Add preview of 1984-2019
The 2020 winners are hinted it, but the source code is not there,
nor are the hints. Someone could look into the HTML and
see what was commented out, but that is not a problem because
WHO won is known and the name of their AWARD is known already.
Moreover a number of the winners have talked about what their
entirs do, so the commented out spoiler text is not a bit problem.
One will find HTML commented out with:
<!-- XXX remove when 2020 is ready
2020 links here
XXX remove when 2020 is ready -->
The "XXX remove when 2020 is ready" string is found at the
start and end comment blocks. Again, soneone could read
what was commented out, but that is OK since most of that
information is known or can be known. The following files
have such HTML comment blocks in them.
./winners.html
./years-spoiler.html
./years.html
In addition, ./all/summary.txt was modified to not
include the 2020 entries for the momemt.
The 2020 winner code AND write-ups are NOT part of this set.
p.s. The all/all.tar.bz2 does not contain the 2020 source code (yet?).
This commit is contained in:
Landon Curt Noll committed
1 parent
28efc67f5d
commit
c0663537cb
1683 files changed
+220064
No files matched your search
Binary file not shown.
Binary file not shown.
Binary file not shown.
Binary file not shown.
Binary file not shown.
Binary file not shown.
@@ -0,0 +1,707 @@
|
||||
#!/usr/bin/env make
|
||||
#
|
||||
# 2019 makefile
|
||||
#
|
||||
# This work by Landon Curt Noll, Simon Cooper, and Leonid A. Broukhis
|
||||
# is licensed under:
|
||||
#
|
||||
# Creative Commons Attribution-ShareAlike 3.0 Unported License.
|
||||
#
|
||||
# See: http://creativecommons.org/licenses/by-sa/3.0/
|
||||
|
||||
#############################
|
||||
# shell used by this Makefile
|
||||
#############################
|
||||
#
|
||||
SHELL= /bin/bash
|
||||
|
||||
#######################
|
||||
# common tool locations
|
||||
#######################
|
||||
#
|
||||
A2P= a2p
|
||||
AR= ar
|
||||
ASA= asa
|
||||
AT= at
|
||||
ATQ= atq
|
||||
ATRM= atrm
|
||||
AWK= awk
|
||||
BANNER= banner
|
||||
BASE64= base64
|
||||
BASENAME= basename
|
||||
BATCH= batch
|
||||
BC= bc
|
||||
BINHEX= binhex
|
||||
BISON= bison
|
||||
BUNZIP2= bunzip2
|
||||
BZCAT= bzcat
|
||||
BZCMP= bzcmp
|
||||
BZDIFF= bzdiff
|
||||
BZEGREP= bzegrep
|
||||
BZFGREP= bzfgrep
|
||||
BZGREP= bzgrep
|
||||
BZIP2= bzip2
|
||||
BZLESS= bzless
|
||||
BZMORE= bzmore
|
||||
C2PH= c2ph
|
||||
C89= c89
|
||||
C99= c99
|
||||
CAL= cal
|
||||
CALC= calc
|
||||
CAT= cat
|
||||
CD= cd
|
||||
CHFLAGS= chflags
|
||||
CHGRP= chgrp
|
||||
CHMOD= chmod
|
||||
CKSUM= cksum
|
||||
CLANG= clang
|
||||
CLANG_PLUSPLUS= clang++
|
||||
CLEAR= clear
|
||||
CMP= cmp
|
||||
COL= col
|
||||
COLLDEF= colldef
|
||||
COLRM= colrm
|
||||
COLUMN= column
|
||||
COMPRESS= compress
|
||||
CP= cp
|
||||
CPIO= cpio
|
||||
CPP= cpp
|
||||
CRC32= crc32
|
||||
CSH= csh
|
||||
CSPLIT= csplit
|
||||
CURL= curl
|
||||
CUT= cut
|
||||
C_PLUSPLUS= c++
|
||||
C_PLUSPLUS_FILT= c++filt
|
||||
DATE= date
|
||||
DC= dc
|
||||
DD= dd
|
||||
DF= df
|
||||
DIFF3= diff3
|
||||
DIFF= diff
|
||||
DIG= dig
|
||||
DIRNAME= dirname
|
||||
ED= ed
|
||||
EGREP= egrep
|
||||
ENV= env
|
||||
EQN= eqn
|
||||
ETAGS= etags
|
||||
EXPECT= expect
|
||||
EXPR= expr
|
||||
FALSE= false
|
||||
FGREP= fgrep
|
||||
FILE= file
|
||||
FIND2PERL= find2perl
|
||||
FIND= find
|
||||
FLEX= flex
|
||||
FLEX_PLUSPLUS= flex++
|
||||
FMT= fmt
|
||||
FOLD= fold
|
||||
FS_USAGE= fs_usage
|
||||
FUNZIP= funzip
|
||||
FUSER= fuser
|
||||
GCC= gcc
|
||||
GDIFFMK= gdiffmk
|
||||
GENCAT= gencat
|
||||
GENSTRINGS= genstrings
|
||||
GETOPT= getopt
|
||||
GETOPTS= getopts
|
||||
GNUMAKE= gnumake
|
||||
GREP= grep
|
||||
GROFF= groff
|
||||
GROFFER= groffer
|
||||
GROG= grog
|
||||
GROPS= grops
|
||||
GROTTY= grotty
|
||||
GUNZIP= gunzip
|
||||
GVIM= gvim
|
||||
GZCAT= gzcat
|
||||
GZEXE= gzexe
|
||||
GZIP_PROG= gzip
|
||||
G_PLUSPLUS= g++
|
||||
H2PH= h2ph
|
||||
H2XS= h2xs
|
||||
HASH= hash
|
||||
HEAD= head
|
||||
HOSTNAME_PROG= hostname
|
||||
ICONV= iconv
|
||||
ID= id
|
||||
INDENT= indent
|
||||
INFO= info
|
||||
JOT= jot
|
||||
KILL= kill
|
||||
KSH= ksh
|
||||
LAST= last
|
||||
LD= ld
|
||||
LESSECHO= lessecho
|
||||
LEX= len
|
||||
LINK= link
|
||||
LN= ln
|
||||
LS= ls
|
||||
M4= m4
|
||||
MAKE= make
|
||||
MAN= man
|
||||
MKDIR= mkdir
|
||||
MKFIFo= mkfifo
|
||||
MKTEMP= mktemp
|
||||
MV= mv
|
||||
NANO= nano
|
||||
NASM= nasm
|
||||
NEQN= neqn
|
||||
NICE= nice
|
||||
NL= nl
|
||||
NM= nm
|
||||
NOHUP= nohup
|
||||
NROFF= nroff
|
||||
NSLOOKUP= nslookup
|
||||
OD= od
|
||||
OPENSSL= openssl
|
||||
PASTE= paste
|
||||
PATCH= patch
|
||||
PATHCHK= pathchk
|
||||
PAX= pax
|
||||
PERL= perl
|
||||
PICO= pico
|
||||
PR= pr
|
||||
PRINTENV= printenv
|
||||
PS= ps
|
||||
PTAR= ptar
|
||||
PTARDIFF= ptardiff
|
||||
PTARGREP= ptargrep
|
||||
PWD= pwd
|
||||
PYDOC= pydoc
|
||||
PYTHON= python
|
||||
PYTHONW= pythonw
|
||||
READLINK= readlink
|
||||
RENICE= renice
|
||||
RESET= reset
|
||||
REV= rev
|
||||
RI= ri
|
||||
RM= rm
|
||||
RMDIR= rmdir
|
||||
RSYNC= rsync
|
||||
RUBY= ruby
|
||||
RVIM= rvim
|
||||
SAY= say
|
||||
SCP= scp
|
||||
SCREEN= screen
|
||||
SCRIPT= script
|
||||
SDIFF= sdiff
|
||||
SED= sed
|
||||
SEQ= seq
|
||||
SFTP= sftp
|
||||
SH= sh
|
||||
SHA1= sha1
|
||||
SHA= sha
|
||||
SHAR= shar
|
||||
SHASUM5_18= shasum5.18
|
||||
SHASUM= shasum
|
||||
SIZE= size
|
||||
SLEEP= sleep
|
||||
SORT= sort
|
||||
SSH= ssh
|
||||
STAT= stat
|
||||
STRIP= strip
|
||||
STTY= stty
|
||||
SUM= sum
|
||||
SYNC= sync
|
||||
TABS= tabs
|
||||
TAIL= tail
|
||||
TAR= tar
|
||||
TEE= tee
|
||||
TEST= test
|
||||
TFTP= tftp
|
||||
TIDY= tidy
|
||||
TIME= time
|
||||
TOP= top
|
||||
TOUCH= touch
|
||||
TPUT= tout
|
||||
TPUT= tput
|
||||
TR= tr
|
||||
TROFF= troff
|
||||
TRUE= true
|
||||
TSET= tset
|
||||
TSORT= tsort
|
||||
UL= ul
|
||||
UNAME= uname
|
||||
UNCOMPRESS= uncompress
|
||||
UNEXPAND= unexpand
|
||||
UNIFDEF= unifdef
|
||||
UNIFDEFALL= unifdefall
|
||||
UNIQ= uniq
|
||||
UNITS= units
|
||||
UNLINK= unlink
|
||||
UNZIP= unzip
|
||||
UNZIPSFX= unzipsfx
|
||||
UPTIME= uptime
|
||||
UUDECODE= uudecode
|
||||
UUENCODE= uuencode
|
||||
UUIDGEN= uuidgen
|
||||
VI= vi
|
||||
VIEW= view
|
||||
VIM= vim
|
||||
VIMDIFF= vimdiff
|
||||
W= w
|
||||
WAIT4PATH= wait4path
|
||||
WAIT= wait
|
||||
WC= wc
|
||||
WHAT= what
|
||||
WHATIS= whatis
|
||||
WHICH= which
|
||||
WHO= who
|
||||
WHOAMI= whoami
|
||||
WHOIS= whois
|
||||
WRTIE= write
|
||||
XAR= xar
|
||||
XARGS= xargs
|
||||
XATTR= xattr
|
||||
XXD= xxd
|
||||
YACC= yacc
|
||||
YES= yes
|
||||
ZCAT= zcat
|
||||
ZCMP= zcmp
|
||||
ZDIFF= zdiff
|
||||
ZEGREP= zegrep
|
||||
ZFGREP= zfgrep
|
||||
ZFORCE= zforce
|
||||
ZGREP= zgrep
|
||||
ZIP= zip
|
||||
ZIPCLOAK= zipcloak
|
||||
ZIPGREp= zipgrep
|
||||
ZIPINFo= zipinfo
|
||||
ZIPNOTE= zipnote
|
||||
ZIPSPLIT= zipsplit
|
||||
ZLESS= zless
|
||||
ZMORE= zmore
|
||||
ZNEW= znew
|
||||
ZPRINT= zprint
|
||||
ZSH= zsh
|
||||
|
||||
# Set X11_LIBDIR to the directory where the X11 library resides
|
||||
#
|
||||
#X11_LIBDIR= /usr/X11R6/lib
|
||||
#X11_LIBDIR= /usr/X11/lib
|
||||
X11_LIBDIR= /opt/X11/lib
|
||||
|
||||
# Set X11_INCLUDEDIR to the directory where the X11 include files reside
|
||||
#
|
||||
#X11_INCDIR= /usr/X11R6/include
|
||||
#X11_INCDIR= /usr/X11/include
|
||||
X11_INCDIR= /opt/X11/include
|
||||
|
||||
# Compiler warnings
|
||||
#
|
||||
#CWARN=
|
||||
#CWARN= -Wall
|
||||
#CWARN= -Wall -Wextra
|
||||
CWARN= -Wall -Wextra -pedantic ${CSILENCE}
|
||||
#CWARN= -Wall -Wextra -Weverything
|
||||
#CWARN= -Wall -Wextra -Weverything -pedantic
|
||||
#CWARN= -Wall -Wextra -Weverything -pedantic ${CSILENCE}
|
||||
|
||||
# Silence warnings that ${CWARN} would normally complain about
|
||||
#
|
||||
#CSILENCE=
|
||||
#CSILENCE= -Wno-implicit-int
|
||||
CSILENCE= -Wno-missing-field-initializers -Wno-for-loop-analysis
|
||||
|
||||
# compiler standard
|
||||
#
|
||||
#CSTD=
|
||||
#CSTD= -ansi
|
||||
CSTD= -std=c11
|
||||
|
||||
# compiler bit architecture
|
||||
#
|
||||
# Some entries require 32-bitness:
|
||||
# ARCH= -m32
|
||||
#
|
||||
# Some entries require 64-bitness:
|
||||
# ARCH= -m64
|
||||
#
|
||||
# By default we assume nothing:
|
||||
#
|
||||
ARCH=
|
||||
|
||||
