1 |
mlosch |
1.1 |
#!/bin/bash |
2 |
|
|
# these are some basic instructions to an optimization of |
3 |
|
|
# MITgcm/verification/tutorial_global_oce_optim with |
4 |
|
|
# MITgcm_offline/mlosch/optim_m1qn3 |
5 |
|
|
# Some tweaking is definitely possible and not described here. |
6 |
|
|
# |
7 |
|
|
# |
8 |
|
|
cd MITgcm/verification |
9 |
|
|
# this is just to compile and run the model for testing. |
10 |
|
|
# I use TAF (just because I have it and OpenAD is a pain to compile on a Mac), |
11 |
|
|
# but that should not make any difference |
12 |
|
|
./testreport -t tutorial_global_oce_optim -adm -j 4 -ncad |
13 |
|
|
# this is the result: |
14 |
|
|
# G D M C A F |
15 |
|
|
# e p a R o d D |
16 |
|
|
# n n k u s G G |
17 |
|
|
# 2 d e n t r r |
18 |
|
|
|
19 |
|
|
# Y Y Y Y 15>16< 7 pass tutorial_global_oce_optim (e=0, w=2) |
20 |
|
|
|
21 |
|
|
# here is the cost function value that I get |
22 |
|
|
# (PID.TID 0000.0001) local fc = 0.620023228182336D+01 |
23 |
|
|
# (PID.TID 0000.0001) global fc = 0.620023228182336D+01 |
24 |
|
|
# now I download and compile optim_m1qn3 |
25 |
|
|
cd ../../ |
26 |
|
|
cvs co MITgcm_contrib/mlosch/optim_m1qn3 |
27 |
|
|
cd MITgcm_contrib/mlosch/optim_m1qn3 |
28 |
|
|
# edit Makefile to adjust to your platform. For me this involves choosing |
29 |
|
|
# the correct CPP command and setting SUFF=for (because MacOS is |
30 |
|
|
# case-insensitive in my case) |
31 |
|
|
make depend |
32 |
|
|
make |
33 |
|
|
# then I get optim.x |
34 |
|
|
cd ../../../MITgcm/verification/tutorial_global_oce_optim/run |
35 |
|
|
cp ../../../../MITgcm_contrib/mlosch/optim_m1qn3/optim.x . |
36 |
|
|
# - turn off the gradient check (in data.pkg: useGrdchk = .FALSE.) |
37 |
|
|
# - tweak the namelist files data.ctrl and data.optim to the compiler needs |
38 |
|
|
# (I need a "/" to terminate a namelist) |
39 |
|
|
# - I would replace fmin with dfminFrac = 0.1 (expected reduction of 10%) |
40 |
|
|
# in data.optim&OPTIM to be independent of the absolute value of the |
41 |
|
|
# cost function. When you run ./mitgcmuv_ad with this you need to comment |
42 |
|
|
# out dfminFrac, because mitgcmuv_ad does not know about this namelist |
43 |
|
|
# parameter (something to be fixed) |
44 |
|
|
# - set numiter=100, nfunc=10, or some other large value. nfunc*numiter is |
45 |
|
|
# the number of simulations that are allowed in total. This number should be |
46 |
|
|
# much larger than numiter, because you may need more than one function call |
47 |
|
|
# (= run of mitgcmuv_ad) per iteration, see m1qn3 docs for details |
48 |
|
|
# - add an empty namelist &M1QN3 to data.optim |
49 |
|
|
# &M1QN3 |
50 |
|
|
# / |
51 |
|
|
./optim.x > opt0.txt |
52 |
|
|
