<-%8n%M1YLNPusxgH87F^+fRTzy{jc42gVId_i+- zwys+~ALuF?39mMgyXLD3JfnqkOVC}^SZAP0f30bK3eCXE`PtbdOWW9G?L>>rB#O`; zvcZ`hMkRdXI7FS-J~C3vtp~gQmiFY~n1wzA{U+SGPML2`BEpZO!w?3L0O!I5GWECofIrN&1tn}5}0NqU6DTeIJK`{hbi7HGN7KO&a zMpCs*2A?RystC)UnlWP3vGSehXVzkI1a;Y(5;q0>08AY>W$_;GqS&I2S^*dReljft;S5<^TXidpz7REBfhh7kl zDOAad97Tv^v~-5x7{4OB`HTD)>V8*32c}~EdU0&@{o{swbaF%A>eN6-{G;(Jdzx)@ z7{Zc;AxYGi7Z^yOlvg<^lCbN%vCq&6AyQI4~jQO9!#96eX6iu8l$G|!Xq#naNBc|)I%Jilw{6?u$ zT&0s6rD>wX(e%65*p7T2)MSy0L1n=knl&PKt>&5wfCH4hZjvnJ)ZOowF;*%WeF4@Y z-;wxz{RRgFN5dbAy~_LUF}8vo8YkO`Ag$+u@8CUEczXR*T>kPTsn_r4Iw zK>_C!y1`Wioi7d*BpFB?j&OCD;`s@t(2o&x7O)lCSQV0mJn`GTr!&VpfL+&R`2)I6 zRymJc8Q9MfVPF_}u{rh2M3Ax#7tJ#YPt^QoD?M45VSib-#URF9s6Jk>1y=y~af~gF zYd&P$#b10gjtU!rz)X<3+)pJjz{}{52ZfND=|cBEsSmK2x*w&jXv}1^O~is#$-5k_ zDK7DQY#m!{%<;!A@n7E*FarJd5vH29`PC`Zav&prZs;TR*P`&Ly#y|k`V}m=sc*zJ zb+g|6*cS8cAiq?B$blZoNhGOwo=AlNu~*h=1_w7_P)Fm76()z-!pdlZ9yL5uBn%0$ zGb=GkB&AqjZusl3g=a+^@J}lm%B-&6IGW|>n%n=l0GGl&CJ6-4-P2I~b8z4x3Xc&V z(Yv+7r3qMcyUx%MeaU~1Akyzdh*PynnE1LZaiTs=;T zmEAxYM#ejqr&Yl&xqfrQKF0rO6F`z$xF8JLb!6fW90hRm#^ZQ@?#}gyD zeTYR)Npf9cgBqC!2EnCkJ4}IsBN_fJf(M;{VyHR7750_pv?TqXV0gHGjvOHLmIvVH zL%u!4~ga;VS7BV#DYb z_|-I3l9k+DCJU_)pH6IKTMDVeL^nza4~6UKr}(fJ7CThw%Q7Hqm|jcs+X6j~wpA#k z)n8DCYN@f2ZBIGg^_ayKHvns6GS`8PG&`ls*wdN9Rq!~`|nuEH#WRhSxtoKSu0yfu4gDUl-!9Ltf#IHAsm zsFOa6_Rv^kw;LZ_EV|(scTD?m`2T8b?0Zv?1LGIe?!+sH1}#%h&*pPaKuYF6Ux7Yx zV0?Y$3B~t+4}UW8 zwX5If%LwfuOv&T}V8Jnq)?6bXGd6pw8QgKN^KfX_$jwDH~;RB8RGjhw|t z??$kFi^UCu_X>U-e;=D!r!VA@pSCDsHCSrYW2|*c!Q#2_)qtdSarQAu-&vhS!H(=n zd_-LbWBf{nMZ+9W+)y9bjaAAG3aQ1J-xGTA5Hd0fh4Jj#6jVta@NSwX7wy4L7)oUF zsIU1@$o{<{ylovOWA^J-yAvH@Q4w7zms*3gDyp$+6WnmC@qt3io}Z)@RyPeaC0XjR&0 zNmJj0rfb8{M|(>0Op4vo1+|jWA(mo!gA9LjObZ#^^}yr$MgE;cCb76yURE^0J05&w zDvznj#XLxYcUHz%viDJEi444UONH41{l;2_Y~uJuVCgQGAQP{p8;bux;W3ZPvS-Tw z6eS{xH*)|l(LxIEX^{L0{sq`{Et5pTGZLjJrP9OpUL<3qxJI=0A@mCZ59C3gG0M1< zjoDzyCEyNeA4^NQoQp}=LMNm5P>eA;1`d`$WNSX*{WKUeqLwoodw1bj2B!IAQx2NW zaW)e*N^fEWm@OPt1O+F4C>V9`=9u+edp%~m+?~4dtp>wFQfI zGyW+|l8+BJ#Bc?D-^w6jKrk_<3Zr!|#m@rRN+?zU%%=ZUwl zh?9|h>e%(wz(3j<&5r2FrIpMf{?-?f%7=L_K{V$8cXpo+(*{`i{I@{w=EfyYbPK>p z_V+D+Y)Smc@AZ8E<`Y|Q%MgL3N!$HDj{1RpnN*8?kK-ZyJ$$#z-2m}Zd-3mAd3ndf zilz%hpEJZa0w~`u>wgcX(eDH6pW+5HN5`N2#GVlV+6oxccdM+6JB^zxFq%;Pic;WD zWRov$l#qdzNn>OgYE%{Dh|TLmtHIGlnTAGSw~~a{XST2l!$q?I5zVx{N;iZVFPQ+| zhogh(8y%z7-w?)0_(*d=GzVncdZiK-HQ$BI=HHuMK}yOr?(ifxDzT5L_v57k<hG&fu>zX7 zt;>wRmPPnI|Bn*`A}=6f(6#DR2@xicSXJ1&W{I_Kp)#))9PTTGfO}+iV{#)PeA$bM zWV31rc80M7+}MsiN;iZ&%rIo+*+($S%AEEHONh;0zKRREQ}0W^20^3_qQk?FOh{RU zCpDTt>tq@0i+w^$k4^mu(5k}SMhvbvbU(E#%80t+yRd@H^ynO9oHj!t0{B^;XyJmH zV>JKXe60ckVB2mQs?JmAS`d7Jd6SO`(BOo8{P9byELv{z+hHmT^`=iE`l61N;ox3f8%kEQh_d8 zVr;iOCKDcxqO{1~|Iu_6FmZHU8+Rz~P~6?!p}0E~Def#3FYfMc#ogVDI~2F#)oX=iYP9V{~y>c1Ri_Wa-mi^wD%V?bnl_DNB3yWrBo30SdSU`>nqN zOPXM`K5`!xHjJB4aFKDzkG6D9)mwG76(K}sMBUgE=8o4Cz&28h9U*CkC?gn0YstKG z4&$-+s+sERAy{CTI!h0kk^8w>?+A{N*2P+l@*K&Tpa<#FM!E(4_98XKZw1az?ub{% z>$JAfuaDb_pY`9kC6DEQ|A89d#(2)1&9*MZy|ue)$Ck1UfX9XG)f?sw%n8&~;7bht zjM{+ao?uFlnklU?+`mZRF0goi`yPW*e2Mor>~n`wc@<3Qq|u#X^mKGhjV276Oao^A zVd=wD!7)2i?GiA+d-8ICzL`(&O*uTYApLgP4+jZ+gZBb}ff4YG6%Yjtof^MCjhuqL z*vz&i3VzJbtIu8mc7A&L`d@$>`T-a*{_@!N1;?&--_|&tbA)RSj?XF?Xdc9sQ#;>Z=O=zgqcYIRBVjfJQ_3iJ7)yvjr|5B2x;*5QN_r* z)2WTX>bK{?B!VE8u%Dm0ip?AK+&wBoW+~1jWT^-aq2RO% z*3v`{b7F0qyX4>7xSPqUNQ!wb14%H7iU%|bodRs|U#c1=tM)VTG=18|5x=;?G!Tgk z$zw84egSOQFc?aY9V8I;hMVxe;GGGKy}HkeI#^xd;zmrN;j(B$ZkC8T64DWTER9wG z;^E*Na8}(IKc7B7k^eQ(Nx`+yqh5r}OFjM`JZij33so@}A_1cJrpMdolmuiohE*`= zdGFi-U~97vLYEL=>!}2(r>*z5d#q7*@Neq;Ur#HvtH2lj1xU8tQ_TlR503}5qKE+w z_vaD+ZFa{K2m(Csb11i3ZXtZ^X9 zutsGYymEz-9v&&V8kEIN#b5>U&M9j%+%nG|~Qkl7`f8PqHA_~#9 z7?Q-{AwrI2#BJ5J#y;#G@DjA#JVE-CHRrY)1aclC|D^UwSA)P#Rwtt?Qo5motyY~A z%9dfjxz=^mZ49>CXf=-P$g7tczCpjnf_-nmBQ}ajjsdm< z0EvSejsHliq#<&h7!Pp?MymPZe6WA~mpl9U!dqunH~;NlnUBj0@HmS%FLM#1N`-*O zbl^zkj{^C=P@bru8k`?ga(u2wC@Ml%FrgBGK=|t#v5=?pMs~JAcatQ<2J%b5j79P+ zZ`GHn@7XaDFVf3God3f3udQr8wmSS4!QVhkb&#}%yk??|GE>ak_a+5AwYp)`J}Yh7 zy0u6A)9rT(rTDH`j=9`5mP;#AwN8Qdu*pb}^3r+8=$C#ufoOa4J*lruOtyQi5B~@Ka zV+^@EQH@iTi?nsMn9deIO)KuC=sfn0AWV{GCn(FW_eicvhIx|iZ2Md$$mh<$)9x6- zAS(aZNK3O=wSOv=-r^MB%v37kFOu>rQY3o>g54GqHmdtqf+dO&CD?{js}iw|11|^f zHXI8KM@hamNukmU6B+sP6RFlu!hEtv|7e%a9C(WgFo6$~gfzO3x5F-E?!jyl^G6PO zosn%MfodF;1gaPC6a_mv{+Ia;YmTeQt6+i!s<)QP)ZvqJ9W7i=Xbd8B34U=tV;tn-K_kC5Y`M!ng=hiiQzbDqMWhGY@VPpWRq_pYnUvV_Vztn zB5%*{rmn5G7W?SZp_~ADQgcqR@5V95LE1H0=A)1eFG>^(?-X{>|2 zSWW^zY|)k*;N4FW5yaLjW6My1B(h2myyEaPe{-RGkgf%W+3RuQ&nHGdVWu{n92wGo zH0}9jPxt6LOn$EZuI{4kmMcrmlr`8*_2%S=a1HOH1Bo(7L?&S(|gaTKc{`StUO2|+IlbC~@L zO0B_wpbO`u(G0b4dsdsrR1Is#gYJZd{|0S<@H=)s3|)1AAU{(l7D z`7BuAkcHG{WdG7XHKVHj{h}c~ZeHedpI2rhU!z^5t)O-;K2=Z#vH%4(iO19+d*%>=!3OVeV&gJr5 z2NOkCyQIl{5hJLKs`lkuEb8UsaWn#7AZ3`Hm_{26^lp z**UP5SVtdCL!yq&foRKMG55f9_MM`gjK}fZg&$9uvRS28W9Y@naIeoSG4s`68@GM? ziAZz)#MpoiONRkFM?k(cjAr<5{9C>!WX6!^FBTJXvEE#D>k=`$CtDXN8I&(sQ3V*1 zlHz$(5wbVr1|Thd`$H0UO{t}qHbDz<100mss83lmXQ-c&^+hTU8*Jl^RMSkyq#G=m z-DE4GCvWzTK@S4jPiTWiS`hnuBW{~Se&;||^=X$3iLiI`#VJ7i&a>_Y+KP#Z%NgNk zkA9Lb#1t#a)wv-Mt->$LxdPH@79=0~+!ZqEqDo1P+@n+|Yw}lltQ@GU<7K=O8U13) zVb*PP5ZG^zQ2lM0KRpeOJ&e*SF`w8Z9!>=Zi`1T?#-1-%?C+W;1j0;35vLqC)wZed zc?^*+LM}NVs(B>F4PfP)BegQXYU~!TPt4+|a-dZSPOtoh78XYqHNeK%*O z@+~q&bQiO4NE~g;ghZm2cJayLFP1SJ?9q+`T~*jhxx_~#%WllweoFinUxRjWHh08~ zdKHi8UoD7*jZwwD5Yxm$7=O^k7Z?=TM-nX7#DX$5(DsUR>0%=!1d%mKg01tz9Mo12 zjXrXqBv4Wvm``xY<#S{#3ok%83RcICMi=>eY%=1&{^Er>oo0w8}xyHr_h&_zTx)YGB=wAR95 z;DVK3)KWA=5HI7~Ku1^kw%n}5Rb1)*Gi9FSPF$OgiE}X056SqHGhA=rFU%=}>5OM} zoT`$LnW3Vi>lS_*I(`9^cW}=85cQh@nsW@1pz&}0{?`$I@otJnwvDOV)B7a`=|*Q{ z5II~GyPT;Hyo{3e@3{||20<*h%JfTWuTQ3&;#?yX8mNj42bqP59a1F$hx(~QZsSq> z3|bE(R?}5UkqM0C8kaP)#fz95|XyKV( zwpQ54=USrK_? z!l@rLI3Ty0VYd~`F>%` zxzLjG&7ND73Uu#yIf|<4L1-`&PR*{C*V(HLdwK+6@Aa4k;Zjxbah6*Pu%sfVSQ=08 zmzAW5>&PiQ^V%GcbJfsQ&85py5XMHy=wU0ImCGmz@zAzP#l+mixUa}yMvkFGV#yCK zD!mysV@4-J@?r<;E5A&1C}{E4;V1wP=G++fad?R5Xdtd<3WIU63~V+>780(-oMt_L zkP>KU;-@fj@_x>1opq1aEtYkqUvwMBvuW7?UeBQ2Um0qZ=$n4FsDYr4eI~JPl zGL>64Vup$=j37g}ll)<3k`Hl?ZuWRe&;1jr)Tqw&WRZR*Uvw7G!rX8Bivm%1l<}Kr ztSY$>xQ>$bttI|Y5q&>g#>*VbVr>7+zY;s>NnhJPnHLhlLBCBi3DOtL$R!DYC}rqF zPdlL_x7)q41G9FKP*c>a)y=OHW0HItjS2H=vsj!7&=EY`XH$rMHVHfip>eA1WxQgZ`2t7$lE<+C~Je!2W z^9-L|b${C_Us07o?6bhV<{R5QGgkEuFGhn)WRp3TqS{g5Y#V#A$ci}%Iu>v0dUx?q zIXaa?zi)){q-+c9!nj)OY1;h0Y7f`b(9l1gOxyA8kt6)W$Gp`+r<}`cKGg{t4){xR zw;i6nnMQq<#@%8z<-B6O^V$gkqFJu{lWjDE{}-?DAJ^RA>v_I_oW z!7^Tdf-jLn`z67oz^9C|*-`0^(KwFX#?w04PeO1`Y^i&UW)b|ErsXrCyRvl&*T-m= zZ*rKNY~!RFpGU8iy%KTFMopw2r`@b6Dj0t2DXkbeTze zupR*t76Svys6x`9Rhp1&xx$O*E}8i|7mTskU$6%X${LPeps2WRP_?M0-GU9W~Q2vlUO=ucYS)t2snu4 zRsFWn%(#&${3WAKx3pqmY^*GH8n(Mfg$)Z(@yBqXCdnES8bpa-cw8jQu@eCn^v6#z zVvy59JY}6dCo-#>gkc|!Cx2?*TdsQZ?f>=Yl zzHe5rJkd~#D`KunUn2ijz|2*Rp8{D}%S>#KG?X@7Y0(y8%qkV~6YqNYHiNzj2l|iJ zYZfA_j`3uD#UxXO=4yBF?1Ytos9lX9RqaQq5Wcv`!t!vc+PXTduwC3Bz6dI1dGOk7 z2dq5@feRXDef-<^d1IR<&J0}dx#r#ER2*ctU}bv*LzEJW6`yK0K%E@IrgazF{AmpC zdQ4mAKU0)eo}ZxhuyV$LqlGH? zwcjT0I6_hqe$m}}jl%ai6C*X149$tRuIsf@%Zgjds3P%64k*SO!Lp zY?ZE3VGZh0a>1;p2f#%!@0EFcx28gfC(YhZP+rvwE+v9LkDM(en#|pn)}xAUOL#!V zFRjeXEwPR;fNaf@k4Ltc<@(tCm%wGC$P^isQ9MPg3|aZY<6NC$K_vYQf%E%-Av({m z6hq`F%Yi^oMYrl-6%FVu;}zQ#(8$gS)#Q%}xK~0@S$0U;Fks5vuZv z|D}lpcxAaOJXk}1QXus4jY$~iJ_K{2zgR!)Ak0tz>fH{oM2l^}Lcv+@nUV2+j@!B+e1(hTrk#;)4t zQ47+1-0QJH*Z#Vi44ut@-1X_TOa4vW@PtEwC*Yyws=9mLzKDUy?{JFAr0Q~YY=2 zYwBw6Cz{5FVs(YoO4okdgWTO~U^91+8_P@#kkKihym)w4vQmpmHC=AEk69aB!9z}l zd-;*!kV}&;%O5m3TuZ#5h;po7X{TkJ~O7a!*DDFZj%fv^}3LGAoQzi1*x@WK7> zUg!BwVktYZ{NvNmSM_k?sGnYLtm zf)aG`KMVFNoGst-_(|nAlDS`)oS2wU5)^6HY=KJU=fI^U=0Sy1f6UoXua(HS(d=wD z*xye{raoO5>78I|5Eg7JV48MT9#=z5HXxCFh@-5WaM89|)~Oq#4^ga!-2_G3RC>Zu z>EaofBFSe+VijZx-~kz;P<9rG5HOFl-KiE&*rr~KLup=vD!$GQ?T zS#hgON<6Obii`{5vX|ClzhsB%Vza4C+{Nw_PD34o_&667l%+#v{oiiD{BL|SZ-g&^ zlhBKO%pde2aj%G#EV2<$H?aEw$4L$aQ2hbLG1!3$(82gUg!^Xz@sa<^G0rFwHh(x` zim|KCN84vaYY?@08Md-4HY{GHs;~qjrm6GnO?gBZCk+kBC6BGeNtzZD1@6R`L$$g> z_#!Q)uDANv#mC2f`iwZqAmf&(Sd!(_Q%gb2^(QWw^!ei`j|7JhNRJ2dbe&E3BCRT4 zPp2q%fwN;abn0jT#^D{?Sr3yXs{RWMvNY_pee{Ey2(*&J#0Lg>F1D%sDFQ=NYvXg(63yrGOQ`a+Ta6fn!Kq{w6a2Tof7dJh1*Lsnt$hj{ zX(E1Bk-1kz=+qh#$fNTn57}T1v7y{Mn2dlgAQWLN9GJNbFk^}NoCA)9FT4PItJ_WY zwt?pQE*Q31SBB-p3d|p-^Edn`bEodpvDEz=mRIE+9`cf=}nY3k5Z zH`t(zSusUOrg(qb{%4oXO=$at&k2T-Eqa*gE`TeDUh>D5Y7VRIddF2hnwj6P8@k^; z1yGgkQEqRHi%udDDD@O4&%JjaW>Bam2YEIv$)x{244qGN87eK!u&w`mX_e91!C9+s zjj;adYUr%4N?@^%h1St}!cZWxv&%X>Ko`$vTwF_|acRiqk##L~(PLy?_i->vEQ~(l zCsl+^mSaN)fF+)gN5t);{1eX!Aqz7d118{}f-g?Q;C*kbF^1Lmj{<;CB;fU4hNeso zvOa85*wp)+6fS3RDOLQY{J~qwdkHZs641UN+168mIzybEo0Ag4XM^i$T|*BePFuz? zJ7T9ugAfQ)EQu?f%9uj$Vxa-mA%mq#Hu~jrnI^k^e!^CCLN1PK^V|iSfG)nq9u+?t zL7;9(m$q{u!V|g}QEWOcUnUU^wVw^Fu{Y>2?nseLebW zikCjMsac%6FeK~RZ;|yvd4V_0LBDNw$tF`N7>(lA`yY=hIBp@y?XofLY88r7!v)Nu zxE6PUQc-cR+_bikrxTn?xgr}7lcAcc%GUegTzWIIc3Nbn=0{M%acyt#SX;D6ppJ|sk`uzPo3CRVggLkPFD(mjJ z4>1UxNPr0Y+2#(#?Whf1)DG*aOW=0`4-vF#Sy-&n@+Vjny2fHGb;03&Ae)PpOuVTQ zYk3^MDVEp@5>lmHZGR+)kWMA+t#aqfMfT9!@7qtD#G4Km99nnvy|!0Ii17 z*a3g|ZQ4nhv?~_cIl}%B!i{dVvZ)H5Yy1V*6Khz0B!$sQSBj(aB#PV$`_+E(hws?t z`#!Fgq0(s#kZ!cYb0W3_X%0%(g(66;62aV~UV?$R9ZVm{{e$XDu}&ru=+lX%E4kp7 zzL9Z@=rj3PR_(sLUA&9nu^Jr!dc$2H83`QLq?s?a`-3dmm6U3W8I(^njB4Zo}z#+R2X-ED2p-I|6G`{h!a`2TlmwEn^|30z{b9IHKo+rR|GN?9RiN7A2_2l zN0qtJ7JAr7uzyW=1KAP3o6yNe#wszGbmeAa#OA|AVAWS*j-0r23}DxgRKjewowT0Z z^LcTCYLK~8+4H-qb+6)<8vOLaPZ@?k?Gt9lE|OTEav;Uvvmh>_Ozo@U-{9lS3DN{a zkL((s2FGhsUH+jv(h4Oqa7&22i1Ah8+FtdJ=Y$^rt@Y?t|6^V;QD4XEI|!-ksb2%< zBwIFDG81QoW|97^iVQuz{Zyb8R6i_PkJ#cou}-E9%%L_2uaR&?2-xMtn}hjhBS$eh z8fK&YtbHJiRI-ae2wy5R_j;et_rc?raoYR?fc=OIzLFJ7woFV+&VN_g{dV#E0zJC- zNrbM)D7z1top!-8QSiC-2xykZCXL@*dhdHKKtO;EbN)>Ut8e5=_YpT(_OV~)cT)BM zWV(#ZzWtrO28VkA6LF$|HsHU!Y;M1j{3o|-FdKcy?t3KidtLF*1;%BOV9hVw4){ZX zbHds--H8|(8M}eEIJ98(Apmyq&hykoc^M0qXZ#2%IT@M&u;iG1xR`l_zDsP{y?Kx@yyDozQ47<2V_j$#BhI0^|qon6Hhrig&(J=`9Db8urNULdSVKOX@L*I9s& z)YxmbeCh_yt#JW#USOJ@vT2=17*Mlwj~LBB!l=q?w4ya+<>yYUZYnh6#v{v=mrd#d z_VBtxtxOY|<#FpGwn4;VsxXyO(8HZ^YQ>%@{B_v{>e34#%l+rC?b~pVWkFl_#dHYRh=mEDG><)IH$xrE^;5O_RGWM^6-1fGDCX zRQ>&4-_YDFm3K@m@U{H-Rx%vz3OJtHE_?ts-KxiPG+v#6>3^i4iMki0wqGvDtqw&k zM=EazDxP!EMvFJSAN(H%y8)1^$f)iH+!TS+sWfaLyEK5-oL&sI!g4dFCl4Ayb@@}N zB!q#vQVcm7^NayTtCa>ylyC>eGzDt7{m`#2HTtaRcx{nWhJtndVj&@kLl-@_HWG~k zgqyvlf|06XV!lzOG!W1sMg0YZtp_f0iVXcDw}9lL71N>+Br+?))R!JPRcxZ!mig0*BpOwoTl&h$+H% zXh|JpaRFv=9GcBaw&vfn_6R~*S?H03Z7z<<5k|)TXryY&<6{k^K(pQQ5Q4!^`Q)Q z#*L2|OmEa0SyT8a+!DQ!JT?>({_9J|c8&MOzN`rozU=nv8Fzk-1lsm{-`U+kfJ*Q&jQrx;vEbtntw=wk(G?N|FpfLE1gfrNFcVQ2 z#@W&`3uEOx(vbt(%nXBMXnHyZn9PsyJr`EY(1CW^WAwP;Q!Djh3#Rx^8A9L zhMz=3a9bVGVG$?Mz}8GrsT>=Cq0feDq9Xz;#I2y!Ww^{v#BeTD!iWuP$zQjSCvz$6 z8m3KWqdy*CFuJBX_Y--(e}JPY=Wp^UZNRuI(D2=QlK@(Yba~I-?IM|nFYL*9y0eBe z5xYV1h{dPvjH>#kdbq7-``#>mJAk7x0ClHKTF4J$%5%4=$(t%FCs6_5GH}AT=*N2Z z%#_qeZ6L@cWpGW7#o9dG5DXTADk6_K_)iarwHN$xKc{{xwwI#b86m}^r8~b1q|{|P z;PG!IjVDHIr&%?9;Kd@ENCUBV4110LK2J5Z)}G4y*g_OsSaiMC8f)Q1aVGJJ8ykbC*|as$ zWqbzRj82=Auy!yi{VpVxwLE2T;;@jkY4p5^96Nn>a_@#11a zCwSQsW5{XJfZBJSF``!2n;q2Cc!j3=!Eu-vVf^gtMx+`QbQMt;;P0SIMOjMgDG%Bp zoqiAQse+J5hCoAx>;o6XZgoB+vp!olo;!i?4?s~Y3)DYAwJQt(XwCbEz&vWe>P7^@ z?}pbud4Dwi226(U-ovTe+5h;?_x?}ieF(gfL9141^7*~+Ir*L%AO-~#eTu*S&sycV zNpv5)Z9(#RT~oyiT_dpitW5&m+Xq09;mg*=E2%j8zXwDkR`5(Wrc45jjO!JG)qI=u z&+F#UY=eWPe6&3_(vTdT3TQ0 zGr5YFIJY3|TZ8_bA2qD4%;GOJ_l&#RZPNi7U7(GR89FXBD?7zb_V znQe|{PP&a69$yidNoO(BqC2=3WV-X+zmpWyW(E4$+Z+8`IZ8B!SF%QTSHd({V%S8f z9D#sa0AqfF(uKhg#6ea@AzcX_Qv^*9+M2K~kFlrlJzhMdcOf}5bxe4_=kf5APGAZY-^$M=U(uTjCRez3Iw;3&)I{f`cuti8b=bmyDT>UR|l zR$iOCxut?jfg^@+zUkVi3mDJCyoB|2<_a~PH*1|w#CT#@lzHS5Hr%vO)_E3ew3?c}{4Q4+KVo9m-{ z(L2Eo^vx<2982*4gUBg?efl0A;>(==(0enwAuk0PjkK*O{)Li!B5ZUu-e+6opd#TE zML$+GT`9^6f5N|BGaX*D!sMjiW#QEino}DcsSoh0nL7fb$XKzkV*X11iHyzm(d5iV zPMqYGXv=(K3rz4xZ)rf-k$2-y?^J_(%#*zb;4h1jzNad{9!&FWI8TuwRgIvxl+&! zx`GX|B>%LhX}wCC$M7A7swUZze)xNkiTB&29D4>c@6~w_dMJ%5HARUojZ{T~;-o4O zK76<#tfXY`>rc{#TFj_D%<~W?hnogS(h7QewO1ajKJ|u>Rwlw*OQ>$js74q=jxY_g zEcv=hY*e$}Dp9Erud8~OXjF`u_&I)QyXh$8Pl&oarL-$*=rK0hv2Dm2jY6O1oArt* zng=E-H8PT{r6HrsrO`LGK)_sJ!)=ovUdqJC4yvf)$eA8Xn^se*59|0vOZZSUr8aq5 z@)g8JnU}v#%|L3>0uh-i9_@K3cu*V1fYFA;3D-aGYIw(7sPutP8qwNbE*OHDC4P-O z(ve8Q>CKPWa^4FtS4^$3#&ym3btqrTzv(t_km-1vn1|xi0ReMY3$?23OIQkVZpH{^ z%Z?cF2A#_c3(N!j9BbSt8R^( zjKh+S9Qvw{I*KKD^dCJp+g`*c-2*8Y36t0jIx1?#ad!4G#HB<(fV;I_Qz?t%M~rMvYi}?qX6|eWkFe2*8t}ZuQlc#V{5B)@W$@Gbedm z|7iM)D|;i|4rWr538Nvklz`ReV>XjO%tSPHd#quhX|-55w-q>;67E!(H-6@z#tgj6 zl*vlnoUSbX)_tL&cQeW((%C(pku0Lx)x}aIMe#&GF}?gMN?gmC#D3j_ z7RkDKTw)&$jorHwu7x;XqDWw~uVq}>(6*>Eb;_^tLHE&v#7@l1mXu>HoHO@YC-EW&J~y=pQ$yqMmq1j$Bd6>$XU`VzRyi%2TVSlh~UIT&lqaEL~}mgbsFjYJqkVu zsiBc1Y(@Ve!%XojV6}E%K)LLRi()9j z>=fy~{XNf3>)7~VISpjc-xhgP91vFBBpYdwWa&p_kj_X%R1{^Xij~xd+i9hzg~75s zBvmg6U(Y0+7JF2jRF#Zp!{7%}oNlCprN%v@P+%B}V51c&(9KCjkw%l}ysBIZMIR&w zwev{~TV>PD5T%EyEyAQ&vnXaz@&_w%H$IB{xrwGJ%@nrkoe0bzI`h~$6{%UH*-Ed$ zH5bS$H`6Ed$b``qL@05CIyme37E}%Ivf*FQh-wdXv?Lc8l*ey~lUo_GFbHo)mcFEj zNP%2;t`uz_M5_?%Dhvf(^(*fF*nVe9+M}?Oj{YfcgUf!~N63VOE|5#Ak+xhzOijUQ$Ah zCgMx#!E=gDW?o!OgViNA3L3Skr?0aj?y!L%Tm%j=!h~k6P~0vCRTF8( zB$9Buwn91)yNv@r)DS4z<(*k)b+^MuHw@7lZou(!gQpN%hiWf?zR_|Ah5u|Ij;Hic z1h4XJ_9I5YPlys?7Qq2XdzLYC?ja<;1fLhO&PqCWGeIql00o)$6s}KLOh+k$-Kc4} z6ifdIgvf9i?tWVL>S+nvs)}=uXP0PgUp7Lt=;Htp*4%Z_>=J2~xz3zpxJ#;x)9)>5 zu5BE@K3{tXH`9FaM)tX$6Iw9*9z-e3~GG^;np#Poym3jqX-B$v0&%Kx|kQbUC|l0t*t ze(`hH3kPVuC8Qf%*O%y*lF|sO=3&SJri6>K4J}Lgna+mBhK)mqImxQ_ zw0+SDrSxjVMc5>J;p%jO{8;s0c0ev@`yAzr16-<<#|eCfCuLAbOc9@^fAI4j;ueib z$7>zT#d6@9{=Iy)zibe4a*@~9AZiejF3`m?tzon$Hc9@jYpZK0+sblw%-Jt7)^-)R zm^N~~)JFMwb|_${kq}MAPa&3(Db^hwjA*8$mz;oO=7Vz>Vm~=oKt5EwRWzDVxW=3c=jW z3_1X;@d8`&fW5ke!?JrBFyvw*LN+%2N2iw;+G0Q%k~uq|0T#OmOVIGw=BgMJYW^m@ z#9OxAfhlaQK9OGL=~@y(P6~EVJxDQ;c1-qfH9@EyMxQ%u_Qf-wZ^ZVW+Syg&EqfKG zELqJevrjq$he4j2s5z8nOt^<*&K*pvSc|A_T76h^)>7WeHnz$7NeELB(;hrrYXOuv z$i%Kpc~dmcD6n)(*CO^;gF))6iGzr0veNBeO#;*5$xZPwr7-{!h#}Czv!RSPqHMIn zCjo}4;qK!$l=8Dp4W7tuw5FJK@(Or?u?^cRs5MYAoReO(r~9)j$dD4L^VXr>iB;~? zZfA+D-(Yai52*is4OHkj+ADRXVdD|>rL&Ej<8ELFWm|Cjs9b>X&13{t9Jzq1%@~3M z^1qF@_5+!-QWXFy0n(9p(Ko)Ts%ay5uy_uj;u*A7>3)j1))1QXrEhvteeWpf_#Agt zS6L+ilD}6Vf;YgirW?kgrTNNz_y=XP{|I>2Cot=`Mr#X+rYE*At=MaF`diN;RI9kF z@`Y#Z7{opgIeB4S*_p-0aPTbUnXG)p%R_ZRi<-owN8ZCYio&N;Xo8jrbM9}VVGy7< zO|&n1au^vF8dXb(>&TYoVsL+OZ+<0Opm%o`tY%WK)m?0?LV6h0cJRBt6l!C#&sE4O zl~P+I2Xg2d!I0J_K44bjJ8p^&vNo>rD$SZj!B!5isz-h@eF($S=Odu|LF?P_^^KTo%j63=l2YCom_{wAIb zaXu{6$rpzh1oY>`|Gz9Z-%bF)1HeOs-X8X=YidCMeyjaAdjUGd0)n)^!Ct5M$akMT zZ9QR^r+*FY=TM!OV?1d0{yZ%Np;k#A0f*Lgq-YSg^v3T3cQRSY+AmR5$W{iYg#DXP^zU zo^snW0j5cdPG0EhAqFkO$&!k>yh{0ld(MTGk)P~$-9$K3n9DDAt{ZTs%Fy1s*GCfs z!r#A=AN#kvopK-Wp~da1vN-wLCu$p6exjdEj@#>O6^ZG_7fMV62(;p&X|$fdGP92!ca&XU+VK3s?~ zn|w6*?4KSpJ%P3)3Tyaxw|f5=LL?Po#|RDb!VJuSIt|#YgxR8pGz2zLz(K>x%q$pa zx?KQn?M~-;tJ$-+W%FY}hos}NU}a{xh%$mML_9=^5Kl^dvHI_WircN)Ca8x(He3Db za(_8mpB4gfyw(Hz$-OC(dR4P}~hIYEL>Oo2J20K6GsO z=0cI(+otc_I?MK?(eR?UFH2FoWw_Y-SqY26DQrFi%BoxvJ(EB`VW9p?;n!nT zD-lgTR9b7l*DOVu8Se0O|71O*BL14)Z&`9G!sUqeT>`Wc5We3N_DYf9Z{TU0)8d2D zrByvgla_s^A0{rP3wZAiMPg}IX~mHUj>Nx=#oHr7E85LC;INSs3{}<6pkC6GO_(Vi zGMy6tR(TRoMQSk7Hq?egZ*C}f(kZt*&>SjTKjAe*&kt)?qF7OXy=FT7cj_kmVb%@U zzLG7NV#*&JWI{|a(iuMw;|Z4e163_y|jx!JLZCkfy5^QHF`?tqNnN~F->789y{tw#~rOPzRc1kCC zgYw+3LDvqWfpK!J><4(JT?BY zgLMKo->F{vw!wrZd21v-J(_L_LxvUmUAIh3Sbr#EE+r=lq?;`vo|1gQbbMy%pr+~* z&viR!qz!S8xC7=%tw46IH9*TK9N=K2rKj_Q_F!(ka{+Yx>esP|hsoP$Gs5@KVRrtB z`=t-~QhVQ`4kj$Lj+v^6N9sW~@%YLKFnyaDuO?=@k-BuU6%Xt@8&w<9`4qI7s;SK^?e$ zzQK-96a(|~n6%0{-{-8sNHyT+c8of8u@TF^J;9u_$Hni(2vdDYpF|&Sk=s}t0+KZr z9bzyuny4;UBb-+W6BClyPuOWv4pOfyBgs4NoE|c-$sVdJc@Qp8r%-50wlIr?DKpDr zUqdxO$Bmt}Ev2JV8nOl$iCQQ(x(5fj0Xu~tejNIFdM(d3(Aoi$^|XBYUOO4Cf?M{5 zrC@=d&Tc$g1OzW*^aYdN5Ej?;t1;`r%lqFRuNZSZ-;2k`+^*p?AMNGk77sUGB_hJk zO{M#N$l1X!Kg@Q$;zt2f1&cS6deqjkd?rYj@G$BQu5~y3b+!Umk423)=c>5 zcgYZ}V!U=9Iu;gM+K0WPer>Kt@kEIP1OynsOU}*xSZAVgD#8cY)3#hM9`m-ZUwho8 z5?j!=_9eK2d0s^sTRFw;1z}njtG@wed1Or3NfSsT=tR?LSOsm2PgM6m$Wmg1gR(T$ zV?!>8{K)Z9dhc}q%@+75(i5&YDt)*ysYl1XuV(TlYGZ|a?GuP=guUd0tM};klsU&3 zFr^6;vx_7!LjevfMHLkdmh_IVe=q^B_U-fs?p9$sC~GsBeBDf&VJx1X@!I;w$o7|$e3plwZ?Q4 zp*xZYrat3_)PSPj-u+YaMmv6s{$@#kls-%R_tq}PlF_?2UKqAPX3U3vqIJjPCsOq& zqGP#yvbjo6Tzvih>n~h%dr?yAHCiqfqFsa%&v)HE%T)vgXm7k&!tABeS?|y4J-V1; zJHIi1?BoA^N#o~+vTo~@i2U+% z{d)o<2S>em>?Z4ZB}_~vQ}(=4rNps}7S+K&wrCnl>bWv8^ycNrVOA7p4VJAj1pWpB z9u^Raf{Z?u2TEAI<|%Fhu0SY>a%csgv^3+9(rS3rwhO{X-7H0ZZ@&pF>e z20vhoXRP(CJLa61meNw|!?R`sCum7Cy>(LVdQkkl{Kg3NB7MVY5hP6*EwPQo<_q>C zqY^&2#jY<`d$wd(MH_iSq_}pfZxH794bIWPmy}KBXzEv2h7nAJK6^I|cg40{%Mxk3 zOJM*)%FDY;5CNTdo;-f2*5ptbZ?4R2Xz|f0J!THp>@=dC&a~wZp946j$1DmdQ0*8E ziv>2-HTH{w#^;@Qp_i2cbs(8wyik;cnffe&w-GK16>E!Ca0OOqxGL=5Yt-RnYHY9= zvJ?5alk~J_Q*MmxethkjU7@~PiUgL2LP|x#wCDJi^xv<@DWCc7bmaeXBEZy@F+}(H z!0j+dp!oMuIqVkyb(1`|vM4++#xs4#*(hCC#u|Sj@~0H9d6$n!!HDnGy#89I7F5$tc0j>0F+GhFMfqlW1cXt7we8Pl1FVNMgA4BpUi1;hpokBw z6f~?>EY}p88^a>0iW;g2TN%HOzAdu)lak4!)J>DTU@(3U*hpp!<9V^8mRx1S%rwFo&C zEnr&l_gCOtl#hMp)I+IQ@iKwcbki8WR#sDePS$sFmL^%dUIG=wrrJK4VMPm?i=g#n zx$361Y6}xKkALSO+d4~IuEgmRNR-bEz^|rx?#yfY$?V_O8`etxeqr;Wn`{|R$3zO5 z!o?%?15-Lmo8g6%_uj}(7Stv@Nfi7HsK2id#+AhV=1}=9(|#&XM7XwIG+G=gU89(P zXY-1zYj2a#uwR+I)(O#f$Om-J$?j@?Yim1XdR3|QLebOAQop)h<8Vq8iTAyA(>^Nf0BDD-(rJMNk5*RdtFtWj97FGRu^UAyMMkxDgJwroFKfkX6 ziFJ{lqMH4bdBMMAayn=Jo#O5dW!Gy|(S6wJG3K`>wkl z5{Yp*yV7G$81!uSkaUFpZ%Z) z*SNYUf2A$b!)nM~s7ugCuJJ2pRhOXa=wL~=iU8-B>B}}U+v{+-#$k+pFo%McC|a#G z(_4*b8h|&5NHtH*uyk9aFxL=d8SWhAf`qUK)`k9J-@P z#D^%bz`C8*{J0tQEy0DBRY?cKD+G&pNOJJJX=>S`FRD>}y~@{9Oazuc_R7TAS`MO& z-~Vg%d@H>eX&p;@J*&+3dJcw<&uY<_$S(V(4wAeHV*KvF#ZHNjM-otXue-4)Ac>ne z_x?dPVbU$BT(*C0XUh7k#y$rQ2#qh3D{9n3ZWCEQHhx)NN%&pN($}7q71!VXxWic- zQp@Iwaf7>T=4Ddnbg@5M_f1X&D^?QeAvO5go_kh#L@0di=TzCJ$smVW<^h4se8 z1ZHH-KME@1-;{DZ21~XP0_lb!3B9ZoHY2U=9kCw zOqAc#CJjk0vN1o>2Q}agnj`+5kL^E9Ff-KKM@%-U!c>~I=vr&<76h_I!=MCcVL{#~mDEEE7Sc7ry*h6FY8fj>@w+{ZQB+3HDhXB|*iqB*p<}Q70 zYZG)@Zw-xc0hlgsPk)X@-kqvbg-o7fq#lm`V)|z*25~chU14Cs9RuANK-n zfdPLt##rNBb}W2KY?ujK(qsu&A3Ay8tgnN`IJEtT;UkZIzix4s(WMfMN=Fz*Br&O* zpta&vcMy)~m0N%RUSCCebvhNkbd2;V zG{wVTxU9R3ET>`VrIjj2E1#~f@kbvsc2(ktNeOqN;ZafWB{PN(4C8&E4JvNRMoy+` zAi(lNt}4UJvQDqARSP$R%4ZF7N{;979R-sOinWV~=(l zRYpnOs`Jbj=y*80yv-}G1mESx&klAU zDDm@F>!Gt5mtEKh)T;+km#}_GEz!qljo-mZE+q+m{yn0D*Mfr2|20Zy3_Gtqc`r@4 z(dM5(B9QV0ls&|OSf#nm>GP6O`NT{HUQ}b*QAGfVTl6PDm zb?LFl;RM6fm~T1+1|i?KJQ*n=8{q9&XGYW~`jhUMSTFnZN48e~QJ`=2!6mjdq2kG| zh!ro3m8Re;s7LyI9@o|&D*ty$4*TCS9I1cA&R;wct~KlG@I@|1me@#^fBKt{Df=Ye zl1x}L-wnGMg7^pV{RXoX?gy_?Q|2qUD7AQ7V&UetHY?pJX6wrH1EV0Xx!p9#6CIEr z$d^qE!DTm;1x0m?m;o<=r>AE&kA3y^`yZ@=f8Jltuz1}1?v?