# defines that are needed to compile
|
||||
#
|
||||
#CDEFINE=
|
||||
#CDEFINE= -DIOCCC=26
|
||||
CDEFINE= ${param}
|
||||
|
||||
# include files that are needed to compile
|
||||
#
|
||||
CINCLUDE=
|
||||
#CINCLUDE= -include stdlib.h
|
||||
#CINCLUDE= -include stdio.h
|
||||
#CINCLUDE= -include stdlib.h -include stdio.h
|
||||
#CINCLUDE= -I ${X11_INCDIR}
|
||||
|
||||
# optimization
|
||||
#
|
||||
# Most compiles will safely use -O2. Some can use only -O1 or -O.
|
||||
# A few compilers have broken optimizers or this entry make break
|
||||
# under those buggy optimizers and thus you may not want anything.
|
||||
# Reasonable compilers will allow for -O3.
|
||||
#
|
||||
#OPT=
|
||||
#OPT= -O
|
||||
#OPT= -O1
|
||||
#OPT= -O2
|
||||
OPT= -O3
|
||||
|
||||
# default flags for ANSI C compilation
|
||||
#
|
||||
CFLAGS= ${CSTD} ${CWARN} ${ARCH} ${CDEFINE} ${CINCLUDE} ${OPT}
|
||||
|
||||
# Libraries needed to build
|
||||
#
|
||||
#LIBS=
|
||||
#LIBS= -L ${X11_LIBDIR}
|
||||
LIBS= -lm
|
||||
|
||||
# ANSI compiler
|
||||
#
|
||||
# Set CC to the name of your ANSI compiler.
|
||||
#
|
||||
# Some entries seem to need gcc. If you have gcc, set
|
||||
# both CC and MAY_NEED_GCC to gcc.
|
||||
#
|
||||
# If you do not have gcc, set CC to the name of your ANSI compiler, and
|
||||
# set MAY_NEED_GCC to either ${CC} (and hope for the best) or to just :
|
||||
# to disable such programs.
|
||||
#
|
||||
CC= cc
|
||||
#CC=clang
|
||||
MAY_NEED_GCC= gcc
|
||||
|
||||
|
||||
##############################
|
||||
# Special flags for this entry
|
||||
##############################
|
||||
#
|
||||
OBJ= prog.o
|
||||
INPUTS_GZ= IOCCC-hints.txt.gz Shakespeare.txt.gz \
|
||||
IOCCC-Rules-Guidelines.txt.gz Eugene_Onegin.txt.gz
|
||||
INPUTS= IOCCC-hints.txt Shakespeare.txt \
|
||||
IOCCC-Rules-Guidelines.txt Eugene_Onegin.txt
|
||||
CPFILES= IOCCC-hints.cp09_1.809 Shakespeare.cp04_1.633 \
|
||||
IOCCC-Rules-Guidelines.cp98_0.175 Eugene_Onegin.cp11_1.188
|
||||
OUTPUTS= IOCCC-hints.output.txt Shakespeare.output.txt \
|
||||
IOCCC-Rules-Guidelines.output.txt Eugene_Onegin.output.txt
|
||||
DATA= ${CPFILES} ${INPUTS_GZ} ${CPFILES}
|
||||
TARGET = lin1 per1 rnn1 rnn2 rnn3 lstm1 lstm2 lstm3 lstmp1 lstmp2 lstmp3 \
|
||||
rlstm1 rlstm2 rlstm3 gru1 gru2 gru3 prog
|
||||
#
|
||||
ALT_OBJ=
|
||||
ALT_TARGET=
|
||||
|
||||
# hyperparameters
|
||||
param = -DRS=.15 -DTR=.95 -DLR=.002 -DLD=.97 -DLE=10 -DCL=5 -DB1=.9 -DB2=.999 \
|
||||
-DEP=1e-8 -DWD=8e-5 -DDI=100 -DSL=200 -DN=50 -DTP=1 \
|
||||
-DPF='"cp%02d_%.3f"'
|
||||
|
||||
##############################################################
|
||||
# Networks
|
||||
#
|
||||
# Note that temp variables have been added to force evaluation
|
||||
# order. This allows checkpoint files to be portable beetween
|
||||
# different compilers.
|
||||
##############################################################
|
||||
|
||||
# linear
|
||||
lin = -DBK='y=x'
|
||||
|
||||
# perceptron
|
||||
per = -DBK='y=T(x)'
|
||||
|
||||
# recurrent neural network
|
||||
rnn = -DHS=128 -DBK=' \
|
||||
hp = I(HS), \
|
||||
t1 = L(HS, x), \
|
||||
h = C(hp, T(A(t1, L(HS, hp)))), \
|
||||
y = h \
|
||||
'
|
||||
|
||||
# long short term memory
|
||||
lstm = -DHS=128 -DBK=' \
|
||||
cp = I(HS), \
|
||||
hp = I(HS), \
|
||||
t1 = L(HS, x), \
|
||||
f = S(A(t1, L(HS, hp))), \
|
||||
t2 = L(HS, x), \
|
||||
i = S(A(t2, L(HS, hp))), \
|
||||
t3 = L(HS, x), \
|
||||
cn = T(A(t3, L(HS, hp))), \
|
||||
t4 = M(f, cp), \
|
||||
c = C(cp, A(t4, M(i, cn))), \
|
||||
t5 = L(HS, x), \
|
||||
o = S(A(t5, L(HS, hp))), \
|
||||
h = C(hp, M(o, T(c))), \
|
||||
y = h \
|
||||
'
|
||||
|
||||
# lstm with passthrough
|
||||
lstmp = -DHS=128 -DBK=' \
|
||||
cp = I(HS), \
|
||||
hp = I(HS), \
|
||||
t1 = L(HS, x), \
|
||||
t2 = A(t1, L(HS, hp)), \
|
||||
f = S(A(CM(cp), t2)), \
|
||||
t3 = L(HS, x), \
|
||||
t4 = A(t3, L(HS, hp)), \
|
||||
i = S(A(CM(cp), t4)), \
|
||||
t5 = L(HS, x), \
|
||||
cn = T(A(t5, L(HS, hp))), \
|
||||
t6 = M(f, cp), \
|
||||
c = C(cp, A(t6, M(i, cn))), \
|
||||
t7 = L(HS, x), \
|
||||
t8 = A(t7, L(HS, hp)), \
|
||||
o = S(A(CM(c), t8)), \
|
||||
h = C(hp, M(o, T(c))), \
|
||||
y = h \
|
||||
'
|
||||
|
||||
# residual lstm
|
||||
rlstm = -DHS=128 -DBK=' \
|
||||
cp = I(HS), \
|
||||
hp = I(HS), \
|
||||
t1 = L(HS, x), \
|
||||
t2 = A(t1, L(HS, hp)), \
|
||||
f = S(A(CM(cp), t2)), \
|
||||
t3 = L(HS, x), \
|
||||
t4 = A(t3, L(HS, hp)), \
|
||||
i = S(A(CM(cp), t4)), \
|
||||
t5 = L(HS, x), \
|
||||
cn = T(A(t5, L(HS, hp))), \
|
||||
t6 = M(f, cp), \
|
||||
c = C(cp, A(t6, M(i, cn))), \
|
||||
t7 = L(HS, x), \
|
||||
t8 = A(t7, L(HS, hp)), \
|
||||
o = S(A(CM(c), t8)), \
|
||||
m = L(HS, T(c)), \
|
||||
h = C(hp, M(o, A(m, L(HS, x)))), \
|
||||
y = h \
|
||||
'
|
||||
|
||||
# gated recurrent unit
|
||||
gru = -DHS=128 -DBK=' \
|
||||
hp = I(HS), \
|
||||
t1 = L(HS, x), \
|
||||
z = S(A(t1, L(HS, hp))), \
|
||||
t2 = L(HS, x), \
|
||||
r = S(A(t2, L(HS, hp))), \
|
||||
t3 = L(HS, x), \
|
||||
c = T(A(t3, L(HS, M(r, hp)))), \
|
||||
zc = OG(1, -1, z), \
|
||||
t4 = M(zc, hp), \
|
||||
h = C(hp, A(t4, M(z, c))), \
|
||||
y = h \
|
||||
'
|
||||
|
||||
# single-layer network
|
||||
one_layer = -DNW=' \
|
||||
x = I(n), \
|
||||
y = L(n, MD(x)) \
|
||||
'
|
||||
|
||||
# two-layer network
|
||||
two_layer = -DNW=' \
|
||||
x = I(n), \
|
||||
y = L(n, MD(MD(x))) \
|
||||
'
|
||||
|
||||
# three-layer network
|
||||
three_layer = -DNW=' \
|
||||
x = I(n), \
|
||||
y = L(n, MD(MD(MD(x)))) \
|
||||
'
|
||||
|
||||
# Enumerated list of networks
|
||||
#
|
||||
lin1 = ${lin} ${one_layer}
|
||||
per1 = ${per} ${one_layer}
|
||||
rnn1 = ${rnn} ${one_layer}
|
||||
rnn2 = ${rnn} ${two_layer}
|
||||
rnn3 = ${rnn} ${three_layer}
|
||||
lstm1 = ${lstm} ${one_layer}
|
||||
lstm2 = ${lstm} ${two_layer}
|
||||
lstm3 = ${lstm} ${three_layer}
|
||||
lstmp1 = ${lstmp} ${one_layer}
|
||||
lstmp2 = ${lstmp} ${two_layer}
|
||||
lstmp3 = ${lstmp} ${three_layer}
|
||||
rlstm1 = ${rlstm} ${one_layer}
|
||||
rlstm2 = ${rlstm} ${two_layer}
|
||||
rlstm3 = ${rlstm} ${three_layer}
|
||||
gru1 = ${gru} ${one_layer}
|
||||
gru2 = ${gru} ${two_layer}
|
||||
gru3 = ${gru} ${three_layer}
|
||||
|
||||
# default network
|
||||
#
|
||||
prog = ${rnn1}
|
||||
|
||||
#################
|
||||
# build the entry
|
||||
#################
|
||||
#
|
||||
all: ${TARGET} ${DATA}
|
||||
@${TRUE}
|
||||
|
||||
${TARGET}: prog.c Makefile ${INPUTS}
|
||||
${CC} ${CFLAGS} -o $@ $< $($@) ${LIBS}
|
||||
|
||||
# alternative executable
|
||||
#
|
||||
alt: ${ALT_TARGET}
|
||||
@${TRUE}
|
||||
|
||||
# data files
|
||||
#
|
||||
data: ${DATA}
|
||||
@${TRUE}
|
||||
|
||||
${INPUTS}: %.txt: %.txt.gz
|
||||
${ZCAT} < $< > $@
|
||||
|
||||
test: ${OUTPUTS}
|
||||
|
||||
# First 1000 lines of generated output
|
||||
#
|
||||
generator = ./lstm2
|
||||
|
||||
test-64bit: IOCCC-hints.output.txt IOCCC-Rules-Guidelines.output.txt Shakespeare.output.txt Eugene_Onegin.output.txt
|
||||
|
||||
IOCCC-hints.output.txt: IOCCC-hints.cp09_1.809 ${generator}
|
||||
${generator} < $< | ${HEAD} -n 1000 > $@
|
||||
|
||||
Shakespeare.output.txt: Shakespeare.cp04_1.633 ${generator}
|
||||
${generator} < $< | ${HEAD} -n 1000 > $@
|
||||
|
||||
IOCCC-Rules-Guidelines.output.txt: IOCCC-Rules-Guidelines.cp98_0.175 ${generator}
|
||||
${generator} < $< | ${HEAD} -n 1000 > $@
|
||||
|
||||
Eugene_Onegin.output.txt: Eugene_Onegin.cp11_1.188 ${generator}
|
||||
${generator} < $< | ${HEAD} -n 1000 > $@
|
||||
|
||||
###############
|
||||
# utility rules
|
||||
###############
|
||||
#
|
||||
everything: all alt
|
||||
|
||||
clean:
|
||||
${RM} -f ${OBJ} ${ALT_OBJ}
|
||||
|
||||
cpclean:
|
||||
${RM} -f cp*_*.*
|
||||
|
||||
clobber: clean cpclean
|
||||
${RM} -f ${TARGET} ${ALT_TARGET}
|
||||
${RM} -f ${INPUTS}
|
||||
${RM} -f ${OUTPUTS}
|
||||
@-if [ -e sandwich ]; then \
|
||||
${RM} -f sandwich; \
|
||||
echo 'ate sandwich'; \
|
||||
fi
|
||||
|
||||
nuke: clobber
|
||||
@${TRUE}
|
||||
|
||||
dist_clean: nuke
|
||||
@${TRUE}
|
||||
|
||||
install:
|
||||
@echo "Surely you are performing, Dr. May!"
|
||||
@${TRUE}
|
||||
|
||||
# backwards compatibility
|
||||
#
|
||||
build: all
|
||||
@${TRUE}
|
||||
|
||||
|
||||
##################
|
||||
# 133t hacker rulz
|
||||
##################
|
||||
#
|
||||
love:
|
||||
@echo 'not war?'