# I get a lot of output in opt0.txt, which is easier to read with |
53 |
|
|
# less -S opt0.out (to truncate long lines) |
54 |
|
|
# essentially I have the same value for the cost function |
55 |
|
|
|
56 |
|
|
# ================================================== |
57 |
|
|
# Large Scale Optimization with off-line capability. |
58 |
|
|
# ================================================== |
59 |
|
|
|
60 |
|
|
# OPTIM_READPARMS: Control options have been read. |
61 |
|
|
# OPTIM_READPARMS: Minimization options have been read. |
62 |
|
|
|
63 |
|
|
# OPTIM_READDATA: Reading cost function and gradient of cost function |
64 |
|
|
# for optimization cycle: 0 |
65 |
|
|
|
66 |
|
|
# OPTIM_READDATA: opened file ecco_cost_MIT_CE_000.opt0000 |
67 |
|
|
# OPTIM_READDATA: nvartype 1 |
68 |
|
|
# OPTIM_READDATA: nvarlength 2315 |
69 |
|
|
# OPTIM_READDATA: yctrlid MIT_CE_000 |
70 |
|
|
# OPTIM_READDATA: filenopt 0 |
71 |
|
|
# OPTIM_READDATA: fileff 6.2002322818233591 |
72 |
|
|
# OPTIM_READDATA: fileiG 1 |
73 |
|
|
# OPTIM_READDATA: filejG 1 |
74 |
|
|
# OPTIM_READDATA: filensx 2 |
75 |
|
|
# OPTIM_READDATA: filensy 2 |
76 |
|
|
# [...] |
77 |
|
|
# OPTIM_READDATA: end of optim_readdata |
78 |
|
|
|
79 |
|
|
|
80 |
|
|
# OPTIM_READPARMS: Iteration number = 0 |
81 |
|
|
# number of control variables = 2315 |
82 |
|
|
# cost function value in ecco_ctrl = 6.2002322818233591 |
83 |
|
|
# expected cost function minimum = 5.5802090536410232 |
84 |
|
|
# expected cost function decrease = 0.62002322818233591 |
85 |
|
|
# Data will be read from the following file: ecco_ctrl_MIT_CE_000.opt0000 |
86 |
|
|
|
87 |
|
|
# OPTIM_SUB: Calling m1qn3_optim for iteration: 0 |
88 |
|
|
# OPTIM_SUB: with nn, REAL_BYTE = 2315 4 |
89 |
|
|
# OPTIM_SUB: read model state |
90 |
|
|
|
91 |
|
|
# OPTIM_READDATA: Reading control vector |
92 |
|
|
# for optimization cycle: 0 |
93 |
|
|
|
94 |
|
|
# OPTIM_READDATA: opened file ecco_ctrl_MIT_CE_000.opt0000 |
95 |
|
|
# OPTIM_READDATA: nvartype 1 |
96 |
|
|
# OPTIM_READDATA: nvarlength 2315 |
97 |
|
|
# OPTIM_READDATA: yctrlid MIT_CE_000 |
98 |
|
|
# OPTIM_READDATA: filenopt 0 |
99 |
|
|
# OPTIM_READDATA: fileff 6.2002322818233591 |
100 |
|
|
# OPTIM_READDATA: fileiG 1 |
101 |
|
|
# OPTIM_READDATA: filejG 1 |
102 |
|
|
# OPTIM_READDATA: filensx 2 |
103 |
|
|
# OPTIM_READDATA: filensy 2 |
104 |
|
|
# OPTIM_READDATA: end of optim_readdata |
105 |
|
|
|
106 |
|
|
|
107 |
|
|
# OPTIM_READDATA: Reading cost function and gradient of cost function |
108 |
|
|
# for optimization cycle: 0 |
109 |
|
|
|
110 |
|
|
# OPTIM_READDATA: opened file ecco_cost_MIT_CE_000.opt0000 |
111 |
|
|
# OPTIM_READDATA: nvartype 1 |
112 |
|
|
# OPTIM_READDATA: nvarlength 2315 |
113 |
