1 |
#!/bin/bash |
2 |
########################################################### |
3 |
# on May, 22, 2019 I copied the content of this directory |
4 |
# github.com/mjlosch/optim_m1qn3.git |
5 |
# I will not update this version any longer, but only the |
6 |
# git-version |
7 |
########################################################### |
8 |
# these are some basic instructions to an optimization of |
9 |
# MITgcm/verification/tutorial_global_oce_optim with |
10 |
# MITgcm_offline/mlosch/optim_m1qn3 |
11 |
# Some tweaking is definitely possible and not described here. |
12 |
# |
13 |
# |
14 |
cd MITgcm/verification |
15 |
# this is just to compile and run the model for testing. |
16 |
# I use TAF (just because I have it and OpenAD is a pain to compile on a Mac), |
17 |
# but that should not make any difference |
18 |
./testreport -t tutorial_global_oce_optim -adm -j 4 -ncad |
19 |
# this is the result: |
20 |
# G D M C A F |
21 |
# e p a R o d D |
22 |
# n n k u s G G |
23 |
# 2 d e n t r r |
24 |
|
25 |
# Y Y Y Y 15>16< 7 pass tutorial_global_oce_optim (e=0, w=2) |
26 |
|
27 |
# here is the cost function value that I get |
28 |
# (PID.TID 0000.0001) local fc = 0.620023228182336D+01 |
29 |
# (PID.TID 0000.0001) global fc = 0.620023228182336D+01 |
30 |
# now I download and compile optim_m1qn3 |
31 |
cd ../../ |
32 |
cvs co MITgcm_contrib/mlosch/optim_m1qn3 |
33 |
cd MITgcm_contrib/mlosch/optim_m1qn3 |
34 |
# edit Makefile to adjust to your platform. For me this involves choosing |
35 |
# the correct CPP command and setting SUFF=for (because MacOS is |
36 |
# case-insensitive in my case) |
37 |
make depend |
38 |
make |
39 |
# then I get optim.x |
40 |
cd ../../../MITgcm/verification/tutorial_global_oce_optim/run |
41 |
cp ../../../../MITgcm_contrib/mlosch/optim_m1qn3/optim.x . |
42 |
# - turn off the gradient check (in data.pkg: useGrdchk = .FALSE.) |
43 |
# - tweak the namelist files data.ctrl and data.optim to the compiler needs |
44 |
# (I need a "/" to terminate a namelist) |
45 |
# - I would replace fmin with dfminFrac = 0.1 (expected reduction of 10%) |
46 |
# in data.optim&OPTIM to be independent of the absolute value of the |
47 |
# cost function. When you run ./mitgcmuv_ad with this you need to comment |
48 |
# out dfminFrac, because mitgcmuv_ad does not know about this namelist |
49 |
# parameter (something to be fixed) |
50 |
# - set numiter=100, nfunc=10, or some other large value. nfunc*numiter is |
51 |
# the number of simulations that are allowed in total. This number should be |
52 |
# much larger than numiter, because you may need more than one function call |
53 |
# (= run of mitgcmuv_ad) per iteration, see m1qn3 docs for details |
54 |
# - add an empty namelist &M1QN3 to data.optim |
55 |
# &M1QN3 |
56 |
# / |
57 |
./optim.x > opt0.txt |
58 |
# I get a lot of output in opt0.txt, which is easier to read with |
59 |
# less -S opt0.out (to truncate long lines) |
60 |
# essentially I have the same value for the cost function |
61 |
|
62 |
# ================================================== |
63 |
# Large Scale Optimization with off-line capability. |
64 |
# ================================================== |
65 |
|
66 |
# OPTIM_READPARMS: Control options have been read. |
67 |
# OPTIM_READPARMS: Minimization options have been read. |
68 |
|
69 |
# OPTIM_READDATA: Reading cost function and gradient of cost function |
70 |
