DMV/CCM – todo-list / progress

http://wiki.python.org/moin/PythonSpeed/PerformanceTips Eg., use +map/reduce/filter/[i for i in [i's]]/(i for i in [i's]) instead of +for-loops; use local variables for globals (global variables or or +functions), etc. +

TODO if loc_h == t, no need to try right-attachment rules OPTIMIZE
** TODO if loc_h == s, no need to try left-attachment rules :OPTIMIZE: @@ -349,7 +353,18 @@ Stick with tag-strings in initialization then switch to numbers for IO-algorithm perhaps? Can probably afford more string-matching in initialization..
DONE inner_dmv() should disregard rules with heads not in sent OPTIMIZE
+
DONE io.debug parameters should not call functions OPTIMIZE
+CLOSED: 2008-06-10 Tue 12:26
+Exchanged all io.debug(str,'level') calls with statements of the form: +
```
+if 'level' in io.DEBUG:
+    print str
+
```
+ +
+and got an almost threefold speed increase on inner(). +
DONE inner_dmv() should disregard rules with heads not in sent OPTIMIZE
CLOSED: 2008-06-08 Sun 10:18
If the sentence is "nn vbd det nn", we should not even look at rules where @@ -465,9 +480,13 @@ In regard to the E/M-step, finding P_STOP, P_CHOOSE.
10 Python-stuff
-
Make those debug statements take a bit less space in emacs: +
Make those debug statements steal a bit less attention in emacs:
```
 (font-lock-add-keywords
+ 'python-mode                   ; not really regexp, a bit slow
+ '(("^\$ *\$\$\\if +'.+' +in +io.DEBUG. *\\(
+\\1    .+$\$+\\)" 2 font-lock-preprocessor-face t)))
+(font-lock-add-keywords
  'python-mode
  '(("\\<\$\\(io\\.\$?debug(.+)\\)" 1 font-lock-preprocessor-face t)))
 
```
@@ -563,6 +582,6 @@ Good tutorial:
Author: Kevin Brubeck Unhammer <K.BrubeckUnhammer at student uva nl >
-
Date: 2008/06/09 13:04:26
+
Date: 2008/06/10 12:33:43

Skrive vha. emacs + org-mode
diff --git a/DMVCCM.org b/DMVCCM.org index 1c412a7..2e7c57c 100755 --- a/DMVCCM.org +++ b/DMVCCM.org @@ -12,13 +12,11 @@ * dmvccm report and project DEADLINE: <2008-06-30 Mon> -But absolute, extended, really-quite-dead-now deadline: August -31... - [[file:src/dmv.py][dmv.py]] - [[file:src/io.py][io.py]] - [[file:src/harmonic.py::harmonic%20py%20initialization%20for%20dmv][harmonic.py]] - +(But absolute, extended, really-quite-dead-now deadline: August 31...) * TODO [#A] P_STOP and P_CHOOSE for IO/EM [[file:src/dmv.py::DMV%20probabilities][dmv-P_STOP]] Remember: The P_{STOP} formula is upside-down (left-to-right also). @@ -146,6 +144,11 @@ The ROOT gets "each argument with equal probability", so in a sentence of three words, 1/3 for each (in [nn vb nn], 