|
|
# OPTIM_READDATA: yctrlid MIT_CE_000 |
114 |
|
|
# OPTIM_READDATA: filenopt 0 |
115 |
|
|
# OPTIM_READDATA: fileff 6.2002322818233591 |
116 |
|
|
# OPTIM_READDATA: fileiG 1 |
117 |
|
|
# OPTIM_READDATA: filejG 1 |
118 |
|
|
# OPTIM_READDATA: filensx 2 |
119 |
|
|
# OPTIM_READDATA: filensy 2 |
120 |
|
|
# OPTIM_READDATA: end of optim_readdata |
121 |
|
|
|
122 |
|
|
# OPTIM_SUB after reading ecco_ctrl and ecco_cost: |
123 |
|
|
# OPTIM_SUB nn = 2315 |
124 |
|
|
# OPTIM_SUB objf = 6.2002322818233591 |
125 |
|
|
# OPTIM_SUB xx(1) = 0.0000000000000000 |
126 |
|
|
# OPTIM_SUB adxx(1) = -6.7879882408306003E-005 |
127 |
|
|
# OPTIM_SUB: cold start, optimcycle = 0 |
128 |
|
|
# OPTIM_SUB: call m1qn3_offline ........ |
129 |
|
|
# OPTIM_SUB: ........................... |
130 |
|
|
# OPTIM_SUB: returned from m1qn3_offline |
131 |
|
|
# OPTIM_SUB: nn = 2315 |
132 |
|
|
# OPTIM_SUB: xx(1) = 0.51934864251896229 0.73000729032300782 |
133 |
|
|
# OPTIM_SUB: adxx(1) = -6.7879882408306003E-005 -9.5413379312958568E-005 |
134 |
|
|
# OPTIM_SUB: omode = -1 |
135 |
|
|
# OPTIM_SUB: niter = 1 |
136 |
|
|
# OPTIM_SUB: nsim = 10000 |
137 |
|
|
# OPTIM_SUB: reverse = 1 |
138 |
|
|
|
139 |
|
|
# OPTIM_SUB: mean(xx) = 0.16365068483545642 |
140 |
|
|
# OPTIM_SUB: max(xx) = 4.6525355815045790 |
141 |
|
|
# OPTIM_SUB: min(xx) = -9.3326211896764324 |
142 |
|
|
# OPTIM_SUB: std(xx) = 15.613450260548481 |
143 |
|
|
|
144 |
|
|
|
145 |
|
|
# OPTIM_STORE_M1QN3: saving the state of m1qn3 in OPWARM.opt0001 |
146 |
|
|
|
147 |
|
|
# OPTIM_SUB: writing 2315 sized control to file ecco_ctrl |
148 |
|
|
|
149 |
|
|
# OPTIM_WRITEDATA: Writing new control vector to file(s) |
150 |
|
|
# for optimization cycle: 1 |
151 |
|
|
|
152 |
|
|
# OPTIM_WRITEDATA: nvartype 1 |
153 |
|
|
# OPTIM_WRITEDATA: nvarlength 2315 |
154 |
|
|
# OPTIM_WRITEDATA: yctrlid MIT_CE_000 |
155 |
|
|
# OPTIM_WRITEDATA: nopt 1 |
156 |
|
|
# OPTIM_WRITEDATA: ff -9999.0000000000000 |
157 |
|
|
# OPTIM_WRITEDATA: iG 1 |
158 |
|
|
# OPTIM_WRITEDATA: jG 1 |
159 |
|
|
# OPTIM_WRITEDATA: nsx 2 |
160 |
|
|
# OPTIM_WRITEDATA: nsy 2 |
161 |
|
|
# OPTIM_WRITEDATA: end of optim_writedata, icvoffset 2315 |
162 |
|
|
|
163 |
|
|
|
164 |
|
|
# ====================================== |
165 |
|
|
# Large Scale Optimization run finished. |
166 |
|
|
# ====================================== |
167 |
|
|
|
168 |
|
|
# ff = -9999 is an intentional dummy value. |
169 |
|
|
|
170 |
|
|
# now you can organize the rest in a loop. In bash, it could look like this: |
171 |
|
|
# first comment out dfminFrac, because mitgcmuv_ad does not know about this, |
172 |
|
|
# dfminFrac is really only needed in the zeroth iteration |
173 |
|
|
# |
174 |
|
