PDF uncertainties with Neural Networks

Transcript

PDF uncertainties with Neural Networks
PDF uncertainties
with Neural Networks
21 novembre
Andrea Piccione (NNPDF)
PDF uncertainties with Neural Networks
Martina Franca, 21/11/07
1/8
un confronto con 20 repliche
Desktop: 1 AMD 64 3000+ (1800 GHz) - 5h
chi2cme total =
chi2cme nmc
=
chi2cme bcd
=
0.770753281
0.709723309
0.825776367
GRID: 5 Intel E2140 (1600 GHz) - 1h
chi2cme total =
chi2cme nmc
=
chi2cme bcd
=
0.762194036
0.698429759
0.819682302
I
a seconda di come raggruppo le repliche cambia il seed del generatore
di numeri pseudo casuali;
I
con 20 repliche la statistica è poca.
Andrea Piccione (NNPDF)
PDF uncertainties with Neural Networks
Martina Franca, 21/11/07
2/8
un confronto con 20 repliche
Desktop: 1 AMD 64 3000+ (1800 GHz) - 5h
chi2cme total =
chi2cme nmc
=
chi2cme bcd
=
0.770753281
0.709723309
0.825776367
GRID: 5 Intel E2140 (1600 GHz) - 1h
chi2cme total =
chi2cme nmc
=
chi2cme bcd
=
0.762194036
0.698429759
0.819682302
I
a seconda di come raggruppo le repliche cambia il seed del generatore
di numeri pseudo casuali;
I
con 20 repliche la statistica è poca.
Andrea Piccione (NNPDF)
PDF uncertainties with Neural Networks
Martina Franca, 21/11/07
2/8
un confronto con 20 repliche
Desktop: 1 AMD 64 3000+ (1800 GHz) - 5h
chi2cme total =
chi2cme nmc
=
chi2cme bcd
=
0.770753281
0.709723309
0.825776367
GRID: 5 Intel E2140 (1600 GHz) - 1h
chi2cme total =
chi2cme nmc
=
chi2cme bcd
=
0.762194036
0.698429759
0.819682302
I
a seconda di come raggruppo le repliche cambia il seed del generatore
di numeri pseudo casuali;
I
con 20 repliche la statistica è poca.
Andrea Piccione (NNPDF)
PDF uncertainties with Neural Networks
Martina Franca, 21/11/07
2/8
una nuova struttura: dag+parametric
Andrea Piccione (NNPDF)
PDF uncertainties with Neural Networks
Martina Franca, 21/11/07
3/8
amga collections
Query> dir
>> /nnpdf/generator
>> collection
>> /nnpdf/evfactinter
>> collection
>> /nnpdf/nnqns
>> collection
>> /nnpdf/runs
>> collection
Andrea Piccione (NNPDF)
PDF uncertainties with Neural Networks
Martina Franca, 21/11/07
4/8
mdcli script #1
generator.sh:
lcg-cr --vo gilda file://$PWD/$LF_NAME -l lfn:$SE_PATH/$LF_NAME -d $SE_NAME
TMP=‘mdcli sequence_next "$MD_PATH"/evfactinter/evfseq‘
mdcli addentry $MD_PATH/generator/$TMP nrep $NREP q2 $Q2 w2 $W2 name_file $LF_NAME
evfactinter.sh:
lcg-cr --vo gilda file://$PWD/$LF_NAME -l lfn:$SE_PATH/$LF_NAME -d $SE_NAME
TMP=‘mdcli sequence_next "$MD_PATH"/evfactinter/evfseq‘
mdcli addentry $MD_PATH/evfactinter/$TMP alpha "0."$ALPHA_I ipt $IPT name_file $LF_NAME
Andrea Piccione (NNPDF)
PDF uncertainties with Neural Networks
Martina Franca, 21/11/07
5/8
mdcli script #1
generator.sh:
lcg-cr --vo gilda file://$PWD/$LF_NAME -l lfn:$SE_PATH/$LF_NAME -d $SE_NAME
TMP=‘mdcli sequence_next "$MD_PATH"/evfactinter/evfseq‘
mdcli addentry $MD_PATH/generator/$TMP nrep $NREP q2 $Q2 w2 $W2 name_file $LF_NAME
evfactinter.sh:
lcg-cr --vo gilda file://$PWD/$LF_NAME -l lfn:$SE_PATH/$LF_NAME -d $SE_NAME
TMP=‘mdcli sequence_next "$MD_PATH"/evfactinter/evfseq‘
mdcli addentry $MD_PATH/evfactinter/$TMP alpha "0."$ALPHA_I ipt $IPT name_file $LF_NAME
Andrea Piccione (NNPDF)
PDF uncertainties with Neural Networks
Martina Franca, 21/11/07
5/8
mdcli script #2
nnqns.sh:
TMPA=‘mdcli selectattr "$MD_PATH"/evfactinter:name_file "’alpha=$ALPHA’"‘
TMPR=‘mdcli selectattr "$MD_PATH"/generator:name_file "’nrep=$NREP’"‘
if [ -z "$TMPA" -o -z "$TMPR" ]
then
echo "ERROR"
else
TMPF=‘mdcli selectattr "$MD_PATH"/nnqns:name_file "’nrep=$NREP and alpha=$ALPHA’"‘
if [ -z "$TMPF" ]
then
NAME_GEN=‘mdcli selectattr "$MD_PATH"/generator:name_file "’nrep=$NREP’"‘
lcg-cp --vo gilda lfn:$SE_PATH/ttmmpp/$NAME_GEN file://$PWD/gen_rep.tgz
NAME_EVK=‘mdcli selectattr "$MD_PATH"/evfactinter:name_file "’alpha=$ALPHA’"‘
lcg-cp --vo gilda lfn:$SE_PATH/ttmmpp/$NAME_EVK file://$PWD/ev_ker.tgz
[ .. ]
Andrea Piccione (NNPDF)
PDF uncertainties with Neural Networks
Martina Franca, 21/11/07
6/8
mdcli script #3
nnqns.sh:
[ .. ]
mdcli addentry $MD_PATH/runs/$NAME_F id $NJOB first_rep $kappap last_rep $NNDIST error 0 status RUNNING
mdcli setattr $MD_PATH/runs/$NAME_F alpha $ALPHA nrep $NREP
checkout.sh:
ALPHA=‘mdcli getattr /nnpdf/runs alpha|head -2|tail -1 ‘
NREP=‘mdcli getattr /nnpdf/runs nrep|head -2|tail -1 ‘
TMP=‘mdcli sequence_next "$MD_PATH"/nnqns/fitseq‘
mdcli addentry $MD_PATH/nnqns/$TMP name_file $LF_NAME"_"$TMP alpha $ALPHA nrep $NREP
Andrea Piccione (NNPDF)
PDF uncertainties with Neural Networks
Martina Franca, 21/11/07
7/8
cosa manca?
una struttura che scriva un articolo pigiando “enter” . . .
in pratica:
I
gestione del job con MyProxy;
I
controlli sui run (per ora solo il minimo sindacale);
I
protezione del database;
I
interfaccia “umana” per amga;
I
riordinare il codice.
Andrea Piccione (NNPDF)
PDF uncertainties with Neural Networks
Martina Franca, 21/11/07
8/8
cosa manca?
una struttura che scriva un articolo pigiando “enter” . . .
in pratica:
I
gestione del job con MyProxy;
I
controlli sui run (per ora solo il minimo sindacale);
I
protezione del database;
I
interfaccia “umana” per amga;
I
riordinare il codice.
Andrea Piccione (NNPDF)
PDF uncertainties with Neural Networks
Martina Franca, 21/11/07
8/8