'nn' gets 2/3). Basically just a frequency count of the corpus... * [#C] Deferred +http://wiki.python.org/moin/PythonSpeed/PerformanceTips Eg., use +map/reduce/filter/[i for i in [i's]]/(i for i in [i's]) instead of +for-loops; use local variables for globals (global variables or or +functions), etc. + ** TODO if loc_h == t, no need to try right-attachment rules :OPTIMIZE: ** TODO if loc_h == s, no need to try left-attachment rules :OPTIMIZE: ** TODO when reestimating P_STOP etc, remove rules with p < epsilon :OPTIMIZE: @@ -185,6 +188,13 @@ Need to clean up the representation. Stick with tag-strings in initialization then switch to numbers for IO-algorithm perhaps? Can probably afford more string-matching in initialization.. +** DONE io.debug parameters should not call functions :OPTIMIZE: + CLOSED: [2008-06-10 Tue 12:26] +Exchanged all io.debug(str,'level') calls with statements of the form: +:if 'level' in io.DEBUG: +: print str + +and got an almost threefold speed increase on inner(). ** DONE inner_dmv() should disregard rules with heads not in sent :OPTIMIZE: CLOSED: [2008-06-08 Sun 10:18] If the sentence is "nn vbd det nn", we should not even look at rules @@ -259,7 +269,11 @@ In regard to the E/M-step, finding P_{STOP}, P_{CHOOSE}. * Python-stuff # <> -Make those debug statements take a bit less space in emacs: +Make those debug statements steal a bit less attention in emacs: +:(font-lock-add-keywords +: 'python-mode ; not really regexp, a bit slow +: '(("^\$ *\$\$\\if +'.+' +in +io.DEBUG. *\\( +:\\1 .+$\$+\\)" 2 font-lock-preprocessor-face t))) :(font-lock-add-keywords : 'python-mode : '(("\\<\$\\(io\\.\$?debug(.+)\\)" 1 font-lock-preprocessor-face t))) diff --git a/src/dmv.py b/src/dmv.py index ae39a98..cfb1908 100755 --- a/src/dmv.py +++ b/src/dmv.py @@ -50,10 +50,8 @@ # reestimate() -# import numpy # numpy provides Fast Arrays, for future optimization -import pprint +#import numpy # numpy provides Fast Arrays, for future optimization import io -import harmonic # non-tweakable/constant "lookup" globals BARS = [0,1,2] @@ -129,7 +127,7 @@ class DMV_Grammar(io.Grammar): self.p_STOP = p_STOP self.p_CHOOSE = p_CHOOSE self.p_ROOT = p_ROOT - self.head_nums = [k for k,v in numtag.iteritems()] + self.head_nums = [k for k in numtag.iterkeys()] class DMV_Rule(io.CNF_Rule): @@ -168,8 +166,8 @@ class DMV_Rule(io.CNF_Rule): return self.p(s == loc_h) elif self.R() == STOP: if not loc_h == s: - io.debug( "(%s given loc_h:%d but s:%d. Todo: optimize away!)" - % (self, loc_h, s) ) + if 'TODO' in io.DEBUG: + print "(%s given loc_h:%d but