%5S#7$$byh;3-_APl znGZN?fy92;ckg=u3v)-%uTzPc@Fubh}+^(j(*3Msl$tUmybX19b;D=mMf{%{a`&<3J8ublq_G?# zb>wftVbRlvVrnL)@AE`1mzEPmHC{xQMc{c)R}Pa~WF+@K*L4yWpj38}dgn zUU120o^bY1NgHJnuw%BqF(Stg!C{L(VC6fCi2LUhn!=7&;@46uG}N031t6pMIKAK-fdwPd+Fq-slT9|rx`ma`uY+EME3r5oh-my zjsMQWNMER1SN?!+C5B9W&70#9NxA4t>oJ7)_eSd9m+r*9Z2=5PN&o8TQv6^;Z{}s> z-NQq=bOkG6%IehH8P}&}+V>~;Jrv-Fq22OeS54^7oi@+r5%Z;TyaN}JBJbzoK7!MLOqE`YXSIj+^D`?_ z|4RJm#c(r(k}Ey|4!@zG7Aof>%C00fjZJWMi*-)CQCC}mU&iRMjLt=x28)y4`@MS` zkyY&IXpN|Bbcb)Q#8DD)tG9R0gsC*XEa?%eWK7sXFgmPe2c_ooNaadjK7T?-5kx3QYzGYF6aKiY07@iLMCY@8Yxah_l~D?Mq6SVMGtp!^ZD*@VO%NeN7GY$J%SiZ-0S4n zqFnUMmfAwTWS*ObeCkG`Y8UNXqZfF4kl_rv{+8YGjLfeZ2r9 z9bojEPpGWNwfHqY8C`wIV9{0-77_Jyf)1ew<3RYXNUa%*rR|tqh=+|;SHNU@mw2e6 z_)by+s^uSFY)?Madd)QBcSpW#e2t>IVfRzU->g+Z2`gvw3Zy3yFOPJB{Ir;9Eb&-i z=xH6!_sH0EK^RadH-ZP^WU-t1GWBG6W4#ermvx?G?mI&c$o3q7Y%&oz{SP$08#~z3PW}R z+jjV=w6Rrw25K`p2f1<{Jx4cmrgx}CF1+UWESaG2kIp;~;ILN{0$u+L&WP}9Vt z@-X?tqg>5u>%{$(b27*?Tu?!7A-a|MhXe#QS>gd`TI7 zdv-v;B=~*`=Heau8P+S66s|N2E2}yKu9||Q&N*ly+*5ay$h*JZrYBFYP$Hzdu`GQXl5rC1cbU?IgL?C9mg@5)@@{Nycw(x zrcZxjC35IYv~`0d3RTg2V|x7Hf1{u0xxrJJl_FN=C+p&(dgD6m7!$o0Y7=K_u-IaG z;`nN(+AHE#wSq5$M!Co~Y7*!>uLJ{c#4 z1F)ID@BSB9T3QN(;sUlYV3H413IJ-@N{@8!9io<3Be^D1I#4E%17(63pl=(gHfRCX zVfcBOt&}E!L>etX+DdOJiA8U(7SdTU{NQX~bGqEU#HqDKA4|;KO#x^+B$3f-kOB=AycKM?jMpODOliZ z;Xqab+-aD2kmxUke>C^4#v*r=;}Y8LmbVukAB}>Ci7E<`y}Zcf)O5MV9EG7~_Zwx^ z(>thwJ$aIH25y-nb~uDK9rR8RsMkuBl-g5vvJLLK6#?N^Z|5m!`x}-`L0lvz48$D z3;u!yDDI1A(g{?Qrp$JlW1SXntG;VdN=z{SuFRrKU*w``?Q{+`M_6qmqh%V$ai0P~ zx+FbI9(loor=C?QXXd8eya!& zBMy6s;d=S5i3tgJ7Jmfs?(OYC;i5!Jy3Y5_D#EO?H(v#b46#=7LNAocyo5nPTt^Pv z7_8LG(~fx#^cc8YqlnctDB&zWzz^Q_fhNN~IcSTeK-H~9n5fd>MyP{jEHC$3l|~2; z>P)?8T5cptl%sHLCxaZMg2hlUz%m+su@Z>t#x&clZEW>zn>m)gDMAP$ep>ENXei`Br6zTiKENj9iC4S3mM z=H(K@Uk5$c|DfA^teq)scHvJZx`r?@Tr`&LV-dEhN5l#0meHispMdIj$hx&sWc_(M zf=@XRV)3HEPBDQ2z?+m$Zu;C)_Ef&hd)$H*8JT$hoI*IHwgJA2(QMV!hd; zXaBQ!R21LO1p9qKrE_1qtRs^cN0X0_FDpO)yM8tHFlXZ48D(^f^Om)xCER#D=I21x zkl<^A;NCYeenV^P2(!tfg6Q=p)hi&jv}xn-8w-&fNYZTcy`j_jVSMn$iF|{1Hopee z>F&dlKR~_3pz0YI>;aaUm$e_SL~krLXk8PNBk!Oi?}s-w-Psunu&>-Ez4SpOe;kq3 z$3Ex?1xEc84_(u@r^{b*bFIMOD%$JH{(6>pPSAX(e~|%+G|Y2u*VEk1T*fhTP zT>y+&PczQoYX)<_$Y^|&S*{gleU`I-m6XMggn9tTS-M6}9*vLv!i zfqeHpexR6lCW)kGQ`Tu=1f&`}-j@Td>QO#Lwo5rhR=AkUN@j@PASja^t>GBS1{g$% zOT-R<d~t%HK{fT5%4U0u@kcKAED(4*+COhHY*oB{4SCVHDfIQ-;sv!=Qzrfo~Lv5_9D)>M}%vEDT|>|#*Po!|MnT7WnA z<>mdG&)EO+57sc?1qaBSZ{~IH4rU{2B$)p&gn$J$U+kq{%o3pRF27Ms(ci-H_mae> zlKwp;-5ded^bfy9-+Nz>Tn7Jpe(XUMt^K~?-Fa%1ra=1_$S4SUZg{h~zgav3ey8Tc zZNJe{f#h$3w;Q6R#drU+tk0>z*4P^%x#jE<$oAIz&$X*j9^tqJ1X*od_Qt{4Oy#@@ zWsDpR=&@W$-sM1%@uHf32lWUf0yn|I&;;7=EJeD^uR;tlzH>*36ehS6}Bq7 z^$X3$c1K>#I3oIYNcAg9rbSW=jk98KATE5@U%vzPCKx%2akHx>5QRW`@tiRmiLKL^w;E7`|;pK3lrGRIU z*)sisTU6K)74L6);J>6>z5t7U%9J$h#so>zs%A^)ZZbic6&E~zM=tr>M2{@<(*3J| zg{~4tE7u0qXNk)Ft$Mzyh~y-6{cxQ+5OX`4&OP=jC2`QM$K`M!qXAR`r)&*M%xQak zQqrCP#%0ibhrwnC1=0h+pd3XMd+0r%PSg00B@RFn{Kmb2gzz6X3m;X_pUzmFeNo>o z>6U*%zenjG-w>f4Tdz)U=nt=p_0QJjfFzFSv>C!WAZ?$!2$WrJQYoPFXu5&!Q(YxM@0(J5enevrd zPS4<~jkzmUR1xlSz=lSgi@0pDy5`p|2hT*$9i6TF7G|Uc$#-{PCwu1Inc6BGH;^<| z`<-xzM!hmZdc|k2f+CamksPIt_zEb?y6byHhha-ii3y?eiQAx^Qs)&kG#f|C1ecD5 z>-yXHk{K6T`iVxn`*_qm=eouvVJtbuR&c~^b}Vn|`j;TG?AcMzp9?DItxYeWS*aMg z_!UlVwhyb$OOLqI%UJHL1KJZHX zg{Hxqo9|PIdznid#?fy!ZA8u!{#^L!lh=uN%orC8+fHpE?_yu2zUk=beA>%OrB}Up zdI55J-#?!? zEnwGsC%TS>g3Gc@9F!}BuVLWVkB~})W1HyOQWM-Ru_eAcjFNX<*2yD$iC4?$9n(^u zn(~*Vj!;Ljl2PZC;Bhj{Sm>S0l}42lDet$aJ9A_Gl}Ey0%6T&V?Mkh=<3CA1FY_M;=mEcUS1*w01OImY zy;0n+D}msYw-4JJ*}0>6^62I#Fb+6+gZ+31-+Koazn(<^d!I}f-FClQF86SmTwRPO z`0_L=z=Mg8-|BTv1a-gRO%CL=Yco@#6~{kxbjp1-nlb5yh)d&cXj7_U!O zdSQ$xyYv1{Ev2GSu*@5V&S0TC+^F;iGXe`dq(9QmW3es(e&U<Uef2jJzc0x3W@2{Fs+lce62|0L|{KCXtJi^`W3sQEzL=esvm8OL$NSY zK;v}q?aqBn-OvNipw2~@j{(g_s!mh0>ZvcyAV0q2B1ly?Y5N3IiC`xJmb#9kg>CER9=5&CMUvBHaHvF-QS9JKfiBHCHyCp0ye!i&XHg?RdwlG(`Bt;G?-r}o+V z&3!TGKeEH}+mKR^{~}B&=0cHap?*8^b?na1$r-3;0WxyD0=ob)46r%SzW|a<%M&3w zOn9Py3f^6}O$R2vWo|(IXC*!Uo_}aT{=C8lqsxH%Fq;1&9~Bk#0O$sz_XK{b0rDfn z2xAi>**Z&Nv^#ja_yR+=w?W0&7|NR$5^$2Tnn;AXbQA8KY9z&R#s5M>{k*}|u$)!= z`cuAdR~b>HuDye|=2Z*a^x5WP29iDeIj{!^v-_INrK_}SdhVW}fn~|eww!h%B!iXm z(Bu!!hM-%>IwbIYb!+=+WL28)uPNiMu*%}a86?#+A2mN@Racl*ZqWI@?DU$r$5xrd zK_+Ho0P8@lp$d<WN;F_s5d4c{_BxlQ&o z-vO8L+|(RZthsC`)){Qi^jWgz-ETzGFBj%g7O<6{&A2Ptd@OovR^ny&5*aSsBRweA zbsbe-r+ltV^&c+TiP>)PoZG^SK#2SdBw8@7CWHdZo^PM~hu@jq6K^0{3c+Re;K%W| z1RRI9X&lw}y}+c`>MbJ0;<@21MC&ca?L&$WLDWv(=^}UqfJwc@>h#{c2ZFWOe{K&{ zBmqK!zKyaf*BLJ}4t8V|0p$r zesEVER8Ovvi&B`}I-v^6E-@ddx~xsHKi|{_sX$OB2}`po<_e6Z=8%|rdwIqtfdM~rh5W4A$kZ-#S8cO4j>8_AcUHSmjsgS>+NF_` zY{S|uEqYa1>*0#J^oqL2E;=P3xvHQ7289E($6ZXu&{i}n9f8yaJ#|sX;f~PwI@d-? z5(8X|L48XG9%S151&k;pyKJ4Fa;|1TQthg6sz7aJV8L0_ZhH}roDwvB?Z6i z0YzQVab^&5h{;2w$u{sd1pFx?hu>ZfeqHu`JvEtrW3pe}wRGpcA(vku-LJ&KkD0;9 zJy%y8Z|0E4$*mXVxnrKUIL-eeDc8c)N11!RR38a?rLrXjX2j(%Y!G!Alt{9l_(k%&` zwnc5z$Zc3iYqot1HJ=B*zR2QGf@wZgUknzr=9KBPu%W~?UEV3F&zqVs_JX=vtkj{h$NZ@*}{iLHxjgK25>Xt!0E|jIKcewBw>=VjS z>|Z~e5`?y75pYH>;oTWUz2bgHd#$cbctKr0HMIn=?B!^6pJk*3E>!QPf~%rnVDiItK;`qBpCR>JeDV8}VNXAGKyK@uZF817a( z(X33~fXChrl@RN9of$0`?r!J(9G*ZmZXE4k126xtppW1b2CKSI=_PLe#{EWlY>kr$ zfM(do@VTjg?plLB0tZC}fv?hl4Hahc{c2KVQ`pXLj&<7tPKF*4_4M^QOb4^z%DY=C zhK@j)+Vo1d6IL)ZZ5khExifySgMs+Jv+6U`gF4k94>z8?RfGZt2Fo=@5M#V%zx}X={mZhGo=ypmNkE4Tg0^Bft#!>#Gq8YSR(0EoS3!GX$yS_ z6OOXWM4C~WFO{+uT$(nx1~u9h$q`Fc$$Ejzstv-(DQKMIPdm(g%ry*t?pj?bDcw6@wXQ*mgNukXikV}M*o z-R2vd_AR>ZJl2D8;%?XknDOIoBp(PPPv#wpLPvS3&OYbOqKR>iU%@KIx0TTSrq@ge zCl^W^vkTZN%F72MepBR|m5B{N30tw^B|oO2#~Ah2zw_20aayIrP^ZnsM9&$?Ed)%6Ji1RA$nYkBdgvL5m|+l_>?9Qy(mz!b2BiV*5%Z zU3D#FdYSYzqs+@tTrF3=x0@O$nY!PEILsldvn&|?LCCs6C+78b&V5;|4$^raHxey8 zxWTp=k1xX-nNRRrakR|gsdFCW zc)9G#cWe&D-|2TDDVXYF=>)$7BjsbKLRPzu@H=t;(u zg)f$zEUS09EYqZ1|VOCe$ zsD_MgfK@dRwOFdc5FuY8M;J+tS_#b}Rog)(En=acHHT&4q|w03(&q+_W@% zO3)QmX}9Cobkiwphq~7U>MrqWGyq1a`Us}PN^YIiJV#4B-m#-$^zl?>R#nSJ;~GD* z=1iq=SF@{TV%ojei7L4MlGp4h(>@;NMYJ+~_L-Hb(P3$A2A_y@*C8oHaH8F!pWyunWY#Q7Y>@~rU-Nkv4o~3x=`V2 z6W}qUe+@5U?6;M#Yb*5G5|$iV(|P}%!*~qts!U((9w$sx#vk|qLrVyy-P;eDo!s^& z~~9jw75 ziQtJdL5pG85ro8L$>U?xlgp$Bf89>Mz7(#g#B)4gJ0*KTF27Qag}Z&?{KVg_38zj7UM5D8$K>ThF!4Qg6Zu3g7Tgga=E6 zl^HjC&RF01;s)G)AEu@?YqP{oaFD~m;CwXk$C9QXaZ8&`bl~&2eRXxUaWV9H0BGtJ z1twxiqw!2un2e|I0|Dt^?>QiK`TBJi$f2akbkGuSeCF#Ibg+!L{J3AbD!;{ICY0tL z?zJ#3;ew@6IAXfftV!x1^0;3+nyC@ck|FYi|7&XE5cjgm zgtcOvt1_~ZO%~_7AKKf^o_(#+3`Mp`ldMoOmtMC6@_*ArVPnJjN9?^K@O}z)>dd(dY*m)$|S5=`!nqWxj@`TBS)v(No;5h)80#XYF=g1X8g=cOPUk z$Slq=W;J(wws2^iSRuB?_I5^N!e=Q_*N(5G3hS{Ba*q)3LgFK38lsbfJ*H`s!AuL$ zf}igbGPu1j_BxqsacJjwD-ZZBVD<4tHkJy+<#|SuQbqaKbw@6{Z{}#nBfEa$0_P0YVfgJq zu;DujmJMp!m|>&0)g*IxjiHsHlQ-jGVlxbq0_S~8O&h8C{>jt~lOqn!DV^_zciAWW53@Kt6sfCMRk&rB#S6`60ulF|EC35>Yxehyig-s z6W$CoiIE83pry-z*9m_uYbM2P19&FO1fuaH~ z{kJT9e6rrUuDZH9*cRcKE>Xs!fPCz&g{E(Q5r&NWpJ(NbX$6f{5GwD?QR5Gv+Z!D@ z&(KFj_tq@hf2p0ArZLz~rdjYKF)wG!(HIpn$fP9$K7uncGD%n)*T7P`vyqlwuMK)T&uYUACIEMT&ueSRutsQFAqnTe>0W;gd_H+5y*dwtHX9@hVU< zr~FeczlVkt13o%6^;vcnj9Pp%@g))QXEREPLZ?9|BIlIA7z4T~~-e z+6!B$?D7<0D*TQa$B+cY#<>qJ9%P3YmQJ0v>EsE<>uVi5=zr7qAfI1{E9??AVyS!@ zG>5NAudg!b_|#)>%R!PSy$6u|=G3JiIK;j+rE}5a}b{J<$5O&Vi#9(ys zq|IPh)FB1P))0GJ27GZnX({Z1!ib3S{?zxwgnALuVO>ubol3lhb@ zF|c>6cjO#uuIJMF4}>3}4@#{iZ)j9G(d|nj8@i!KUK+gr+f8cD`3qm5!qp|b=S`pv zxTt_ZWfo^hxC^^O^_vtI@Z7c5j00P}fr$yURj}7L_7QPC=I?0$7aJucELN!bF~riJ zF67Gx-FI^=LK_`qI-a`-H8yvA3hkOqM)k2o8b~VJZmb+=aa7+~OK)ncAO^Y2L_;J6 zGd_$69+w(DONRI@lpldR%H>@&RM$^?z;BF0%p3ct0w3q!GfrZ5qx~%JJQ8)6vd62- z`GwqIpu-#15sviiwsXiTQ)7{K#C}~aW`nf%O4|k3gQO7N#^9+WNHIHK5kl0bc$`~h z)llSX+o*D)r-NoC{E_*4OKR`By;5~llvVx!DJi1@o^)Fw5>xD_ut|ps46n=sq{x9Q z!E`QMCp}zys+mOeBdQTG=j-iCa{x2dn2u|F!nO_cSY4OoI6b2OeKm~BIRD3>(h5}a z3XK0dUzeAPSY2HWy?^}18B$%F&-@p<{>9{3^;F(~<~Q3veHFT)PK~AWIy`C1F?f_% z{1mp0IE{Bj3_2ufb)a{6%dzJJCHLk78n{H4dV8LNJ>!4D>RCyPY*r+Din8T7L9V^n zeBZt2JjUQxy|T&R?tN;n4UcSRev6*MrK68@dEuv{&RDZQ&T3_AFh=IbQD(0rGB@@* zI`=igX=WF{gZfm!)A(<@R;z~TEsYfjn>B9iR3ItIx|-SgU?nd z8T?y^#2E+mwd%AZJ!sw=HfPOft2#Z7QzLlKo8L~UY}GUNe^QnA<%1>sj?JN9q z0eM@!5sIuMl{3z`!D3)N5O5c;+0uCww~ znvx&P(EK!)I4bH(oDLx#Gn0sO6P2c|q+I^?<9oqwbMS?953;y7L#|DA$FJEb`b&LG zt%zHAv;Bz8R7|R#H2Ix1d~8%R7Tl&<{12k|0PO%pa!3TR9OQx7fJM|;qmhW`p#F6{ zJ$c>$zj+RhY15T6&VvN*-txmVllG0UrNw&(xh?H9l+mEfi{+g6&qY6q{51Er5kiaP zfK?qpcC;YO`rl`Z0sq5)X5hDBCI?!Z+X0Hrbe`iQl>l13b`mtn{#b_-^x4AU_z0J* zFpO~moVs1+pa)2pgB@1!H^eF>g#+40NXcavP<9k_5H~|B-edgyR ze#-Klj@Bb$$-8U8g{uorOLJ*YZe>PCik8fj?4OD}BrR6|@(?;zMU1=q4Is^EQWBb{ z+Mu2x*0_1<*gAD-Oh!IQGipryKxy@sIxZOcPmD5>Iin1D{K>98NtA1tc`T@HGkq;M z8~p;Y$jVti$6%zwEI`m-eFV+Ztu zOphX7(DzCIvv1L0oiR$svOohyg>UDlP}2A29`n91l*N1LuU}ex<}M>}=64jO(EBUa z&~@bgF0$Oxr&MpVB-~&XD)e&iS+vkS!HUD zo*JC({|6waR-ASN0OWy3IfiK|3L51K6-?!K7kr&-2gaNmJbcY{Z8H<39OLK>gFZb~ z%Wx5n9GnA@E|}mhJ~T4-qbCe;5G;3SKXf=E$4W#UWi2uLhSo-j25W@|W0NIQx_{nz zwwOxZL~3^0nnH<_2v$W3*G+)U-6uL%B$LI)(^YX0>N2a(BQ9t3=zr$53zN&f{6I5&O?nuXfk~5qkfWD3k zRk__{TV8k&x=OiQ!}w?z(hbeXyM#7YKDEbOGBg9jg-}@BqpjIgT=vbpcb#bM~tRb1EuC$am1|FV0q#G1Z8i6R_G$_!H?6!<*rNW#$hfYH<7r1@I4}ScO$SfFU$wh>A zy3?B_%V|K^Rg*qCBP}c}tP!$@p*MG=Q#EuB^K)7;d2+Z!n3N+9f8$659(MQDeHAJ! z!b%{?W)RZu=N!u#^SxFK&Pq>W-mekpYxYKZ?yZ06wIx5CM9DV|b20md`erFh+d``h zQFMrJmjI{ z+X-a>#Z>fRrWopJaP(@yq17nLYc`>&D0BJhx+{AygF?iSIaS@-%z@6i3qdsOCc~7W zQg=kroQ~ztpIn@h7M6+POH(#y#o#yYnQM9F8om!2aA9eT4Kl|XXj|Ubfrv)@uH>S8 z<2v0$48o9LuMTBMr690?8NC!o@i`Ts2=R~Dh=;l(63!gwQTO)d#4)#T&24Zz5 z3~gW(cE)+19B#fV$CYr&2W^}DAwm(a&p3l@!b>P-QrM)+(JYB177(=_91`sW^ZyQO zO9vI<;7v3^D|x%K@)p7=|C)-~DoZKeIM3xrr>WGkHsK0qIQoEj04`@*_@FFdkYI4b zvG2;Y0tU#^no{RoXMjmeZ=3-QD&5uX`Q(SUgi9N2F z!$tG>1vI-Bh)zP}>HDO-w_uM4V5x!Yq*yw^kb+E!dq# zApQZFvP`QI>j%Hyy&n!uv#)B)spb7u{Rf(8PR*O(-6kFD@3b&WbuRY65!5pjv;>DLP!10b z!Hx34HRHjK7F^IEg&ndJEBp@w6nxZNh#w7x1)jK*O#Pab30>0D`h=W$>7kEua`%#N zF6wsnRxu$t`KgbpL&BZ_pzX}a2lFYGX#A&jK<5G_4I7M@y(hbjTR)SfpOVv#^@9v~?9c{={yqQzWCyrEclyp~JmUex%T zk<>;spJI8qy=kenChlBdfW={cm~(T9SQiS<#6^(a6@`vIee~O<`f;jS4np=8 z>5t8n)!#kd5RHqP50>+NL%xEfI4h_KmGE)g)L|v*hXTVNt7p59;b`2#nkr+-#Kkq_mHm{tK+Ex#w2=oF zN6fs%zE`ss{&W#N^(-voCu@FlH$O4lTI=fp52amlgf7bxw;xF6%48Dj18`IP;3L88 z5$SW1fePHBPaS>BCkN;kj~UL78KjS4XBj9y+^m22xDRC@_Hv{1*EIQT-n@@c`m9F9 zX~oU95gU03cRW8el}z_{bF{@zPHNNE|L}Lwt)0HWnSx|wiS(l*ANnMhg1t$8M4)v?+0C9Bg#8?OhMB-E|gobX_V{c&}S!`DaZ`F}K>gI}L-yvDPQW!qS` zy=B|xvTb{{tYzEQT6Qhl%eCq}-`_dsFX+|teV+UN;JV&dNW5r|(agC~=^%<#Qa7pr z!#Zm@N2?$aUO}7!eE$iq8JG4WPMFl55J36gov*YW>A3EXYE2HHf|!)x#z=~ldhL>^ z@P-eLgnW-`1J)W2y6E62 z0s&sUcC`jbgIrv`v=QWEH6YZsz-|8I5 zRh)Xv!+1=#NZd=?j2a0XMyA=p$hQCf6rNu|kU`1OA&8-8_{+zA50{dL>wVwjL+Lq! zHq(PEMr+F*-BAf^KZB7|p@}xxuFERw#YYnIMr)7Tvk2Q`P^lq^JPC?DoWh zAKJ82hbl=wM_bt@Wl`kILs8286SF(yy0GD~kzr4|P6JC#B{etA^P3)Ijg|7|rwqS= z7G1q}@`W{wzzmb%B#)?0lIUIY6J81dkNx#M)yK8hN;DKnLk=P7TIUbkmWGwo=j3Dt z>uQPfDrvg}37A(n|A~HoR;7O}>n*M!824X^@vr+l`dHwa(3m!qtMPcuQ-(uFRE>~q z6l6mvWzi)zHnK>2Op;fjo-xVM(%X{o`wL?G=kgP1jQ+~SMwcmJQ~+6O)02hUr3ZT~ z2q@-z#juPR;=u)}hv)fyezFZ3rymB`3Dzd~2rNHyhgO%J{x(g26Sl)#FS0->wKmD3 z78e%>9k=skirjO*j0r{^+#-NSy!E*Jg(>^+lX#W~uCbJnFh-viW6M>`t3y9Ze+3VyJnxyQXBu{61A`SD; zpiODv!y#&^ozBeab~)kCrC4K02XL9s5PxQ6-G#=)82vlofvyhjVl~pbr!9bmWLCon zzNyMLvR0F1&s4S|`3oPk5$d22nZGCfV7qx`pSA5N_|C-a=pWHpv}xKB@<7|G5*-u- z4gpS@^?25k!KgK>Vs{-lEu9GaX!EfQqcu z5zvkAWNMp#d41}MIO9uliLkJY49BZW-ZzZ5x0i-ipp~s*67UdW9_w0&^OYzJExd1@ ziKL>8xU^lmmbQIz8?UKazOY91#c}QP2x(lQ4|nKm#kN5{PafA%*`peyScIP^J>K03 zG`7*xH9EkK@mu)?hmIz_45KJURZy;jzdXak8C74w!-XmS-0vqnCW1{!9|cOu2sqN~oDbutA&c!b=5TUy##l6Dn(WQ_Hy{-#&0f%*dG2>IW$u3+UB|EafZ z?%)t_yIceC2s}5>3`UQSt(NN_<%=C!M=p+5EZFf3{GZK%m<&gOfcOuVp4$_AQ6oM< z%|u^8tK1;Q?3p50e=Run_umq4g+G_h5Cyl&G^$(R1iW{!qU}mQWNGD4%BFZSBs)p+ z9ij+ddJ?WWGM7mW@6aP_l*Q?w#l~`B2U%j$`Byy^lsX#H#v5f(y2aJcUmL8pMpN^Z zJh|G);d0#sbR|9>pvWzE(~w=l3+>#O9wZCQ&}3aJrzYf<(Q(s981R^q25Kf5oMz3C z>oD;(C55iUPCJIJI<#$#v_$Wi)^;T1r<{}%bui)R>WJbVpjE^MKJmc)Fjq<+(X;Fk z^q1K)R99P5S<(^brKz)Z+Il0t`bR}@#%%YeF2!Oj+?bVzXKZ`hX!*XWEd5*i2GPBM z-B#|K4A>yt_9lkxrXx^w@JCJ=guc-S2~O_;dJw&C?=;{N1gTS2RI!D1^$aQ}xz_}) zn_viMTug&TE$a1dY1h>SmiwE2Av z4M}gKhm!mI3Rgd$;+W(^tcp&3*G1S$dE%NP6|QR6_VV!fDGw z8=reE3)tG^E4@VNXzCas#I60WEo}wf4>z5KSKv432yp4B#?2aRa~>=NQP-jA-@J*X z>J2*s7i^kCxSZDefMZ$QZXf#iRp|Ei7Et{FCp!o@3mDZ1{FOaYeqLlOxU|9H6&SHwMvai5@FgQyR|i&NTG zuCte7x`u&y$s+u1RQF@uTSDZK4C=dDvadeia}=QsJ5PU@>kvgCaEpjvCyf9Lu}J%Y zeiT}*t^c(}0`;|IF82uI7IEwq@rf?ZX{zRKih+KYGsHz!nX@44E~#{gHVZ37=ug2+ zI`slQ_Xn~uetr(==`td_?RK#U-gYy*SqYS)^&EU$nXBR)&H!sjlN-nB5h87NmL0;^ zEhW#z-?#SO)%7+P(3DP367$^GGl3SY4BDM!QN!C0$g&Q~Dhwfp2?D6PS%yzrlY~OE z-#o4mHjsq^b#DX!7V4iS&{-e|;Tc%H69{%m=8aBL%?a3~4gvRx3|;%YeneM7l>NK zLu&m{Mw!A>#dj>aTdOd&HmtrN;@K&(goq&Ii3%rl?(W16(QckKuWUpn2RB6RAy%Z1 z#D_%N`l?dc2X!R%wd4!z7j$kY7g_2p$EoA?)+Gx*?D@{1TIlA3!x!i&QPtHwH; z$3X_xGX{_?+{*aVyE^N%Jaby{qm{n=rlzvZ7Oh2m-4>z^oomkvcZtaTRv`90w6OoJ z?n4>qJ}&-ywrdnN01XNU2T+prKl%ajY@_@8`=xom9#kvbTSu_P08j%+p9|>8V&mig z?x!faSC7IwPq6aZX8!8XN19w}r;CS!thI5f-{{QIQ-QO{IJNmo6noT8JS^H{Be@NRz4O{kgL z=#3nkU8TD9M!D95^XcqhTLa!K^MgrnG>$kGRPGZA#f`+s*k)tpnHkcUX4*6G8?wWXMzZJy{k^B@b$(2lWCm6&o+O9-P z+ZAmlU~l4yF%L6(F3>uGgw;>>90QE4w&To5h%MUtd zZ?8$Ay97Z!pTq)^H$Z6s!l?jN9w77t01PIRhAo?vHn07#tW_kP0H6|JGXSDLSL{jG z!PeoFU!Ruw3>*;`z*s4afd2)6GORGV!dj%aP5USgHN5pN?VYKv!4Ndd;d5B z%b8yCeL6ZnJ2H`$D#{`~0oi73jnPnZ1rBLmLPpE`4d08h2TE{zHs8kIN{Wc^+qXsX zE829^pp#}*5?w^m*-#U${u3E3PCRs0rCAwdsb7Occh|dTdRO>MX1CEPJx0Y|&U1y+Zm1wu>Jz)uL=-t1VnMFs;aK1e_WNxY<`WeY^Kin3gBAAf&5m$1O=F>YF8J$ zyaYZSDNaI1Z?=?u^@j~L>*MrdBkAfQzJbKeNS0@qYI3BYTPeDfahq3W1x)P4|%U5+?#NnvAwA zin{G9nt1iM-HyzN@^~jH1^W2C%X0-k^MPo-R?OzC2K157cbJ+91;Doq%rs84yvlqDNIID|-M<E)zT^LjS^dgmcc4s2Om z3$hJ(`>Qo<$s&eL>RF}-CCKY-Ev5iA8Zb@zubQjt=#T@1WFv6|fQdf+gzm5F)&FY& zPELq!^E%g~bojY!Bj%BnzM8N_BF6xQa} z8GGfWXrd>#-8`7tk|sk?K@8irI=c-8t~&pKA#=5ZU3R8;^UIcC+mlyG@4KwOU?Uqd zSnn@;(kVmRC)>(Tyht$8!Jx)8Y70(N7i}TfI}1vLBjzv0`)lN-k;uHx6CYx(O_?(F zUXy#!c)f3P23DxuPGcXH7gvw5d@LFin>fh1C7~^pi@9+047Cd!j-h%s#wPB^9sWt(osgi@YtW@VtvmZh0^1=-Ta!I72&dA{e1MJ@psipH0hgQmSK1LW*UZ^B$l1pXvhqLqo`zxp=f$1cbbKp zzFxI#$SHmkj^Ye2bersx~iu~F#mliS;*W|Aj5S-?&BM< zXkA6Hm;z~-*kWhScbF%4B_Ep8`vjek5l5ZMg~2d#F}Se63h3<|MKB{a!w}i#20h3)!=f1BuIYC@s_|DJuT42~%P%c8x&~cHVkgjsz*gh3 ztZ_*<4ci&8YTnj04|=y|(ca8kDUjx@u4&pS$jEi1q`JCZvv#_kTG8t9JO8>XUp0W` z5fGqh6b*{n2FU<>>u%k$1X@8-An^ICR<0|;|K9vmEM9H z0~M5$r+`%T(ur+mUQHsu>ZRwVR{fZ@tT~Z9aZyzt9Jvb{L!H}A#g}j_;@U&a5@%-) zo!`(uD3lmb84)z3>`b7Cj|%}d9#NQr zGa?BvkE0)U`(9S-T#Gi%vnsWE%(Z7;K#iQitiKD`@BmNHXD+Um%a@OB)c3%xm+CF@ zoUfqq{%ul%Zz$lEs5(0O*kySS28BF=+V9(=>07|+kPXT)we2!w{Nx3wd?rCbv+UoI zKuO(}0S}hXQF%{MBbNV}qoWjkgOBn8a>WYzPET#*TWdO@flf0+8PE*_yI^3T2;ik| z?&cM_bG6*(HBnoZEuX$7tg~pJEzg;E2B$eT!aT`IUEj|nnCI}4$^9EJgbJp!zb@=< z9v`~X%~+S&o@FU!XYG(ky{ydIh$eVi8+|<8pK%N6B^bWJSfpxZEIF4$fp%(QhVq)> zrmfb6*yhg8 z=+&IS_{tYBtQ>(V8^FWYXgU2CWJrJi_n)sd767CRf~+kEfQRTvoa6<~(Q8uI6PPgq zV?vqHm;a3IkiyrHpbWcx{@&aFP5AsaEjW5REj$fS)xi3Im_y?C9g_DTQ2P4LLb>X{ zi>-GikO`O9{Tz4eUVjT33Ou^C=f1k@zS{8sARugA5)avn9B(GN^+X%5bLrRuZWzD< z|9|n~0F8HDRL#J12il|_7la0%wRh6b=>2%@J&JSm-HWGhbhq8WG4w?rWNialk7fNw z0F*$G&iuIXZMq|!8AMJ6X44=rjQsk(^%SZhOQKtdEMTURD`6KGuRdEy3V`sEX7xgn z6Q#D_ebLQUD23flKjDb3B^vGLn6k%FRm==dYZ>REf*O&B|3^$j!;#}oKTElhjPJ z=>gA=*tZo}Lizt24fruTtDNh(3lAiFw*wX!;792KfCJluB=4-CoE8vh1oT6oLM`ts zEMN)Weik0|*?RpG&$e|F@1aDW_5|{Zj@W`UrJ!6&lDA`Na8L#&Ku`G74=Z#XG-3&| zm3sJ&dR(9NzMnDe`CWd?+gr8F6O07w(NU^g1i)R!z%n2z3Le-6aJwS=4PHuv2~Mvo z7k-KZyQEx-0ek702H}x_;{oXR9py7Fq`=b9d|&g2^Qtw@YF*cin=pHiX~HXDVst43 zz)qq|4;Sl*2nJbkj~hcqrL>$m#cT@W!)PX`TbHf5sq<&|GFti^xm|FhReiTCt!sYO z(csoG{dEscE^_|MriTgyq8Q1kYLyX43C~!?XJKpfUQ)2(;>x(wR&uib2v*TwPvO@m z4`>K>f;Jg8S%qNHIcf!JC8RGV{l%o~LY0zF5QYiLUiQ5MFDiB(>#6oX(_jLp6lMy5aw@TXQ-$JPbb zwCb!gw8`p1p-Vw&1Li~xcve-5JARkYp;v{?C^ju4L?BmenHju zWj^n4)c<^x2l$Xd@v&{b&+H)5Myt4w`w^(3#|7*+n)Qq^Wx3uJm8w2iopehp@aFdX z0|Y}3XnhZRXw9zZVoo1#kKgvkl6o9f`0GC$iMc4C(D2iAllXc?Y}Y!zNT=<)xHn>d z`@$3drC!6-Y-}VFPQV?aCcG$|QZ<8(Bsn zUZtxmw%=>DkZKpFrL`Q7Ws6VtR=2H#5>Cg?v(yti(*6xsvNk|OaXBlU)*TAJKjLGe zItmA2VYE6xG5(snUgusL_|bhD>>8s4x8NI6t*oCdaS)AG@j&R$Yn^qW^>cGZ_- zzaUbB2`rcBR2Fg$nL3w7sOru0&tlbJho~f$PhU&lenMQ2RFgvIHeyHVVARqM9-iS@ z4fKJ*MvP(0dOS6PsC|)*N8&e=Z_xJ8$~#{EJg}r=dO~ioHtZHIQ&Q#D*(~Dpj)FWO zU-lPE(-iBK(m#AoMsZtuF4tD$YNPT9PCCZ18C<)`OVFToU?fuD8(bkBn`(X}hh4J4~=&`(O=`8v)bybK9ZXjD~> zB2yZJt83x6NBnW|PGLWZ-p3255*%J6!?QFEln+Omh}s7r=?9nB5y!_E^1u284}@7%4IWto&oGEM*T6sWJ@zH22A4!pQU zY!(8$FtiH~{l^-#k`3AO6o5jyPeE1dCryDQk#&2g=+P*tVCBx!4TIq1ZN+vSpSXX4EqN!*mbyHyr})!gW0 z*NrT?qBu+P+gV?8NV?q`tgZ-W*);?G*v&-Xopt#c5ntABip2uUlNKc_Y)_5!2b7q2 z1FhHgmO^sc()fBUrf>k!F6{kS9t-{ui`@psx4{dz7CK&t>9>NxJJl2`!Js`Jbre1rvNhXV1$(si>55D~tWP(l@d2 z6I`g8%qpHU&k2J8BfJUIKI@a=D?UEmmsza{m1r~eD3u(;;RDdj_k_}EWHwe_^d*yMKrqERq*bk8G|FuaBQz@=T;vP;!1;x$U? zLK+<))!Q_$shYtDnNNbkx^G{Wdvihda#X)#ageZo4%7t1g`!|S###s#>1u+z%S_ey3WP2Cj+-ztct1W|$_dFLs)_ih z16?y_TOdnzw3c=&E_3LSTX5qkjkH9B7-*{y5C~$cIe#buzOC%+e;T*J<<3*paPy{>EvGEbhp*0^3VP$G`3(Ul=u6|c zOwO}gJW&rN1UrvX@Q!-W>;VJDBj*4P9DHp+*<|BKcB8VFIpWe8!j6~0;4~9uB+y`{ zP98`hwCKw5a)AG#Q#p9`-y?SY_vB0r5FDf!epdtQhV?axdXistkKdr2y1n*5q&C3M z`tP*%ogSEVN)vm9f*1hE1t3t@r2l_)(=KM-wx<7qrsv1&*7Wo3(epx?(Bd^1)YCqV z@Hi;>-TyekKhlnC57A?CP4D z&Fz@{wan{oR@%dk-_p{O?o$j1P(|R+^9^w2`&?-O3uB*&UCRdsUTXJu3RHYNfi`zvDu39&*0+Jq*TNp>o30oXM;qNbRLXL z5L!qwITG-cCZh6lbn{rDf3QNwFzCim7AI!~Q(T{Jf}dx_S<74!#n-!cd+`5^Gtw3M zcoJ}7%9K=fHX$bpw_3#Qxg0;R_OW`N8#qUDVe}p6p6+xD*_-2SA z)tJAr{ohn~$UE@Rkfy?6G@;83Iik`i*%eG%GZSkAVlg;*Qo$#&;h^WhYL{NV2Vo2; zJpKN#`F6Q}rKiKd!JptB3KxK?8$P^u6QAnxbrk9LXj`5ZKx7$29VrlGky_HKo&<^(B7fpGX=9XWz1GX${ceS$-kB`{7D@)?^d zpB(S(=Hs63$Qg(stTT z8M|f==A~jOScQm=4zXnJYf|UQLO%}gjF1;|mzxl`80ELP{e@Lx zL>xg?&-?<5s|GB+B;Aa3o=VJt`jFFzGWm(yOQbNdc?-Ts?pjC)3@lG$C=R{AXdV&vH5~VHDoi<&Zl!U2$j+`Rok~AyDU= zWSZdxUL==3jRpPb_Dgum3W|oUR5D0PjZUtr9e?5ygBqR?s#2?2<9le z{3c!Usfr4MlV?G-8>1q8L{Al$s>r@-DMr$(FytTf4;IRRoY2{dfh69LS+7+YL&V-1 zv@JM%nC0y!h_2@heS^oMYQ$A4EnnF>4hrpd`Y{%G6r!cBlZ;J!^vlgav|vwxuYmCL zXThA#9?w@qUdB%`j4UibIGQtInX-dv7WHU65x{@^`|K-T@7o%w@=$Trv$|kM zFsOutRNUfbI>paQ2_;!A3*V2&=r-MJX~c4v^d_z;6^PsDP?f}y;yUYY_I5lx(pu5>d%^0xL-nwW2CD(5ZRu6#a$p_?^-&}yAubOQ7u5H z&Yr?2Y&T~8MCOm2(4%vW7l|*?TI1k^Onqd;L+5#^Xdf@U_WafWV{{s9aXN4*5c-Es zLXz&{q6MgTa99i|^QOP;4grG0LX~n>VPVA*l>&`Q&QPsQ8`Rcy?|Q>?x6K`Yte{ZN zBH+`l-UDs3-9JOLX+qTjqfTHbzq@O)mCg*v;)?zFJK+VmdD|s7vai@e`%695h!DBxCrSxttjRzE(efTDVdYhJv3(**WN5N5_O+@kR*tvev(M6 zX4hc$4ss)-Y^DBesUaj&drK!<+loQ!^;Bnf$(w$B6IO>oaqM!e<*=#-$~aSZ_jDlK z^e<3=yWLYGNz%<8?*RY3cq(~`9h}md>g0{*zj%qMmwWzL3y!52LNQAlFJUP#afp>k zm_smv{~n!zEPCI}+yS6nQTY1RVPejld#a>}d*qsW*>Rzk=y4Q#RAJnnW@U+@re8k3 z?^ds3QDhrEX4qp0io0aHsYt7xz8nQ%;yG^O^;KF1#kAB6BSnEu^p>nc8?#-9org

