-
Notifications
You must be signed in to change notification settings - Fork 0
/
EXP_IDS_FOR_MANUSCRIPT
139 lines (102 loc) · 4.73 KB
/
EXP_IDS_FOR_MANUSCRIPT
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
Full list of experiments for manuscript results and figures:
------------------------------------------------------------
*exp_id numbers* refer to the leading digits of the enumerated .yml files found
in root/experiments, root/experiments_DA and root/experiments_parametrization.
exp_id numbers were given to experiments chronologically during the development
of this repository.
*exp_id folder names* refer to the name of the folder containing the experiment
results in root/results/emulators/L96/models/*exp_id folder name* resp.
root/results/emulators/L96/results/data_assimilation/*exp_id folder name*,
and root/results/emulators/L96/results/parametrization/*exp_id folder name*.
##################################################
# root/experiments: Network-training experiments #
##################################################
Training of deepNets on K=40, F=8 and different dataset sizes N :
- Resilts for evluation of emulators (Fig. 2), but some networks used throughout the study!
exp_id number # exp_id folder name
77 # deepNet_predictState_J_initRand_RK_multiTrial01 *best of three seeds
85 # deepNet_predictState_J_initRand_RK_multiTrial01_seed2
86 # deepNet_predictState_J_initRand_RK_multiTrial01_seed3
97 # deepNet_predictState_J_initRand_RK_multiTrial07 *best of three seeds
98 # deepNet_predictState_J_initRand_RK_multiTrial07_seed2
99 # deepNet_predictState_J_initRand_RK_multiTrial07_seed3
74 # deepNet_predictState_J_initRand_RK_multiTrial1
89 # deepNet_predictState_J_initRand_RK_multiTrial1_seed2
90 # deepNet_predictState_J_initRand_RK_multiTrial1_seed3 * best of three seeds
71 # deepNet_predictState_J_initRand_RK_multiTrial10 * best of three seeds
87 # deepNet_predictState_J_initRand_RK_multiTrial10_seed2
88 # deepNet_predictState_J_initRand_RK_multiTrial10_seed3
Local training of cheap emulators on parts of whole space with in total K = 640 :
- Provides results for local training of emulators (Fig. A3).
exp_id folder name:
minimalNet_predictState_J_initRand_RK_multiTrial_thorough_local***_bs**
exp_id number # batch-size, local region size
49 # bs=1, K_local=640
50 # bs=1, K_local=160
51 # bs=1, K_local=40
52 # bs=1, K_local=10
53 # bs=4, K_local=640
54 # bs=4, K_local=160
55 # bs=4, K_local=40
56 # bs=4, K_local=10
57 # bs=16, K_local=640
58 # bs=16, K_local=160
59 # bs=16, K_local=40
60 # bs=16, K_local=10
61 # bs=64, K_local=640
62 # bs=64, K_local=160
63 # bs=64, K_local=40
64 # bs=64, K_local=10
Training of deepNet on K=36, F=10, dt=0.01, N=1200 for parametrization learning:
- Provides emulator for parametrization learning (Fig. 4).
exp_id number # exp_id folder name
104 # deepNet_predictState_J_K36_initRand_RK_multiTrial01
Dummy ‘training’ file for bilinear network with analytic weights (notice learning rate = 0.):
- Bilinear network with analytical parameters acts as ground-truth L96 implementation in pytorch.
exp_id number # exp_id folder name
35 # bilinearNet_predictState_J_analytic_RK_multiTrial_thorough_local
Training of small networks (bilinearNet, squareNet) on K=40, F=8 and various dataset sizes N:
- Results for problem-tailored neural netwoks (Fig. A2).
exp_id number # exp_id folder name
75 # minimalNet_predictState_J_initRand_RK_multiTrial01
72 # minimalNet_predictState_J_initRand_RK_multiTrial1
69 # minimalNet_predictState_J_initRand_RK_multiTrial10
76 # bilinearNet_predictState_J_initRand_RK_multiTrial01
73 # bilinearNet_predictState_J_initRand_RK_multiTrial1
70 # bilinearNet_predictState_J_initRand_RK_multiTrial10
######################################################
# root/experiments_DA: Data Assimilation experiments #
######################################################
Provide results of Data Assimilation with emulators (Fig. 3).
Experiments with trained emulator:
exp_id folder name
4DVar_oneLevelK40_deepEmulator_Twin_**_N01_overlap_0_long_B1_0_slowDynamics
exp_id number # integration window length (in steps)
92 # Twin=8
91 # Twin=16
90 # Twin=24
89 # Twin=32
88 # Twin=40
87 # Twin=48
86 # Twin=56
84 # Twin=64
85 # Twin=72
Experiments with true Lorenz-96 model (here: bilinear network with analytic weights):
exp_id folder name
4DVar_oneLevelK40_analyticEmulator_Twin_**_overlap_0_long_B1_0_slowDynamics
exp_id number # integration window length (in steps)
76 # Twin=8,
72 # Twin=16
68 # Twin=24
64 # Twin=32
60 # Twin=40
56 # Twin=48
52 # Twin=56
48 # Twin=64
80 # Twin=72
##########################################################################
# root/experiments_parametrization: Parametrization learning experiments #
##########################################################################
Provides final results of Parametrization learning with emulators (Fig. 4).
exp_id number # exp_id folder name
22 # parametrization_deepNet_nnParam_fastDynamics_T5_nrollout10_cheaperNet