|
||||
@${TRUE}
|
||||
|
||||
haste:
|
||||
$(MAKE) waste
|
||||
@${TRUE}
|
||||
|
||||
waste:
|
||||
@echo 'haste'
|
||||
@${TRUE}
|
||||
|
||||
make:
|
||||
@echo 'Attend a maker faire'
|
||||
@${TRUE}
|
||||
|
||||
easter_egg:
|
||||
@echo you expected to often mis-understand this $${RANDOM} magic
|
||||
@echo chongo '<was here>' "/\\oo/\\"
|
||||
@echo Eggy
|
||||
|
||||
fabricate fashion form frame manufacture produce: make
|
||||
@${TRUE}
|
||||
|
||||
sandwich:
|
||||
@if [ `id -u` -eq 0 ]; then \
|
||||
echo 'Okay.'; \
|
||||
echo $${RANDOM}`date +%s`$${RANDOM} > $@; \
|
||||
else \
|
||||
echo 'What? Make it yourself.'; \
|
||||
exit 1; \
|
||||
fi
|
||||
|
||||
# Understand the history of the Homebrew Computer Club
|
||||
# as well as the West Coast Computer Faire and
|
||||
# you might be confused different. :-)
|
||||
#
|
||||
supernova: nuke
|
||||
@-if [ -r .code_anal ]; then \
|
||||
${RM} -f .code_anal_v6; \
|
||||
else \
|
||||
echo "planet deniers, like some members of the IAU, are so cute when they try to defend their logic"; \
|
||||
fi
|
||||
@echo A $@ helps ${MAKE} the elements that help form planets
|
||||
@${TRUE}
|
||||
|
||||
deep_magic:
|
||||
@-if [ -r .code_anal ]; then \
|
||||
ccode_analysis --deep_magic fc6b6af226b7cc598de3d48fc20ed0bd54f7b4f7f0f651a32b9cf8345b2a3b3f --FNV1a_hash_512_bit "prog"; \
|
||||
else \
|
||||
echo "Wrong! Do it again!"; \
|
||||
sleep 2; \
|
||||
echo "Wrong! Do it again!"; \
|
||||
sleep 2; \
|
||||
echo "Wrong! Do it again!"; \
|
||||
fi
|
||||
@${TRUE}
|
||||
|
||||
magic: deep_magic
|
||||
@-if [ -r .code_anal ]; then \
|
||||
ccode_analysis --level 391581 --mode 216193 --FNV1a_hash_512_bit "prog"; \
|
||||
else \
|
||||
echo "If you don't eat yer meat, you can't have any pudding!"; \
|
||||
echo "How can you have any pudding if you don't eat yer meat?!"; \
|
||||
fi
|
||||
@${TRUE}
|
||||
|
||||
# The IOCCC resident astronomer states
|
||||
#
|
||||
charon: supernova
|
||||
@echo $@ is a dwarf planet
|
||||
@echo dwarf is a type of planet
|
||||
@echo therefore IAU, $@ is a planet
|
||||
|
||||
pluto: supernova
|
||||
${MAKE} charon
|
||||
@echo $@ is a dwarf planet
|
||||
@echo dwarf is a type of planet
|
||||
@echo therefore, $@ is a planet
|
||||
@echo get used to having lots of planets because a $< can ${MAKE} a lot of them
|
||||
Binary file not shown.
Binary file not shown.
@@ -0,0 +1,659 @@
|
||||
<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01 Transitional//EN" "http://www.w3.org/TR/html4/loose.dtd">
|
||||
<html>
|
||||
<head>
|
||||
<meta name="GENERATOR" content="mkd2html 2.2.0">
|
||||
<meta http-equiv="Content-Type"
|
||||
content="text/html; charset=utf-8"> <link rel="stylesheet"
|
||||
type="text/css"
|
||||
href="../hint.css">
|
||||
<title>Most in need to be tweeted</title>
|
||||
</head>
|
||||
<body>
|
||||
<h1>Most in need to be tweeted</h1>
|
||||
|
||||
<p>Christopher Mills<br>
|
||||
Twitter: @MisterXopher</p>
|
||||
|
||||
<h2>Judges' comments:</h2>
|
||||
|
||||
<h3>To use:</h3>
|
||||
|
||||
<pre><code>make
|
||||
|
||||
make cpclean
|
||||
# Let this run for about about an hour and then kill it
|
||||
./prog Shakespeare.txt
|
||||
./prog < $(< ls -1tr cp* | tail -1) | head -100
|
||||
</code></pre>
|
||||
|
||||
<h3>Try:</h3>
|
||||
|
||||
<pre><code># If you are on a 64-bit system:
|
||||
|
||||
make test-64bit
|
||||
|
||||
less IOCCC-Rules-Guidelines.output.txt
|
||||
|
||||
less IOCCC-hints.output.txt
|
||||
|
||||
less Eugene_Onegin.output.txt
|
||||
|
||||
# However, as the binary model files used to produce the output are in an implementation-specific format,
|
||||
# your mileage may vary.
|
||||
</code></pre>
|
||||
|
||||
<h3>Selected Judges Remarks:</h3>
|
||||
|
||||
<p>Can a machine learn?</p>
|
||||
|
||||
<p>Some say so.</p>
|
||||
|
||||
<p>But can a machine learn to write like Shakespeare,</p>
|
||||
|
||||
<p>or tweet like <a href="http://trumptwitterarchive.com">Trump</a>?</p>
|
||||
|
||||
<p>Can it write rules and guidelines for the IOCCC?</p>
|
||||
|
||||
<p>You decide. :-)</p>
|
||||
|
||||
<h2>Author’s comments:</h2>
|
||||
|
||||
<h1>Welcome to OMLET! 🍳</h1>
|
||||
|
||||
<p>OMLET is the <em>Obfuscated Machine Learning Environment Toolkit</em>, a
|
||||
micro-framework for experimenting with <a href="https://en.wikipedia.org/wiki/Recurrent_neural_network">recurrent neural networks</a>. OMLET
|
||||
lets you build, train and evaluate <a href="https://en.wikipedia.org/wiki/Deep_learning">deep neural networks</a>. Why invest
|
||||
hours reading documentation and megabytes of disk space on a full-featured DNN
|
||||
framework like <a href="https://www.tensorflow.org">TensorFlow</a> or <a href="http://torch.ch">Torch</a> when you can have full RNN
|
||||
functionality in less than 4 KB!</p>
|
||||
|
||||
<p>OMLET has the following features:</p>
|
||||
|
||||
<ul>
|
||||
<li>User-programmable network configurations and hyperparameters</li>
|
||||
<li>Support for various types of recurrent and feed-forward neural networks
|
||||
including vanilla RNNs, LSTMs and GRUs of depths of up to 99 layers</li>
|
||||
<li>No limit on parameter size (except for those imposed by the system)</li>
|
||||
<li>Training and inference modes, with periodic checkpointing</li>
|
||||
<li>Advanced <a href="https://arxiv.org/abs/1412.6980">Adam</a> optimizer with <a href="https://www.fast.ai/2018/07/02/adam-weight-decay/">weight decay</a> for simplified
|
||||
training</li>
|
||||
<li>Hyperparamer support for batch sizing, learning rate schedule, weight
|
||||
decay and gradient clipping</li>
|
||||
<li>Easily extensible (requires some expertise in the C programming language)</li>
|
||||
<li>Friendly markdown documentation</li>
|
||||
</ul>
|
||||
|
||||
|
||||
<p>OMLET is based on <a href="https://cs.stanford.edu/people/karpathy/">Andrej Karpathy’s</a> <a href="https://github.com/karpathy/char-rnn">character-level language model</a>
|
||||
as described in his blog post <a href="http://karpathy.github.io/2015/05/21/rnn-effectiveness/"><em>The Unreasonable Effectiveness of Recurrent
|
||||
Neural Networks</em></a>. I’ve included a small sample dataset to use for
|
||||
training, but you can have even more fun by downloading some larger datasets:</p>
|
||||
|
||||
<ul>
|
||||
<li><a href="https://cs.stanford.edu/people/karpathy/char-rnn/linux_input.txt">Linux kernel source</a></li>
|
||||
<li><a href="https://cs.stanford.edu/people/karpathy/char-rnn/warpeace_input.txt"><em>War and Peace</em></a></li>
|
||||
<li><a href="https://github.com/ryanmcdermott/trump-speeches">Donald Trump’s 2016 campaign speeches</a></li>
|
||||
</ul>
|
||||
|
||||
|
||||
<h2>Getting started with OMLET</h2>
|
||||
|
||||
<p>OMLET has three operating modes:</p>
|
||||
|
||||
<ul>
|
||||
<li><em>Training</em>, which trains a network from scratch</li>
|
||||
<li><em>Continuation</em>, which trains starting from an existing checkpoint</li>
|
||||
<li><em>Inference</em>, which uses an existing trained network to make predictions</li>
|
||||
</ul>
|
||||
|
||||
|
||||
<p>For your first OMLET experiment, we will try training a simple single-level
|
||||
RNN to write some <a href="https://en.wikipedia.org/wiki/William_Shakespeare">Shakespere</a> plays. Start by typing</p>
|
||||
|
||||
<pre><code>make
|
||||
</code></pre>
|
||||
|
||||
<p>After it builds, train it using</p>
|
||||
|
||||
<pre><code>./prog shakespere.txt
|
||||
</code></pre>
|
||||
|
||||
<p>This will immediately start outputting gibberish to the output, e.g.</p>
|
||||
|
||||
<pre><code>./prog shakespere.txt
|
||||
sins ohennAu
|
||||
T-teooclelp tiThoWy
|
||||
|
||||
g
|
||||
nlakuafy
|
||||
e
|
||||
sselW usnsofueB Aoee pasfUsuslhe ooM ot Wou moy
|
||||
me neltAl -no IoyI mhuyakse inT-l chu ghenn ffo? fnsoe yhyye
|
||||
ue nnfrlass heUthole saounlcesyee pee
|
||||
t,
|
||||
T0:0% 3.210888
|
||||
o',,vU
|
||||
An ,hTf lnm Far rur:s moilt WoEgrv wonds mith Aog thernw
|
||||
Rni So
|
||||
co Nnd :
|
||||
For an bImy pgafoun:
|
||||
Wf'r hom wortiverita
|
||||
int fod mous Eheledet,
|
||||
Tho he theket nonS wnu-ang dorlaMSp
|
||||
nrocWiSe tflg 'o.
|
||||
T0:0% 2.995950
|
||||
d whecedhencrysesil yr bn,
|
||||
we hh y thiwt
|
||||
|
||||
|
||||
hut fithlot,
|
||||
Fmdy s he alt
|
||||
Vh th no dh foud bobt werw:s Aotnf Fhwi't whe, eusu
|
||||
|
||||
lhh thele wewcond ary soupfy wind tDont couc ths:
|
||||
er fucwald oncli hen bos, f
|
||||
T0:1% 2.878945
|
||||
</code></pre>
|
||||
|
||||
<p>The gibberish is the networks attempt to write a play. So far, it’s not very
|
||||
successful! Between the chunks of gibberish are training progress reports
|
||||
that look like</p>
|
||||
|
||||
<pre><code>T0:1% 2.878945
|
||||
</code></pre>
|
||||
|
||||
<p>This tells you that we are 1% through training epoch 0, and that the training
|
||||
<a href="https://en.wikipedia.org/wiki/Loss_function">loss</a> is about 2.88. As the training continues, the training loss will
|
||||
reduce and the generated snippets will improve quickly:</p>
|
||||
|
||||
<pre><code>ses, kuth
|
||||
LAs of the wish,
|
||||
As, I nos you,
|
||||
Yov not to nalll,
|
||||
Tr tot wonds.
|
||||
|
||||
First Sondy, llt lrte, our, tw.
|
||||
|
||||
First.
|
||||
|
||||
BRUTUS:
|
||||
Helsting the kith gops of hoch Whay, fars surd what to,
|
||||
The cownens golt te.
|
||||
T0:12% 2.259114
|
||||
eplerrotrur tandans one wiok thy or thach and cullice ded yourssting
|
||||
And wours:
|
||||
Whed ur surt.
|
||||
|
||||
SINENIUS:
|
||||
On we lain bith
|
||||
rerytund: tich lon hyivetetgor.
|
||||
|
||||
VOLUONIA:
|
||||
He brich nom dove worthan then wise,
|
||||
T0:13% 2.254926
|
||||
</code></pre>
|
||||
|
||||
<p>It’s already started to figure out things about Shakespere plays – how to
|
||||
spell short English words, how long lines tend to be, and that characters take
|
||||
turns speaking with their names capitalized. The training loss has dropped to
|
||||
2.25 and the improvement is noticeable.</p>
|
||||
|
||||
<p>Eventually, we will finish with the training data set and move on to a
|
||||
validation cycle:</p>
|
||||
|
||||
<pre><code>y fath onother,
|
||||
I sucess,
|
||||
For I me the west crare.