s:%d. Todo: optimize away!)" % (self, loc_h, s) return 0.0 else: return self.p(t == loc_h) @@ -179,8 +177,8 @@ class DMV_Rule(io.CNF_Rule): anything before.''' if self.LHS() == self.L(): if not loc_h == s: - io.debug( "(%s given loc_h (loc_L):%d but s:%d. Todo: optimize away!)" - % (self, loc_h, s) ) + if 'TODO' in io.DEBUG: + print "(%s given loc_h (loc_L):%d but s:%d. Todo: optimize away!)" % (self, loc_h, s) return 0.0 else: return self.p(r == loc_h) @@ -269,15 +267,16 @@ def inner(s, t, LHS, loc_h, g, sent, chart): return "\t"*n_t if (s, t, LHS, loc_h) in chart: - io.debug("%s*= %.4f in chart: s:%d t:%d LHS:%s loc:%d" - %(tab(),chart[(s, t, LHS, loc_h)], s, t, - DMV_Rule.bar_str(LHS), loc_h)) + if 'INNER' in io.DEBUG: + print "%s*= %.4f in chart: s:%d t:%d LHS:%s loc:%d" % (tab(),chart[(s, t, LHS, loc_h)], s, t, + DMV_Rule.bar_str(LHS), loc_h) return chart[(s, t, LHS, loc_h)] else: if s == t: if not loc_h == s: # terminals are always F,F for attachment - io.debug("%s*= 0.0 (wrong loc_h)" % tab()) + if 'INNER' in io.DEBUG: + print "%s*= 0.0 (wrong loc_h)" % tab() return 0.0 elif (LHS, O(s)) in g.p_terminals: prob = g.p_terminals[LHS, O(s)] # "b[LHS, O(s)]" in Lari&Young @@ -285,17 +284,22 @@ def inner(s, t, LHS, loc_h, g, sent, chart): # todo: assuming this is how to deal w/lacking # rules, since we add prob.s, and 0 is identity prob = 0.0 - io.debug( "%sLACKING TERMINAL:" % tab()) + io.debug( "%sLACKING TERMINAL:" % tab(),'INNER') # todo: add to chart perhaps? Although, it _is_ simple lookup.. - io.debug( "%s*= %.4f (terminal: %s -> %s_%d)" - % (tab(),prob, DMV_Rule.bar_str(LHS), O(s), loc_h) ) + if 'INNER' in io.DEBUG: + print "%s*= %.4f (terminal: %s -> %s_%d)" % (tab(),prob, DMV_Rule.bar_str(LHS), O(s), loc_h) return prob else: p = 0.0 # "sum over j,k in a[LHS,j,k]" for rule in g.sent_rules(LHS, sent_nums): - io.debug( "%ssumming rule %s s:%d t:%d loc:%d" % (tab(),rule,s,t,loc_h) ) + if 'INNER' in io.DEBUG: + print "%ssumming rule %s s:%d t:%d loc:%d" % (tab(),rule,s,t,loc_h) L = rule.L() R = rule.R() + if loc_h == t and rule.LHS() == L: + break # 25% faster with these cut-offs + if loc_h == s and rule.LHS() == R: + break # if it's a STOP rule, rewrite for the same range: if (L == STOP) or (R == STOP): if L == STOP: @@ -303,12 +307,11 @@ def inner(s, t, LHS, loc_h, g, sent, chart): elif R == STOP: pLR = e(s, t, L, loc_h, n_t+1) p += rule.p_STOP(s, t, loc_h) * pLR - io.debug( "%sp= %.4f (STOP)" % (tab(), p) ) + if 'INNER' in io.DEBUG: + print "%sp= %.4f (STOP)" % (tab(), p) else: # not a STOP, an attachment rewrite: - for r in