#sV$13Ux!auoan*k1{`_0m^_X5JAdCWd!2sf?!1P{{ zG_m%wyEvUBecI_-%lzXT%JAQb3Aqqj2xMhB?%}^VHq|X@9fP!@jA?Q!E&3jGbyo<= za>6{3U=AceF&1%S^Eud|)zu5&4tA~y+i@H^EP#3EMx1d`xa&r}G$%Z#lcgkB+QiEB zm#hP&WLHq1={HvB4GL^qZmLNK)+I4w7*nP58qw7?Z!87!F|#F{Vyp_bo|DYO7yNwA z!P11WSFDMkFEM*OD{j->b>eckUUoDQJtPZ~9W^oL!*z@?m=Pj|8T-)<2V=Zcv-3E0 zBwZZ4VeAu83S%{aMcEkLy!gk|oWYKP{F6!J{doxfoHuIv3 zcli-$M%LFAZ%ljbIn9Cbfeqzsgh91ghto_N%!POtF=7$K{`S=yv3xfKy`X1!My5a! zL;Aw4gcZ9Lz7rj&bK8fMcl+G)(K)hR3=#f%I-|^-p{=iffhO)s5#dqcw{D?axd_^X zjGM9l5Dv)Q8%?-+c1IwK95xv;VUv=Q;``&M$>>3G(tU2%!XS6!mshD->#{cz=T=ml zja?ENuqFHXybgK8An)o8WbBH7>($CVy6}_}k&BvVDy~NcmX$ydz>Pq{>t$zcWyz#Q zoqj_{ZeNimLRt4yCqUz?s}li&3;FPXKn5Tjv|cOLBd<3jOc#=Ixh128m=kK0A-aS& zs1bZpyK-P3jnQEt+LS00FvA_McD5dT6UVqe>2n%oe{-Fpyp+BnJmbsnaz&-Th(f=!pnw3Bd49moR8 z{+!-yd{|R(w}sg)6clX+^4_hSJ=8Vf0Fl#8OF)Pv=+*^53W|dFSoK=x@eYF;##z*Z zg9IseRyMqLML4+;4zq)CV>4@OI)gxII7+5RqUu6uefe)fFnIVQU&+G`vsJ^|axI2m_P)B}e={yb-@TK1Xyw%V zlhCK;_=2OXUbz@7lA4m?$ar|G%Jnq%*T?p}yVE$c22)U4(kvHOsvp(Qkk>iibUwd8^TPycW5zPlQP;>7ldO&ku{;I|C`=*lwvQF- z2Gn385PwKUl)c|RbV@RmaM^8rqN0k4i9rsM_?dbkh2lXWrV8vz0gV!1w}2(~G?Lek zy40|fm@C2q{GfB!e6cOFw!%2T%LB}aK!-?LRN|0Vc zNIbzD5J6J{ve1=go$egm*6fl=V72BU(14q)UhQU>BH7vC(Tzf^8>E6j1_kKtu{J>aN?;D+;qd7KIV~iX?lIZsLUqGuny>uKbP5OUEav z*Ix$k5^98a_hrjp@qCvKI9sq5M&{x+vbXaMmpTYNV@#K9xo%&1Lh!C$uF6A#NyKKxTkNsG}Ck7+4V3 z2c;AG+RT;=&EqbfNDmVtPjStx79QT!M;O}rLBjdF%lT2a`MxMME_v|a-~fgfYKbZ$ zn~Xg7$Qdwk*BJKx26kK^{_JJu+=D})N`<8q1@J<~ZCf;IM-Q9q323Dc154{7AE`Uv z4eMc-*4v)djzI>^8ekEIqSE)D=G64O+4viGmE}o;ZaPz6IF!@p>h+drr!B{=o+-d! z)c*1Qs_Tn{JZNNP6 zU!Kb~prkvs2jGI#Aj*TI%)Hf9-*uAHmKT&J%Gb+|OXK96g7Na6b zH;hb#63ag(nvfZBAfTMMz}y(|^9IK#34B66T<N$7h^HDfbkV1gcdZwjF zw_e@qNS(5?RIQsj8I*0_)avMC^H~m~VU-aO6o}CcgJk)DgotQ;^exaT+8-d>7exLB zj#of`6crWqDL8ZbNVaSqu*QYcmaOx(rA*xTpP}e4AAsg(&nzDYy-qs=VLgzru*V?d zq3L-#-knBd)>+3!n{xWxam=0H4)vSTyX<`(kKxJvr`7DQ*LuU5-?{g`X$?UO*u>IOZOoceR zgu7-6&Cs@k??Dp>##J9DzsF@5yL@H+an>WY z3`0@+-M@PJ@gFLIDM@Olq^vA|$Lh~eu}I)UHYrl6Tm%e1K%EVoXp75_e-~pH7uA^! zx+1N!n4*sjj>xb|TyOc2bhNVsu_ry`2(f-6f-(tc0{+r~=v9vIhmM|y-z`zbwz&S= z++}y;dQ2jDf0kgfIY|bbuE1sf1)|O1jBEOq&qmZ80#Km>nWQ;6!yHq{ai*O+!o;do zD_O$Xj>H+pRGBSQ^52pfgJ6HCB>z^3xr`*Nkw$S`q%495qO8OoK4E4;7?>j1P>Ggz zd@io7AJ%Eq_-Q;2Ay*>;bXz~9=_Oq|XtH;`IB-iw>qYd29uc(u1? zM-f3KzWvjiNac;y!9b#elEam=33g^19qxF|G>55p7EQq1 z&iIF0K&w4seNvoUv(5qV3T||HVFJcFA|k51>04liiU&lO2t2GsjaSk!5rW4SZf(jL5MW&3E3# zv+fY1|Ds5}t<`st%A(j#sN>=XE)?c>t#$NBy?7AGz)!$Ik??}oED&u96gCHlg=k|B z(L)@-B}0!FG{ODKwqE*B#tCRL&kFgV*hOuGk#>L|Nr!UDvE>u0YQO%P;rNc1L1abO z&9eO}1Vjpe3><)Pf!O!DKmMihUf`kd*?0H zrf+ZF(1s&>Y2jR*f-`D$E=p$xr3L^*^=xwCPD!xACHveXPAFpjFaNkOJneAOvAXuM zx4A!g*p1cD^0KIrv5rTUL~>l|9B@W!= z<3eh>Cil+Vt-3;KXaib^9mG)7h;+(vv2I02t2;u;j5lsV5jZ5+k)xe!UnPy7&5CW! zv$g_U=eOIasn6+o4>Cv_UIE?RV#ms5&42XQlse4kHWSF3?7XQ9 zz()3liCV`A(tVHQRDhpboHStnZJdW4-e;slEgR4~T> z!-_S&4<;uok{E3r-W-DS4MB^IqN!&)JTGm8;QKi*#$>^?S$SEXi(xBO&Y(OM`#%LutsmjO~ zdslmvj2@l@tC}of{=+!wZ@yLo9^{Th$15RIx(C0( z9#4FRCn6d|5Ti&L7B0kQ*)KgSp8^qiEtodDF9jv6Sm9X53GniWNK9zn0Zg)VC7(;a zTqWuuxT>^>HF|DgTy>hsF8TL7q8`wW{=LB;PbTU%Wa-{rN%oCTD8A*Le)JL|>D^M4 zmq*;(+*E%2nf*IZ`@(HrqpDQ?C~}xI)Ll0_JNxA9tf-+Od3SeL@?K}=At3xw*KwsT z2uM~*1JQWRC)00ASp1^i-~B;3Ejl_nxIna&wD1eM#~&}t*cN~5B(y*_>NUZ@dF_rL zW4F1qJ1N~~OVxu8tr0a73B-c<(P;8tzoNhov=6DmlN|$pt)v7xWqb&rq6=Mh?Y3Vx zyIr^b$+wqpIwpqPiX74^1Bb%08Ooc;#7sYed!c4bBTmrIj^%mCx(VwdUPMgI-};0_ z<)ySXho^*u>5Ee2^)f#GDO1|$+;Nm)F&Vt!7VFpe_I;m7J!%SXTNA;xDCw^%9WE%< z11Tz{0XdHbh;maH6?>TdI`eb2N{PdKLXMo=uR=G?<_3C*HSLF4vU#dSMMZ3D9VN12U?Z7U z>R@D^B~N9-Z`N}18P90&;6J&ylX7wdtyiUO*(*IT7Dk$_E5qsN=vFQsB9T3Rr5)fo z$zuPjA>*5_ed_m?2xs$>yD7uCR6Se6ic>t>>0 zo-8Tra&WKVSfXXBw$PLdF}*Cf`75y!DKxwxx9Sf2Fk)a_yct>U(QMjcU3N9&B$Co^Wug}A^nDG@L@ zs{j;mXaR)S?~+zgS|~EbapRQyNMkah&@}OgJdL{Kor(kdIp-(bURyxA&2~tz!}njN z0k(8>ZGsk(6et<+TzY4gd-pE`j$Diy&a|O$t6vTd4_$uq9!WI8a)0{8MI~8HSOH1r zt=h1rtD>TE+~{m%6!`Mu&Cbrg(&crVJkd);LUQ!}dSwaR@B();;u|iLOf=PTii|RF zeYlwAv~5oE85I@uC?rCJHw>lEaD7%1tzniK-rWXkYiqa!1Vul8Mh*-NB8oXG1yR3rxi5D5p#F_qH!c4m0 zy+c;53YLpQ_Xb70E3Zelfw~Z_yen6ha5BZKNLoMG*CcIuoqSce?bWFL)KE|y6$V2q z@&eBbOeedusP+RNwS)`x$*H8B9=PA^x*Xnt=l4sJutTo~8Z!bVBBY~(ugic1@XZR+ zmGWB3$-+M_tHP=*9CO7CW7zD`(bDGE)v*F_ktxQlFcG-^2onD>;Km!DnAih#!V1A7yxa)wf%{9GF<&joJqL08e!Lw}AK_vZ!AQz#1K? z9A~REIf-%1c|4XVYyBMo{JS?J-#E5VA&?HYxt_YWaki8CvnhkIg)6})pCoo*N8eue!OR{fD7}K! zg?YiElBhM^pr}S4#VUY!nW)pi+L7$6iYE9f#!V|g6YuC&)-u`Cs%h)Xyr)T zs(y6*`kgdbUf_zQSurG&P9mFp?GqyCVbB+F?wH1FHnf}UHkD3AMSIM+*H2{_!v6MZ zHsDfA^0(8xf#ZY{2GR^0F4_i5!%^jDH>t=exwUpyCC~ZQRh}}Tw*WElm{_x?>{jZ? z$OXsbiE5YS#l;~3$M?d*|Iqc;VO?%r*C--L3P`sgDGkyR0)ikd9n#VbA`OxfN-5pa zptOKANQ;1gNOwzjfAjY|-+P_+T-P~&?Y)5w_wQb7%{k^6W6m|Au<;V7MnyGF$kWqP zlB9e%tX(@|M=BJ@Y5Mh87Bl8f!vPMo(?_izr~acyF(OQi7)%<#>$YWkIU?Z6`{Hq) zr;?p(%AXhXzkQ@3r||Oe&F{HuH+u3xEhIiaODl&@9Dy3q!6PRdl9K!(D(Vpzm;dnk zi0CbA)P-bZf=w;i;fXfwNR?>Hf3%e!o(tOI%V5!8FnHQARq4@s-a*A*6r_A>YF-A#% zd&6qm(tBu%Kio-(zWKN>LK*QnEue)=fb4nKaNDay#7V)x>Cue#L3p?~j>FRl=SA1-%8#g{T}&N9y68iBGL zPcjl#j9a9pD<-!ek;s!A1@iqSepDo`fBJi0@@b7wno@E~3KXOZAla+bWv*D zvvF!EKcJ-u@QL^?tCCtgXB|)GGGI%S7XO9DyVX}5 zgDd}u=G&ixs#tbCb8(laxkfAY$ya}QzPcIam@X&^aiAztcW+?L2zb}ry`H;jOYl7eRHG>WhmDoo}{e8X6kR;XDsn{yk0jlt4}$6$)vh zM3c3;x|)!X@aLvS6-oJg!9CB30han{>Tr8CwHJB#p-dG|C#R-<-W8aFhBG{`f8_mB27#?yt35MknKtN*B5 zO)vI(O^KCP{Gs65SSdg5piv3uBq^rBm(o<`ju$p|Fz%B?>C2a@4=RmRO;^|<13MkL%^re; zLlfwtjmIR_0=lBaO8II64AKr(9xP4;%$Nh>HzJ5TmTnIGu9e|}iJ_+yPf`)7!u&gC zruF*o-@hfDoq6CeXT;^o^P*;cZy?yZ74_?v(!}JX1?BF$XI0n60wir^RODWnh0n6H zjfdpM5@RqdNFj4ky3dtJAIiw zM~9tOpWaA(zNcx1Y5SVzm(E7Pr3Pl8UF%i9PVO9OEU9e*8RiobuGlZpLY7D#ZZU7o z^Pm3l4;tdUUy2zX*61UPJWhxd7s-afvVUSXjFI`WE5ZJ+iH>t%mcBVtwzuj);ZYR9gsckwe8uA$s2n?&G=m+?H-P)1AH( z+Ad*jy%Fhae?Y5=K|RFCMD?pM#HA<@)6!BQyKt-}6iu;;vfM~6guZX-*1IP(c>ANb zqdzcskeapAD~m>r1&%FtnDJ0mz8>Njma2Hx=J&L1`@?ol<}_oBRsq)tr&=PbiZFxi z%gT$iJbDRFV(f3dOC{>Nk@DF0W60Fz3Xl>|d>^*Vv7q*F9k`H-Mhh{=EhW1fwsNNs zlbJ~&N&8F0M+e7z5(=RBidpvk0ZEP5a1g`w?mSIn(K}(GrjN>=n#G^sv?2W|H7sSN zUTIgD@tp8&?_k6U#ydPb92^;WYG6RUn(Cr7rQD)KlC^qr!IbkE&d$N{q{Hy&Xz;W> zlc;DKqK~Ixw+*YrpfMTYbJKSnlIH#!f#Ug}l+YKd;QmQz;P%7>Rq<`==$g`w!MSvepzo{XyrYM31r@j#&!-Fd|`E#t12>6Q5wvd|Z6R z7ubrI8h&&kwZJvbrkW!gNgex8J8N)~;_B*Ak4kJ27k}t;(+M8b`A;gB0VR^HvI&uT zE%#76KOF76&u2MIK-=vfM5Vc)^^s*972CU_yQhF1bC*UAGaL_HuFW=9J03IRLfwDB zUiE_pXOzHnsMna)d*4w%jb^zoL#63oR7BpzoGl!jSz1|T<>xakiIu7Rp{rmUV8*Hb zROF5y%8?KO?~)P+4LTiIJUl#9uY9oH4q4LE(FI0EV*e6c#HN=s0S+HT)R+j)rbNIj zka;C*fVK%jiln3Gp6m`ssbs`MpF6p*cs&bL|HkH~(AD8^K6q+55vdkJ;?l~mc9CsT z&$veG-(izmIA`lt4ikyr9Q!_wjHpjDJjfrGgQB{sKLka)~+!;Mv?{GcCiX>2L9#cIR)68OF(K(8QCQ0Xco+of7<o_S|kRl=)*;sA;U#4vF% z=2vnVmY7gG^dT zjEi^P?WecLL>A`VJGTi}m{HDOb%rvEF$yUf@nBF&si}w2cmGt|tuyvkVDze!RBL`( z^9+-rzU~E)OgQrCdt`_3Cb>rsEPtj7(a^H)(+}let9~>sU9F;T6|!Myy3`47L$z)F zokFSQKpkcAqVu(TyARustjVpstgKtbqt9QzJ_<0md6r@sE|t(H{qiLh0Cy&DZt5t+ zgoiB$zZCJydE9@0%L7?U8+f+g7Z6Y~A*xVY6bS9khij_za>;+-1pb?C&Zwy1!NkH^ zJrrzQ<3DoCF4(h$!lS84tVGkMTbEHJOiE8wf8$Esr+s%<{!v1D(44ln-k?dRiQMli zm=;ymOK)tGblzTDrkz#)43msKR11PdvTu8t#7ENwE(Yo7xCf4I-KY75`()SwuYv>S zxiTblh*2)*PUI7RauNDdqk4BSI^LEvu2Px%iK)4H$@o^eOYG;rfS4k}dw~l;&L2ZM34;&c_dpkwKTRuwY_mX9s|wqN7u{cHp(z z=G|co2@J-YR(<>SLR(u_Lt`|+ESi&068(5a2CG(x--m3bXgnBdKV#{ctN-;N$;Mhzz8)t5^IWbkNmx<7)Tp(p{wW^xMaegpq1b1I_5AItwE09M7A3o+xv9 zjTvV=D6$JT{^*|_o&4_sAI-dItc>4ZuJDrcd`ol``aAUu`+bBV!e)f`Ybj}ho?7#x zL=J8QJALpt!f<3@#U^=(qR?uiP8%lY<9Fx28BU*>&?{t_XxATEcmZaC2^*iJ^!JDe zidh!~(&~tVFcW&&_!4EW;Ia43jkut9dE*x`RctaOs&TwbwuH^Ye2- z2umIth53KWjN2olINV-;{2woX*YYQ=(+#cW-h`$@2B4LW22V}`?qeUV?AN`$MT}=eVZH22U)x_twtwxEXE}TKTf$L}@=Fb-f&(+>XMTPD| zrKIiNp>pNPeRR*)I;{x_8-;L+THyH;+J_I<6X!<#K4fF>3iQ?Hb?1#{!?l)yFBN{< zv@G%B$b~G|AReXHJ&cjB#w&F5*D@r4GOSR5x@Q{jGR~_5S_);cvpP zU0of=UcT8YRQ7p(vYMdpjKSG-YO;32Bv0`I0%DS*Jjp1%hEJgKr;>XZ=0!->!5I@6 zcwbPW2*`E^eFsC7f30gsS22m%9mZ(VbedY87ZL@OxRy->-r#CVH@Y_(D3}vGeB}P? zQ~ukd-G3LyNHTXA;!9Q5Hq*`>Vm{J$4cZrwJv_ zNjQyC%$u}ytdGnhRl@CfL}R6pC3KGsXWCu>SM(5XK=gzX;c=gZz9G|!oU7BtC>N+y zo%;F4kaqu74(xYQH~>)LbOa>4DF&%__o&n_4LlM4KpCSIj|#7g$;rvN*niT+L|~Vs z_qh>&w1s{<6eKx&eWseuQ{2^V^7c^27s)FMx18P@Ez08OeTmN9lPx8R(YYb@ho;>l zpY?L}u`_d%B2jaQ)#Y|YS9ndwNrw=*u)R8$#r+J zbmRTg!FESs)Oo=7%!rGHIsAveNv1MC&C@|HU0Q#3!#RCMb#~H{=g(>IPT8aX#dp6! zGPr$9(KFQW)@Kwx2Fz6l%63~pxD&hJ>sK9T#ghYLmN>&LFEc>BU%zs6c6D*5^w;Wh z%=q2-uo3rt6bs#RyG4XWC_TS6^cfK@_4R*lj22ro)ZQ(n3=yBT?Jz#&4*RqE2i~K= zH~Z=EAxpEH>#LHN_52IXru?u|#Lm~~fI3MZJTcQ*$&TPKfp4pcZNo*%x6aPaybn@l zm(t$M3^fkuIW$W9`ij9l0Q!!ts0gAj&Nd_d0o1mrQ}zP){Vw-PfGpT~Hc>`)h&l(g zZHH7qm+SU%QSf_BN*o*5?3W4wQF*xd46b&67`S|#hu(5uws^X