|
||||
|
||||
TRANIO:
|
||||
Whow and'd have not to had you you one in my lapteny she very ame come me a gut and shourd aghir you as ignested; shend to make I strem
|
||||
To h
|
||||
T0:99% 1.924063
|
||||
notnce gaud and is nicked thou day, ha the dusing you disaid: in thim, you things in ere thee thus erile Iht that tare theme my hast thesp thou shay: thou not eaten-or-ho-bess resing: I the but had d
|
||||
T0:99% 1.923128
|
||||
V0:0% 1.678907
|
||||
V0:9% 1.700527
|
||||
V0:18% 1.733179
|
||||
V0:27% 1.714891
|
||||
V0:36% 1.716672
|
||||
V0:45% 1.782946
|
||||
V0:54% 1.835629
|
||||
V0:63% 1.876108
|
||||
V0:72% 1.906814
|
||||
V0:81% 1.924096
|
||||
V0:90% 1.954492
|
||||
V0:99% 1.969287
|
||||
serfs you'll alliencseard:
|
||||
We
|
||||
got you? before
|
||||
I say.
|
||||
|
||||
Farstred dentlentecaly, sir, I it one bosticield
|
||||
All me the backnour mino,
|
||||
Whith capitaned mid! but stell the ifvemion
|
||||
Willerity.
|
||||
|
||||
First Cumfol of
|
||||
T1:0% 1.885619
|
||||
</code></pre>
|
||||
|
||||
<p>Validation cycles are used to test the network to see if it has learned to
|
||||
generalize – how well performs on data it hasn’t seen before (as opposed to
|
||||
the training data that the network will see many times as it trains). Progress
|
||||
on the validation set is also displayed with a validation progress report that
|
||||
looks like</p>
|
||||
|
||||
<pre><code>V0:36% 1.716672
|
||||
</code></pre>
|
||||
|
||||
<p>which means we are 36% of the way through the validation for epoch 0, and the
|
||||
validation loss is about 1.72. Comparing the validation loss and the
|
||||
training loss will give you an idea of how well the network is learning and
|
||||
can let you know if the network is <a href="https://en.wikipedia.org/wiki/Overfitting">overfitting or underfitting</a>.</p>
|
||||
|
||||
<p>As part of the training process, the data set (which for OMLET is the
|
||||
<code>shakespere.txt</code> file you gave on the command line) is divided into
|
||||
<a href="https://en.wikipedia.org/wiki/Training,_validation,_and_test_sets">training and validation sets</a> (by default, 95% of the data is used for
|
||||
training, but like most OMLET parameters, you can change this at compile-time).</p>
|
||||
|
||||
<p>At the end of the validation run, OMLET writes out a checkpoint file with a
|
||||
name like <code>cp01_1.970</code>. This saves the state of the run at the start of the
|
||||
epoch 1, after computing a validation loss of 1.970. The checkpoint is
|
||||
helpful if you need to stop and restart training. You can stop training by
|
||||
doing <code>Control-C</code>.</p>
|
||||
|
||||
<p>You can continue training from a previous checkpoint by providing the
|
||||
checkpoint file name as the second parameter, for example:</p>
|
||||
|
||||
<pre><code>./prog shakespere.txt cp01_1.970
|
||||
</code></pre>
|
||||
|
||||
<p>After the validation cycle finishes, OMLET begins the next epoch by restarting
|
||||
training at the beginning of the training set. Training continues forever,
|
||||
until you quit it with <code>Control-C</code>. You should monitor the checkpoints to
|
||||
see that the validation loss continues to drop. If it rises, the network has
|
||||
probably started to overfit on the training data.</p>
|
||||
|
||||
<p>Once you’ve trained the validation loss as low as it will go, you can use
|
||||
OMLET to run the network in inference mode which uses the frozen checkpoint
|
||||
parameters to generate data. Inference mode takes the checkpoint file as
|
||||
<em>standard input</em> (not on the command line) and hence must be run with a
|
||||
command like</p>
|
||||
|
||||
<pre><code>./prog < cp55_1.807
|
||||
</code></pre>
|
||||
|
||||
<p>Running it produces an infinite amount of generated output, until you hit
|
||||
<code>Control-C</code> to stop it.</p>
|
||||
|
||||
<p>Note that if you decide to change networks or use a different input file, you
|
||||
will want to delete all the checkpoint files because the format depends on
|
||||
both the network and the input – using the wrong checkpoint is likely to
|
||||
cause a crash.</p>
|
||||
|
||||
<h2>Experimenting with different networks</h2>
|
||||
|
||||
<p>The default network for OMLET (the one you get if you type <code>make</code> with the
|
||||
executable name <code>prog</code>) is the simplest recurrent neural network. It looks
|
||||
like</p>
|
||||
|
||||
<pre><code>h = tanh(Wxh * x + Whh * h' + Bh)
|
||||
y = Why * h + By
|
||||
</code></pre>
|
||||
|
||||
<p>where</p>
|
||||
|
||||
<ul>
|
||||
<li><code>x</code> is the <em>input vector</em></li>
|
||||
<li><code>y</code> is the <em>output vector</em></li>
|
||||
<li><code>h</code> is the <em>hidden state vector</em></li>
|
||||
<li><code>h'</code> is the previous value of <code>h</code></li>
|
||||
<li><code>Wxh</code>, <code>Whh</code>, and <code>Why</code> are <em>weight matrices</em></li>
|
||||
<li><code>Bh</code> and <code>By</code> are <em>bias vectors</em>.</li>
|
||||
<li><code>tanh</code> is the <a href="https://en.wikipedia.org/wiki/Hyperbolic_function">hyperbolic tangent function</a></li>
|
||||
</ul>
|
||||
|
||||
|
||||
<p>The <code>W</code>’s and <code>B</code>’s are the trainable parameters of the network, and the
|
||||
process of training is optimizing the values of these parameters to minimize
|
||||
the loss of the network across the training set.</p>
|
||||
|
||||
<p>It is the presence of the hidden state vector that allows the RNN to
|
||||
“remember” the past. We can see what would happen if we removed this hidden
|
||||
state. If you type</p>
|
||||
|
||||
<pre><code>make lin1
|
||||
</code></pre>
|
||||
|
||||
<p>OMLET will create an <a href="https://en.wikipedia.org/wiki/ADALINE">ADALINE</a> network that does</p>
|
||||
|
||||
<pre><code>y = Wxy * x + By
|
||||
</code></pre>
|
||||
|
||||
<p>This simple <a href="https://en.wikipedia.org/wiki/Linear_map">linear</a> <a href="https://en.wikipedia.org/wiki/Feedforward_neural_network">feed-forward network</a>. You can run it with</p>
|
||||
|
||||
<pre><code>./lin1 shakespeare.txt
|
||||
</code></pre>
|
||||
|
||||
<p>The linear network won’t be able to get past the gibberish stage, because it
|
||||
lacks history:</p>
|
||||
|
||||
<pre><code>./lin1 shakespeare.txt
|
||||
UERond w,
|
||||
|
||||
|
||||
Gir:
|
||||
KINof s, mesther s thouth.
|
||||
E:
|
||||
KINTret, at fu,
|
||||
GOMy t, as sth kesewit sooos atse ang k, ck,
|
||||
Sotheouserivesthecowhet been's, t he, h nre; t and, har wiread of pincer cedst sur has, ut:
|
||||
|
||||
T14:67% 2.465115
|
||||
UKESpan,
|
||||
NGaromy soreate e m esewfoure pamitherarjulthengeoly tl.
|
||||
NG s at e! w.
|
||||
|
||||
WAllinoully?
|
||||
Wamisw ofilem:
|
||||
I'delandinarrstath har aksubly s cath Whern t Is, weciss:
|
||||
|
||||
GLat s; llde.
|
||||
Y aterit dsthence
|
||||
T14:67% 2.465404
|
||||
</code></pre>
|
||||
|
||||
<p>It is able to guess at what character is likely to follow the current one
|
||||
(by doing a <a href="https://en.wikipedia.org/wiki/Linear_regression">linear regression</a>), but it lacks any history beyond that to
|
||||
guide it.</p>
|
||||
|
||||
<p>You might be wondering about the role of the <code>tanh</code> function in the RNN.
|
||||
<code>tanh</code> acts as an <a href="https://en.wikipedia.org/wiki/Activation_function">activation function</a> which adds <a href="https://en.wikipedia.org/wiki/Nonlinear_system">nonlinearity</a>
|
||||
to the network and allows it to solve complicated problems. Without
|
||||
nonlinearity, all of the linear functions would fold together into a single
|
||||
matrix-vector multiply and you’d effectively regress to the linear network
|
||||
above. Alas, even adding a nonlinearity to the feed-forward network
|
||||
(creating a <a href="https://en.wikipedia.org/wiki/Perceptron">perceptron</a>) does not improve the performance because we
|
||||
still lack the history provided by the hidden state vector (although if you
|
||||
want to try it yourself, you can do so with <code>make per1</code>).</p>
|
||||
|
||||
<h2>Going deeper</h2>
|
||||
|
||||
<p>We can try to improve the RNN’s performance by stacking RNN modules atop each
|
||||
other:</p>
|
||||
|
||||
<pre><code>h1 = RNN(h1', x)
|
||||
h2 = RNN(h2', h1)
|
||||
y = Why * h2 + By
|
||||
</code></pre>
|
||||
|
||||
<p>with <code>RNN(h, x)</code> defined as above. Each RNN module has its own set of
|
||||
parameters and its own hidden state vector. This will improve the network’s
|
||||
performance, at the cost of a much larger parameter space</p>
|
||||
|
||||
<blockquote><p><strong>IMPORTANT NOTE</strong>: Since OMLET uses the system stack for network storage,
|
||||
larger networks may cause OMLET to crash (typically with a message like
|
||||
<code>Segmentation fault</code>) unless the system stack size is first increased.