range(s, t): - # if loc_h == t, no need to try right-attachments, - # if loc_h == s, no need to try left-attachments... todo + for r in xrange(s, t): p_h = rule.p_ATTACH(r, loc_h, s=s) if rule.LHS() == L: locs_L = [loc_h] @@ -316,19 +319,21 @@ def inner(s, t, LHS, loc_h, g, sent, chart): elif rule.LHS() == R: locs_L = locs(head(L), sent_nums, s, r, loc_h) locs_R = [loc_h] - # see http://tinyurl.com/4ffhhw + # see http://tinyurl.com/4ffhhw , and + # http://tinyurl.com/6o5nj7 p += sum([e(s, r, L, loc_L, n_t+1) * p_h * e(r+1, t, R, loc_R, n_t+1) for loc_L in locs_L for loc_R in locs_R]) - io.debug( "%sp= %.4f (ATTACH)" % (tab(), p) ) + if 'INNER' in io.DEBUG: + print "%sp= %.4f (ATTACH)" % (tab(), p) chart[(s, t, LHS, loc_h)] = p return p # end of e-function inner_prob = e(s,t,LHS,loc_h, 0) - if 1 in io.DEBUG: + if 'INNER' in io.DEBUG: print debug_chart(g,sent,chart) return inner_prob # end of dmv.inner(s, t, LHS, loc_h, g, sent, chart) @@ -346,11 +351,50 @@ def debug_chart(g,sent,chart): def inner_sent(loc_h, g, sent, chart): return inner(0, len(sent)-1, ROOT, loc_h, g, sent, chart) +def prune2(s,t,LHS,loc_h, g, sent_nums, chart): + '''Removes unused subtrees with positive probability from the + chart. + + Unused := any and all mothers (or grandmothers etc.) have + probability 0.0''' + def prune_helper(keep,s,t,LHS,loc_h): + keep = keep and chart[(s,t,LHS,loc_h)] > 0.0 + for rule in g.sent_rules(LHS, sent_nums): + L = rule.L() + R = rule.R() + if R==STOP: + if (s,t,L,loc_h) in chart: + prune_helper(keep, s,t, L,loc_h) + elif L==STOP: + if (s,t,R,loc_h) in chart: + prune_helper(keep, s,t, R,loc_h) + else: + for r in xrange(s,t): + for loc_L in locs(head(L), sent_nums, s, r): + if (s,r,rule.L(),loc_L) in chart: + prune_helper(keep, s ,r,rule.L(),loc_L) + for loc_R in locs(head(R), sent_nums, r+1, t): + if (r+1,t,rule.R(),loc_R) in chart: + prune_helper(keep,r+1,t,rule.R(),loc_R) + + if not (s,t,LHS,loc_h) in keepchart: + keepchart[(s,t,LHS,loc_h)] = keep + else: # eg. if previously some parent rule had 0.0, but then a + # later rule said "No, I've got a use for this subtree" + keepchart[(s,t,LHS,loc_h)] += keep + return None + + keepchart = {} + prune_helper(True,s,t,LHS,loc_h) + for (s,t,LHS,loc_h),v in keepchart.iteritems(): + if not v: + chart.pop((s,t,LHS,loc_h)) + def prune(s,t,LHS,loc_h, g, sent_nums, chart): '''Removes unused subtrees with positive probability from the chart. - Unused := all parent subtrees have (eventually, looking upwards) + Unused := any and all mothers (or grandmothers etc.) have probability 0.0''' def