# for optimization cycle: 0 |
71 |
|
72 |
# OPTIM_READDATA: opened file ecco_cost_MIT_CE_000.opt0000 |
73 |
# OPTIM_READDATA: nvartype 1 |
74 |
# OPTIM_READDATA: nvarlength 2315 |
75 |
# OPTIM_READDATA: yctrlid MIT_CE_000 |
76 |
# OPTIM_READDATA: filenopt 0 |
77 |
# OPTIM_READDATA: fileff 6.2002322818233591 |
78 |
# OPTIM_READDATA: fileiG 1 |
79 |
# OPTIM_READDATA: filejG 1 |
80 |
# OPTIM_READDATA: filensx 2 |
81 |
# OPTIM_READDATA: filensy 2 |
82 |
# [...] |
83 |
# OPTIM_READDATA: end of optim_readdata |
84 |
|
85 |
|
86 |
# OPTIM_READPARMS: Iteration number = 0 |
87 |
# number of control variables = 2315 |
88 |
# cost function value in ecco_ctrl = 6.2002322818233591 |
89 |
# expected cost function minimum = 5.5802090536410232 |
90 |
# expected cost function decrease = 0.62002322818233591 |
91 |
# Data will be read from the following file: ecco_ctrl_MIT_CE_000.opt0000 |
92 |
|
93 |
# OPTIM_SUB: Calling m1qn3_optim for iteration: 0 |
94 |
# OPTIM_SUB: with nn, REAL_BYTE = 2315 4 |
95 |
# OPTIM_SUB: read model state |
96 |
|
97 |
# OPTIM_READDATA: Reading control vector |
98 |
# for optimization cycle: 0 |
99 |
|
100 |
# OPTIM_READDATA: opened file ecco_ctrl_MIT_CE_000.opt0000 |
101 |
# OPTIM_READDATA: nvartype 1 |
102 |
# OPTIM_READDATA: nvarlength 2315 |
103 |
# OPTIM_READDATA: yctrlid MIT_CE_000 |
104 |
# OPTIM_READDATA: filenopt 0 |
105 |
# OPTIM_READDATA: fileff 6.2002322818233591 |
106 |
# OPTIM_READDATA: fileiG 1 |
107 |
# OPTIM_READDATA: filejG 1 |
108 |
# OPTIM_READDATA: filensx 2 |
109 |
# OPTIM_READDATA: filensy 2 |
110 |
# OPTIM_READDATA: end of optim_readdata |
111 |
|
112 |
|
113 |
# OPTIM_READDATA: Reading cost function and gradient of cost function |
114 |
# for optimization cycle: 0 |
115 |
|
116 |
# OPTIM_READDATA: opened file ecco_cost_MIT_CE_000.opt0000 |
117 |
# OPTIM_READDATA: nvartype 1 |
118 |
# OPTIM_READDATA: nvarlength 2315 |
119 |
# OPTIM_READDATA: yctrlid MIT_CE_000 |
120 |
# OPTIM_READDATA: filenopt 0 |
121 |
# OPTIM_READDATA: fileff 6.2002322818233591 |
122 |
# OPTIM_READDATA: fileiG 1 |
123 |
# OPTIM_READDATA: filejG 1 |
124 |
# OPTIM_READDATA: filensx 2 |
125 |
# OPTIM_READDATA: filensy 2 |
126 |
# OPTIM_READDATA: end of optim_readdata |
127 |
|
128 |
# OPTIM_SUB after reading ecco_ctrl and ecco_cost: |
129 |
# OPTIM_SUB nn = 2315 |
130 |
# OPTIM_SUB objf = 6.2002322818233591 |
131 |
# OPTIM_SUB xx(1) = 0.0000000000000000 |
132 |
# OPTIM_SUB adxx(1) = -6.7879882408306003E-005 |
133 |
# OPTIM_SUB: cold start, optimcycle = 0 |
134 |
# OPTIM_SUB: call m1qn3_offline ........ |
135 |
# OPTIM_SUB: ........................... |
136 |
# OPTIM_SUB: returned from m1qn3_offline |
137 |
# OPTIM_SUB: nn = 2315 |
138 |
# OPTIM_SUB: xx(1) = 0.51934864251896229 0.73000729032300782 |
139 |
# OPTIM_SUB: adxx(1) = -6.7879882408306003E-005 -9.5413379312958568E-005 |
140 |
# OPTIM_SUB: omode = -1 |
141 |
# OPTIM_SUB: niter = 1 |
142 |
# OPTIM_SUB: nsim = 10000 |
143 |
# OPTIM_SUB: reverse = 1 |
144 |
|
145 |
# OPTIM_SUB: mean(xx) = 0.16365068483545642 |
146 |
# OPTIM_SUB: max(xx) = 4.6525355815045790 |
147 |
# OPTIM_SUB: min(xx) = -9.3326211896764324 |
148 |
# OPTIM_SUB: std(xx) = 15.613450260548481 |
149 |
|
150 |
|
151 |
# OPTIM_STORE_M1QN3: saving the state of m1qn3 in OPWARM.opt0001 |
152 |
|
153 |
# OPTIM_SUB: writing 2315 sized control to file ecco_ctrl |
154 |
|
155 |
# OPTIM_WRITEDATA: Writing new control vector to file(s) |
156 |