g2tn(^#??EF-GxzF#V$zh}~+IqhPnSv>Fd6KK9iL4GSo{0J2lm6w-yO>AZQ&FnIw z2FUL#gZ%duNFTF3s=Kh;j>LZckp#`rmDF+czTQe&*D&U-I#xon zTkT!3;_B>N1y2%?d&}A5Q8Lj}$&fo}jQ9O*qq!2F9Qko~Bg@6^T|5yYo=Ew$lv`qy zn5?m7mK$ZvOMZ4*dJ7;K^TUU-uTg|948+7;Lwo7 zvu9|qc&RSGgw(a!Fa2(AMkXzJo^b!>-g{$`laIl}&Nab4q~(xPCfmyQpkL>bEHYq< zsMTn?6^D$M-CGQGg3uAE0!hk$aG+~Au$?Lv<&H~P4PF59kQw3WC4Q`bwr#794T|2i zlCj6zTispOd%5l3(E@rebI|b5%{vFg%o#6E`H66t#;U zg$0!_UzLf{1T=(uO#~U;>eJdo;hK%-`bV09og^Gm?K)0nOvuiHNdhyQQo)c^|<#3d!Lh`L(yet{7*K!_a zX7nR!sBv@$z}!0ur&U!~$E2o;%$rpd;XTKkx13$FgTk!mipTUbQ?vpS;R>c1ZscJv zww3RnG4?pauSTLQStwAd^b`jC6nUG|GIx)Rp(I+gAlvV4} z-ErB1Xe(ZHvrMkrzr+|?;+xgFM+vQc#P0X^_a7b~ztGivZe_KeGd5)#)UB>oz9$?^ zh9UFp8EndD@yiA1iR(q1;n@npj1dgN%B;y+Uv#BCJkGKU&3?&HO71S4zy-coX=!O@ zuU_r#_i-M;)>Bbcl>j}dX2z2h`hhHnJrAT3`^}-+9)N9n26z8bPFy#7oj4!#3n+j0 zQ&d!BHj`Ke?>ZMSsI1NpzRv$A5reg}S^ z>2IzjEA2tG{GMkcV)fzYt8Ybh1klUTGcd$_{K#u{&Uu-$>MFdxz1<1orzh8_jPn-8 z&eZU4!3lwKAEQ@{wHvRdD)G!c6CepKL~zggF@o>oR(D~1)X+n zeLbtLF6rk_1%8Jah3DbEgyRa5X5YiRUF9evMp1N#`XRu6Oh{0B@uFf5S*9MF?3X^1 zX?v6$Z^t}XK0|Tu-&s?eQDu@mC3ZhBrrsX6EKV?Mv!3!2;e} zcM?%WU?;eO+`;jWjzzg~`yID`f8T;Wq=;(xdQ_uCyF38oe$#SoL6}^?Bw{d)c60&7 z4xM6KKW`QzEA zIHU)jaSHwPd`iR$Wdb}(r&)_Bcb2G=$A2hw2>(!nS2!>C$`Ml)m6uK~cCF6%3q%JE z(WsidDDzFFEQ#-@&Av-R_cGcThvNW8&*IJR`xG&<&q1ftDm$VkiKy$i^|J?G^46w* z&=`BnCr;8K1q?l+HEa~=%y5*}74en~w%sMn$j-UgG6p1Roq#V|XF8%D(u}pDrlycf z=74LjG`|Z}#DUnlvevz!WB;NZB0JoIA%Pz7JdE8 z5-GpZoE&;Q&hP7LJb5U|WuY(h@YPu5?i^$w1m?QPgR}+&rY%l!A`_ zu#zU#x(``c78W8nSKzXEk%yS%Va&D6S;DOYA#Y}V5;itA`NBtd>?AC9p{1iKfgr3@YGw(z)ya@G4RgsXM~+- zce@@;29RK?k+GxEO#}pg3!r@4dYjJYpfdEWk%NImOn}6_QxSao(5KEJMzt5Gp(Q9w z3CxGzdR;weyPXF!mB%I zXzTy}{n$J*Dc9b;xQZ*Ho!5xToj?;LQCFMJ%VRXt$@~7@Q^`q}|M8`LO-y`XS4TWe zq7{YM)aRMVJHh{WId47V@l7Bb&%75J%bVYu)k@#pNATB^{*pSZK?9R+8sGQk$Epe; zscY*y?3uEJ29|4V#_nz{(gXj6J94rmZ1<*IBcdeJG%09N8XFsV1q9-N&{0%;Z~vN^ zdhmQbcs5mOH9#+~fP~V{&W@0n_=S!R+3SmJ_a8ea%6RJ&dhV|26JD0(+G{&I`SAA4 z=Br1Fd_i%Va_XXLI9o{(K};6evN6IZl4B^|-(1~f8!im*B=HG<@=0BFtGREvYCsj!7BvV#VU8W;K(b^|UNItG;|GN-Jbho@J4g z-6&y}Z_?G-Sx{U|hlgHf-k)qm@KDLT4Yj^+*~HcsJAW9XY%TuWA`mY?{CY2H%KM-n z5O#dV^EE%er_li2|nq!cx3y$sWfmH~2R0TfKGFE<0Zl9`+2)+wBv zGVil}ve`(_A-f?wUjUllzI_|?e?)jAJ8e%d_xmNI9(vqz@RN%g zy>b&&A|j%Dlch5pVQ;RZ-k#s?3O17X`oy_?4ZE*l0xM)Ug#|~EQ9FLIOVmVSkAho| z-Z@?#o0?A{kyw?G)Ty-$*ZU5-g6coT3iOHutY{;58s_UFYeM_)xXRaMpC=qUIuP(c~Y;QdlE zvyxm)$@zE74I$7P_#9~)8gB6JfM*&GjTq_e&mcA4?4Yu;6Ru^{gD99;syC@@bs`Dh0gpZO>s z4KkJAMBNb(_Q1A=8y7IoqMb7?hLaZ6|>%-__w#4d#(LW ziWX(3>4F2{Q~}rdQU{6Q$WfCXI)Yn(iAw!2DCRDT;UeN&@p|@2{9PwI*V{X_#}hVE zLQO?Tgpw+tl0hZc^YYC!DvFl?7AB_qdTs(EsR+}c#kzq(7pYvVIXlwbJMF1FUUcjs zBXX_r@x<77q(-1dPC3LH*s7k~Iy^d3%BNSOJPKu)31Z+mus~UnNm9)EO!2tt4 z{R1|i87FZ2Z%#S9djCeERf> z-Lk!Q2%<||lvW0>Q78J{HQcZ&=p1In%o`5=t>fWomy8cjeq0@HOidT5g+v zG0s53BSkKAPKubSl(7WkY29jnKPk@c{(k;Xi+5EG$+V zelshXo9T3#R}M{0fo&YeBHhEhK-#pZiaeo@sKP7R`iy6$_+N~PySrf3(^{+HJTj(z zi*{Fo%wg%(I(?2Hyn9i%j7!uQe zFxgXPenHx}>;6>gG#J;tM7NrSL%C(*Z(*UcSGBwug#N6D;?-JA4c6}2T2m9J-@w|F zm}Hdl({a3C%x#BjXqxBXx}a@fh?9GSx}g~_TCar5K8{pScuz>go|H`3{2}jpU8c#G zxdQ=@hhIOwrR?l7o)x>j)6A)=vA#0Aio0-h?j+brMoQYU?uw6&!j+_OdgTkXJnVe> z$B!vW$0$B+9KUFvcT|DeYI19(YFH`dqmyt-RdOMVy4u~(L0)e;G+}c9Mu+WO?s;Ub z`BlG;=At@!$itvC0a4WZUDWt3!bA>!cndq(|GuQ#eQ*dMpu7Hc%0z38>PXAuRV{^C zPRJQ!^nJ=;tLN9ct4ZZJ;A4!JPNAF+iw%_%STc``_yhz1pe{VX8nJ!r*D%`kwM+5v z^;!gzPROMsYFCTXJX&-rhpbY@+Is%NL0k9kQgdJ{5}yUyX`A>(H&%+ybGf$zeqt>Q z36B3vOtPCJSPVdsf}jdPj9biCq@E;k%t?q05prv4VB*$2_!vEqQ-t!D8^+ZOsz_eQ zv-3NEN(E9D6l4p%*j`OT9`-d9{;jPN2A=B-y}i8%kyQM8Q5@-Sjl=yEemmXWtE6yP z@}j6su$16U{+XQ}`2Kn`yPQJ)!w)>!j{s3BpP%4Kf}R*OZJ(c=9RO3mBV`t|DR#Bl z))Wnj4C|-25^pM+Lth_wXx@VQhl#!| z1y&@1S?_}!TD+P`MS1zYpdBk0@8Y5&*m;CRL{7ltI<2O?L6)G%xxBpmf!EC@r;cR$ zPIxXRDX9k>4flOdY2QAf$-a#%iGJtK9dIvo8gZ?0P^2kXme#fp#DW`lvl+TWiw%1p8J`EAQ@PER{b8AH65qj!fjFOe!R2?5?+)16O(K>x~6 z<1#63?kCo5Z@1#85h3Eo!FZs8|Ni?|_Qi|eRYFC&UYU-0x*KnV(+J|v+k<6G7&+;J z$2N1Jmb<>bMtxZ^!t>I%)ZnR}yR89G^P>D%L-suXHJ1VbU^`Lu(7wnw<_Q)4b0sAN zqm`SVA4+22{#M!Vd>Hsc@bJM$$^(MM9>pQY#rqZl1yg&j(L&~x;^zV&;C_nEzH$`tlKYaLrp@H{>OTGEeRZc&G783crI<7W>wdwmxLXkCAco2 zHSudaTeZR){i<@0iC3@khlO$z(pfB()T=JM#Sj>jk}{{L$Xd>K`(bz+$iCetO(&vF z=cA?j7qd5Fu1=pmetZhT6iB<=GreH|`ACc=zu5tPMeweG`M}m<3X%y1Zi`N=Hu492 zPx%(6*tsmNwBIYyMQUanhns6_%y4^57D|*m!lto8KAZRQ0oIA_fxAc?K5CU=S0@R! z@FoM!wc+jl`=ly{*^(`5?}A$C1yAICIU0EUov2xhCvGKWXFq^yiYks_q7b`BBkt8Z z+#%{;r>E`($`yHe=BoLb3eu-s&ubHA!v5yMGy({d$}%!{;IZXjphLQJ999%?0vvLE zHG7kWm?R~?%@=EAjDV>3&`}@)JaPRzIkGNReLgff88L0&WyFPtjv^^9IN#TUWgJ)W z7s_R@he^{n-8Cmk$bVx?TqL|))jIzMoAR$s%VMyaIMrN3AM#JFgrdGi#<(!RMv1uH z3c-KOgc?>=h_*cY;)nFVs}ly(J^KJPE3w)R2D>d{8NjRumC?4RE8SbX zfa&@-E?rb_90AJgkEkh2ca-wA_8_>_uqQPB1FL$`|FKHK=L2t2)~la|MIqKZA7`BR zUzGbKvJXW%3`?<292Kce_`=vM()U|0KF+mtd*jyiEoTD}IW#mRJ!@I2Wu?`unZR)r z|55peF0}fnYH5Yy1@PJYQQ#p{;u`0WF3q;#BJJ+&eIP38)3&FBNw5?c7&zs7vDI*O zJb4LYfaVbc=ZL*CR$><~CTvRgV5si!@81ES(}64g@lHb=LhN(J_Y#q zZ>5E@`UzdD6FxH6b3oNWGP86#d&AjOyV0+bKvvhmbuZ}SQp$G;z*9w)b#|`!DyaWl zBP=-5A%%V3$zviejXWV=i!U-bru2VURwZL&S_sLB{VS-`A;I+C>o2?_4}V*al;)Gq zyclK%)&3#te+crG_NNgTQC+}j--b{Nq)uFFDy8R6!A(flO0c1N$Hi(AUVy-id!U|u zUGQBB{@tCM>m33%W#aIi213DX= z6WoV_qAA%ZLm(f1ehM5Xn5Kbcd7+`v0Rl$h*ROd{8_|y_J;k#|69JBwnS~|1yGM#M zP@VNj_N%WAgO-bv@D{SEXNfV|(5S0!em(O~MTaK_y2OZ=959SfVX>yCdDDiIQ=R^I z&4*w|g9#Y1a4CtM-BuM@6$WEARV=vHbY#uevc9Z^%vO=8dlF4+*K~uA76x_g>;eLa zNAUbMU+WmnGs+Ko_U^92M6M;5j0;L!u~Qm6LjMWdSrjZ~*lrl$ZGrr9_wHRVpZ^BC z4ZOPla*-APa;mCaB?v^k#b)W}v#zm%ibCcWxYCOx z0X0Dl1hp3NC&y8#fo?5j|22)ftGc(G{_Ab$;gN*6s!$?FOfXOMvVQs(G~|nA;2Omu zg&jq!G%W797XbSX8&ys1(Zf5o96?RdEK4Ko(L>Cq<0vI58A&;x0u`_6i^P*u!GK3Zbs8ary5){_d ziAoPLFexnxe+oN2HU&93WZ-O|kf4UR^li2u;YsX}d-VrBs})DlinRE!OV^5|FoBZ;lbo84j>n73s|Ue&0pMkW z7uY>8kY7}UsT2P?A+8$vsl%LcxyB?SCV-a&;zUn>e=fMNtx4h~-+_r)oWr`>r^xst z!4V3=uMS4xMYvmSB`cJA0qUi&v*0Jn%DAkWC3A?8x8n^SE^8MMy%!9UK(b?N?5C{GCBHb?i8>R@V1sm*n(29)TF5yFN(zNK_;yNe)b zkg>4@uC{cv4Ob74T7bJP(W~JYvCp)K%T{zBJXW-dQWg2J%uWv$F9ew7PZ=*Wmipmb zV-e1SZMZ!qQ<3ijQygf~U?;oCB_GZVxzUgd5M?5N<*765l^~45q5^~n*3|^V56Pw? zY5+eMqRZ4BC?rXtrusKKV5WK-5D5pwc~7$Y2P$ z!@Q#)qSCMZSJB7J#+Cytmfhqx0lPJ}z`GDw!dru*N?09{)s-~;kE6>@8($aV`(+5`zu5D~Q5E4pCPA0z|yE!-1 z_dQ#FQr33ypK+1m*QDSqQ&qi(NUM;U<}egt2UB3&N~HN=QbPWVa5gW`eoc`XiZhjs z&K~kl`OKIvU%mvtKx|x`6vAGGNy5jWTv|k-{Hbsqq&wJ;kHo~Ndv{)9u3@}6%G(+q zE&lgwx$$#av_X@%5MT)T{YLmD|6@hgGVox4ap9If->(v+PC?)cAPeg%GNxacT!BBf zl@w00=kx%|8n8jC5(dL7t1YSx>RJxg$^-aO^{FuUNlH+KNe3q2)<*i^2<2xqVS!Xx_D1!&hINX5f?$){g% z%5OL7Smetu>Qqa5{s}s95V`p1NY%jLSCK|B@u+{Nf)(FY)T>v^l3OnV&>=QrWJ~fEg@7>IW?$3A+hoVTAmtz_ME%oi^v(rb z{h^YPJj1(gBE!j{e5K1O))stWdv)L~cK^TJ72ESYO8F7r_W5~Zh|U1M;Seqyc`{X* zD!ekT{QG0Rr3DG0N%}#J<$gSFaTe-#8j3m_YH4=*|4*J~uX!mtrUoG?l+FFDwoXh) zKu9`*u4{M>O=pa7sGU51-&MvPJf=kcE+3D@hiX$fj=&Q;Cu%QWDjFCVsQVchF-J><0yn6!ZK7Jfb6S^8SPjU14{4nlmx;ct0;rw$_3$pn-b zut5_1xI3@W4mosp)9@x4X4eKmfe$TD+~n^gtR=}9pDnK|wr)TceZVrXd4w>+LpMim zZZ6V_0?aKV1arFVmbjj3PNauXjkc5;- zLQ$}#9MJ9J?8?TZO(+{1G-R0rgbA?0h}aY!5dlkf-KJxa##d;4+QXDl=5?}No2K$- zJHhh>ra+a3vb}ZnAXKVjV**)*)?%t269zs16<<>ev=!g7_2uN|DsC_kNrHVs$;<08 zG?AE@nL+YLw5{;>Iu^=Q^rLNM{`4MJiIg#2+bNYE@cM0C{d;=)ecy62o(1XF;^**O z1Qvv0JMh7pRbRZoFJH~Th46&%v5a;F1Rz72&ht(+2-%svxthHV8I8pA!|tlOU%b&? zG;Hk!NcZ@N*o~MZHHh|FQm_39Ph(JQ7Z-?tp&5R)Wd#p$ zV{AkkU;xkrfxSv~>LiQHQ?n2jgR14AT2|*LXa&$L7>O59`Ae8}A&@7Q-k3iHd{1$A zSPwr$B{SwEcPhU!el#Y*5!&902#+Aw%OjpbQDC~NfVCZX1Vkfe7WVAp?b^Wo{r!iJ zyb0J0@a0=D;@Mob_(}jZPMf^sOZ~80(BQZuLPJ=d5FgLKDiIo&^n~UKBx}(Af!GYM zsHhOqD#@#mRbAXa{1?3NN`)DpIb2dtkK)+}O5@QY4XBK|Mty_3Dy&%Fxg1SX2}~pRPzxPSzC0FqhlWc55NjhZmFVBGgkc&5ppX;m zjUu+?;GjPg$=|;xUtW4o+o=>&&l5>iv*gMw2XnG^sV&Gv>_8#E}k-a)W+uC=wwm(Aydj)BfLwlm{|C2x!kX1S9$d zFR^Ij9`z>yF)>7`1go3%<9#k5?^UifDguw`X1ryW64PUoJ72z0wzye?lU>rnpw)5wC)M{j`kJEl(1mbuu`0&g3 zLL(rgPGg=Vh?$Bb9FiuAG$f3Q!9L!=y#OC9! z(q@A*2(1x-VBopRtVfF&^e})sn+u_^pM#f-XkdZfl?D;^C3botDDpgdw7Qi>GQb?(HsE*df`FSG_5XNaru{Yh0(D5`JX93iQKHABV2DZe zJzqy?;J_zKoHQZu62fT#XaL#h>ih^Xo|$0V4@&P!xN{97Y7!j6&;b`V1INtV+M)4?_|23P8n|lAA=6HI?KJB99F1aCwqh& z6+00UJ6j@vG1K?=z!^(rq^ig~Ks;Kd1=vFbjsJ3XJt*~?ArJKYK=cO`Na8pUf5*2u z1JFC1x`%9RU^tQ%08D628Yb8toKspFmEzEpgy7E+ljRZfs3UtW%%KX=2S1L;#lK41 zsoEH*@OO(1ufjV!W%dsan5J(aPYbW6dvQSL5e~G+Bj#t%Uakm$n=+AzRTD|#)?ZK< zfJ-Iy)Cabjo?r`w2^VDDSJ%))xeMdYlVBv-0eB`~$=v!oGxMu3KfM>xC~u<71oanu@h7PiTrkarSH6g2Gm-lu=s7#ei0nTONjbV0WI+VCA1dD z{$HULQBbDu?Lp)6KNvj)ZOC80Z?FJI@}ka`9*~^C(l|qlo@7$8QCFpF9bAL7+j3>v`(t#t&(X07X{hbfx1l_yE8F z3M@iFNl8v^t^2Q1)J8%6BO5#Y^z`O=Nt_jITwE6rA@-pz--ipd4iIX3Ug?D_y*ngR zSc)_FLI+}e{D0}2BNf*fR1*CbDO}N~N=iy2#|#qQm4}_P#6mz`tbz+TdAE^W@0;VK zT}_8&1U{QNaYX67_i>*|xly^cBop))%Ru0%V1pthywwODM41hMh;Dz1 zsSW;jP)n9Z^IrfslagvLNnd2%n*pL$s^ud#>j~GmN;C zs%tFRZJcBb4KubW{J@ow(D!N0_&MG^dbu`B-WMPO!|y0~j&!ckjzEq;9Cy7M9Rvjx zU?MmH5HJxYKZK(+(uS^Q?^oS$NEatMju~IlrFgbrw16wp?Sc$$wt4tJBYM`n#8uMhPCV6*ddcO zK-SfV2{fsG*Jr>A0GxFDVNyx(!$0??FrG#tK&U*iCtsJiLDbn)ZI_&3Wl14E9M1D%t{*^lP?puz)v0S~J){3)D+X!9>WgY#cyFaPn^XA| zL^5R{eF27#{Q1Mt!>08gN7W^`yiFlW0;dd^5TS5En~Y98f*#S*cS4=Lxjr#%tfK|j zcL!qd6JmA|zwKW{f$_J|O91F(Xex08i%dCG#{hA0TIixg+5lE`A*#_|q9;>>2L&X2 z|DV=Jum=aC;l_Ua_ycUy0}whfTM+3Aq;g>H>BLW4aii&hgF_xApIGU@twf8fGo>gR+XTU9Rdh~Z4fz~ z3xxC-yxk06ZaV9C*hk>O`Oy0nK&k?3j|c;`;C%xYb`gR=7Ery2q^IxP!$2wG9t|ua zf(mz>^*Mn-H0XFmd7v-C7%>TGn*m$H+=s{8a~s8_D+s$&)lDRZIgRSANFGlu89d5$ zKfu1V(9R6{^&BF$>?R@_5Fn8u>dEVq<_xX?CCA@*27Dq=&eywbEwsG74Q8T;uweT) z6ZHoL1>3ARbF;Tlq)|Rf|Dd@|j7RcBU1hM1n=pJL=ntusV&dm8xY`$)Kl9H^LINuZ z?pSLV$A?7khY~ICb@;2QA`9M=h_5w{HZG3jVxxQ1eBC-f8Y4|;PMUfl6#DsKrSX1q zC|1ISZ?o53Z0r%YQ0F&G&K?QUrJtr*qPRO>U4`J@|KPOP`Oy2oZxjj*lktjIKnBVH zP)~TtD@)$3;IScP6&vIf?+~5Egv5dc;F!72j=E7yn!w<+^MOlGT1%TrQoUU zX|1vv1_|?~_c)rGnfZ2s+04w$8b_zn^XD6f$ZDn;m;F-ZJ7)akV?4b6UDrQl zP((pX1_sOx1C9p^>@&BjrOtGK33&O{)azSjg*KZEEOq{u{d5DHzj6LGFiZ;m+x>9u z9wimkV|8-(Pe7eLpryS90--iA-PxZ%8}@Wf_Hy~0y9PJQKQjOM#MYJ#(k^XWkCB5z zsiQC023$7LS5{U`z*}$%=W%ashz{;XMzYiS36nujt`wuGKO;pmzkmARZDr-7cs88? z1d|}z^tHB@G$bTsYV+E6#aX^E$-3ahpFi6N8Y28X=!i&Jo`tXs11zM_b4eN0XiS}&x=ca88M$C)XT#*f7f0lW_;kIg8Z!ty=*O@1H?TlB4Kml=^#pcviV%b7;lBKk% zDIE)E*;xnzMGI*ljL<`D3ZSpCse})Mtyk`%IM-1c`Wz8K-id%4ya(Xs=*TwQn4&6+ z0cvkxM)AU)I=Z=aZco?OdtWv{_&^-0%}d@_y@a-t-xTM)J4qpg`ukiQX901To}T`< z{tf59fB(+Cw;>R#x%(V#X#5=?H{M_BiS98gC@N}$GDEmgQ%j44CrM#m8}cFu2logF zZo`djNOn>FVV;>t(FIf_BqZ?2Mhic}{? zkSTyjK?d10F)=Z-u8s^m!?kaAsK-LS6qn3D*~Hx2otN@HoeyPZW`^lrnD7cOwg#Xg zLY<$VmsL;*0v9dJNt=VI5bzJB4B&LkXn=aArluwpapwXX{4GO6L&yq7higMf6TS8S zCaST}a47xXzrPFaY&2LQI9SJLpSOTn2nr1~h29dd`Lx69x<^5Q<>lqY20l#*35j>_ z-*>Z4RBk^%H{n|3~d7Xy@gXE+h7zsdf_mBO!hPi563SFp zON@a4hb%ZWIEeJ@*|Wr?q}PJ0>-@ILV*-?xU-p-rg+5S=hQE7QrWPxX}96=!=4_H_>>c;xac(7rh6b1$cIwmHvDK;>v@7}#@0dg0qs#*k5 ziUU$0{O)xaUsEH%sE~*R7r%g435JN!2?~nTY-0_3ofNgcSHk-G`scRSOC736 zvg+APw~QzO`?9?$Rm;Oe=()VSF~}cYUKb=B58Fut z)mgS&L!H40KU z2QUPFpKa2kqa!W>fv$4XE-r@|5tuWH29Laj#|Cm!Yexs_-rnBxI+AR%wR>EuHum=R zwXW+g{FDRnkT&Z6l!4X(i4(jpKXT;ajI6ECf1ki~XEagKMjXUyc!Xw_iR&(NcRHRQ zTEUBN>+in{ErHg> z7LBTnWR{BEbs$sD<01na=D42Sxz^i=r-X5{aELW^8I|PaH|rqM;$UZtu_!|z4}SC! zu7Zt;i4lCgL1hQ!@GU{}1a$I>h=`QIv;bedi_XQp%;I7Me*jxZf4=dT z?y{)op=_3hfyL{)8Frt#NTtlp83O_XjcsfqAceF;IXhZ&t`D{oFr;i(C0rbAB*I}K z;O6EQ^t%>GO--#`+#1u+)MUA&(bv(D+gz0`h28KGE)sf{RptVoOaGfU5+FrB0r&h# zlRAs^)2Ap9DBytM!mt+@INvn9Fc{Q;X~1if7#kaq;df021216+eW(ia570eEoJ02` zs(J|`I^0ipX~$SNq3sXmIKk(Z0QBQ~I*$fG1_$1mb-la}99OkG1qnFAHFs1v66jz( zCtt{9_#Bf#o^m?=r=yT2tT8VNy&^E0lm-|8tW#=n-$>{!q!kp@_7ITbMTVpOR1VfhkHxw-u^N9n1W5=D*IaN(F$yZGF~q&UO+Js2jNqai?Pgz5sNsovUy5GU z`oI_7g@vgy-*#E-yqo*+Ep2izkAD!e9IU3Kq$HQ!1TW-m6PUTPdvqjiY;2sba}Nn| zwJh@O+vPAt>}QGqE(p-*=;%*iSPG+3YQe%|nvErh369V$h_Iz8T*rovB0Lapa ziNlaXJ7Ib7h=@?BsHom`m_Tji$c{9r7hV_#7f=8_N)G!cRF*yqWxPegH`Xj`=hA zAifuqvu>WhV2{(r$nHW*flF7Oat010CwfN4BI}|hqDO!RR#sO{*MH|4Ma}~NRKJPC zW5~v)p=EhM9C0gCD(rneJcW}D!z3K%|I^#G$3vO6;n%Ec6q;mPQJHtL6A3%jCg+TC zSZUXxwb~V&-#6(c>zc0@Q++Dd{0NF+8*bgAx!jrPjezcctDpzl zZDUH8>_1w1kfJmbzYy4O+=i1^%z%@|9UOiA`o*6f=UfbmkNFFNek(c`@cs8j*=GKS z5ZBb<5GIzDh1A)laRu{hEgg?AJw!B3Ux^v#qQ%?2ZZEpT`S@+!*{TfqZH=Z3{U#(o ziE0-LHY_9H&liu0S`rgeX%SBJ8ylmiEb!f%65 zm-UWm+v0m!`*-ZVB=?yT)~7@dP3~|n+8{?GRJmsJyIkILaxi6H@$x;Hs2B)Vhc%b! z)C0jOhlV;UH`SqG1bpsN9&a9^6uft0L4iJ?IV}ED*OF-t0XT6CEvWX-QqP zPogpeXvHsLVIYbWZLKY}B_nh3p}A{9b+xp#s7pc1&Q?^gbam6p%gc9EJ{$jwIcJW8 zk58_A40pK)Pn|V8$u-ZMzG0scujcWyM8YC9&e;DzPdWn5JbaNvZ3NQnu$w=OylKrc z#-2=j^vDOY*8t$y=9ZR$8p#TJaAcrD38PM`ar^rE${k0SWw`=Kw!P_|(b0N$a&nSw zY<%V$5t1zOd-t?}=6x=0&EzX%`ET`R;C!~AgpyuxKsh{o(Bp*DPJLutyDYFnv2cZ7 zY;-i+!eZgSAmC8Ec##fLO&1efOLWP(-mn34H8vH5-TV>H7TGJDu=9{Oblr&k1iD~8 zTxM9An@d6A3^iQ*6?%IG21c%}5y}{FtLxNO4eFcSpNIUe8YaSduZ8Q|50WkDs~m|f zK<%CPPg`2la6_wVYRDEGLyR}k+s^#nVc|M{if46{fXvJW#@&zQNRb%RN-gOeA{>T? zhip$*-ClpkRor6@7%G4?himswL13f+zNS!5w`UC(cm2cLpd@5@En<}f=x3pNz5XJF zap|nyi6?gij|ocwNLx}?wv^4L)^&0vIE=QNV>^*~0k2L{GeZ}7Y6BWD5d4j&mo4^) zuL;Y2`0yc8N^)?usbH$VcFWjW2z|$k&dYOw5v3TV4OF9^Dv+BvMDMzoy>WSxyP+6g za7n->Vtx^SR~YhI&F$?CsAdg+*7`BwO~mL&2^;x!++J~Ha;ucZ-{@qT0~+(`I>>HJ zH(&ehSqw6!s4Pr6e0VWLc1FoYK z!^0Y2Hc@X#N%3hyk?4Max|x~Tj?FKjiA*H#P*bx2EwpnjBDnBfyj?GRzJ8)bKPjdp z5?h0uHkC_RnbF7GxDIF+0~{Uj!w(($US4x#GT9`Uj>PpjrZ?Q#6MZf7*IHU~fT5v3 z^ypDnqXnyeT`J_im1Jjo?^n#F;E7>i!ck-$mz%YF>&ui^^>giN!xx18T4fQxh3PzRFR%N<{n;&n6`+(c|L4=E^uSs4 zV@cR>Y;`Kct?A_dh63CCe2a+E&TWWMS{E%^gfRk3^49L;SnYXxDJ}TSO^^~kE-uW8 zZ@Gtg7&Kag@U!1g@bVfk`kzquA8*(f%DMyI|vr1--JwS+ME#NO|!BV@3pAG3 zwYnX_Er*MCp{HI@iffT^ZcOThnq2H+;F-r$^N1q* zd_D9klD216moM*!?WNe*?kkU~iyG7Sxft*fr(S+n4MmyfDQ)COX7{4w7;a4hFB$qP zG*eZKD$!JHG#-gc;8vvgmhBi9gSx4#Gy}3pXlM`+Nf8b*aNFtk3`q_LUF;FNxEN@) z=kiY6kASj3`+_HhDBpRL5g*ZlnUKG`rJ>Gkl5kHCBy9=NeQ@XVUup1svd&JubRLNY5j|2J58qbnleV!SQ}71FvtpQTy-cQpVv%U{ zRVYI40b-8nx1N$EBxpcr1RW;vXUM*x%53`h!}y{zXAT&7m6E8&3ImHWG3z6UTWv@7 zug6R{4WLL8-zSOCuZ9O_t?cb3ERSh;TT9D6v;st0di3FUVJZ}~v15?QN=XNX>gvfi zzxyi{3aZe=!7$4e{Ej?GHR8t~Ntp6>sFFM-gsdTVjwhkjUu5L-f@({&ZnzZ|H?BuX zj~IE4qBxU=p3Xn+ub}-ACM#AL8b3&dTZZ}tvDn9m%L!D(!2r-yx%40l!p_x%>@N&G z>>MjG!6*9pX%@LNkr5iX(-R}odUd+!rMCFBN~6v<6<2RaFXCh4Yn)zz`lLE`QNUSz zHcNQ7*b(O{$;Chp?uQqGIai^1u22#t$H!AUrT}%wkiGBBbMx~L(hkW&&@5He)z5Cf z!?CooUTHA6!88YL^iSTO2yV~a2Z~4xlD!F!rEF}|ss9?>)V(9Yja84n?N2F~bsSL{ z1ss5S26AR3d^|wt92eWy~;ju=6RHCO<$YAZ#GUKk>aLU zC)QSDmGQ|j(a3o(2!SAym0LI@tJ@2Og>oHiQ()eGSRNH#tw7E#eM8=0.5.0 +sphinx +sphinx-autobuild +sphinx-markdown-tables +sphinx_rtd_theme \ No newline at end of file diff --git a/docs/source/asr/deepspeech_architecture.md b/docs/source/asr/models_introduction.md similarity index 81% rename from docs/source/asr/deepspeech_architecture.md rename to docs/source/asr/models_introduction.md index be9471d9..ab2b8bac 100644 --- a/docs/source/asr/deepspeech_architecture.md +++ b/docs/source/asr/models_introduction.md @@ -1,6 +1,5 @@ -# Deepspeech2 -## Streaming - +# Models introduction +## Streaming DeepSpeech2 The implemented arcitecure of Deepspeech2 online model is based on [Deepspeech2 model](https://arxiv.org/pdf/1512.02595.pdf) with some changes. The model is mainly composed of 2D convolution subsampling layer and stacked single direction rnn layers. @@ -14,8 +13,8 @@ In addition, the training process and the testing process are also introduced. The arcitecture of the model is shown in Fig.1.