prune_helper(keep,s,t,LHS,loc_h): keep = keep and chart[(s,t,LHS,loc_h)] > 0.0 @@ -364,7 +408,7 @@ def prune(s,t,LHS,loc_h, g, sent_nums, chart): if (s,t,R,loc_h) in chart: prune_helper(keep, s,t, R,loc_h) else: - for r in range(s,t): + for r in xrange(s,t): for loc_L in locs(head(L), sent_nums, s, r): if (s,r,rule.L(),loc_L) in chart: prune_helper(keep, s ,r,rule.L(),loc_L) @@ -383,6 +427,8 @@ def prune(s,t,LHS,loc_h, g, sent_nums, chart): prune_helper(True,s,t,LHS,loc_h) for (s,t,LHS,loc_h),v in keepchart.iteritems(): if not v: + if 'PRUNE' in io.DEBUG: + print "popping s:%d t:%d LHS:%s loc_h:%d" % (s,t,LHS,loc_h) chart.pop((s,t,LHS,loc_h)) # end prune(s,t,LHS,loc_h, g, sent_nums, chart) @@ -408,37 +454,48 @@ def reestimate(g, corpus): P_STOP(-STOP|...) = 1 - P_STOP(STOP|...) ''' f = reestimate_zeros(g.head_nums) for sent in corpus: - chart = {} + #print sent sent_nums = [g.tagnum(w) for i,w in enumerate(sent)] - for loc_h,h in zip(range(len(sent)), sent_nums): + chart = {} + for loc_h,h in zip(xrange(len(sent)), sent_nums): + #print loc_h inner_sent(loc_h, g, sent, chart) - io.debug( debug_chart (g,sent,chart) ,'reest_chart') - prune_sent(loc_h, g, sent_nums, chart) - io.debug( debug_chart( g,sent,chart) ,'reest_chart') - for s in range(loc_h): # s 0.0: r.probN = f[('STOP','num',h)] / f[('STOP','den',h)] - io.debug( "p(STOP|%s,L,N): %s / %s = %s"%(h,f[('STOP','num',h)],f[('STOP','den',h)],f[('STOP','num',h)]/f[('STOP','den',h)]) , 'reest') + if 'reest' in io.DEBUG: + print "p(STOP|%d=%s,L,N): %s / %s = %s"%(h,g.numtag(h),f[('STOP','num',h)],f[('STOP','den',h)],f[('STOP','num',h)]/f[('STOP','den',h)]) else: - io.debug( "p(STOP|%s,L,N): %s / 0 = %s"%(h,f[('STOP','num',h)]), 'reest') + if 'reest' in io.DEBUG: + print "p(STOP|%d=%s,L,N): %s / 0"%(h,g.numtag(h),f[('STOP','num',h)]) @@ -463,9 +520,13 @@ def testreestimation(): 'det nn vbd', 'det nn vbd c vbd det pp', 'det nn vbd', 'det nn vbd c vbd pp', 'det nn vbd pp', 'det nn vbd det', ]] + import harmonic g = harmonic.initialize(corpus) + corpus = [s.split() for s in ['det nn vbd det nn' ]] reestimate(g, corpus) + + def testgrammar_h(): # Non, Adj _h_ = DMV_Rule((LRBAR,0), STOP, ( RBAR,0), 1.0, 1.0) # LSTOP h_S = DMV_Rule(( RBAR,0),(NOBAR,0), STOP, 0.4, 0.3) # RSTOP @@ -480,6 +541,12 @@ def testgrammar_h(): # Non, Adj return DMV_Grammar([ rh, _h_, h_S, h_A, h ],b2,0,0,0, {0:'h'}, {'h':0}) +def testreestimation_h(): + io.DEBUG=['reest'] + g = testgrammar_h() + reestimate(g,['h h h'.split()]) + + def testprune(): g = testgrammar_h() sent = 'h h h'.split() @@ -490,16 +557,12 @@ def