|
||||
The exact command for doing so depends on your shell and your system’s
|
||||
hard limits. On <code>sh</code>/<code>ksh</code>/<code>bash</code> shells, you can view the hard limit
|
||||
with <code>ulimit -Hs</code> and set it with <code>ulimit -s 65532</code> (replacing <code>65532</code>
|
||||
with the actual hard limit). On <code>csh</code>/<code>tcsh</code> shells, you can view
|
||||
the hard limit with <code>limit -h stacksize</code> and set it with
|
||||
<code>limit stacksize 65532</code> (replacing <code>65532</code> with the actual hard limit).</p></blockquote>
|
||||
|
||||
<p>You can try the deeper network by doing</p>
|
||||
|
||||
<pre><code>make rnn2
|
||||
</code></pre>
|
||||
|
||||
<p>(or even <code>make rnn3</code> if you want a three-layer RNN) and train it with</p>
|
||||
|
||||
<pre><code>./rnn2 shakespeare.txt
|
||||
</code></pre>
|
||||
|
||||
<p>The additional depth should allow the network to make better predictions (it
|
||||
can represent more complicated history), but it may take a long time to train
|
||||
– both because the network (being larger) now requires more time to train and
|
||||
because of the <a href="https://en.wikipedia.org/wiki/Vanishing_gradient_problem">vanishing and exploding gradient problem</a>, which might
|
||||
keep it from ever reaching its potential.</p>
|
||||
|
||||
<h2>LSTMs and GRUs</h2>
|
||||
|
||||
<p>RNNs are particularly hard to train because the they are trained using
|
||||
<a href="https://en.wikipedia.org/wiki/Backpropagation_through_time">bankpropagation through time</a>. The RNN is trained by effectively
|
||||
converting it into a non-recurrent network by making many copies of it
|
||||
and propagating the hidden state through the copies. During training,
|
||||
the backpropagation through many clones of the network amplifies the
|
||||
gradient, worsening the exploding and vanishing gradient problem.</p>
|
||||
|
||||
<p><a href="https://en.wikipedia.org/wiki/Long_short-term_memory">Long Short Term Memory</a> networks (also called LSTMs) were developed to
|
||||
solve this problem. Christopher Olah gives a good description of them
|
||||
at his <a href="http://colah.github.io/posts/2015-08-Understanding-LSTMs/">blog posting</a>. You can build a two-level LSTM by doing</p>
|
||||
|
||||
<pre><code>make lstm2
|
||||
</code></pre>
|
||||
|
||||
<p>and train with it with</p>
|
||||
|
||||
<pre><code>./lstm2 shakespeare.txt
|
||||
</code></pre>
|
||||
|
||||
<p>The LSTM is much easier to train because it explicitly decides how to update
|
||||
its hidden state via “gates”. These gates are called</p>
|
||||
|
||||
<ul>
|
||||
<li><em>input gate</em>, which decides what part of the input to pay attention to</li>
|
||||
<li><em>forget gate</em>, which decides what part of the hidden state to forget</li>
|
||||
<li><em>output gate</em>, which decides what part of the hidden state is used to
|
||||
produce the output</li>
|
||||
</ul>
|
||||
|
||||
|
||||
<p>The basic LSTM equations are</p>
|
||||
|
||||
<pre><code>f = sigmoid(Wxf * x + Whf * h' + Bf)
|
||||
i = sigmoid(Wxi * x + Whi * h' + Bi)
|
||||
o = sigmoid(Wxo * x + Who * h' + Bo)
|
||||
c = f * c' + i * tanh(Wxc * x + Whc * h' + Bc)
|
||||
h = o * tanh(c)
|
||||
</code></pre>
|
||||
|
||||
<p>Where</p>
|
||||
|
||||
<ul>
|
||||
<li><code>x</code> is the input vector</li>
|
||||
<li><code>h</code> is the hidden state (and the output to the next layer)</li>
|
||||
<li><code>c</code> is the <em>cell state</em> which represents the “memory” of the LSTM</li>
|
||||
<li><code>h'</code> and <code>c</code>‘ are the previous values of <code>h</code> and <code>c</code> respectively</li>
|
||||
<li><code>f</code> is the <em>forget gate</em> that tells the LSTM what portion of the hidden
|
||||
state to forget</li>
|
||||
<li><code>i</code> is the <em>input gate</em> that tells the LSTM what portion of the input
|
||||
vector to pay attention to</li>
|
||||
<li><code>o</code> is the <em>output gate</em> that tells the LSTM what portion of the cell
|
||||
state to use to generate the hidden state</li>
|
||||
<li><code>Wxf</code>, <code>Whf</code>, <code>Wxi</code>, <code>Whi</code>, <code>Wxo</code>, <code>Who</code>, <code>Wxc</code>, and <code>Whc</code> are trainable
|
||||
parameter matrices</li>
|
||||
<li><code>Bf</code>, <code>Bi</code>, <code>Bo</code>, and <code>Bc</code> are trainable bias vectors</li>
|
||||
<li><code>tanh</code> is the hyperbolic tangent function</li>
|
||||
<li><code>sigmoid</code> is the <a href="https://en.wikipedia.org/wiki/Logistic_function">logistic function</a></li>
|
||||
</ul>
|
||||
|
||||
|
||||
<p>There are several LSTM variants (see <a href="http://colah.github.io/posts/2015-08-Understanding-LSTMs/">C. Olah’s blog post</a> for more
|
||||
examples). One important one is the <a href="https://en.wikipedia.org/wiki/Gated_recurrent_unit">gated recurrent unit</a>. GRUs are
|
||||
simplified versions of an LSTM which combine the gates together, meaning they
|
||||
require fewer learned parameters. This allows them to train faster than a
|
||||
generic LSTM. You can build a two-layer GRU with</p>
|
||||
|
||||
<pre><code>make gru2
|
||||
</code></pre>
|
||||
|
||||
<h2>Building your own networks</h2>
|
||||
|
||||
<p>The OMLET <code>Makefile</code> comes with one-, two- and three-layer RNNs, LSTMs and
|
||||
GRUs, along with simpler feed-forward networks like multi-layer perceptrons and
|
||||
a linear network. This isn’t the limit of OMLET’s power – you can create your
|
||||
own networks by modifying the <code>Makefile</code>. Networks are passed in on the
|
||||
compiler’s command-line by using <code>-D</code> directives. The network is defined by
|
||||
a <code>-DNW='...'</code> command which consists of a series of comma-separated
|
||||
assignments. For example, the simple one-layer RNN could be defined like</p>
|
||||
|
||||
<pre><code>-DNW=` x = I(n), hp = I(128), \
|
||||
h = C(hp, T(A(L(128, x), L(128, hp)))), \
|
||||
y = L(n, h)'
|
||||
</code></pre>
|
||||
|
||||
<p>The network declares <code>x</code> as an input vector (there must be a declaration for
|
||||
<code>x</code>). It is declared as <code>I(n)</code>, which is an input vector of size <code>n</code>, which
|
||||
is the number of characters of the input alphabet (OMLET computes this from
|
||||
the input file at the start of training). OMLET will arrange to present the
|
||||
input character as a <a href="https://en.wikipedia.org/wiki/One-hot">one-hot</a> vector based on the current input
|
||||
character.</p>
|
||||
|
||||
<p>The second declaration, <code>hp</code>, declares the previous hidden state vector (what
|
||||
we called <code>h'</code> above). We declare this to be of size 128 – an arbitrary
|
||||
choice. A larger state vector can (theoretically) carry more state, but at
|
||||
a cost of larger parameter matrices and longer training time. You can
|
||||
experiment with increasing the hidden vector size and see.</p>
|
||||
|
||||
<p>The third line is the core of the RNN. It sets <code>h</code>, the hidden vector output
|
||||
to be the sum of two linear elements specified by <code>L</code>. The <code>L</code> function
|
||||
takes two parameters – the output vector size (which must match the size of
|
||||
<code>h</code>) and the input vector. <code>L</code> will compute <code>y = W * x + B</code> where each <code>L</code>
|
||||
has its own <code>W</code> (weight) and <code>B</code> (bias) training parameters. Both <code>x</code> and
|
||||
<code>hp</code> are sent through <code>L</code> and the result passed through the <code>A</code> function,
|
||||
which does vector addition. That result is passed through <code>T</code> which does
|
||||
element-wise <code>tanh</code> activation.</p>
|
||||
|
||||
<p>Next, we wrap the whole thing with the <code>C</code> function. <code>C</code> connects <code>hp</code>
|
||||
with <code>h</code>, causing the new value of <code>h</code> to be passed to the <code>hp</code> vector on
|
||||
the next iteration of the algorithm (allowing the RNN to retain state in <code>h</code>).</p>
|
||||
|
||||
<p>Finally, the whole result is passed through another instance of <code>L</code>, this time
|
||||
producing a vector of size <code>n</code>, which will have the negative log
|
||||
<a href="https://en.wikipedia.org/wiki/Likelihood_function">likelihood</a> function. This is assigned to <code>y</code>, which is the output of
|
||||
the network (and hence must also be declared).</p>
|
||||
|
||||
<p>OMLET will take the <code>y</code> result and pass it through the <a href="https://en.wikipedia.org/wiki/Softmax_function">softmax</a> function,
|
||||
which converts the log probabilities into a probability distribution. In
|
||||
inference mode, this is used to select the next character to emit. In
|
||||
training mode, this is used to generate the loss which is backpropagated.</p>
|
||||
|
||||
<p>As an example of a more complicated network, we can look at a two-layer
|
||||
GRU network:</p>
|
||||
|
||||
<pre><code>-DHS=128, \
|
||||
-DNW=' x = I(n), \
|
||||
y = L(n, MD(MD(x)))' \
|
||||
-DBK=' hp = I(HS), \
|
||||
z = S(A(L(HS, x), L(HS, hp))), \
|
||||
r = S(A(L(HS, x), L(HS, hp))), \
|
||||
c = T(A(L(HS, x), L(HS, hp))), \
|
||||
zc = OG(1, -1, z), \
|
||||
h = C(hp, A(M(zc, hp), M(z, c))), \
|
||||
y = h'
|
||||
</code></pre>
|
||||
|
||||
<p>We are using a few new tricks here – first, we are defining <code>HS</code> as the size
|
||||
of the hidden and cell vectors. There’s nothing special about this name, its
|
||||
just convenient to specify it so we don’t have a bunch of constants in the
|
||||
code. Second, the network itself is very simple – it declares <code>x</code> and has
|
||||
the matrix that converts the <code>HS</code>-sized hidden vector back to the <code>n</code>-sized
|
||||
alphabet vector… but it now calls <code>MD</code>, which is the user-definable
|
||||
module (here we are using it twice, to have two cascaded GRU blocks). The
|
||||
<code>MD</code> function performs the sub-network defined by the <code>BK</code> compile-time
|
||||
parameter (specified in the <code>-DBK='...'</code> setting). This sub-module again
|
||||
takes an <code>x</code> parameter and produces a <code>y</code> output. Inside it, we declare
|
||||
<code>hp</code> and <code>h</code>, the previous and current state vector, plus equations for the
|
||||
various GRU gates (these use <code>S</code> for the sigmoid activation function).
|
||||
One final new call is <code>OG</code> which does offset and gain, performing
|
||||
<code>y = offset + gain * x</code> where <code>offset</code> (the first parameter) and <code>gain</code> (the
|
||||
second) are constants. We are using this here to compute <code>(1 - z)</code> for the
|
||||
GRU’s linear interpolator.</p>
|
||||
|
||||
<p>The full set of available function blocks follows:</p>
|
||||
|
||||
<ul>
|
||||
<li><code>I(s)</code>: declares a vector (input or state) of size <code>s</code></li>
|
||||
<li><code>L(s, x)</code>: learnable linear function <code>y = W * x + B</code> with an output vector
|
||||
size of <code>s</code></li>
|
||||
<li><code>CM(x)</code>: learnable element-wise gain function <code>y = W * x</code></li>
|
||||
<li><code>A(a, b)</code>: element-wise add: <code>y = a + b</code></li>
|
||||
<li><code>M(a, b)</code>: element-wise multiply: <code>y = a * b</code></li>
|
||||
<li><code>S(x)</code>: sigmoid activation function <code>y = sigmoid(x)</code></li>
|
||||
<li><code>T(x)</code>: hyperbolic tangent activation function <code>y = tanh(x)</code></li>
|
||||
<li> <code>C(xp, x)</code>: copy <code>x</code> to <code>xp</code> in the next time step (propagate through
|
||||
time)</li>
|
||||
<li> <code>OG(o, g, x)</code>: apply a constant offset and gain: <code>y = o + g * x</code></li>
|
||||
<li> <code>MD(x)</code>: apply the sub-network specified by <code>BK</code></li>
|
||||
</ul>
|
||||
|
||||
|
||||
<p>Note: even if you don’t use <code>MD</code> in your network, you should still define
|
||||
<code>BK</code> by adding <code>-DBK='y=x'</code> to the command line, otherwise you will get a
|
||||
compile-time error.</p>
|
||||
|
||||
<h2>Hyperparameters</h2>
|
||||
|
||||
<p>OMLET has a large number of training and inference parameters which can be
|
||||
changed by the user. All of these are set by <code>-D</code> on the compile command line.
|
||||
The list of hyperparameters follows:</p>
|
||||
|
||||
<ul>
|
||||
<li><code>TP</code>: Temperature parameter for use in inference mode. This divides the
|
||||
log probabilities before softmax. A low temperature makes the model choose
|
||||
safer but more boring choices. A high temperature takes more risks but
|
||||
makes more mistakes. Default is 1.0, which uses the computed
|
||||
probabilities.</li>
|
||||
<li>‘N’: Batch size. This is the number of times the RNN is unrolled, so it
|
||||
controls how far back in the past the RNN can see. The default is 50.
|
||||
Larger batch sizes update the weights less frequently and allow the RNN to
|
||||
see farther back in time, but at a cost of proportionally more memory.</li>
|
||||
<li><code>TR</code>: The percentage of batches in the input data set that will be used
|
||||
for training. The default value of 0.95 sets this as 95%.</li>
|
||||
<li><code>LR</code>: The initial learning rate, The default is 0.002.</li>
|
||||
<li><code>LE</code>: The epoch where the learning rate will start decaying. Defaults to
|
||||
epoch 10.</li>
|
||||
<li><code>LD</code>: Learning rate decay, per epoch (after <code>LE</code> epochs). The learning
|
||||
rate is scaled by this number. Default is 0.97</li>
|
||||
<li><code>WD</code>: Weight-decay parameter, to promote <a href="https://en.wikipedia.org/wiki/Regularization_(mathematics)">regularizaiton</a>. The default
|
||||
is 0.00008.</li>
|
||||
<li><code>RS</code>: The random scale for weight initialization. Weight parameters will
|
||||
be initialized to be between <code>-RS</code> to <code>+RS</code>. The default is 0.15.</li>
|
||||
<li><code>CL</code>: Clamp value for gradients. Gradients will be limited to the range
|
||||
<code>-CL</code> to <code>+CL</code>. Default is 5.</li>
|
||||
<li><code>B1</code>: Momentum mean parameter for Adam optimizer. Set to 0.9.</li>
|
||||
<li><code>B2</code>: Momentum variance parameter for Adam optimizer, Set to 0.999.</li>
|
||||
<li><code>EP</code>: Epsilon parameter for Adam, to provide numerical stability. Set to
|
||||
0.00000001.</li>
|
||||
<li><code>DI</code>: How often to print a training or validation progress message and
|
||||
inference snippet. The default prints every 100 training batches.</li>
|
||||
<li><code>SL</code>: The number of characters to print when doing an inference snippet.
|
||||
Default is 200.</li>
|
||||
<li><code>PF</code>: Format string for the checkpoint filename. The default is
|
||||
<code>"cp%02d_%.3f"</code> which includes the epoch and validation loss. You may
|
||||
wish to add a subdirectory to the name to keep checkpoint files out of
|
||||
the current directory.</li>
|
||||
</ul>
|
||||
|
||||
|
||||
<hr />
|
||||
|
||||
<!--
|
||||
(c) Copyright 1984-2019, [Leo Broukhis, Simon Cooper, Landon Curt Noll][judges] - All rights reserved
|
||||
This work is licensed under a [Creative Commons Attribution-ShareAlike 3.0 Unported License][cc].