- -
Fig.1 The Arcitecture of deepspeech2 online model + +
Fig.1 The Arcitecture of deepspeech2 online model

@@ -23,13 +22,13 @@ The arcitecture of the model is shown in Fig.1. #### Vocabulary For English data, the vocabulary dictionary is composed of 26 English characters with " ' ", space, \ and \. The \ represents the blank label in CTC, the \ represents the unknown character and the \ represents the start and the end characters. For mandarin, the vocabulary dictionary is composed of chinese characters statisticed from the training set and three additional characters are added. The added characters are \, \ and \. For both English and mandarin data, we set the default indexs that \=0, \=1 and \= last index. ``` - # The code to build vocabulary - cd examples/aishell/s0 - python3 ../../../utils/build_vocab.py \ - --unit_type="char" \ - --count_threshold=0 \ - --vocab_path="data/vocab.txt" \ - --manifest_paths "data/manifest.train.raw" "data/manifest.dev.raw" +# The code to build vocabulary +cd examples/aishell/s0 +python3 ../../../utils/build_vocab.py \ + --unit_type="char" \ + --count_threshold=0 \ + --vocab_path="data/vocab.txt" \ + --manifest_paths "data/manifest.train.raw" "data/manifest.dev.raw" # vocabulary for aishell dataset (Mandarin) vi examples/aishell/s0/data/vocab.txt @@ -41,29 +40,29 @@ vi examples/librispeech/s0/data/vocab.txt #### CMVN For CMVN, a subset or the full of traininig set is chosed and be used to compute the feature mean and std. ``` - # The code to compute the feature mean and std +# The code to compute the feature mean and std cd examples/aishell/s0 python3 ../../../utils/compute_mean_std.py \ - --manifest_path="data/manifest.train.raw" \ - --spectrum_type="linear" \ - --delta_delta=false \ - --stride_ms=10.0 \ - --window_ms=20.0 \ - --sample_rate=16000 \ - --use_dB_normalization=True \ - --num_samples=2000 \ - --num_workers=10 \ - --output_path="data/mean_std.json" + --manifest_path="data/manifest.train.raw" \ + --spectrum_type="linear" \ + --delta_delta=false \ + --stride_ms=10.0 \ + --window_ms=20.0 \ + --sample_rate=16000 \ + --use_dB_normalization=True \ + --num_samples=2000 \ + --num_workers=10 \ + --output_path="data/mean_std.json" ``` #### Feature Extraction - For feature extraction, three methods are implemented, which are linear (FFT without using filter bank), fbank and mfcc. - Currently, the released deepspeech2 online model use the linear feature extraction method. - ``` - The code for feature extraction - vi deepspeech/frontend/featurizer/audio_featurizer.py - ``` +For feature extraction, three methods are implemented, which are linear (FFT without using filter bank), fbank and mfcc. +Currently, the released deepspeech2 online model use the linear feature extraction method. +``` +The code for feature extraction +vi deepspeech/frontend/featurizer/audio_featurizer.py +``` ### Encoder The encoder is composed of two 2D convolution subsampling layers and a number of stacked single direction rnn layers. The 2D convolution subsampling layers extract feature representation from the raw audio feature and reduce the length of audio feature at the same time. After passing through the convolution subsampling layers, then the feature representation are input into the stacked rnn layers. For the stacked rnn layers, LSTM cell and GRU cell are provided to use. Adding one fully connected (fc) layer after the stacked rnn layers is optional. If the number of stacked rnn layers is less than 5, adding one fc layer after stacked rnn layers is recommand. @@ -84,11 +83,11 @@ vi deepspeech/models/ds2_online/deepspeech2.py vi deepspeech/modules/ctc.py ``` -## Training Process +### Training Process Using the command below, you can train the deepspeech2 online model. ``` - cd examples/aishell/s0 - bash run.sh --stage 0 --stop_stage 2 --model_type online --conf_path conf/deepspeech2_online.yaml +cd examples/aishell/s0 +bash run.sh --stage 0 --stop_stage 2 --model_type online --conf_path conf/deepspeech2_online.yaml ``` The detail commands are: ``` @@ -127,11 +126,11 @@ fi By using the command above, the training process can be started. There are 5 stages in "run.sh", and the first 3 stages are used for training process. The stage 0 is used for data preparation, in which the dataset will be downloaded, and the manifest files of the datasets, vocabulary dictionary and CMVN file will be generated in "./data/". The stage 1 is used for training the model, the log files and model checkpoint is saved in "exp/deepspeech2_online/". The stage 2 is used to generated final model for predicting by averaging the top-k model parameters based on validation loss. -## Testing Process +### Testing Process Using the command below, you can test the deepspeech2 online model. - ``` - bash run.sh --stage 3 --stop_stage 5 --model_type online --conf_path conf/deepspeech2_online.yaml - ``` +``` +bash run.sh --stage 3 --stop_stage 5 --model_type online --conf_path conf/deepspeech2_online.yaml +``` The detail commands are: ``` conf_path=conf/deepspeech2_online.yaml @@ -139,7 +138,7 @@ avg_num=1 model_type=online avg_ckpt=avg_${avg_num} - if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then +if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then # test ckpt avg_n CUDA_VISIBLE_DEVICES=2 ./local/test.sh ${conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} ${model_type}|| exit -1 fi @@ -156,19 +155,16 @@ fi ``` After the training process, we use stage 3,4,5 for testing process. The stage 3 is for testing the model generated in the stage 2 and provided the CER index of the test set. The stage 4 is for transforming the model from dynamic graph to static graph by using "paddle.jit" library. The stage 5 is for testing the model in static graph. - -## Non-Streaming +## Non-Streaming DeepSpeech2 The deepspeech2 offline model is similarity to the deepspeech2 online model. The main difference between them is the offline model use the stacked bi-directional rnn layers while the online model use the single direction rnn layers and the fc layer is not used. For the stacked bi-directional rnn layers in the offline model, the rnn cell and gru cell are provided to use. The arcitecture of the model is shown in Fig.2.