# for optimization cycle: 1 |
157 |
|
158 |
# OPTIM_WRITEDATA: nvartype 1 |
159 |
# OPTIM_WRITEDATA: nvarlength 2315 |
160 |
# OPTIM_WRITEDATA: yctrlid MIT_CE_000 |
161 |
# OPTIM_WRITEDATA: nopt 1 |
162 |
# OPTIM_WRITEDATA: ff -9999.0000000000000 |
163 |
# OPTIM_WRITEDATA: iG 1 |
164 |
# OPTIM_WRITEDATA: jG 1 |
165 |
# OPTIM_WRITEDATA: nsx 2 |
166 |
# OPTIM_WRITEDATA: nsy 2 |
167 |
# OPTIM_WRITEDATA: end of optim_writedata, icvoffset 2315 |
168 |
|
169 |
|
170 |
# ====================================== |
171 |
# Large Scale Optimization run finished. |
172 |
# ====================================== |
173 |
|
174 |
# ff = -9999 is an intentional dummy value. |
175 |
|
176 |
# now you can organize the rest in a loop. In bash, it could look like this: |
177 |
# first comment out dfminFrac, because mitgcmuv_ad does not know about this, |
178 |
# dfminFrac is really only needed in the zeroth iteration |
179 |
# |
180 |
# tabula rasa |
181 |
rm ecco_c* OPWARM.* m1qn3_output.txt |
182 |
myiter=0 |
183 |
cat > data.optim <<EOF # |
184 |
# ******************************** |
185 |
# Off-line optimization parameters |
186 |
# ******************************** |
187 |
&OPTIM |
188 |
optimcycle=${myiter}, |
189 |
numiter=100, |
190 |
nfunc=100, |
191 |
#dfminFrac = 0.1, |
192 |
iprint=10, |
193 |
nupdate=8, |
194 |
/ |
195 |
|
196 |
&M1QN3 |
197 |
/ |
198 |
EOF |
199 |
|
200 |
while (( $myiter < 20 )) |
201 |
do |
202 |
# formatter iteration count |
203 |
it=`echo $myiter | awk '{printf "%03i",$1}'` |
204 |
echo "iteration ${myiter}" |
205 |
# comment out dfminFrac from data.optim |
206 |
sed -i '' 's/.*dfminFrac.*/#dfminFrac = 0.1,/' data.optim |
207 |
# increment counter in data.optim |
208 |
sed -i .${it} "s/.*optimcycle.*/ optimcycle=${myiter},/" data.optim |
209 |
./mitgcmuv_ad > output${it}.txt |
210 |
# add dfminFrac = 0.1, |
211 |
sed -i '' 's/#dfminFrac.*/ dfminFrac = 0.1,/' data.optim |
212 |
./optim.x > opt${it}.txt |
213 |
# increase counter for next iteration |
214 |
((myiter++)) |
215 |
done |
216 |
|
217 |
# grep "fc " output001.txt |
218 |
# (PID.TID 0000.0001) early fc = 0.000000000000000D+00 |
219 |
# (PID.TID 0000.0001) local fc = 0.132325873958879D+02 |
220 |
# (PID.TID 0000.0001) global fc = 0.132325873958879D+02 |
221 |
|
222 |
# grep "global fc " output???.txt |
223 |
# output000.txt:(PID.TID 0000.0001) global fc = 0.620023228182336D+01 |
224 |
# output001.txt:(PID.TID 0000.0001) global fc = 0.132325873958879D+02 |
225 |
# output002.txt:(PID.TID 0000.0001) global fc = 0.615567811433600D+01 |
226 |
# output003.txt:(PID.TID 0000.0001) global fc = 0.615556878869932D+01 |
227 |
# output004.txt:(PID.TID 0000.0001) global fc = 0.615547009131471D+01 |
228 |
|
229 |
# as you can see, the improvement is not very good after the initial steps. |
230 |
# and the optimization will not be successful in the end, i.e. satisfy |
231 |
# the tolerance set by epsg. |
232 |
# This is likely because the number of timesteps in data is small (20) |
233 |
# for this test. Try longer intergations (e.g. 1 year) and wait longer |
234 |
|
235 |
# check out m1qn3_output.txt, which records the output of m1qn3 |
236 |
# the python script plotfc.py greps the cost function values out of |
237 |
# m1qn3_ouput.txt and plots them and the number of simulations per iteration |
238 |
# which typically increase with decreasing cost function |
239 |
|
240 |
|