|
||||
|
||||
[judges]: http://www.ioccc.org/judges.html
|
||||
[cc]: http://creativecommons.org/licenses/by-sa/3.0/
|
||||
-->
|
||||
|
||||
<TABLE><TR>
|
||||
<TD><a rel="license" href="https://creativecommons.org/licenses/by-sa/3.0/"><img alt="Creative Commons License" style="border-width:0" src="/png/by-sa-3.0-88x31.png" /></a></TD>
|
||||
<TD><P>© Copyright 1984-2019,
|
||||
<A HREF="/judges.html">Leo Broukhis, Simon Cooper, Landon Curt Noll</A>
|
||||
- All rights reserved<BR>
|
||||
This work is licensed under a <a rel="license" href="https://creativecommons.org/licenses/by-sa/3.0/">Creative Commons Attribution-ShareAlike 3.0 Unported License</a>.</P></TD>
|
||||
<TD> <!--<a href="https://validator.w3.org/check?uri=referer"><img src="https://www.w3.org/Icons/valid-html401" alt="Valid HTML 4.01 Transitional" height="31" width="88"></a>--></TD>
|
||||
</TR></TABLE>
|
||||
</body>
|
||||
</html>
|
||||
@@ -0,0 +1,636 @@
|
||||
# Most in need to be tweeted
|
||||
|
||||
Christopher Mills
|
||||
Twitter: @MisterXopher
|
||||
|
||||
|
||||
## Judges' comments:
|
||||
### To use:
|
||||
|
||||
make
|
||||
|
||||
make cpclean
|
||||
# Let this run for about about an hour and then kill it
|
||||
./prog Shakespeare.txt
|
||||
./prog < $(< ls -1tr cp* | tail -1) | head -100
|
||||
|
||||
### Try:
|
||||
|
||||
# If you are on a 64-bit system:
|
||||
|
||||
make test-64bit
|
||||
|
||||
less IOCCC-Rules-Guidelines.output.txt
|
||||
|
||||
less IOCCC-hints.output.txt
|
||||
|
||||
less Eugene_Onegin.output.txt
|
||||
|
||||
# However, as the binary model files used to produce the output are in an implementation-specific format,
|
||||
# your mileage may vary.
|
||||
|
||||
### Selected Judges Remarks:
|
||||
|
||||
Can a machine learn?
|
||||
|
||||
Some say so.
|
||||
|
||||
But can a machine learn to write like Shakespeare,
|
||||
|
||||
or tweet like [Trump](http://trumptwitterarchive.com)?
|
||||
|
||||
Can it write rules and guidelines for the IOCCC?
|
||||
|
||||
You decide. :-)
|
||||
|
||||
|
||||
## Author's comments:
|
||||
|
||||
Welcome to OMLET! 🍳
|
||||
=================
|
||||
|
||||
OMLET is the _Obfuscated Machine Learning Environment Toolkit_, a
|
||||
micro-framework for experimenting with [recurrent neural networks][1]. OMLET
|
||||
lets you build, train and evaluate [deep neural networks][2]. Why invest
|
||||
hours reading documentation and megabytes of disk space on a full-featured DNN
|
||||
framework like [TensorFlow][3] or [Torch][4] when you can have full RNN
|
||||
functionality in less than 4 KB!
|
||||
|
||||
OMLET has the following features:
|
||||
|
||||
+ User-programmable network configurations and hyperparameters
|
||||
+ Support for various types of recurrent and feed-forward neural networks
|
||||
including vanilla RNNs, LSTMs and GRUs of depths of up to 99 layers
|
||||
+ No limit on parameter size (except for those imposed by the system)
|
||||
+ Training and inference modes, with periodic checkpointing
|
||||
+ Advanced [Adam][42] optimizer with [weight decay][43] for simplified
|
||||
training
|
||||
+ Hyperparamer support for batch sizing, learning rate schedule, weight
|
||||
decay and gradient clipping
|
||||
+ Easily extensible (requires some expertise in the C programming language)
|
||||
+ Friendly markdown documentation
|
||||
|
||||
OMLET is based on [Andrej Karpathy's][40] [character-level language model][41]
|
||||
as described in his blog post [_The Unreasonable Effectiveness of Recurrent
|
||||
Neural Networks_][44]. I've included a small sample dataset to use for
|
||||
training, but you can have even more fun by downloading some larger datasets:
|
||||
|
||||
+ [Linux kernel source][30]
|
||||
+ [_War and Peace_][31]
|
||||
+ [Donald Trump's 2016 campaign speeches][32]
|
||||
|
||||
|
||||
Getting started with OMLET
|
||||
--------------------------
|
||||
|
||||
OMLET has three operating modes:
|
||||
|
||||
+ _Training_, which trains a network from scratch
|
||||
+ _Continuation_, which trains starting from an existing checkpoint
|
||||
+ _Inference_, which uses an existing trained network to make predictions
|
||||
|
||||
For your first OMLET experiment, we will try training a simple single-level
|
||||
RNN to write some [Shakespere][5] plays. Start by typing
|
||||
|
||||
make
|
||||
|
||||
After it builds, train it using
|
||||
|
||||
./prog shakespere.txt
|
||||
|
||||
This will immediately start outputting gibberish to the output, e.g.
|
||||
|
||||
./prog shakespere.txt
|
||||
sins ohennAu
|
||||
T-teooclelp tiThoWy
|
||||
|
||||
g
|
||||
nlakuafy
|
||||
e
|
||||
sselW usnsofueB Aoee pasfUsuslhe ooM ot Wou moy
|
||||
me neltAl -no IoyI mhuyakse inT-l chu ghenn ffo? fnsoe yhyye
|
||||
ue nnfrlass heUthole saounlcesyee pee
|
||||
t,
|
||||
T0:0% 3.210888
|
||||
o',,vU
|
||||
An ,hTf lnm Far rur:s moilt WoEgrv wonds mith Aog thernw
|
||||
Rni So
|
||||
co Nnd :
|
||||
For an bImy pgafoun:
|
||||
Wf'r hom wortiverita
|
||||
int fod mous Eheledet,
|
||||
Tho he theket nonS wnu-ang dorlaMSp
|
||||
nrocWiSe tflg 'o.
|
||||
T0:0% 2.995950
|
||||
d whecedhencrysesil yr bn,
|
||||
we hh y thiwt
|
||||
|
||||
|
||||
hut fithlot,
|
||||
Fmdy s he alt
|
||||
Vh th no dh foud bobt werw:s Aotnf Fhwi't whe, eusu
|
||||
|
||||
lhh thele wewcond ary soupfy wind tDont couc ths:
|
||||
er fucwald oncli hen bos, f
|
||||
T0:1% 2.878945
|
||||
|
||||
The gibberish is the networks attempt to write a play. So far, it's not very
|
||||
successful! Between the chunks of gibberish are training progress reports
|
||||
that look like
|
||||
|
||||
T0:1% 2.878945
|
||||
|
||||
This tells you that we are 1% through training epoch 0, and that the training
|
||||
[loss][6] is about 2.88. As the training continues, the training loss will
|
||||
reduce and the generated snippets will improve quickly:
|
||||
|
||||
ses, kuth
|
||||
LAs of the wish,
|
||||
As, I nos you,
|
||||
Yov not to nalll,
|
||||
Tr tot wonds.
|
||||
|
||||
First Sondy, llt lrte, our, tw.
|
||||
|
||||
First.
|
||||
|
||||
BRUTUS:
|
||||
Helsting the kith gops of hoch Whay, fars surd what to,
|
||||
The cownens golt te.
|
||||
T0:12% 2.259114
|
||||
eplerrotrur tandans one wiok thy or thach and cullice ded yourssting
|
||||
And wours:
|
||||
Whed ur surt.
|
||||
|
||||
SINENIUS:
|
||||
On we lain bith
|
||||
rerytund: tich lon hyivetetgor.
|
||||
|
||||
VOLUONIA:
|
||||
He brich nom dove worthan then wise,
|
||||
T0:13% 2.254926
|
||||
|
||||
It's already started to figure out things about Shakespere plays -- how to
|
||||
spell short English words, how long lines tend to be, and that characters take
|
||||
turns speaking with their names capitalized. The training loss has dropped to
|
||||
2.25 and the improvement is noticeable.
|
||||
|
||||
Eventually, we will finish with the training data set and move on to a
|
||||
validation cycle:
|
||||
|
||||
y fath onother,
|
||||
I sucess,
|
||||
For I me the west crare.
|
||||
|
||||
TRANIO:
|
||||
Whow and'd have not to had you you one in my lapteny she very ame come me a gut and shourd aghir you as ignested; shend to make I strem
|
||||
To h
|
||||
T0:99% 1.924063
|
||||
notnce gaud and is nicked thou day, ha the dusing you disaid: in thim, you things in ere thee thus erile Iht that tare theme my hast thesp thou shay: thou not eaten-or-ho-bess resing: I the but had d
|
||||
T0:99% 1.923128
|
||||
V0:0% 1.678907
|
||||
V0:9% 1.700527
|
||||
V0:18% 1.733179
|
||||
V0:27% 1.714891
|
||||
V0:36% 1.716672
|
||||
V0:45% 1.782946
|
||||
V0:54% 1.835629
|
||||
V0:63% 1.876108
|
||||
V0:72% 1.906814
|
||||
V0:81% 1.924096
|
||||
V0:90% 1.954492
|
||||
V0:99% 1.969287
|
||||
serfs you'll alliencseard:
|
||||
We
|
||||
got you? before
|
||||
I say.
|
||||
|
||||
Farstred dentlentecaly, sir, I it one bosticield
|
||||
All me the backnour mino,
|
||||
Whith capitaned mid! but stell the ifvemion
|
||||
Willerity.
|
||||
|
||||
First Cumfol of
|
||||
T1:0% 1.885619
|
||||
|
||||
Validation cycles are used to test the network to see if it has learned to
|
||||
generalize -- how well performs on data it hasn't seen before (as opposed to
|
||||
the training data that the network will see many times as it trains). Progress
|
||||
on the validation set is also displayed with a validation progress report that
|
||||
looks like
|
||||
|
||||
V0:36% 1.716672
|
||||
|
||||
which means we are 36% of the way through the validation for epoch 0, and the
|
||||
validation loss is about 1.72. Comparing the validation loss and the
|
||||
training loss will give you an idea of how well the network is learning and
|
||||
can let you know if the network is [overfitting or underfitting][7].
|
||||
|
||||
As part of the training process, the data set (which for OMLET is the
|
||||
`shakespere.txt` file you gave on the command line) is divided into
|
||||
[training and validation sets][8] (by default, 95% of the data is used for
|
||||
training, but like most OMLET parameters, you can change this at compile-time).
|
||||
|
||||
At the end of the validation run, OMLET writes out a checkpoint file with a
|
||||
name like `cp01_1.970`. This saves the state of the run at the start of the
|
||||
epoch 1, after computing a validation loss of 1.970. The checkpoint is
|
||||
helpful if you need to stop and restart training. You can stop training by
|
||||
doing `Control-C`.
|
||||
|
||||
You can continue training from a previous checkpoint by providing the
|
||||
checkpoint file name as the second parameter, for example:
|
||||
|
||||
./prog shakespere.txt cp01_1.970
|
||||
|
||||
After the validation cycle finishes, OMLET begins the next epoch by restarting
|
||||
training at the beginning of the training set. Training continues forever,
|
||||
until you quit it with `Control-C`. You should monitor the checkpoints to
|
||||
see that the validation loss continues to drop. If it rises, the network has
|
||||
probably started to overfit on the training data.
|
||||
|
||||
Once you've trained the validation loss as low as it will go, you can use
|
||||
OMLET to run the network in inference mode which uses the frozen checkpoint
|
||||
parameters to generate data. Inference mode takes the checkpoint file as
|
||||
*standard input* (not on the command line) and hence must be run with a
|
||||
command like
|
||||
|
||||
./prog < cp55_1.807
|
||||
|
||||
Running it produces an infinite amount of generated output, until you hit
|
||||
`Control-C` to stop it.
|
||||
|
||||
Note that if you decide to change networks or use a different input file, you
|
||||
will want to delete all the checkpoint files because the format depends on
|
||||
both the network and the input -- using the wrong checkpoint is likely to
|
||||
cause a crash.
|
||||
|
||||
Experimenting with different networks
|
||||
-------------------------------------
|
||||
|
||||
The default network for OMLET (the one you get if you type `make` with the
|
||||
executable name `prog`) is the simplest recurrent neural network. It looks
|
||||
like
|
||||
|
||||
h = tanh(Wxh * x + Whh * h' + Bh)
|
||||
y = Why * h + By
|
||||
|
||||
where
|
||||
|
||||
* `x` is the *input vector*
|
||||
* `y` is the *output vector*
|
||||
* `h` is the *hidden state vector*
|
||||
* `h'` is the previous value of `h`
|
||||
* `Wxh`, `Whh`, and `Why` are *weight matrices*
|
||||
* `Bh` and `By` are *bias vectors*.