- -
Fig.2 The Arcitecture of deepspeech2 offline model + +
Fig.2 The Arcitecture of deepspeech2 offline model

- - For data preparation and decoder, the deepspeech2 offline model is same with the deepspeech2 online model. The code of encoder and decoder for deepspeech2 offline model is in: @@ -180,7 +176,7 @@ The training process and testing process of deepspeech2 offline model is very si Only some changes should be noticed. For training and testing, the "model_type" and the "conf_path" must be set. - ``` +``` # Training offline cd examples/aishell/s0 bash run.sh --stage 0 --stop_stage 2 --model_type offline --conf_path conf/deepspeech2.yaml diff --git a/docs/source/asr/getting_started.md b/docs/source/asr/quick_start.md similarity index 83% rename from docs/source/asr/getting_started.md rename to docs/source/asr/quick_start.md index 478f3bb3..ecce0743 100644 --- a/docs/source/asr/getting_started.md +++ b/docs/source/asr/quick_start.md @@ -1,5 +1,4 @@ -# Getting Started - +# Quick Start of Speech-To-Text Several shell scripts provided in `./examples/tiny/local` will help us to quickly give it a try, for most major modules, including data preparation, model training, case inference and model evaluation, with a few public dataset (e.g. [LibriSpeech](http://www.openslr.org/12/), [Aishell](http://www.openslr.org/33)). Reading these examples will also help you to understand how to make it work with your own data. Some of the scripts in `./examples` are not configured with GPUs. If you want to train with 8 GPUs, please modify `CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7`. If you don't have any GPU available, please set `CUDA_VISIBLE_DEVICES=` to use CPUs instead. Besides, if out-of-memory problem occurs, just reduce `batch_size` to fit. @@ -11,68 +10,52 @@ Let's take a tiny sampled subset of [LibriSpeech dataset](http://www.openslr.org ```bash cd examples/tiny ``` - Notice that this is only a toy example with a tiny sampled subset of LibriSpeech. If you would like to try with the complete dataset (would take several days for training), please go to `examples/librispeech` instead. - - Source env - ```bash source path.sh ``` - **Must do this before starting do anything.** - Set `MAIN_ROOT` as project dir. Using defualt `deepspeech2` model as default, you can change this in the script. - + **Must do this before you start to do anything.** + Set `MAIN_ROOT` as project dir. Using defualt `deepspeech2` model as `MODEL`, you can change this in the script. - Main entrypoint - ```bash bash run.sh ``` - This just a demo, please make sure every `step` is work fine when do next `step`. + This is just a demo, please make sure every `step` works well before next `step`. More detailed information are provided in the following sections. Wish you a happy journey with the *DeepSpeech on PaddlePaddle* ASR engine! ## Training a model -The key steps of training for Mandarin language are same to that of English language and we have also provided an example for Mandarin training with Aishell in ```examples/aishell/local```. As mentioned above, please execute ```sh data.sh```, ```sh train.sh```, ```sh test.sh``` and ```sh infer.sh``` to do data preparation, training, testing and inference correspondingly. We have also prepared a pre-trained model (downloaded by local/download_model.sh) for users to try with ```sh infer_golden.sh``` and ```sh test_golden.sh```. Notice that, different from English LM, the Mandarin LM is character-based and please run ```local/tune.sh``` to find an optimal setting. +The key steps of training for Mandarin language are same to that of English language and we have also provided an example for Mandarin training with Aishell in ```examples/aishell/local```. As mentioned above, please execute ```sh data.sh```, ```sh train.sh```, ```sh test.sh```and ```sh infer.sh```to do data preparation, training, testing and inference correspondingly. We have also prepared a pre-trained model (downloaded by local/download_model.sh) for users to try with ```sh infer_golden.sh```and ```sh test_golden.sh```. Notice that, different from English LM, the Mandarin LM is character-based and please run ```local/tune.sh```to find an optimal setting. ## Speech-to-text Inference An inference module caller `infer.py` is provided to infer, decode and visualize speech-to-text results for several given audio clips. It might help to have an intuitive and qualitative evaluation of the ASR model's performance. - ```bash CUDA_VISIBLE_DEVICES=0 bash local/infer.sh ``` - We provide two types of CTC decoders: *CTC greedy decoder* and *CTC beam search decoder*. The *CTC greedy decoder* is an implementation of the simple best-path decoding algorithm, selecting at each timestep the most likely token, thus being greedy and locally optimal. The [*CTC beam search decoder*](https://arxiv.org/abs/1408.2873) otherwise utilizes a heuristic breadth-first graph search for reaching a near global optimality; it also requires a pre-trained KenLM language model for better scoring and ranking. The decoder type can be set with argument `decoding_method`. ## Evaluate a Model - To evaluate a model's performance quantitatively, please run: - ```bash CUDA_VISIBLE_DEVICES=0 bash local/test.sh ``` - The error rate (default: word error rate; can be set with `error_rate_type`) will be printed. -For more help on arguments: - ## Hyper-parameters Tuning - The hyper-parameters $\alpha$ (language model weight) and $\beta$ (word insertion weight) for the [*CTC beam search decoder*](https://arxiv.org/abs/1408.2873) often have a significant impact on the decoder's performance. It would be better to re-tune them on the validation set when the acoustic model is renewed. `tune.py` performs a 2-D grid search over the hyper-parameter $\alpha$ and $\beta$. You must provide the range of $\alpha$ and $\beta$, as well as the number of their attempts. - - ```bash CUDA_VISIBLE_DEVICES=0 bash local/tune.sh ``` - The grid search will print the WER (word error rate) or CER (character error rate) at each point in the hyper-parameters space, and draw the error surface optionally. A proper hyper-parameters range should include the global minima of the error surface for WER/CER, as illustrated in the following figure.

- -
An example error surface for tuning on the dev-clean set of LibriSpeech + +
An example error surface for tuning on the dev-clean set of LibriSpeech