testprune(): prune( 0,2,ROOT,2, g,[0,0,0],chart) print debug_chart(g,sent,chart) -def testreestimation_h(): - io.DEBUG=['reest'] - g = testgrammar_h() - reestimate(g,['h h h'.split()]) if __name__ == "__main__": io.DEBUG = [] import timeit - timeit.Timer("dmv.testreestimation_h()",'''import dmv -reload(dmv)''').timeit(1) + print timeit.Timer("dmv.testreestimation()",'''import dmv +reload(dmv)''').timeit(100) g_dup = testgrammar_h() diff --git a/src/dmv.pyc b/src/dmv.pyc dissimilarity index 62% index 0864141ee5e1e752e89b4c986dbda98a4b24c09b..fc606e71aa3b77a0b1800a72f6e654c1f4ff05e0 100644 GIT binary patch delta 4951 zcwU`YeQaA-6+idAXD83Goy2zR_#=t)Z9A>ortP{l3F)?UX-AuMPPZ;;yDZ6jN!+G( z96MD8$2e8xj&_LgJ52Oz;mv0{#G@2_bC*A;dH`IOjau ziL*@z62Ig7`Mc+wd%m7;ToE5UCd|#hxg1a5GjePOihmA#j^L9&tpFJK5pWS;0vie# zDi|8fsW2g^J)ux{LZzt*4NR5R&tlsKrUr(M+jcOko>~j9SA(I$!-%AV;oyb?j2bWj z(ls!PcupF;1~5M11jEG>E}n3gC)~`0;7=_)j2zT~QO98QVAONN1BQniUNF4e@PXlj zb!s($(ZEA~F#OyIfDzzEBN&a`2!avhMhJ`$H^N|qxe-B8z_7yw1+q;-%mFbu*;LX} zGfXst*$n9xFd|?y!8HZa&0sXcH393x$IK8(F%gCGMzg-CRg-my>OyKp!DxYWD=nT! z!K+}R4GbJ?1+xwOc=)0Vgwe*m;bR#A;uCPx9XRU@)DA{F1GR4fb#Si&NZSUaF;HhE zLmgmrfzb_S#}-bEd&9?arxZb0sFJZ*94#sRZ~T*wDIbaef zCpK@>UQorf{8QCyVqQL9{gSx2`C0uIAu{q2$D3kVrfPoP`~X>`LiVZ5R2Qw(j?e8h zZpuT>6PgRwbomA6T!iM+1#{`5`9Qv)V`tu+GA1+2m+~=(tapvbp9Vr=O2C*>2KkWE9C*yMLSSM>s&c6xMZRPOfP)tJOi%1b1Kq+mlbbo92nhvhTg z!@FJ}p6W<;)hhzRD_o*Z_(iocLl~upIsb@cuvV+a9%cH*j zhz0?i#ocJQV?1X=#!27#aru2uuYAjwu%-53U%u~~7fCtYaJ%Q_EjEJ|i}KZmS@C1M z?X>vG=7!%U;t3M52T81oT68(7S|XmZ0%$gskUtD`ir3|D18MQ*<{gcrLcAfL2~LaO z$oGRkuou_76553WzYjImbO#>M7vG3szSjD`wS(Rv=qDyz+O^(0!J^HY_cQk|H$-j%8+5VGWYG z{L(TP;CqPe0ItKnW&xLP0JWSsBZpgW?rrAYnw>Y;DCiVdA_j()N+`PGD*0k-*Pw;7 zYDb(6g%PS?#ZWeEkhT+_8-jKZoj;|lRYL_v3u08u4_gBRMT}}lSJ8=UI#9cs+YVUO z0e!0I^Q(?xCwm<`G$Tjb!X6z0*YM~X$RvwvLgf$r>`mE{F;j=TnuEfy@zjT zPQaQQ3izo7WEYJd^@O`IVFil3A>Xx-8MO3bAvhSJ2Lx}FbL|OHC!cQbH7z+;UEtf$ zw!@$h9@qfx;{u;2H>T!}s}I)EiJvI8^J8LvRT21{$z7FA9j(G82Rkk|=w+)`IC2~x z89ADgH#*+yCfnwSlG;mz93pjuNIMjxjA2eM&&lh%8yrnEL@vG}|I#^vK7F|B%j=l- zp_<^Q*=cIEAo_Ki2#Q9u|0SLzphQkTwgK>QWtI9Os%bq^0SU)JY^U3TY{KM}+`#ygN~ImU>5tjLR=}ch?Y$ zsk2u9Mt67pKggHe1mvn!6H0hk{K`}PfS7vYj&TUs3g7V$CTTIH2;@=a`%I_w|lbHviQyC*VwXhJqlr7Ag zOZjLlyA+*U!qL)zS(r)2qw`aj&CR}^&nse7zPkJ1$rMxs#u_4Lh}=sY(Cx8Q7fo|< zW`1g^z^oLUNEVltGv?&Hxv*$1rN(KCPwtERY*?tkyi6tUL=L}}>`XH6KIWas{e%l8 z@YGCMJi`+2y`-_^Qjd$q03Fj0zr@H~iXw{2%SfQ+(oq|@lt&FS1@VKEuTH85?t|M|o{G2g3y%H1uzwSe$BvAJZFD=>JL7|KYpX`A((o9yywCC5>hm7t@)!X#Q}g5iJlo zact~xXFj@+otd1+{)h!