|
||||
* `tanh` is the [hyperbolic tangent function][9]
|
||||
|
||||
The `W`'s and `B`'s are the trainable parameters of the network, and the
|
||||
process of training is optimizing the values of these parameters to minimize
|
||||
the loss of the network across the training set.
|
||||
|
||||
It is the presence of the hidden state vector that allows the RNN to
|
||||
"remember" the past. We can see what would happen if we removed this hidden
|
||||
state. If you type
|
||||
|
||||
make lin1
|
||||
|
||||
OMLET will create an [ADALINE][10] network that does
|
||||
|
||||
y = Wxy * x + By
|
||||
|
||||
This simple [linear][11] [feed-forward network][12]. You can run it with
|
||||
|
||||
./lin1 shakespeare.txt
|
||||
|
||||
The linear network won't be able to get past the gibberish stage, because it
|
||||
lacks history:
|
||||
|
||||
./lin1 shakespeare.txt
|
||||
UERond w,
|
||||
|
||||
|
||||
Gir:
|
||||
KINof s, mesther s thouth.
|
||||
E:
|
||||
KINTret, at fu,
|
||||
GOMy t, as sth kesewit sooos atse ang k, ck,
|
||||
Sotheouserivesthecowhet been's, t he, h nre; t and, har wiread of pincer cedst sur has, ut:
|
||||
|
||||
T14:67% 2.465115
|
||||
UKESpan,
|
||||
NGaromy soreate e m esewfoure pamitherarjulthengeoly tl.
|
||||
NG s at e! w.
|
||||
|
||||
WAllinoully?
|
||||
Wamisw ofilem:
|
||||
I'delandinarrstath har aksubly s cath Whern t Is, weciss:
|
||||
|
||||
GLat s; llde.
|
||||
Y aterit dsthence
|
||||
T14:67% 2.465404
|
||||
|
||||
It is able to guess at what character is likely to follow the current one
|
||||
(by doing a [linear regression][13]), but it lacks any history beyond that to
|
||||
guide it.
|
||||
|
||||
You might be wondering about the role of the `tanh` function in the RNN.
|
||||
`tanh` acts as an [activation function][14] which adds [nonlinearity][15]
|
||||
to the network and allows it to solve complicated problems. Without
|
||||
nonlinearity, all of the linear functions would fold together into a single
|
||||
matrix-vector multiply and you'd effectively regress to the linear network
|
||||
above. Alas, even adding a nonlinearity to the feed-forward network
|
||||
(creating a [perceptron][16]) does not improve the performance because we
|
||||
still lack the history provided by the hidden state vector (although if you
|
||||
want to try it yourself, you can do so with `make per1`).
|
||||
|
||||
Going deeper
|
||||
------------
|
||||
|
||||
We can try to improve the RNN's performance by stacking RNN modules atop each
|
||||
other:
|
||||
|
||||
h1 = RNN(h1', x)
|
||||
h2 = RNN(h2', h1)
|
||||
y = Why * h2 + By
|
||||
|
||||
with `RNN(h, x)` defined as above. Each RNN module has its own set of
|
||||
parameters and its own hidden state vector. This will improve the network's
|
||||
performance, at the cost of a much larger parameter space
|
||||
|
||||
> **IMPORTANT NOTE**: Since OMLET uses the system stack for network storage,
|
||||
> larger networks may cause OMLET to crash (typically with a message like
|
||||
> `Segmentation fault`) unless the system stack size is first increased.
|
||||
> The exact command for doing so depends on your shell and your system's
|
||||
> hard limits. On `sh`/`ksh`/`bash` shells, you can view the hard limit
|
||||
> with `ulimit -Hs` and set it with `ulimit -s 65532` (replacing `65532`
|
||||
> with the actual hard limit). On `csh`/`tcsh` shells, you can view
|
||||
> the hard limit with `limit -h stacksize` and set it with
|
||||
> `limit stacksize 65532` (replacing `65532` with the actual hard limit).
|
||||
|
||||
You can try the deeper network by doing
|
||||
|
||||
make rnn2
|
||||
|
||||
(or even `make rnn3` if you want a three-layer RNN) and train it with
|
||||
|
||||
./rnn2 shakespeare.txt
|
||||
|
||||
The additional depth should allow the network to make better predictions (it
|
||||
can represent more complicated history), but it may take a long time to train
|
||||
-- both because the network (being larger) now requires more time to train and
|
||||
because of the [vanishing and exploding gradient problem][17], which might
|
||||
keep it from ever reaching its potential.
|
||||
|
||||
LSTMs and GRUs
|
||||
--------------
|
||||
|
||||
RNNs are particularly hard to train because the they are trained using
|
||||
[bankpropagation through time][18]. The RNN is trained by effectively
|
||||
converting it into a non-recurrent network by making many copies of it
|
||||
and propagating the hidden state through the copies. During training,
|
||||
the backpropagation through many clones of the network amplifies the
|
||||
gradient, worsening the exploding and vanishing gradient problem.
|
||||
|
||||
[Long Short Term Memory][19] networks (also called LSTMs) were developed to
|
||||
solve this problem. Christopher Olah gives a good description of them
|
||||
at his [blog posting][20]. You can build a two-level LSTM by doing
|
||||
|
||||
make lstm2
|
||||
|
||||
and train with it with
|
||||
|
||||
./lstm2 shakespeare.txt
|
||||
|
||||
The LSTM is much easier to train because it explicitly decides how to update
|
||||
its hidden state via "gates". These gates are called
|
||||
|
||||
* _input gate_, which decides what part of the input to pay attention to
|
||||
* _forget gate_, which decides what part of the hidden state to forget
|
||||
* _output gate_, which decides what part of the hidden state is used to
|
||||
produce the output
|
||||
|
||||
The basic LSTM equations are
|
||||
|
||||
f = sigmoid(Wxf * x + Whf * h' + Bf)
|
||||
i = sigmoid(Wxi * x + Whi * h' + Bi)
|
||||
o = sigmoid(Wxo * x + Who * h' + Bo)
|
||||
c = f * c' + i * tanh(Wxc * x + Whc * h' + Bc)
|
||||
h = o * tanh(c)
|
||||
|
||||
Where
|
||||
|
||||
* `x` is the input vector
|
||||
* `h` is the hidden state (and the output to the next layer)
|
||||
* `c` is the _cell state_ which represents the "memory" of the LSTM
|
||||
* `h'` and `c`' are the previous values of `h` and `c` respectively
|
||||
* `f` is the _forget gate_ that tells the LSTM what portion of the hidden
|
||||
state to forget
|
||||
* `i` is the _input gate_ that tells the LSTM what portion of the input
|
||||
vector to pay attention to
|
||||
* `o` is the _output gate_ that tells the LSTM what portion of the cell
|
||||
state to use to generate the hidden state
|
||||
* `Wxf`, `Whf`, `Wxi`, `Whi`, `Wxo`, `Who`, `Wxc`, and `Whc` are trainable
|
||||
parameter matrices
|
||||
* `Bf`, `Bi`, `Bo`, and `Bc` are trainable bias vectors
|
||||
* `tanh` is the hyperbolic tangent function
|
||||
* `sigmoid` is the [logistic function][21]
|
||||
|
||||
There are several LSTM variants (see [C. Olah's blog post][20] for more
|
||||
examples). One important one is the [gated recurrent unit][22]. GRUs are
|
||||
simplified versions of an LSTM which combine the gates together, meaning they
|
||||
require fewer learned parameters. This allows them to train faster than a
|
||||
generic LSTM. You can build a two-layer GRU with
|
||||
|
||||
make gru2
|
||||
|
||||
Building your own networks
|
||||
--------------------------
|
||||
|
||||
The OMLET `Makefile` comes with one-, two- and three-layer RNNs, LSTMs and
|
||||
GRUs, along with simpler feed-forward networks like multi-layer perceptrons and
|
||||
a linear network. This isn't the limit of OMLET's power -- you can create your
|
||||
own networks by modifying the `Makefile`. Networks are passed in on the
|
||||
compiler's command-line by using `-D` directives. The network is defined by
|
||||
a `-DNW='...'` command which consists of a series of comma-separated
|
||||
assignments. For example, the simple one-layer RNN could be defined like
|
||||
|
||||
-DNW=` x = I(n), hp = I(128), \
|
||||
h = C(hp, T(A(L(128, x), L(128, hp)))), \
|
||||
y = L(n, h)'
|
||||
|
||||
The network declares `x` as an input vector (there must be a declaration for
|
||||
`x`). It is declared as `I(n)`, which is an input vector of size `n`, which
|
||||
is the number of characters of the input alphabet (OMLET computes this from
|
||||
the input file at the start of training). OMLET will arrange to present the
|
||||
input character as a [one-hot][23] vector based on the current input
|
||||
character.
|
||||
|
||||
The second declaration, `hp`, declares the previous hidden state vector (what
|
||||
we called `h'` above). We declare this to be of size 128 -- an arbitrary
|
||||
choice. A larger state vector can (theoretically) carry more state, but at
|
||||
a cost of larger parameter matrices and longer training time. You can
|
||||
experiment with increasing the hidden vector size and see.
|
||||
|
||||
The third line is the core of the RNN. It sets `h`, the hidden vector output
|
||||
to be the sum of two linear elements specified by `L`. The `L` function
|
||||
takes two parameters -- the output vector size (which must match the size of
|
||||
`h`) and the input vector. `L` will compute `y = W * x + B` where each `L`
|
||||
has its own `W` (weight) and `B` (bias) training parameters. Both `x` and
|
||||
`hp` are sent through `L` and the result passed through the `A` function,
|
||||
which does vector addition. That result is passed through `T` which does
|
||||
element-wise `tanh` activation.
|
||||
|
||||
Next, we wrap the whole thing with the `C` function. `C` connects `hp`
|
||||
with `h`, causing the new value of `h` to be passed to the `hp` vector on
|
||||
the next iteration of the algorithm (allowing the RNN to retain state in `h`).
|
||||
|
||||
Finally, the whole result is passed through another instance of `L`, this time
|
||||
producing a vector of size `n`, which will have the negative log
|
||||
[likelihood][24] function. This is assigned to `y`, which is the output of
|
||||
the network (and hence must also be declared).
|
||||
|
||||
OMLET will take the `y` result and pass it through the [softmax][25] function,
|
||||
which converts the log probabilities into a probability distribution. In
|
||||
inference mode, this is used to select the next character to emit. In
|
||||
training mode, this is used to generate the loss which is backpropagated.
|
||||
|
||||
As an example of a more complicated network, we can look at a two-layer
|
||||
GRU network:
|
||||
|
||||
-DHS=128, \
|
||||
-DNW=' x = I(n), \
|
||||
y = L(n, MD(MD(x)))' \
|
||||
-DBK=' hp = I(HS), \
|
||||
z = S(A(L(HS, x), L(HS, hp))), \
|
||||
r = S(A(L(HS, x), L(HS, hp))), \
|
||||
c = T(A(L(HS, x), L(HS, hp))), \
|
||||
zc = OG(1, -1, z), \
|
||||
h = C(hp, A(M(zc, hp), M(z, c))), \
|
||||
y = h'
|
||||
|
||||
We are using a few new tricks here -- first, we are defining `HS` as the size
|
||||
of the hidden and cell vectors. There's nothing special about this name, its
|
||||
just convenient to specify it so we don't have a bunch of constants in the
|
||||
code. Second, the network itself is very simple -- it declares `x` and has
|
||||
the matrix that converts the `HS`-sized hidden vector back to the `n`-sized
|
||||
alphabet vector... but it now calls `MD`, which is the user-definable
|
||||
module (here we are using it twice, to have two cascaded GRU blocks). The
|
||||
`MD` function performs the sub-network defined by the `BK` compile-time
|
||||
parameter (specified in the `-DBK='...'` setting). This sub-module again
|
||||
takes an `x` parameter and produces a `y` output. Inside it, we declare
|
||||
`hp` and `h`, the previous and current state vector, plus equations for the
|
||||
various GRU gates (these use `S` for the sigmoid activation function).
|
||||
One final new call is `OG` which does offset and gain, performing
|
||||
`y = offset + gain * x` where `offset` (the first parameter) and `gain` (the
|
||||
second) are constants. We are using this here to compute `(1 - z)` for the
|
||||
GRU's linear interpolator.
|
||||
|
||||
The full set of available function blocks follows:
|
||||
|
||||
* `I(s)`: declares a vector (input or state) of size `s`
|
||||
* `L(s, x)`: learnable linear function `y = W * x + B` with an output vector
|
||||
size of `s`
|
||||
* `CM(x)`: learnable element-wise gain function `y = W * x`
|
||||
* `A(a, b)`: element-wise add: `y = a + b`
|
||||
* `M(a, b)`: element-wise multiply: `y = a * b`
|
||||
* `S(x)`: sigmoid activation function `y = sigmoid(x)`
|
||||
* `T(x)`: hyperbolic tangent activation function `y = tanh(x)`
|
||||
* `C(xp, x)`: copy `x` to `xp` in the next time step (propagate through
|
||||
time)
|
||||
* `OG(o, g, x)`: apply a constant offset and gain: `y = o + g * x`
|
||||
* `MD(x)`: apply the sub-network specified by `BK`
|
||||
|
||||
Note: even if you don't use `MD` in your network, you should still define
|
||||
`BK` by adding `-DBK='y=x'` to the command line, otherwise you will get a
|
||||
compile-time error.