Usually, as the figure shows, the variation of language model weight ($\alpha$) significantly affect the performance of CTC beam search decoder. And a better procedure is to first tune on serveral data batches (the number can be specified) to find out the proper range of hyper-parameters, then change to the whole validation set to carray out an accurate tuning. diff --git a/docs/source/asr/released_model.md b/docs/source/asr/released_model.md deleted file mode 100644 index 50670aaf..00000000 --- a/docs/source/asr/released_model.md +++ /dev/null @@ -1,28 +0,0 @@ -# Released Models - -## Acoustic Model Released in paddle 2.X -Acoustic Model | Training Data | Token-based | Size | Descriptions | CER | WER | Hours of speech -:-------------:| :------------:| :-----: | -----: | :----------------- |:--------- | :---------- | :--------- -[Ds2 Online Aishell Model](https://deepspeech.bj.bcebos.com/release2.1/aishell/s0/aishell.s0.ds_online.5rnn.debug.tar.gz) | Aishell Dataset | Char-based | 345 MB | 2 Conv + 5 LSTM layers with only forward direction | 0.0824 |-| 151 h -[Ds2 Offline Aishell Model](https://deepspeech.bj.bcebos.com/release2.1/aishell/s0/aishell.s0.ds2.offline.cer6p65.release.tar.gz)| Aishell Dataset | Char-based | 306 MB | 2 Conv + 3 bidirectional GRU layers| 0.065 |-| 151 h -[Conformer Online Aishell Model](https://deepspeech.bj.bcebos.com/release2.1/aishell/s1/aishell.chunk.release.tar.gz) | Aishell Dataset | Char-based | 283 MB | Encoder:Conformer, Decoder:Transformer, Decoding method: Attention + CTC | 0.0594 |-| 151 h -[Conformer Offline Aishell Model](https://deepspeech.bj.bcebos.com/release2.1/aishell/s1/aishell.release.tar.gz) | Aishell Dataset | Char-based | 284 MB | Encoder:Conformer, Decoder:Transformer, Decoding method: Attention | 0.0547 |-| 151 h -[Conformer Librispeech Model](https://deepspeech.bj.bcebos.com/release2.1/librispeech/s1/conformer.release.tar.gz) | Librispeech Dataset | Word-based | 287 MB | Encoder:Conformer, Decoder:Transformer, Decoding method: Attention |-| 0.0325 | 960 h -[Transformer Librispeech Model](https://deepspeech.bj.bcebos.com/release2.1/librispeech/s1/transformer.release.tar.gz) | Librispeech Dataset | Word-based | 195 MB | Encoder:Conformer, Decoder:Transformer, Decoding method: Attention |-| 0.0544 | 960 h - -## Acoustic Model Transformed from paddle 1.8 -Acoustic Model | Training Data | Token-based | Size | Descriptions | CER | WER | Hours of speech -:-------------:| :------------:| :-----: | -----: | :----------------- | :---------- | :---------- | :--------- -[Ds2 Offline Aishell model](https://deepspeech.bj.bcebos.com/mandarin_models/aishell_model_v1.8_to_v2.x.tar.gz)|Aishell Dataset| Char-based| 234 MB| 2 Conv + 3 bidirectional GRU layers| 0.0804 |-| 151 h| -[Ds2 Offline Librispeech model](https://deepspeech.bj.bcebos.com/eng_models/librispeech_v1.8_to_v2.x.tar.gz)|Librispeech Dataset| Word-based| 307 MB| 2 Conv + 3 bidirectional sharing weight RNN layers |-| 0.0685| 960 h| -[Ds2 Offline Baidu en8k model](https://deepspeech.bj.bcebos.com/eng_models/baidu_en8k_v1.8_to_v2.x.tar.gz)|Baidu Internal English Dataset| Word-based| 273 MB| 2 Conv + 3 bidirectional GRU layers |-| 0.0541 | 8628 h| - - - -## Language Model Released - -Language Model | Training Data | Token-based | Size | Descriptions -:-------------:| :------------:| :-----: | -----: | :----------------- -[English LM](https://deepspeech.bj.bcebos.com/en_lm/common_crawl_00.prune01111.trie.klm) | [CommonCrawl(en.00)](http://web-language-models.s3-website-us-east-1.amazonaws.com/ngrams/en/deduped/en.00.deduped.xz) | Word-based | 8.3 GB | Pruned with 0 1 1 1 1;
About 1.85 billion n-grams;
'trie' binary with '-a 22 -q 8 -b 8' -[Mandarin LM Small](https://deepspeech.bj.bcebos.com/zh_lm/zh_giga.no_cna_cmn.prune01244.klm) | Baidu Internal Corpus | Char-based | 2.8 GB | Pruned with 0 1 2 4 4;
About 0.13 billion n-grams;
'probing' binary with default settings -[Mandarin LM Large](https://deepspeech.bj.bcebos.com/zh_lm/zhidao_giga.klm) | Baidu Internal Corpus | Char-based | 70.4 GB | No Pruning;
About 3.7 billion n-grams;
'probing' binary with default settings diff --git a/docs/source/conf.py b/docs/source/conf.py index 7e32a22c..fb69e599 100644 --- a/docs/source/conf.py +++ b/docs/source/conf.py @@ -20,9 +20,15 @@ # If extensions (or modules to document with autodoc) are in another directory, # add these directories to sys.path here. If the directory is relative to the # documentation root, use os.path.abspath to make it absolute, like shown here. +import os +import sys + import recommonmark.parser import sphinx_rtd_theme +sys.path.insert(0, os.path.abspath('../..')) +autodoc_mock_imports = ["soundfile", "librosa"] + # -- Project information ----------------------------------------------------- project = 'paddle speech' @@ -46,10 +52,10 @@ pygments_style = 'sphinx' extensions = [ 'sphinx.ext.autodoc', 'sphinx.ext.viewcode', - 'sphinx_rtd_theme', + "sphinx_rtd_theme", 'sphinx.ext.mathjax', - 'sphinx.ext.autosummary', 'numpydoc', + 'sphinx.ext.autosummary', 'myst_parser', ] @@ -76,6 +82,7 @@ smartquotes = False # relative to this directory. They are copied after the builtin static files, # so a file named "default.css" will overwrite the builtin "default.css". html_static_path = ['_static'] +html_logo = '../images/paddle.png' # -- Extension configuration ------------------------------------------------- # numpydoc_show_class_members = False diff --git a/docs/source/index.rst b/docs/source/index.rst index 3c196d2d..06bc2f3f 100644 --- a/docs/source/index.rst +++ b/docs/source/index.rst @@ -10,34 +10,44 @@ Contents .. toctree:: :maxdepth: 1 :caption: Introduction - - asr/deepspeech_architecture + introduction .. toctree:: :maxdepth: 1 - :caption: Getting_started - - asr/install - asr/getting_started - + :caption: Quick Start + install + asr/quick_start + tts/quick_start + .. toctree:: :maxdepth: 1 - :caption: More Information + :caption: Speech-To-Text + asr/models_introduction asr/data_preparation asr/augmentation asr/feature_list - asr/ngram_lm - + asr/ngram_lm .. toctree:: :maxdepth: 1 - :caption: Released_model + :caption: Text-To-Speech - asr/released_model + tts/basic_usage + tts/advanced_usage + tts/zh_text_frontend + tts/models_introduction + tts/gan_vocoder + tts/demo + tts/demo_2 +.. toctree:: + :maxdepth: 1 + :caption: Released Models + + released_model .. toctree:: :maxdepth: 1 @@ -45,3 +55,8 @@ Contents asr/reference + + + + + diff --git a/docs/source/asr/install.md b/docs/source/install.md similarity index 92% rename from docs/source/asr/install.md rename to docs/source/install.md index 8cecba12..0c27a4db 100644 --- a/docs/source/asr/install.md +++ b/docs/source/install.md @@ -8,7 +8,7 @@ To avoid the trouble of environment setup, [running in Docker container](#runnin ## Setup (Important) -- Make sure these libraries or tools installed: `pkg-config`, `flac`, `ogg`, `vorbis`, `boost`, `sox, and `swig`, e.g. installing them via `apt-get`: +- Make sure these libraries or tools installed: `pkg-config`, `flac`, `ogg`, `vorbis`, `boost`, `sox`, and `swig`, e.g. installing them via `apt-get`: ```bash sudo apt-get install -y sox pkg-config libflac-dev libogg-dev libvorbis-dev libboost-dev swig python3-dev @@ -44,6 +44,14 @@ bash setup.sh source tools/venv/bin/activate ``` +## Simple Setup + +```python +git clone https://github.com/PaddlePaddle/DeepSpeech.git +cd DeepSpeech +pip install -e . +``` + ## Running in Docker Container (optional) Docker is an open source tool to build, ship, and run distributed applications in an isolated environment. A Docker image for this project has been provided in [hub.docker.com](https://hub.docker.com) with all the dependencies installed. This Docker image requires the support of NVIDIA GPU, so please make sure its availiability and the [nvidia-docker](https://github.com/NVIDIA/nvidia-docker) has been installed. diff --git a/docs/source/introduction.md b/docs/source/introduction.md new file mode 100644 index 00000000..2f71b104 --- /dev/null +++ b/docs/source/introduction.md @@ -0,0 +1,33 @@ +# PaddleSpeech + +## What is PaddleSpeech? +PaddleSpeech is an open-source toolkit on PaddlePaddle platform for two critical tasks in Speech - Speech-To-Text (Automatic Speech Recognition, ASR) and Text-To-Speech Synthesis (TTS), with modules involving state-of-art and influential models. + +## What can PaddleSpeech do? + +### Speech-To-Text +(An introduce of ASR in PaddleSpeech is needed here!) + +### Text-To-Speech +TTS mainly consists of components below: +- Implementation of models and commonly used neural network layers. +- Dataset abstraction and common data preprocessing pipelines. +- Ready-to-run experiments. + +PaddleSpeech TTS provides you with a complete TTS pipeline, including: +- Text FrontEnd + - Rule based Chinese frontend. +- Acoustic Models + - FastSpeech2 + - SpeedySpeech + - TransformerTTS + - Tacotron2 +- Vocoders + - Multi Band MelGAN + - Parallel WaveGAN + - WaveFlow +- Voice Cloning + - Transfer Learning from Speaker Verification to Multispeaker Text-To-Speech Synthesis + - GE2E + +Text-To-Speech helps you to train TTS models with simple commands. diff --git a/docs/source/released_model.md b/docs/source/released_model.md new file mode 100644 index 00000000..3b60f15a --- /dev/null +++ b/docs/source/released_model.md @@ -0,0 +1,55 @@ +# Released Models + +## Speech-To-Text Models +### Acoustic Model Released in paddle 2.X +Acoustic Model | Training Data | Token-based | Size | Descriptions | CER | WER | Hours of speech +:-------------:| :------------:| :-----: | -----: | :----------------- |:--------- | :---------- | :--------- +[Ds2 Online Aishell Model](https://deepspeech.bj.bcebos.com/release2.1/aishell/s0/aishell.s0.ds_online.5rnn.debug.tar.gz) | Aishell Dataset | Char-based | 345 MB | 2 Conv + 5 LSTM layers with only forward direction | 0.0824 |-| 151 h +[Ds2 Offline Aishell Model](https://deepspeech.bj.bcebos.com/release2.1/aishell/s0/aishell.s0.ds2.offline.cer6p65.release.tar.gz)| Aishell Dataset | Char-based | 306 MB | 2 Conv + 3 bidirectional GRU layers| 0.065 |-| 151 h +[Conformer Online Aishell Model](https://deepspeech.bj.bcebos.com/release2.1/aishell/s1/aishell.chunk.release.tar.gz) | Aishell Dataset | Char-based | 283 MB | Encoder:Conformer, Decoder:Transformer, Decoding method: Attention + CTC | 0.0594 |-| 151 h +[Conformer Offline Aishell Model](https://deepspeech.bj.bcebos.com/release2.1/aishell/s1/aishell.release.tar.gz) | Aishell Dataset | Char-based | 284 MB | Encoder:Conformer, Decoder:Transformer, Decoding method: Attention | 0.0547 |-| 151 h +[Conformer Librispeech Model](https://deepspeech.bj.bcebos.com/release2.1/librispeech/s1/conformer.release.tar.gz) | Librispeech Dataset | Word-based | 287 MB | Encoder:Conformer, Decoder:Transformer, Decoding method: Attention |-| 0.0325 | 960 h +[Transformer Librispeech Model](https://deepspeech.bj.bcebos.com/release2.1/librispeech/s1/transformer.release.tar.gz) | Librispeech Dataset | Word-based | 195 MB | Encoder:Transformer, Decoder:Transformer, Decoding method: Attention |-| 0.0544 | 960 h + +### Acoustic Model Transformed from paddle 1.8 +Acoustic Model | Training Data | Token-based | Size | Descriptions | CER | WER | Hours of speech +:-------------:| :------------:| :-----: | -----: | :----------------- | :---------- | :---------- | :--------- +[Ds2 Offline Aishell model](https://deepspeech.bj.bcebos.com/mandarin_models/aishell_model_v1.8_to_v2.x.tar.gz)|Aishell Dataset| Char-based| 234 MB| 2 Conv + 3 bidirectional GRU layers| 0.0804 |-| 151 h| +[Ds2 Offline Librispeech model](https://deepspeech.bj.bcebos.com/eng_models/librispeech_v1.8_to_v2.x.tar.gz)|Librispeech Dataset| Word-based| 307 MB| 2 Conv + 3 bidirectional sharing weight RNN layers |-| 0.0685| 960 h| +[Ds2 Offline Baidu en8k model](https://deepspeech.bj.bcebos.com/eng_models/baidu_en8k_v1.8_to_v2.x.tar.gz)|Baidu Internal English Dataset| Word-based| 273 MB| 2 Conv + 3 bidirectional GRU layers |-| 0.0541 | 8628 h| + +### Language Model Released + +Language Model | Training Data | Token-based | Size | Descriptions +:-------------:| :------------:| :-----: | -----: | :----------------- +[English LM](https://deepspeech.bj.bcebos.com/en_lm/common_crawl_00.prune01111.trie.klm) | [CommonCrawl(en.00)](http://web-language-models.s3-website-us-east-1.amazonaws.com/ngrams/en/deduped/en.00.deduped.xz) | Word-based | 8.3 GB | Pruned with 0 1 1 1 1;
About 1.85 billion n-grams;
'trie' binary with '-a 22 -q 8 -b 8' +[Mandarin LM Small](https://deepspeech.bj.bcebos.com/zh_lm/zh_giga.no_cna_cmn.prune01244.klm) | Baidu Internal Corpus | Char-based | 2.8 GB | Pruned with 0 1 2 4 4;
About 0.13 billion n-grams;
'probing' binary with default settings +[Mandarin LM Large](https://deepspeech.bj.bcebos.com/zh_lm/zhidao_giga.klm) | Baidu Internal Corpus | Char-based | 70.4 GB | No Pruning;
About 3.7 billion n-grams;
'probing' binary with default settings + +## Text-To-Speech Models +### Acoustic Models +Model Type | Dataset| Example Link | Pretrained Models +:-------------:| :------------:| :-----: | :----- +Tacotron2|LJSpeech|[tacotron2-vctk](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/ljspeech/tts0)|[tacotron2_ljspeech_ckpt_0.3.zip](https://paddlespeech.bj.bcebos.com/Parakeet/tacotron2_ljspeech_ckpt_0.3.zip) +TransformerTTS| LJSpeech| [transformer-ljspeech](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/ljspeech/tts1)|[transformer_tts_ljspeech_ckpt_0.4.zip](https://paddlespeech.bj.bcebos.com/Parakeet/transformer_tts_ljspeech_ckpt_0.4.zip) +SpeedySpeech| CSMSC | [speedyspeech-csmsc](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/csmsc/tts2) |[speedyspeech_nosil_baker_ckpt_0.5.zip](https://paddlespeech.bj.bcebos.com/Parakeet/speedyspeech_nosil_baker_ckpt_0.5.zip) +FastSpeech2| CSMSC |[fastspeech2-csmsc](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/csmsc/tts3)|[fastspeech2_nosil_baker_ckpt_0.4.zip](https://paddlespeech.bj.bcebos.com/Parakeet/fastspeech2_nosil_baker_ckpt_0.4.zip) +FastSpeech2| AISHELL-3 |[fastspeech2-aishell3](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/aishell3/tts3)|[fastspeech2_nosil_aishell3_ckpt_0.4.zip](https://paddlespeech.bj.bcebos.com/Parakeet/fastspeech2_nosil_aishell3_ckpt_0.4.zip) +FastSpeech2| LJSpeech |[fastspeech2-ljspeech](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/ljspeech/tts3)|[fastspeech2_nosil_ljspeech_ckpt_0.5.zip](https://paddlespeech.bj.bcebos.com/Parakeet/fastspeech2_nosil_ljspeech_ckpt_0.5.zip) +FastSpeech2| VCTK |[fastspeech2-csmsc](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/vctk/tts3)|[fastspeech2_nosil_vctk_ckpt_0.5.zip](https://paddlespeech.bj.bcebos.com/Parakeet/fastspeech2_nosil_vctk_ckpt_0.5.zip) + + +### Vocoders + +Model Type | Dataset| Example Link | Pretrained Models +:-------------:| :------------:| :-----: | :----- +WaveFlow| LJSpeech |[waveflow-ljspeech](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/ljspeech/voc0)|[waveflow_ljspeech_ckpt_0.3.zip](https://paddlespeech.bj.bcebos.com/Parakeet/waveflow_ljspeech_ckpt_0.3.zip) +Parallel WaveGAN| CSMSC |[PWGAN-csmsc](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/csmsc/voc1)|[pwg_baker_ckpt_0.4.zip.](https://paddlespeech.bj.bcebos.com/Parakeet/pwg_baker_ckpt_0.4.zip) +Parallel WaveGAN| LJSpeech |[PWGAN-ljspeech](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/ljspeech/voc1)|[pwg_ljspeech_ckpt_0.5.zip](https://paddlespeech.bj.bcebos.com/Parakeet/pwg_ljspeech_ckpt_0.5.zip) +Parallel WaveGAN| VCTK |[PWGAN-vctk](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/vctk/voc1)|[pwg_vctk_ckpt_0.5.zip](https://paddlespeech.bj.bcebos.com/Parakeet/pwg_vctk_ckpt_0.5.zip) + +### Voice Cloning +Model Type | Dataset| Example Link | Pretrained Models +:-------------:| :------------:| :-----: | :----- +GE2E| AISHELL-3, etc. |[ge2e](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/other/ge2e)|[ge2e_ckpt_0.3.zip](https://paddlespeech.bj.bcebos.com/Parakeet/ge2e_ckpt_0.3.zip) +GE2E + Tactron2| AISHELL-3 |[ge2e-tactron2-aishell3](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/aishell3/vc0)|[tacotron2_aishell3_ckpt_0.3.zip](https://paddlespeech.bj.bcebos.com/Parakeet/tacotron2_aishell3_ckpt_0.3.zip) diff --git a/docs/source/tts/advanced_usage.md b/docs/source/tts/advanced_usage.md index 52980044..297f274f 100644 --- a/docs/source/tts/advanced_usage.md +++ b/docs/source/tts/advanced_usage.md @@ -1,6 +1,5 @@ - # Advanced Usage -This sections covers how to extend parakeet by implementing your own models and experiments. Guidelines on implementation are also elaborated. +This sections covers how to extend TTS by implementing your own models and experiments. Guidelines on implementation are also elaborated. For the general deep learning experiment, there are several parts to deal with: 1. Preprocess the data according to the needs of the model, and iterate the dataset by batch. @@ -8,7 +7,7 @@ For the general deep learning experiment, there are several parts to deal with: 3. Write out the training process (generally including forward / backward calculation, parameter update, log recording, visualization, periodic evaluation, etc.). 5. Configure and run the experiment. -## Parakeet's Model Components +## PaddleSpeech TTS's Model Components In order to balance the reusability and function of models, we divide models into several types according to its characteristics. For the commonly used modules that can be used as part of other larger models, we try to implement them as simple and universal as possible, because they will be reused. Modules with trainable parameters are generally implemented as subclasses of `paddle.nn.Layer`. Modules without trainable parameters can be directly implemented as a function, and its input and output are `paddle.Tensor`. @@ -68,11 +67,11 @@ There are two common ways to define a model which consists of several modules. ``` When a model is a complicated and made up of several components, each of which has a separate functionality, and can be replaced by other components with the same functionality, we prefer to define it in this way. -In the directory structure of Parakeet, modules with high reusability are placed in `parakeet.modules`, but models for specific tasks are placed in `parakeet.models`. When developing a new model, developers need to consider the feasibility of splitting the modules, and the degree of generality of the modules, and place them in appropriate directories. +In the directory structure of PaddleSpeech TTS, modules with high reusability are placed in `parakeet.modules`, but models for specific tasks are placed in `parakeet.models`. When developing a new model, developers need to consider the feasibility of splitting the modules, and the degree of generality of the modules, and place them in appropriate directories. -## Parakeet's Data Components +## PaddleSpeech TTS's Data Components Another critical componnet for a deep learning project is data. -Parakeet uses the following methods for training data: +PaddleSpeech TTS uses the following methods for training data: 1. Preprocess the data. 2. Load the preprocessed data for training. @@ -154,7 +153,7 @@ def _convert(self, meta_datum: Dict[str, Any]) -> Dict[str, Any]: return example ``` -## Parakeet's Training Components +## PaddleSpeech TTS's Training Components A typical training process includes the following processes: 1. Iterate the dataset. 2. Process batch data. @@ -164,7 +163,7 @@ A typical training process includes the following processes: 6. Write logs, visualize, and in some cases save necessary intermediate results. 7. Save the state of the model and optimizer. -Here, we mainly introduce the training related components of Parakeet and why we designed it like this. +Here, we mainly introduce the training related components of TTS in Pa and why we designed it like this. ### Global Repoter When training and modifying Deep Learning models，logging is often needed, and it has even become the key to model debugging and modifying. We usually use various visualization tools，such as , `visualdl` in `paddle`, `tensorboard` in `tensorflow` and `vidsom`, `wnb` ,etc. Besides, `logging` and `print` are usuaally used for different purpose. @@ -245,7 +244,7 @@ def test_reporter_scope(): In this way, when we write modular components, we can directly call `report`. The caller will decide where to report as long as it's ready for `OBSERVATION`, then it opens a `scope` and calls the component within this `scope`. - The `Trainer` in Parakeet report the information in this way. + The `Trainer` in PaddleSpeech TTS report the information in this way. ```python while True: self.observation = {} @@ -269,7 +268,7 @@ We made an abstraction for these intermediate processes, that is, `Updater`, whi ### Visualizer Because we choose observation as the communication mode, we can simply write the things in observation into `visualizer`. -## Parakeet's Configuration Components +## PaddleSpeech TTS's Configuration Components Deep learning experiments often have many options to configure. These configurations can be roughly divided into several categories. 1. Data source and data processing mode configuration. 2. Save path configuration of experimental results. @@ -293,28 +292,26 @@ The following is the basic `ArgumentParser`: 3. `--output-dir` is the dir to save the training results.（if there are checkpoints in `checkpoints/` of `--output-dir` , it's defalut to reload the newest checkpoint to train) 4. `--device` and `--nprocs` determine operation modes，`--device` specifies the type of running device, whether to run on `cpu` or `gpu`. `--nprocs` refers to the number of training processes. If `nprocs` > 1, it means that multi process parallel training is used. (Note: currently only GPU multi card multi process training is supported.) -Developers can refer to the examples in `Parakeet/examples` to write the default configuration file when adding new experiments. +Developers can refer to the examples in `examples` to write the default configuration file when adding new experiments. -## Parakeet's Experiment template +## PaddleSpeech TTS's Experiment template -The experimental codes in Parakeet are generally organized as follows: +The experimental codes in PaddleSpeech TTS are generally organized as follows: ```text -├── conf -│ └── default.yaml (defalut config) -├── README.md (help information) -├── batch_fn.py (organize metadata into batch) -├── config.py (code to read default config) -├── *_updater.py (Updater of a specific model) -├── preprocess.py (data preprocessing code) -├── preprocess.sh (script to call data preprocessing.py) -├── synthesis.py (synthesis from metadata) -├── synthesis.sh (script to call synthesis.py) -├── synthesis_e2e.py (synthesis from raw text) -├── synthesis_e2e.sh (script to call synthesis_e2e.py) -├── train.py (train code) -└── run.sh (script to call train.py) +. +├── README.md (help information) +├── conf +│ └── default.yaml (defalut config) +├── local +│ ├── preprocess.sh (script to call data preprocessing.py) +│ ├── synthesize.sh (script to call synthesis.py) +│ ├── synthesize_e2e.sh (script to call synthesis_e2e.py) +│ └──train.sh (script to call train.py) +├── path.sh (script include paths to be sourced) +└── run.sh (script to call scripts in local) ``` +The `*.py` files called by above `*.sh` are located `${BIN_DIR}/` We add a named argument. `--output-dir` to each training script to specify the output directory. The directory structure is as follows, It's best for developers to follow this specification: ```text @@ -330,4 +327,4 @@ exp/default/ └── test/ (output dir of synthesis results) ``` -You can view the examples we provide in `Parakeet/examples`. These experiments are provided to users as examples which can be run directly. Users are welcome to add new models and experiments and contribute code to Parakeet. +You can view the examples we provide in `examples`. These experiments are provided to users as examples which can be run directly. Users are welcome to add new models and experiments and contribute code to PaddleSpeech. diff --git a/docs/source/tts/basic_usage.md b/docs/source/tts/basic_usage.md deleted file mode 100644 index fc2a5bad..00000000 --- a/docs/source/tts/basic_usage.md +++ /dev/null @@ -1,115 +0,0 @@ -# Basic Usage -This section shows how to use pretrained models provided by parakeet and make inference with them. - -Pretrained models in v0.4 are provided in a archive. Extract it to get a folder like this: -``` -checkpoint_name/ -├──default.yaml -├──snapshot_iter_76000.pdz -├──speech_stats.npy -└──phone_id_map.txt -``` -`default.yaml` stores the config used to train the model. -`snapshot_iter_N.pdz` is the chechpoint file, where `N` is the steps it has been trained. -`*_stats.npy` is the stats file of feature if it has been normalized before training. -`phone_id_map.txt` is the map of phonemes to phoneme_ids. - -The example code below shows how to use the models for prediction. -## Acoustic Models (text to spectrogram) -The code below show how to use a `FastSpeech2` model. After loading the pretrained model, use it and normalizer object to construct a prediction object，then use fastspeech2_inferencet(phone_ids) to generate spectrograms, which can be further used to synthesize raw audio with a vocoder. - -```python -from pathlib import Path -import numpy as np -import paddle -import yaml -from yacs.config import CfgNode -from parakeet.models.fastspeech2 import FastSpeech2 -from parakeet.models.fastspeech2 import FastSpeech2Inference -from parakeet.modules.normalizer import ZScore -# Parakeet/examples/fastspeech2/baker/frontend.py -from frontend import Frontend - -# load the pretrained model -checkpoint_dir = Path("fastspeech2_nosil_baker_ckpt_0.4") -with open(checkpoint_dir / "phone_id_map.txt", "r") as f: - phn_id = [line.strip().split() for line in f.readlines()] -vocab_size = len(phn_id) -with open(checkpoint_dir / "default.yaml") as f: - fastspeech2_config = CfgNode(yaml.safe_load(f)) -odim = fastspeech2_config.n_mels -model = FastSpeech2( - idim=vocab_size, odim=odim, **fastspeech2_config["model"]) -model.set_state_dict( - paddle.load(args.fastspeech2_checkpoint)["main_params"]) -model.eval() - -# load stats file -stat = np.load(checkpoint_dir / "speech_stats.npy") -mu, std = stat -mu = paddle.to_tensor(mu) -std = paddle.to_tensor(std) -fastspeech2_normalizer = ZScore(mu, std) - -# construct a prediction object -fastspeech2_inference = FastSpeech2Inference(fastspeech2_normalizer, model) - -# load Chinese Frontend -frontend = Frontend(checkpoint_dir / "phone_id_map.txt") - -# text to spectrogram -sentence = "你好吗？" -input_ids = frontend.get_input_ids(sentence, merge_sentences=True) -phone_ids = input_ids["phone_ids"] -flags = 0 -# The output of Chinese text frontend is segmented -for part_phone_ids in phone_ids: - with paddle.no_grad(): - temp_mel = fastspeech2_inference(part_phone_ids) - if flags == 0: - mel = temp_mel - flags = 1 - else: - mel = paddle.concat([mel, temp_mel]) -``` - -## Vocoder (spectrogram to wave) -The code below show how to use a ` Parallel WaveGAN` model. Like the example above, after loading the pretrained model, use it and normalizer object to construct a prediction object，then use pwg_inference(mel) to generate raw audio (in wav format). - -```python -from pathlib import Path -import numpy as np -import paddle -import soundfile as sf -import yaml -from yacs.config import CfgNode -from parakeet.models.parallel_wavegan import PWGGenerator -from parakeet.models.parallel_wavegan import PWGInference -from parakeet.modules.normalizer import ZScore - -# load the pretrained model -checkpoint_dir = Path("parallel_wavegan_baker_ckpt_0.4") -with open(checkpoint_dir / "pwg_default.yaml") as f: - pwg_config = CfgNode(yaml.safe_load(f)) -vocoder = PWGGenerator(**pwg_config["generator_params"]) -vocoder.set_state_dict(paddle.load(args.pwg_params)) -vocoder.remove_weight_norm() -vocoder.eval() - -# load stats file -stat = np.load(checkpoint_dir / "pwg_stats.npy") -mu, std = stat -mu = paddle.to_tensor(mu) -std = paddle.to_tensor(std) -pwg_normalizer = ZScore(mu, std) - -# construct a prediction object -pwg_inference = PWGInference(pwg_normalizer, vocoder) - -# spectrogram to wave -wav = pwg_inference(mel) -sf.write( - audio_path, - wav.numpy(), - samplerate=fastspeech2_config.fs) -``` diff --git a/docs/source/tts/demo.rst b/docs/source/tts/demo.rst index a6f18f88..948fc056 100644 --- a/docs/source/tts/demo.rst +++ b/docs/source/tts/demo.rst @@ -11,7 +11,7 @@ The main processes of TTS include: When training ``Tacotron2``、``TransformerTTS`` and ``WaveFlow``, we use English single speaker TTS dataset `LJSpeech `_ by default. However, when training ``SpeedySpeech``, ``FastSpeech2`` and ``ParallelWaveGAN``, we use Chinese single speaker dataset `CSMSC `_ by default. -In the future, ``Parakeet`` will mainly use Chinese TTS datasets for default examples. +In the future, ``PaddleSpeech TTS`` will mainly use Chinese TTS datasets for default examples. Here, we will display three types of audio samples: @@ -441,7 +441,7 @@ Audio samples generated by a TTS system. Text is first transformed into spectrog Chinese TTS with/without text frontend -------------------------------------- -We provide a complete Chinese text frontend module in ``Parakeet``. ``Text Normalization`` and ``G2P`` are the most important modules in text frontend, We assume that the texts are normalized already, and mainly compare ``G2P`` module here. +We provide a complete Chinese text frontend module in ``PaddleSpeech TTS``. ``Text Normalization`` and ``G2P`` are the most important modules in text frontend, We assume that the texts are normalized already, and mainly compare ``G2P`` module here. We use ``FastSpeech2`` + ``ParallelWaveGAN`` here. diff --git a/docs/source/tts/demo_2.rst b/docs/source/tts/demo_2.rst new file mode 100644 index 00000000..37922fcb --- /dev/null +++ b/docs/source/tts/demo_2.rst @@ -0,0 +1,7 @@ +Audio Sample (PaddleSpeech TTS VS Espnet TTS) +================== + +This is an audio demo page to contrast PaddleSpeech TTS and Espnet TTS, We use their respective modules (Text Frontend, Acoustic model and Vocoder) here. +We use Espnet's released models here. + +FastSpeech2 + Parallel WaveGAN in CSMSC diff --git a/docs/source/tts/gan_vocoder.md b/docs/source/tts/gan_vocoder.md new file mode 100644 index 00000000..4931f307 --- /dev/null +++ b/docs/source/tts/gan_vocoder.md @@ -0,0 +1,9 @@ +# GAN Vocoders +This is a brief introduction of GAN Vocoders, we mainly introduce the losses of different vocoders here. + +Model | Generator Loss |Discriminator Loss +:-------------:| :------------:| :----- +Parallel Wave GAN| adversial loss
Feature Matching | Multi-Scale Discriminator | +Mel GAN |adversial loss
Multi-resolution STFT loss | adversial loss| +Multi-Band Mel GAN | adversial loss
full band Multi-resolution STFT loss
sub band Multi-resolution STFT loss |Multi-Scale Discriminator| +HiFi GAN |adversial loss
Feature Matching
Mel-Spectrogram Loss | Multi-Scale Discriminator
Multi-Period Discriminato | diff --git a/docs/source/tts/index.rst b/docs/source/tts/index.rst deleted file mode 100644 index 74abe60d..00000000 --- a/docs/source/tts/index.rst +++ /dev/null @@ -1,45 +0,0 @@ -.. parakeet documentation master file, created by - sphinx-quickstart on Fri Sep 10 14:22:24 2021. - You can adapt this file completely to your liking, but it should at least - contain the root `toctree` directive. - -Parakeet -==================================== - -``parakeet`` is a deep learning based text-to-speech toolkit built upon ``paddlepaddle`` framework. It aims to provide a flexible, efficient and state-of-the-art text-to-speech toolkit for the open-source community. It includes many influential TTS models proposed by `Baidu Research `_ and other research groups. - -``parakeet`` mainly consists of components below. - -#. Implementation of models and commonly used neural network layers. -#. Dataset abstraction and common data preprocessing pipelines. -#. Ready-to-run experiments. - -.. toctree:: - :maxdepth: 1 - :caption: Introduction - - introduction - -.. toctree:: - :maxdepth: 1 - :caption: Getting started - - install - basic_usage - advanced_usage - cn_text_frontend - released_models - -.. toctree:: - :maxdepth: 1 - :caption: Demos - - demo - - -Indices and tables -================== - -* :ref:`genindex` -* :ref:`modindex` -* :ref:`search` diff --git a/docs/source/tts/install.md b/docs/source/tts/install.md deleted file mode 100644 index 24e44b17..00000000 --- a/docs/source/tts/install.md +++ /dev/null @@ -1,47 +0,0 @@ -# Installation -## Install PaddlePaddle -Parakeet requires PaddlePaddle as its backend. Note that 2.1.2 or newer versions of paddle is required. - -Since paddlepaddle has multiple packages depending on the device (cpu or gpu) and the dependency libraries, it is recommended to install a proper package of paddlepaddle with respect to the device and dependency library versons via `pip`. - -Installing paddlepaddle with conda or build paddlepaddle from source is also supported. Please refer to [PaddlePaddle installation](https://www.paddlepaddle.org.cn/install/quick?docurl=/documentation/docs/zh/install/pip/linux-pip.html) for more details. - -Example instruction to install paddlepaddle via pip is listed below. - -### PaddlePaddle with GPU -```python -# CUDA10.1 的 PaddlePaddle -python -m pip install paddlepaddle-gpu==2.1.2.post101 -f https://www.paddlepaddle.org.cn/whl/linux/mkl/avx/stable.html -# CUDA10.2 的 PaddlePaddle -python -m pip install paddlepaddle-gpu -i https://mirror.baidu.com/pypi/simple -# CUDA11.0 的 PaddlePaddle -python -m pip install paddlepaddle-gpu==2.1.2.post110 -f https://www.paddlepaddle.org.cn/whl/linux/mkl/avx/stable.html -# CUDA11.2 的 PaddlePaddle -python -m pip install paddlepaddle-gpu==2.1.2.post112 -f https://www.paddlepaddle.org.cn/whl/linux/mkl/avx/stable.html -``` -### PaddlePaddle with CPU -```python -python -m pip install paddlepaddle==2.1.2 -i https://mirror.baidu.com/pypi/simple -``` -## Install libsndfile -Experimemts in parakeet often involve audio and spectrum processing, thus `librosa` and `soundfile` are required. `soundfile` requires a extra C library `libsndfile`, which is not always handled by pip. - -For Windows and Mac users, `libsndfile` is also installed when installing `soundfile` via pip, but for Linux users, installing `libsndfile` via system package manager is required. Example commands for popular distributions are listed below. -```bash -# ubuntu, debian -sudo apt-get install libsndfile1 -# centos, fedora -sudo yum install libsndfile -# openSUSE -sudo zypper in libsndfile -``` -For any problem with installtion of soundfile, please refer to [SoundFile](https://pypi.org/project/SoundFile/). -## Install Parakeet -There are two ways to install parakeet according to the purpose of using it. - - 1. If you want to run experiments provided by parakeet or add new models and experiments, it is recommended to clone the project from github (Parakeet), and install it in editable mode. - ```python - git clone https://github.com/PaddlePaddle/Parakeet - cd Parakeet - pip install -e . - ``` diff --git a/docs/source/tts/introduction.md b/docs/source/tts/introduction.md deleted file mode 100644 index d350565c..00000000 --- a/docs/source/tts/introduction.md +++ /dev/null @@ -1,27 +0,0 @@ -# Parakeet - PAddle PARAllel text-to-speech toolKIT - -## What is Parakeet? -Parakeet is a deep learning based text-to-speech toolkit built upon paddlepaddle framework. It aims to provide a flexible, efficient and state-of-the-art text-to-speech toolkit for the open-source community. It includes many influential TTS models proposed by Baidu Research and other research groups. - -## What can Parakeet do? -Parakeet mainly consists of components below: -- Implementation of models and commonly used neural network layers. -- Dataset abstraction and common data preprocessing pipelines. -- Ready-to-run experiments. - -Parakeet provides you with a complete TTS pipeline, including: -- Text FrontEnd - - Rule based Chinese frontend. -- Acoustic Models - - FastSpeech2 - - SpeedySpeech - - TransformerTTS - - Tacotron2 -- Vocoders - - Parallel WaveGAN - - WaveFlow -- Voice Cloning - - Transfer Learning from Speaker Verification to Multispeaker Text-To-Speech Synthesis - - GE2E - -Parakeet helps you to train TTS models with simple commands. diff --git a/docs/source/tts/released_models.md b/docs/source/tts/models_introduction.md similarity index 90% rename from docs/source/tts/released_models.md rename to docs/source/tts/models_introduction.md index 7899c1c5..87b65c51 100644 --- a/docs/source/tts/released_models.md +++ b/docs/source/tts/models_introduction.md @@ -1,12 +1,12 @@ -# Released Models -TTS system mainly includes three modules: `text frontend`, `Acoustic model` and `Vocoder`. We introduce a rule based Chinese text frontend in [cn_text_frontend.md](./cn_text_frontend.md). Here, we will introduce acoustic models and vocoders, which are trainable models. +# Models introduction +TTS system mainly includes three modules: `Text Frontend`, `Acoustic model` and `Vocoder`. We introduce a rule based Chinese text frontend in [cn_text_frontend.md](./cn_text_frontend.md). Here, we will introduce acoustic models and vocoders, which are trainable models. The main processes of TTS include: 1. Convert the original text into characters/phonemes, through `text frontend` module. 2. Convert characters/phonemes into acoustic features , such as linear spectrogram, mel spectrogram, LPC features, etc. through `Acoustic models`. 3. Convert acoustic features into waveforms through `Vocoders`. -A simple text frontend module can be implemented by rules. Acoustic models and vocoders need to be trained. The models provided by Parakeet are acoustic models and vocoders. +A simple text frontend module can be implemented by rules. Acoustic models and vocoders need to be trained. The models provided by PaddleSpeech TTS are acoustic models and vocoders. ## Acoustic Models ### Modeling Objectives of Acoustic Models @@ -92,7 +92,7 @@ At present, there are two mainstream acoustic model structures.
-You can find Parakeet's tacotron2 example at `Parakeet/examples/tacotron2`. +You can find PaddleSpeech TTS's tacotron2 with LJSpeech dataset example at [examples/ljspeech/tts0](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/ljspeech/tts0). ### TransformerTTS **Disadvantages of the Tacotrons:** @@ -146,7 +146,7 @@ Transformer TTS is a seq2seq acoustic model based on Transformer and Tacotron2. - The ability to perceive local information is weak, and local information is more related to pronunciation. - Stability is worse than Tacotron2. -You can find Parakeet's Transformer TTS example at `Parakeet/examples/transformer_tts`. +You can find PaddleSpeech TTS's Transformer TTS with LJSpeech dataset example at [examples/ljspeech/tts1](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/ljspeech/tts1). ### FastSpeech2 @@ -212,7 +212,7 @@ FastSpeech2 is similar to FastPitch but introduces more variation information of
-You can find Parakeet's FastSpeech2/FastPitch example at `Parakeet/examples/fastspeech2`, We use token-averaged pitch and energy values introduced in FastPitch rather than frame level ones in FastSpeech2. +You can find PaddleSpeech TTS's FastSpeech2/FastPitch with CSMSC dataset example at [examples/csmsc/tts3](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/csmsc/tts3), We use token-averaged pitch and energy values introduced in FastPitch rather than frame level ones in FastSpeech2. ### SpeedySpeech [SpeedySpeech](https://arxiv.org/abs/2008.03802) simplify the teacher-student architecture of FastSpeech and provide a fast and stable training procedure. @@ -226,7 +226,7 @@ You can find Parakeet's FastSpeech2/FastPitch example at `Parakeet/examples/fast
-You can find Parakeet's SpeedySpeech example at `Parakeet/examples/speedyspeech/baker`. +You can find PaddleSpeech TTS's SpeedySpeech with CSMSC dataset example at [examples/csmsc/tts2](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/csmsc/tts2). ## Vocoders In speech synthesis, the main task of the vocoder is to convert the spectral parameters predicted by the acoustic model into the final speech waveform. @@ -276,7 +276,7 @@ Here, we introduce a Flow-based vocoder WaveFlow and a GAN-based vocoder Paralle - It is a small-footprint flow-based model for raw audio. It has only 5.9M parameters, which is 15x smalller than WaveGlow (87.9M). - It is directly trained with maximum likelihood without probability density distillation and auxiliary losses as used in [Parallel WaveNet](https://arxiv.org/abs/1711.10433) and [ClariNet](https://openreview.net/pdf?id=HklY120cYm), which simplifies the training pipeline and reduces the cost of development. -You can find Parakeet's WaveFlow example at `Parakeet/examples/waveflow`. +You can find PaddleSpeech TTS's WaveFlow with LJSpeech dataset example at [examples/ljspeech/voc0](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/ljspeech/voc0). ### Parallel WaveGAN [Parallel WaveGAN](https://arxiv.org/abs/1910.11480) trains a non-autoregressive WaveNet variant as a generator in a GAN based training method. @@ -286,10 +286,10 @@ You can find Parakeet's WaveFlow example at `Parakeet/examples/waveflow`. - Use non-causal convolution instead of causal convolution. - The input is random Gaussian white noise. - The model is non-autoregressive both in training and prediction, which is fast -- Multi-resolution STFT loss. +- Multi-resolution STFT loss.