_5}72j<<+xva5XXazu#w5=uz@+k_YVJupNES1&yK#{V^yf z<6|An%!qGS|r{SC@w;LYR$at_2#T2&EqF^C~%EZWJ?Zyw< z6{iA4A((r0N?V003i-w240D-gRYe%gb+)@OhH=N%JXS>pClt%_R;~*?0|A&lDKLkv zJSJb>*RYzckUb3Ns1>aq<&4 z1rMqH)|vZ?hqmLPcKFowlsTRiT$Rfsy$#i+Q{{Ru){%0sH_%vy+=hs7G}#!z?M=m^ ze*2o;*w^K^_P4GVv(ww?hi}7~+#7IarJfF-BYoK^=jRqMYZY+Lk8_<|S-Y(05D=7@ zU66KjrCrNu^&?aW4%L;a4of*zMdbvsAh@7G_SvG^yu5P2-@l^~X30DDrtONGo zcX#=_P8<{t+~Ks*Dlb+EAs(H=`hkChAbuL4_(Jm|Bpe|46}$8kgv{J@*S&_DUc4^~ z$|mTwX@#zd;)`bOu+eV>Y=B-$J6K6!%t`Yv37*^&@Su$%Q9r3O>FSTlQv<{MbEFh9wD~Ch zTXB8M0p1R918N59Z}$qsU=Ca z7dYJmu2h#&DzXsBDeOu_)Qe7$_v46h>2k7Q<_qKsh4iJVLOPp?#d#`yX)(K0h~f&z zlDUwbGGf?_rzsuDBcGA+&ayZMiLe>7L4U5y9NTi2bPl$i;3AIlvz4yzEA!5NF;|iz zb%970N~)os5-d$UMC2-wy>fZ5T`bG544%|Z6F2V+wj8`)*L79z))V+$t@rC5-L2d8 U8W)SKi+V*@bdf;8UCI{y4~>puX$-Q1hRX%gG< zO$p4#K$u!h+J^KB_G1+j(>7@mng)?3G))trN$kg9f2K_n`?LK3P5fYN`!NZIcFudR zofg_H&UXnapL`SFcpB?F{|Nb~dQYBM0QWCQ?I^X%MFqg5e}qc_xh9X0 z!v?QFP63ntRWLPJQX!`@pVPRT^YBzo2g}3j7m4?RrGx2_ybmk`%(@OM3KR`M(+kH> ztRpc#c#aPAgK0=60H$9uK`;YgArwNeNFoGza2sGQ*8*lpPPEF2mgYp86b|KoD?CS% zb%5C>!NOp+6H|v=1k4UO)Cp!-GEp!il8J%YDVaE!QOR_H8Iw#mm~qMUfY~LP1eo2D z>6Mx}iX{-2fMuKC(=b{e$A;S`^{pP3J=YJ-eOB1Du_Z0K zC{1qlgV_s(0baaJ>UtqJ2qqmu*n<$Ivr8eeWm5JW!`=i(t=|l22+V#7G_(seAbU3e zY1aX15@+y6oD`S@l>#fZi!&^H$4`_WQ;@K{NG3U(9@()K590meN%c|usu+zX#EWWJ z=-QX?)Xp8+4^%Y8AHA>Rig@1l3a;%OF%BS>#N+g zNW;3YLeF4!=Y!BE>cC43$jE?R3@8g4aojwO!>$OeBQu=?tP;M;*051th5Knpg{th7 z1^{gh3Ld@@w9uJYR3XY&fNXlUye(=S-5!;Uqls%B&x(=onfN3fkPS2GCSvoMrgd?1 zNjx7uSpOlP-o}!tVT@rILug<-`feiQcZslgGEA63W*R~&EKWm6g_UrHloK}yDI7I5 zl3qL}e?@*T;F~)l z>Op;swHy~O#d`H|X2-;vu@wGVd=M+(yE|v%rxD)~FLzzU_k`B{T4T*GyT?c0V|2gt zP#inyr(-(A(<4mMOh%bZiuRtICqwL&ol;L(!4G%-+IvkwkEkaT_!Dt0nGVd71*CaQ zMf@o_h1;UFKfZrNWvv}g7Ju%zRhqAqXs)Lr4;$676!?FHPF-HgV?qQY5`cEa2JVR4D{4@hxq`n#5H!0 z-X1_pTAvqpo6melN zawG(H3y?^JMJA!uR6b(v`6OE_tl#4(FSpP_)^3bU%^St5gBffUe;&Mj&e5eBg2+|v zdKfym1FnMX(}GAxIii;R>Qile;&rM8Zl;QF@Mw z#q`ISZ1P5ZIb)}d=2i!fiC;WD6~|$b48-;9IT|@5-WpvVI?r>Q$+CxrZ0^q0)nqBZVXfQA%H~CT!?G&L<@{wUnYJ#|;AVbx^@;K1YO#2! zu)dVsEIppzFe{@wuZ(?N!IL6Aap&z1GIj?OnTEN{WFKbo2n)itk)`C4WtA3|^BZ<{ zPJCygQC8DOqhTxGGRpfRJy7sGwU0fY*xyU&LQ07u+Md*eVv~xw{!^J&!Mb9*0Lbp20cimZ` zTc72wHpI09{dG=V?p)c6v*izAcB`FtEwTA@&h_PG%4#;umZLQHk@R;9C%&2Ue*wu^P(CMFJ^+b`c6m4m5D(q^(+TsW9AWisO|U}xJoL>76Cmk7NA$Bv{b*(LGD z!LfRRV-id@gpbQ(uiuCU!e=djnZZ%%lW#I_kMOLF5WqGc>YB$ z6US%_)K>IT5zuiTCLl=3J|SC(#-h}Esenx4gi8;jA9N+?;n^^D;|N8k{5WL0bb|YW z%%HFjTXlNc78LdX=L?lTy^f#QdWFM^k^n`#jJoe}-l)pg?zR-0@5Qs^eT`Q?3AUHN z&g9Z%;N0SqVXUwVDX-868|74b39(~o!@axkUy++`yBU7W#jnbcoA-n4mEWgJ@jhMN z*Qj0L3j(v0I?txtclI9HhkGlnY}ouIstc9knw7K*YgU1MRB8XBi&K?!P&{{cCNEvW z-_1^CkXK1Mv>$>0T!54WBv%&r!wEQGjh+55i{c5qXRf5K0Z3% z`#$4EB;FAHGW6R+NKx#

7 [#C] Deferred

10 Python-stuff