|
||||
|
||||
Hyperparameters
|
||||
---------------
|
||||
|
||||
OMLET has a large number of training and inference parameters which can be
|
||||
changed by the user. All of these are set by `-D` on the compile command line.
|
||||
The list of hyperparameters follows:
|
||||
|
||||
* `TP`: Temperature parameter for use in inference mode. This divides the
|
||||
log probabilities before softmax. A low temperature makes the model choose
|
||||
safer but more boring choices. A high temperature takes more risks but
|
||||
makes more mistakes. Default is 1.0, which uses the computed
|
||||
probabilities.
|
||||
* 'N': Batch size. This is the number of times the RNN is unrolled, so it
|
||||
controls how far back in the past the RNN can see. The default is 50.
|
||||
Larger batch sizes update the weights less frequently and allow the RNN to
|
||||
see farther back in time, but at a cost of proportionally more memory.
|
||||
* `TR`: The percentage of batches in the input data set that will be used
|
||||
for training. The default value of 0.95 sets this as 95%.
|
||||
* `LR`: The initial learning rate, The default is 0.002.
|
||||
* `LE`: The epoch where the learning rate will start decaying. Defaults to
|
||||
epoch 10.
|
||||
* `LD`: Learning rate decay, per epoch (after `LE` epochs). The learning
|
||||
rate is scaled by this number. Default is 0.97
|
||||
* `WD`: Weight-decay parameter, to promote [regularizaiton][27]. The default
|
||||
is 0.00008.
|
||||
* `RS`: The random scale for weight initialization. Weight parameters will
|
||||
be initialized to be between `-RS` to `+RS`. The default is 0.15.
|
||||
* `CL`: Clamp value for gradients. Gradients will be limited to the range
|
||||
`-CL` to `+CL`. Default is 5.
|
||||
* `B1`: Momentum mean parameter for Adam optimizer. Set to 0.9.
|
||||
* `B2`: Momentum variance parameter for Adam optimizer, Set to 0.999.
|
||||
* `EP`: Epsilon parameter for Adam, to provide numerical stability. Set to
|
||||
0.00000001.
|
||||
* `DI`: How often to print a training or validation progress message and
|
||||
inference snippet. The default prints every 100 training batches.
|
||||
* `SL`: The number of characters to print when doing an inference snippet.
|
||||
Default is 200.
|
||||
* `PF`: Format string for the checkpoint filename. The default is
|
||||
`"cp%02d_%.3f"` which includes the epoch and validation loss. You may
|
||||
wish to add a subdirectory to the name to keep checkpoint files out of
|
||||
the current directory.
|
||||
|
||||
|
||||
[1]: https://en.wikipedia.org/wiki/Recurrent_neural_network
|
||||
[2]: https://en.wikipedia.org/wiki/Deep_learning
|
||||
[3]: https://www.tensorflow.org
|
||||
[4]: http://torch.ch
|
||||
[5]: https://en.wikipedia.org/wiki/William_Shakespeare
|
||||
[6]: https://en.wikipedia.org/wiki/Loss_function
|
||||
[7]: https://en.wikipedia.org/wiki/Overfitting
|
||||
[8]: https://en.wikipedia.org/wiki/Training,_validation,_and_test_sets
|
||||
[9]: https://en.wikipedia.org/wiki/Hyperbolic_function
|
||||
[10]: https://en.wikipedia.org/wiki/ADALINE
|
||||
[11]: https://en.wikipedia.org/wiki/Linear_map
|
||||
[12]: https://en.wikipedia.org/wiki/Feedforward_neural_network
|
||||
[13]: https://en.wikipedia.org/wiki/Linear_regression
|
||||
[14]: https://en.wikipedia.org/wiki/Activation_function
|
||||
[15]: https://en.wikipedia.org/wiki/Nonlinear_system
|
||||
[16]: https://en.wikipedia.org/wiki/Perceptron
|
||||
[17]: https://en.wikipedia.org/wiki/Vanishing_gradient_problem
|
||||
[18]: https://en.wikipedia.org/wiki/Backpropagation_through_time
|
||||
[19]: https://en.wikipedia.org/wiki/Long_short-term_memory
|
||||
[20]: http://colah.github.io/posts/2015-08-Understanding-LSTMs/
|
||||
[21]: https://en.wikipedia.org/wiki/Logistic_function
|
||||
[22]: https://en.wikipedia.org/wiki/Gated_recurrent_unit
|
||||
[23]: https://en.wikipedia.org/wiki/One-hot
|
||||
[24]: https://en.wikipedia.org/wiki/Likelihood_function
|
||||
[25]: https://en.wikipedia.org/wiki/Softmax_function
|
||||
[26]: https://en.wikipedia.org/wiki/Dropout_(neural_networks)
|
||||
[27]: https://en.wikipedia.org/wiki/Regularization_(mathematics)
|
||||
|
||||
[30]: https://cs.stanford.edu/people/karpathy/char-rnn/linux_input.txt
|
||||
[31]: https://cs.stanford.edu/people/karpathy/char-rnn/warpeace_input.txt
|
||||
[32]: https://github.com/ryanmcdermott/trump-speeches
|
||||
|
||||
[40]: https://cs.stanford.edu/people/karpathy/
|
||||
[41]: https://github.com/karpathy/char-rnn
|
||||
[42]: https://arxiv.org/abs/1412.6980
|
||||
[43]: https://www.fast.ai/2018/07/02/adam-weight-decay/
|
||||
[44]: http://karpathy.github.io/2015/05/21/rnn-effectiveness/
|
||||
|
||||
|
||||
--------------------------------------------------------------------------------
|
||||
<!--
|
||||
(c) Copyright 1984-2019, [Leo Broukhis, Simon Cooper, Landon Curt Noll][judges] - All rights reserved
|
||||
This work is licensed under a [Creative Commons Attribution-ShareAlike 3.0 Unported License][cc].
|
||||
|
||||
[judges]: http://www.ioccc.org/judges.html
|
||||
[cc]: http://creativecommons.org/licenses/by-sa/3.0/
|
||||
-->
|
||||
@@ -0,0 +1,73 @@
|
||||
|
||||
|
||||
#/*void*/include<math.h>
|
||||
#include/*#pragma _Atomic int*/<stdio.h>
|
||||
/* { */struct{ int o,s; } typedef V ;
|
||||
struct{ int x; V o,i,p; }typedef R; int
|
||||
d,w,a,i,b, u,g,n,j,F,h; R*Q,B [99],*E=B
|
||||
; V I(int s ){ V o={ d,s} ; d +=s; return
|
||||
#define O ; { } } if(! n --) { { { ; } }
|
||||
o; } float*p,ll=1,* q, *e,s,*r,*G,*W,y=1,
|
||||
# define _ V i ){ V o= I(i .s); * E++=( R) {
|
||||
*x,l,*z,m; V A(V p,_ 4,o ,p,i} ; return o; } V
|
||||
X,Y; double drand48(); V C (V i,V o){ *E++=(R)
|
||||
#define OO ){ P(Q->o,a=Q->i. o-n; b=Q->p .o-n\
|
||||
; r=b+W+n){ { n=Q->x O//struct volatile union {
|
||||
{ 3,o,i} ; return o; } unsigned char c[256],t[
|
||||
sizeof c+N]; V M(V p,_ 5,o,p,i} ; return o ; }
|
||||
struct{ float r ; short k ,p; char unsigned d[
|
||||
sizeof c]; } k={ LR} ; V S(_ 2,o,i} ; return o
|
||||
; } void Z(float*p,int n){ while(n--) *p++=0; }
|
||||
V L(int s,V i){ V o=I(s),p={ w,s+i.s*s } ; w+=p
|
||||
#define P(v,i)n=v.o; e=G+n; p=e+v.s; q= p+d; i\
|
||||
; while(--q,p-->e)// _Alignas _Imaginary _Bool;
|
||||
.s; *E++=(R){0,o,i,p}; return o; } V CM(V i){ V
|
||||
o=I(a=i.s),p={ w,a} ; w+=a; *E++=(R){ 7, o,i,
|
||||
p} ; return o; } void J(int c){ Z(e=G+X.o, X.s)
|
||||
; c[e]++ ; for(Q=B; Q<E; ++Q OO*p=*r++; n=Q ->i
|
||||
.s; while(n--)*p+=*r++*e[a+n]O*p=tanh(a[p]) O*p
|
||||
=1/(exp(-a[p])+1)O p[a+u]=*p O*p=p[a]+b[p]O* p=
|
||||
a[p]*p[b]O*p=a[p ]*Q->p.s+Q->p.o O*p=a[p]**r ++
|
||||
; } } } Z(G+d,d); P(Y,m=*e)m=fmax(m,*p); P(Y, s
|
||||
=0)s+=*q=exp((*p-m)/TP); P(Y, )*q/=s; } V T(_ 1
|
||||
,o,i}; return o; } int K(int K ){ J(K); P(Y,s =
|
||||
drand48( ))if((s-=*q)<0)break; return j=p-e; }
|
||||
V OG(int a,int b,_ 6,o,(V){ b,a} } ; return o;
|
||||
} void D(char unsigned*s){ float V[2*d]; int R
|
||||
=u; u=V-G; J(c[*s++ ]); n=c[*s]; l-=log(n[++q]
|
||||
--)/N; G+=u; p=d+G; if(1[s])D(s); else while(p
|
||||
-->G){ p[x-G]=*p; d[p]=0; } G-=u; Q=E; while(F
|
||||
<i&&Q-->B OO r++[w ]+=*q; n=Q->i .s; z=e+a +n;
|
||||
while(n--){ w[r ]+=*q**--z; d[z]+=*q**r++; } O
|
||||
a[q]+=*q*(1-*p**p)O q[a]+=*q*(1-*p)**p O*q+=q [
|
||||
a+u]O a[q]+=*q; q[b]+=*q O q[a]+=b[p]**q; b[q ]
|
||||
+=p[a ]**q O a[q ]+=Q->p.s**q O a[q ]+=*q**r; r
|
||||
++[w]+=*q**p; } } } u=R; } V MD( V x){ V BK; //
|
||||
return y; } int main(int Q,char**E){ FILE*f, *o
|
||||
=stdin; if(--Q){ o=fopen(2[E],"r"); f=fopen (1[
|
||||
E],"r"); while(w>-(n=getc(f))){ ++h; if(!n[ t])
|
||||
{ n[t]++; k.d[k.k]=n; n[c]=k .k++; } } } o &&f\
|
||||
read(&k,sizeof k,1,o); n=k.k; { V NW; X=x ; Y=
|
||||
y; } float R[d+d],B[4*w],q[d*2]; G=x=R; W =B; Z
|
||||
(W,4*w); p=W+w; while(p-->W)*p=2*drand48 ()*RS-
|
||||
RS; o&&fread(W,1,sizeof B,o); Z(G,d); h /=N; i=
|
||||
h*TR; h-=i; for(t[1+N]=m; Q; fclose(o) ){ rewi\
|
||||
nd(f ); while(fread(t,1 +N,1,f) ){ G= q; p=d+G;
|
||||
while(p-->G)*p=p[x-G]; D(t); ungetc (j=N[t],f)
|
||||
; if(F++<i){ a=w*3; b=a- w; p=W+w; while(p-->W
|
||||
){ m=fmax(-CL,fmin(CL,w[p ])); ll *=B1; p[a]=m+
|
||||
(a[p]-m)*B1; y*=B2; m*=m; b[p]= m+(p[b]-m)*B2;
|
||||
*p+=(a[p]/sqrt(p[b]/(1-y)+EP)/ (ll-1)+WD**p)*k.
|
||||
r; p[w]=0; } if(1>F%DI){ G=R ; z=SL+W; while(z
|
||||
-->B)putchar(k.d[K(j)]); } } if(1>F%DI ){ b=F;
|
||||
a=i; j=b>a; if(j){ b-=a; a= h; } printf(&j[//;
|
||||
"\n%c%d:%d%% %f\n"],j[ "TV"],k.p,100*b/a,l/b
|
||||
); } l*=F!=i; } /** */ char p[999]; if(LE
|
||||
<=++k .p)k.r*=LD; /*if*/ sprintf(p,PF,k.p
|
||||
,l/h); fwrite( &k,sizeof k, 1,o =fopen(p,
|
||||
"w")); F=l =0; fwrite(B,w,/** */sizeof*
|
||||
W,o); } while(putchar(k.d[K(j)]) ){ } }
|
||||
/* typedef volatile long double not ;
|
||||
register restrict unsigned volatile void
|
||||
_Complex struct bitand*/
|
||||
|
||||
Reference in new issue
Block a user