|
# tabula rasa |
175 |
|
|
rm ecco_c* OPWARM.* m1qn3_output.txt |
176 |
|
|
myiter=0 |
177 |
|
|
cat > data.optim <<EOF # |
178 |
|
|
# ******************************** |
179 |
|
|
# Off-line optimization parameters |
180 |
|
|
# ******************************** |
181 |
|
|
&OPTIM |
182 |
|
|
optimcycle=${myiter}, |
183 |
|
|
numiter=100, |
184 |
|
|
nfunc=100, |
185 |
|
|
#dfminFrac = 0.1, |
186 |
|
|
iprint=10, |
187 |
|
|
nupdate=8, |
188 |
|
|
/ |
189 |
|
|
|
190 |
|
|
&M1QN3 |
191 |
|
|
/ |
192 |
|
|
EOF |
193 |
|
|
|
194 |
|
|
while (( $myiter < 20 )) |
195 |
|
|
do |
196 |
|
|
# formatter iteration count |
197 |
|
|
it=`echo $myiter | awk '{printf "%03i",$1}'` |
198 |
|
|
echo "iteration ${myiter}" |
199 |
|
|
# comment out dfminFrac from data.optim |
200 |
|
|
sed -i '' 's/.*dfminFrac.*/#dfminFrac = 0.1,/' data.optim |
201 |
|
|
# increment counter in data.optim |
202 |
|
|
sed -i .${it} "s/.*optimcycle.*/ optimcycle=${myiter},/" data.optim |
203 |
|
|
./mitgcmuv_ad > output${it}.txt |
204 |
|
|
# add dfminFrac = 0.1, |
205 |
|
|
sed -i '' 's/#dfminFrac.*/ dfminFrac = 0.1,/' data.optim |
206 |
|
|
./optim.x > opt${it}.txt |
207 |
|
|
# increase counter for next iteration |
208 |
|
|
((myiter++)) |
209 |
|
|
done |
210 |
|
|
|
211 |
|
|
# grep "fc " output001.txt |
212 |
|
|
# (PID.TID 0000.0001) early fc = 0.000000000000000D+00 |
213 |
|
|
# (PID.TID 0000.0001) local fc = 0.132325873958879D+02 |
214 |
|
|
# (PID.TID 0000.0001) global fc = 0.132325873958879D+02 |
215 |
|
|
|
216 |
|
|
# grep "global fc " output???.txt |
217 |
|
|
# output000.txt:(PID.TID 0000.0001) global fc = 0.620023228182336D+01 |
218 |
|
|
# output001.txt:(PID.TID 0000.0001) global fc = 0.132325873958879D+02 |
219 |
|
|
# output002.txt:(PID.TID 0000.0001) global fc = 0.615567811433600D+01 |
220 |
|
|
# output003.txt:(PID.TID 0000.0001) global fc = 0.615556878869932D+01 |
221 |
|
|
# output004.txt:(PID.TID 0000.0001) global fc = 0.615547009131471D+01 |
222 |
|
|
|
223 |
|
|
# as you can see, the improvement is not very good after the initial steps. |
224 |
|
|
# and the optimization will not be successful in the end, i.e. satisfy |
225 |
|
|
# the tolerance set by epsg. |
226 |
|
|
# This is likely because the number of timesteps in data is small (20) |
227 |
|
|
# for this test. Try longer intergations (e.g. 1 year) and wait longer |
228 |
|
|
|
229 |
|
|
# check out m1qn3_output.txt, which records the output of m1qn3 |
230 |
|
|
# the python script plotfc.py greps the cost function values out of |
231 |
|
|
# m1qn3_ouput.txt and plots them and the number of simulations per iteration |
232 |
|
|
# which typically increase with decreasing cost function |
233 |
|
|
|
234 |
|
|
|