-You can find Parakeet's Parallel WaveGAN example at `Parakeet/examples/parallelwave_gan/baker`. +You can find PaddleSpeech TTS's Parallel WaveGAN with CSMSC example at [examples/csmsc/voc1](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/csmsc/voc1). diff --git a/docs/source/tts/quick_start.md b/docs/source/tts/quick_start.md new file mode 100644 index 00000000..f5d16bbf --- /dev/null +++ b/docs/source/tts/quick_start.md @@ -0,0 +1,193 @@ +# Quick Start of Text-To-Speech +The examples in PaddleSpeech are mainly classified by datasets, the TTS datasets we mainly used are: +* CSMCS (Mandarin single speaker) +* AISHELL3 (Mandarin multiple speaker) +* LJSpeech (English single speaker) +* VCTK (English multiple speaker) + +The models in PaddleSpeech TTS have the following mapping relationship: +* tts0 - Tactron2 +* tts1 - TransformerTTS +* tts2 - SpeedySpeech +* tts3 - FastSpeech2 +* voc0 - WaveFlow +* voc1 - Parallel WaveGAN +* voc2 - MelGAN +* voc3 - MultiBand MelGAN +* vc0 - Tactron2 Voice Clone with GE2E + +## Quick Start + +Let's take a FastSpeech2 + Parallel WaveGAN with CSMSC dataset for instance. (./examples/csmsc/)(https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/csmsc) + +### Train Parallel WaveGAN with CSMSC +- Go to directory + ```bash + cd examples/csmsc/voc1 + ``` +- Source env + ```bash + source path.sh + ``` + **Must do this before you start to do anything.** + Set `MAIN_ROOT` as project dir. Using `parallelwave_gan` model as `MODEL`. + +- Main entrypoint + ```bash + bash run.sh + ``` + This is just a demo, please make sure source data have been prepared well and every `step` works well before next `step`. +### Train FastSpeech2 with CSMSC +- Go to directory + ```bash + cd examples/csmsc/tts3 + ``` +- Source env + ```bash + source path.sh + ``` + **Must do this before you start to do anything.** + Set `MAIN_ROOT` as project dir. Using `fastspeech2` model as `MODEL`. +- Main entrypoint + ```bash + bash run.sh + ``` + This is just a demo, please make sure source data have been prepared well and every `step` works well before next `step`. + +The steps in `run.sh` mainly include: +- source path. +- preprocess the dataset, +- train the model. +- synthesize waveform from metadata.jsonl. +- synthesize waveform from text file. (in acoustic models) +- inference using static model. (optional) + +For more details , you can see `README.md` in examples. + +## Pipeline of TTS +This section shows how to use pretrained models provided by TTS and make inference with them. + +Pretrained models in TTS are provided in a archive. Extract it to get a folder like this: +**Acoustic Models:** +```text +checkpoint_name +├── default.yaml +├── snapshot_iter_*.pdz +├── speech_stats.npy +├── phone_id_map.txt +├── spk_id_map.txt (optimal) +└── tone_id_map.txt (optimal) +``` +**Vocoders:** +```text +checkpoint_name +├── default.yaml +├── snapshot_iter_*.pdz +└── stats.npy +``` +- `default.yaml` stores the config used to train the model. +- `snapshot_iter_*.pdz` is the chechpoint file, where `*` is the steps it has been trained. +- `*_stats.npy` is the stats file of feature if it has been normalized before training. +- `phone_id_map.txt` is the map of phonemes to phoneme_ids. +- `tone_id_map.txt` is the map of tones to tones_ids, when you split tones and phones before training acoustic models. (for example in our csmsc/speedyspeech example) +- `spk_id_map.txt` is the map of spkeaker to spk_ids in multi-spk acoustic models. (for example in our aishell3/fastspeech2 example) + +The example code below shows how to use the models for prediction. +### Acoustic Models (text to spectrogram) +The code below show how to use a `FastSpeech2` model. After loading the pretrained model, use it and normalizer object to construct a prediction object，then use `fastspeech2_inferencet(phone_ids)` to generate spectrograms, which can be further used to synthesize raw audio with a vocoder. + +```python +from pathlib import Path +import numpy as np +import paddle +import yaml +from yacs.config import CfgNode +from parakeet.models.fastspeech2 import FastSpeech2 +from parakeet.models.fastspeech2 import FastSpeech2Inference +from parakeet.modules.normalizer import ZScore +# examples/fastspeech2/baker/frontend.py +from frontend import Frontend + +# load the pretrained model +checkpoint_dir = Path("fastspeech2_nosil_baker_ckpt_0.4") +with open(checkpoint_dir / "phone_id_map.txt", "r") as f: + phn_id = [line.strip().split() for line in f.readlines()] +vocab_size = len(phn_id) +with open(checkpoint_dir / "default.yaml") as f: + fastspeech2_config = CfgNode(yaml.safe_load(f)) +odim = fastspeech2_config.n_mels +model = FastSpeech2( + idim=vocab_size, odim=odim, **fastspeech2_config["model"]) +model.set_state_dict( + paddle.load(args.fastspeech2_checkpoint)["main_params"]) +model.eval() + +# load stats file +stat = np.load(checkpoint_dir / "speech_stats.npy") +mu, std = stat +mu = paddle.to_tensor(mu) +std = paddle.to_tensor(std) +fastspeech2_normalizer = ZScore(mu, std) + +# construct a prediction object +fastspeech2_inference = FastSpeech2Inference(fastspeech2_normalizer, model) + +# load Chinese Frontend +frontend = Frontend(checkpoint_dir / "phone_id_map.txt") + +# text to spectrogram +sentence = "你好吗？" +input_ids = frontend.get_input_ids(sentence, merge_sentences=True) +phone_ids = input_ids["phone_ids"] +flags = 0 +# The output of Chinese text frontend is segmented +for part_phone_ids in phone_ids: + with paddle.no_grad(): + temp_mel = fastspeech2_inference(part_phone_ids) + if flags == 0: + mel = temp_mel + flags = 1 + else: + mel = paddle.concat([mel, temp_mel]) +``` + +### Vocoder (spectrogram to wave) +The code below show how to use a ` Parallel WaveGAN` model. Like the example above, after loading the pretrained model, use it and normalizer object to construct a prediction object，then use `pwg_inference(mel)` to generate raw audio (in wav format). + +```python +from pathlib import Path +import numpy as np +import paddle +import soundfile as sf +import yaml +from yacs.config import CfgNode +from parakeet.models.parallel_wavegan import PWGGenerator +from parakeet.models.parallel_wavegan import PWGInference +from parakeet.modules.normalizer import ZScore + +# load the pretrained model +checkpoint_dir = Path("parallel_wavegan_baker_ckpt_0.4") +with open(checkpoint_dir / "pwg_default.yaml") as f: + pwg_config = CfgNode(yaml.safe_load(f)) +vocoder = PWGGenerator(**pwg_config["generator_params"]) +vocoder.set_state_dict(paddle.load(args.pwg_params)) +vocoder.remove_weight_norm() +vocoder.eval() + +# load stats file +stat = np.load(checkpoint_dir / "pwg_stats.npy") +mu, std = stat +mu = paddle.to_tensor(mu) +std = paddle.to_tensor(std) +pwg_normalizer = ZScore(mu, std) + +# construct a prediction object +pwg_inference = PWGInference(pwg_normalizer, vocoder) + +# spectrogram to wave +wav = pwg_inference(mel) +sf.write( + audio_path, + wav.numpy(), + samplerate=fastspeech2_config.fs) +``` diff --git a/docs/source/tts/cn_text_frontend.md b/docs/source/tts/zh_text_frontend.md similarity index 96% rename from docs/source/tts/cn_text_frontend.md rename to docs/source/tts/zh_text_frontend.md index 8a025179..26255ad3 100644 --- a/docs/source/tts/cn_text_frontend.md +++ b/docs/source/tts/zh_text_frontend.md @@ -1,5 +1,5 @@ # Chinese Rule Based Text Frontend -TTS system mainly includes three modules: `text frontend`, `Acoustic model` and `Vocoder`. We provide a complete Chinese text frontend module in Parakeet, see exapmle in `Parakeet/examples/text_frontend/`. +A TTS system mainly includes three modules: `Text Frontend`, `Acoustic model` and `Vocoder`. We provide a complete Chinese text frontend module in PaddleSpeech TTS, see exapmle in [examples/other/text_frontend/](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/other/text_frontend). A text frontend module mainly includes: - Text Segmentation diff --git a/examples/other/text_frontend/README.md b/examples/other/text_frontend/README.md index d5c6dd4d..0ec4ec70 100644 --- a/examples/other/text_frontend/README.md +++ b/examples/other/text_frontend/README.md @@ -20,5 +20,73 @@ Run the command below to get the results of test. ./run.sh ``` The `avg WER` of g2p is: 0.027495061517943988 +```text + + SYSTEM SUMMARY PERCENTAGES by SPEAKER + + ,------------------------------------------------------------------------. + | ./exp/g2p/text.g2p | + |------------------------------------------------------------------------| + | SPKR | # Snt # Wrd | Corr Sub Del Ins Err S.Err | + |------+-----------------+-----------------------------------------------| + | bak | 9996 299181 | 290969 8198 14 14 8226 5249 | + |========================================================================| + | Sum | 9996 299181 | 290969 8198 14 14 8226 5249 | + |========================================================================| + | Mean |9996.0 299181.0 |290969.0 8198.0 14.0 14.0 8226.0 5249.0 | + | S.D. | 0.0 0.0 | 0.0 0.0 0.0 0.0 0.0 0.0 | + |Median|9996.0 299181.0 |290969.0 8198.0 14.0 14.0 8226.0 5249.0 | + `------------------------------------------------------------------------' + + SYSTEM SUMMARY PERCENTAGES by SPEAKER + + ,--------------------------------------------------------------------. + | ./exp/g2p/text.g2p | + |--------------------------------------------------------------------| + | SPKR | # Snt # Wrd | Corr Sub Del Ins Err S.Err | + |--------+-----------------+-----------------------------------------| + | bak | 9996 299181 | 97.3 2.7 0.0 0.0 2.7 52.5 | + |====================================================================| + | Sum/Avg| 9996 299181 | 97.3 2.7 0.0 0.0 2.7 52.5 | + |====================================================================| + | Mean |9996.0 299181.0 | 97.3 2.7 0.0 0.0 2.7 52.5 | + | S.D. | 0.0 0.0 | 0.0 0.0 0.0 0.0 0.0 0.0 | + | Median |9996.0 299181.0 | 97.3 2.7 0.0 0.0 2.7 52.5 | + `--------------------------------------------------------------------' +``` The `avg CER` of text normalization is: 0.006388318503308237 +```text + + SYSTEM SUMMARY PERCENTAGES by SPEAKER + + ,----------------------------------------------------------------. + | ./exp/textnorm/text.tn | + |----------------------------------------------------------------| + | SPKR | # Snt # Wrd | Corr Sub Del Ins Err S.Err | + |------+--------------+------------------------------------------| + | utt | 125 2254 | 2241 2 11 2 15 4 | + |================================================================| + | Sum | 125 2254 | 2241 2 11 2 15 4 | + |================================================================| + | Mean |125.0 2254.0 |2241.0 2.0 11.0 2.0 15.0 4.0 | + | S.D. | 0.0 0.0 | 0.0 0.0 0.0 0.0 0.0 0.0 | + |Median|125.0 2254.0 |2241.0 2.0 11.0 2.0 15.0 4.0 | + `----------------------------------------------------------------' + + SYSTEM SUMMARY PERCENTAGES by SPEAKER + + ,-----------------------------------------------------------------. + | ./exp/textnorm/text.tn | + |-----------------------------------------------------------------| + | SPKR | # Snt # Wrd | Corr Sub Del Ins Err S.Err | + |--------+--------------+-----------------------------------------| + | utt | 125 2254 | 99.4 0.1 0.5 0.1 0.7 3.2 | + |=================================================================| + | Sum/Avg| 125 2254 | 99.4 0.1 0.5 0.1 0.7 3.2 | + |=================================================================| + | Mean |125.0 2254.0 | 99.4 0.1 0.5 0.1 0.7 3.2 | + | S.D. | 0.0 0.0 | 0.0 0.0 0.0 0.0 0.0 0.0 | + | Median |125.0 2254.0 | 99.4 0.1 0.5 0.1 0.7 3.2 | + `-----------------------------------------------------------------' +``` From 670a68ad95ff86bec4022fd3782640db274cf833 Mon Sep 17 00:00:00 2001 From: TianYuan Date: Thu, 28 Oct 2021 09:16:33 +0000 Subject: [PATCH 2/2] fix textfrontend readme, fix imgs link --- README.md | 39 +++++++++-------- docs/source/asr/models_introduction.md | 4 +- docs/source/asr/quick_start.md | 2 +- docs/source/tts/models_introduction.md | 22 +++++----- examples/other/text_frontend/README.md | 58 +------------------------- examples/tiny/s0/README.md | 1 - parakeet/data/batch.py | 4 +- parakeet/exps/tacotron2/ljspeech.py | 9 ++-- 8 files changed, 43 insertions(+), 96 deletions(-) diff --git a/README.md b/README.md index e0769720..468f42a6 100644 --- a/README.md +++ b/README.md @@ -9,34 +9,34 @@ English | [简体中文](README_ch.md)

- + ------------------------------------------------------------------------------------ ![License](https://img.shields.io/badge/license-Apache%202-red.svg) ![python version](https://img.shields.io/badge/python-3.7+-orange.svg) ![support os](https://img.shields.io/badge/os-linux-yellow.svg) -**PaddleSpeech** is an open-source toolkit on [PaddlePaddle](https://github.com/PaddlePaddle/Paddle) platform for two critical tasks in Speech - **Automatic Speech Recognition (ASR)** and **Text-To-Speech Synthesis (TTS)**, with modules involving state-of-art and influential models. +**PaddleSpeech** is an open-source toolkit on [PaddlePaddle](https://github.com/PaddlePaddle/Paddle) platform for two critical tasks in Speech - **Automatic Speech Recognition (ASR)** and **Text-To-Speech Synthesis (TTS)**, with modules involving state-of-art and influential models. Via the easy-to-use, efficient, flexible and scalable implementation, our vision is to empower both industrial application and academic research, including training, inference & testing module, and deployment. Besides, this toolkit also features at: - **Fast and Light-weight**: we provide a high-speed and ultra-lightweight model that is convenient for industrial deployment. -- **Rule-based Chinese frontend**: our frontend contains Text Normalization (TN) and Grapheme-to-Phoneme (G2P, including Polyphone and Tone Sandhi). Moreover, we use self-defined linguistic rules to adapt Chinese context. -- **Varieties of Functions that Vitalize Research**: +- **Rule-based Chinese frontend**: our frontend contains Text Normalization (TN) and Grapheme-to-Phoneme (G2P, including Polyphone and Tone Sandhi). Moreover, we use self-defined linguistic rules to adapt Chinese context. +- **Varieties of Functions that Vitalize Research**: - *Integration of mainstream models and datasets*: the toolkit implements modules that participate in the whole pipeline of both ASR and TTS, and uses datasets like LibriSpeech, LJSpeech, AIShell, etc. See also [model lists](#models-list) for more details. - *Support of ASR streaming and non-streaming data*: This toolkit contains non-streaming/streaming models like [DeepSpeech2](http://proceedings.mlr.press/v48/amodei16.pdf), [Transformer](https://arxiv.org/abs/1706.03762), [Conformer](https://arxiv.org/abs/2005.08100) and [U2](https://arxiv.org/pdf/2012.05481.pdf). - -Let's install PaddleSpeech with only a few lines of code! + +Let's install PaddleSpeech with only a few lines of code! >Note: The official name is still deepspeech. 2021/10/26 @@ -44,7 +44,7 @@ Let's install PaddleSpeech with only a few lines of code! # 1. Install essential libraries and paddlepaddle first. # install prerequisites sudo apt-get install -y sox pkg-config libflac-dev libogg-dev libvorbis-dev libboost-dev swig python3-dev libsndfile1 -# `pip install paddlepaddle-gpu` instead if you are using GPU. +# `pip install paddlepaddle-gpu` instead if you are using GPU. pip install paddlepaddle # 2.Then install PaddleSpeech. @@ -109,7 +109,7 @@ If you want to try more functions like training and tuning, please see [ASR gett PaddleSpeech ASR supports a lot of mainstream models, which are summarized as follow. For more information, please refer to [ASR Models](./docs/source/asr/released_model.md). @@ -125,7 +125,7 @@ The current hyperlinks redirect to [Previous Parakeet](https://github.com/Paddle - + @@ -199,7 +199,7 @@ PaddleSpeech TTS mainly contains three modules: *Text Frontend*, *Acoustic Model - @@ -292,11 +292,11 @@ PaddleSpeech TTS mainly contains three modules: *Text Frontend*, *Acoustic Model
Acoustic Model Aishell2 Conv + 5 LSTM layers with only forward direction 2 Conv + 5 LSTM layers with only forward direction Ds2 Online Aishell Model
Text Frontend   + chinese-fronted
-## Tutorials +## Tutorials Normally, [Speech SoTA](https://paperswithcode.com/area/speech) gives you an overview of the hot academic topics in speech. If you want to focus on the two tasks in PaddleSpeech, you will find the following guidelines are helpful to grasp the core ideas. -The original ASR module is based on [Baidu's DeepSpeech](https://arxiv.org/abs/1412.5567) which is an independent product named [DeepSpeech](https://deepspeech.readthedocs.io). However, the toolkit aligns almost all the SoTA modules in the pipeline. Specifically, these modules are +The original ASR module is based on [Baidu's DeepSpeech](https://arxiv.org/abs/1412.5567) which is an independent product named [DeepSpeech](https://deepspeech.readthedocs.io). However, the toolkit aligns almost all the SoTA modules in the pipeline. Specifically, these modules are * [Data Prepration](docs/source/asr/data_preparation.md) * [Data Augmentation](docs/source/asr/augmentation.md) @@ -318,4 +318,3 @@ PaddleSpeech is provided under the [Apache-2.0 License](./LICENSE). ## Acknowledgement PaddleSpeech depends on a lot of open source repos. See [references](docs/source/asr/reference.md) for more information. - diff --git a/docs/source/asr/models_introduction.md b/docs/source/asr/models_introduction.md index ab2b8bac..c99093bd 100644 --- a/docs/source/asr/models_introduction.md +++ b/docs/source/asr/models_introduction.md @@ -13,7 +13,7 @@ In addition, the training process and the testing process are also introduced. The arcitecture of the model is shown in Fig.1.

- +
Fig.1 The Arcitecture of deepspeech2 online model

@@ -160,7 +160,7 @@ The deepspeech2 offline model is similarity to the deepspeech2 online model. The The arcitecture of the model is shown in Fig.2.

- +
Fig.2 The Arcitecture of deepspeech2 offline model

diff --git a/docs/source/asr/quick_start.md b/docs/source/asr/quick_start.md index ecce0743..da1620e9 100644 --- a/docs/source/asr/quick_start.md +++ b/docs/source/asr/quick_start.md @@ -54,7 +54,7 @@ CUDA_VISIBLE_DEVICES=0 bash local/tune.sh The grid search will print the WER (word error rate) or CER (character error rate) at each point in the hyper-parameters space, and draw the error surface optionally. A proper hyper-parameters range should include the global minima of the error surface for WER/CER, as illustrated in the following figure.

- +
An example error surface for tuning on the dev-clean set of LibriSpeech

diff --git a/docs/source/tts/models_introduction.md b/docs/source/tts/models_introduction.md index 87b65c51..b1329758 100644 --- a/docs/source/tts/models_introduction.md +++ b/docs/source/tts/models_introduction.md @@ -27,14 +27,14 @@ At present, there are two mainstream acoustic model structures. - Acoustic decoder (N Frames - > N Frames).
-
+
- Sequence to sequence acoustic model: - M Tokens - > N Frames.
-
+
### Tacotron2 @@ -54,7 +54,7 @@ At present, there are two mainstream acoustic model structures. - CBHG postprocess. - Vocoder: Griffin-Lim.
-
+
**Advantage of Tacotron:** @@ -89,7 +89,7 @@ At present, there are two mainstream acoustic model structures. - The alignment matrix of previous time is considered at the step `t` of decoder.
-
+
You can find PaddleSpeech TTS's tacotron2 with LJSpeech dataset example at [examples/ljspeech/tts0](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/ljspeech/tts0). @@ -118,7 +118,7 @@ Transformer TTS is a combination of Tacotron2 and Transformer. - Positional Encoding.
-
+
#### Transformer TTS @@ -138,7 +138,7 @@ Transformer TTS is a seq2seq acoustic model based on Transformer and Tacotron2. - Uniform scale position encoding may have a negative impact on input or output sequences.
-
+
**Disadvantages of Transformer TTS:** @@ -184,14 +184,14 @@ Instead of using the encoder-attention-decoder based architecture as adopted by • Can be generated in parallel (decoding time is less affected by sequence length)
-
+
#### FastPitch [FastPitch](https://arxiv.org/abs/2006.06873) follows FastSpeech. A single pitch value is predicted for every temporal location, which improves the overall quality of synthesized speech.
-
+
#### FastSpeech2 @@ -209,7 +209,7 @@ Instead of using the encoder-attention-decoder based architecture as adopted by FastSpeech2 is similar to FastPitch but introduces more variation information of speech.
-
+
You can find PaddleSpeech TTS's FastSpeech2/FastPitch with CSMSC dataset example at [examples/csmsc/tts3](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/csmsc/tts3), We use token-averaged pitch and energy values introduced in FastPitch rather than frame level ones in FastSpeech2. @@ -223,7 +223,7 @@ You can find PaddleSpeech TTS's FastSpeech2/FastPitch with CSMSC dataset example - Describe a simple data augmentation technique that can be used early in the training to make the teacher network robust to sequential error propagation.
-
+
You can find PaddleSpeech TTS's SpeedySpeech with CSMSC dataset example at [examples/csmsc/tts2](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/csmsc/tts2). @@ -289,7 +289,7 @@ You can find PaddleSpeech TTS's WaveFlow with LJSpeech dataset example at [examp - Multi-resolution STFT loss.
-
+
You can find PaddleSpeech TTS's Parallel WaveGAN with CSMSC example at [examples/csmsc/voc1](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/examples/csmsc/voc1). diff --git a/examples/other/text_frontend/README.md b/examples/other/text_frontend/README.md index 0ec4ec70..0bf6e72d 100644 --- a/examples/other/text_frontend/README.md +++ b/examples/other/text_frontend/README.md @@ -21,72 +21,18 @@ Run the command below to get the results of test. ``` The `avg WER` of g2p is: 0.027495061517943988 ```text - - SYSTEM SUMMARY PERCENTAGES by SPEAKER - - ,------------------------------------------------------------------------. - | ./exp/g2p/text.g2p | - |------------------------------------------------------------------------| - | SPKR | # Snt # Wrd | Corr Sub Del Ins Err S.Err | - |------+-----------------+-----------------------------------------------| - | bak | 9996 299181 | 290969 8198 14 14 8226 5249 | - |========================================================================| - | Sum | 9996 299181 | 290969 8198 14 14 8226 5249 | - |========================================================================| - | Mean |9996.0 299181.0 |290969.0 8198.0 14.0 14.0 8226.0 5249.0 | - | S.D. | 0.0 0.0 | 0.0 0.0 0.0 0.0 0.0 0.0 | - |Median|9996.0 299181.0 |290969.0 8198.0 14.0 14.0 8226.0 5249.0 | - `------------------------------------------------------------------------' - - SYSTEM SUMMARY PERCENTAGES by SPEAKER - ,--------------------------------------------------------------------. - | ./exp/g2p/text.g2p | - |--------------------------------------------------------------------| - | SPKR | # Snt # Wrd | Corr Sub Del Ins Err S.Err | + | | # Snt # Wrd | Corr Sub Del Ins Err S.Err | |--------+-----------------+-----------------------------------------| - | bak | 9996 299181 | 97.3 2.7 0.0 0.0 2.7 52.5 | - |====================================================================| | Sum/Avg| 9996 299181 | 97.3 2.7 0.0 0.0 2.7 52.5 | - |====================================================================| - | Mean |9996.0 299181.0 | 97.3 2.7 0.0 0.0 2.7 52.5 | - | S.D. | 0.0 0.0 | 0.0 0.0 0.0 0.0 0.0 0.0 | - | Median |9996.0 299181.0 | 97.3 2.7 0.0 0.0 2.7 52.5 | `--------------------------------------------------------------------' ``` The `avg CER` of text normalization is: 0.006388318503308237 ```text - - SYSTEM SUMMARY PERCENTAGES by SPEAKER - - ,----------------------------------------------------------------. - | ./exp/textnorm/text.tn | - |----------------------------------------------------------------| - | SPKR | # Snt # Wrd | Corr Sub Del Ins Err S.Err | - |------+--------------+------------------------------------------| - | utt | 125 2254 | 2241 2 11 2 15 4 | - |================================================================| - | Sum | 125 2254 | 2241 2 11 2 15 4 | - |================================================================| - | Mean |125.0 2254.0 |2241.0 2.0 11.0 2.0 15.0 4.0 | - | S.D. | 0.0 0.0 | 0.0 0.0 0.0 0.0 0.0 0.0 | - |Median|125.0 2254.0 |2241.0 2.0 11.0 2.0 15.0 4.0 | - `----------------------------------------------------------------' - - SYSTEM SUMMARY PERCENTAGES by SPEAKER - ,-----------------------------------------------------------------. - | ./exp/textnorm/text.tn | - |-----------------------------------------------------------------| - | SPKR | # Snt # Wrd | Corr Sub Del Ins Err S.Err | + | | # Snt # Wrd | Corr Sub Del Ins Err S.Err | |--------+--------------+-----------------------------------------| - | utt | 125 2254 | 99.4 0.1 0.5 0.1 0.7 3.2 | - |=================================================================| | Sum/Avg| 125 2254 | 99.4 0.1 0.5 0.1 0.7 3.2 | - |=================================================================| - | Mean |125.0 2254.0 | 99.4 0.1 0.5 0.1 0.7 3.2 | - | S.D. | 0.0 0.0 | 0.0 0.0 0.0 0.0 0.0 0.0 | - | Median |125.0 2254.0 | 99.4 0.1 0.5 0.1 0.7 3.2 | `-----------------------------------------------------------------' ``` diff --git a/examples/tiny/s0/README.md b/examples/tiny/s0/README.md index 7dc16dc3..11118dc4 100644 --- a/examples/tiny/s0/README.md +++ b/examples/tiny/s0/README.md @@ -37,4 +37,3 @@ ```bash bash local/export.sh ckpt_path saved_jit_model_path ``` - diff --git a/parakeet/data/batch.py b/parakeet/data/batch.py index 5e7ac399..515074d1 100644 --- a/parakeet/data/batch.py +++ b/parakeet/data/batch.py @@ -53,8 +53,8 @@ def batch_text_id(minibatch, pad_id=0, dtype=np.int64): peek_example = minibatch[0] assert len(peek_example.shape) == 1, "text example is an 1D tensor" - lengths = [example.shape[0] for example in - minibatch] # assume (channel, n_samples) or (n_samples, ) + lengths = [example.shape[0] for example in minibatch + ] # assume (channel, n_samples) or (n_samples, ) max_len = np.max(lengths) batch = [] diff --git a/parakeet/exps/tacotron2/ljspeech.py b/parakeet/exps/tacotron2/ljspeech.py index 59c855eb..20dc29d3 100644 --- a/parakeet/exps/tacotron2/ljspeech.py +++ b/parakeet/exps/tacotron2/ljspeech.py @@ -67,16 +67,19 @@ class LJSpeechCollector(object): # Sort by text_len in descending order texts = [ - i for i, _ in sorted( + i + for i, _ in sorted( zip(texts, text_lens), key=lambda x: x[1], reverse=True) ] mels = [ - i for i, _ in sorted( + i + for i, _ in sorted( zip(mels, text_lens), key=lambda x: x[1], reverse=True) ] mel_lens = [ - i for i, _ in sorted( + i + for i, _ in sorted( zip(mel_lens, text_lens), key=lambda x: x[1], reverse=True) ]