-
Notifications
You must be signed in to change notification settings - Fork 0
/
assimilate.py
263 lines (221 loc) · 9.16 KB
/
assimilate.py
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
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
#!/bin/env python
# -*- coding: utf-8 -*-
#
# Created on 05/06/2023
# Created for 2022_ddim_for_attractors
#
# @author: Tobias Sebastian Finn, [email protected]
#
# Copyright (C) {2023} {Tobias Sebastian Finn}
# System modules
import logging
from collections.abc import MutableMapping
# External modules
import torch
import pytorch_lightning as pl
import hydra
from hydra.utils import instantiate
import wandb
from omegaconf import DictConfig
from tqdm import tqdm
# Internal modules
from dyn_ddim.dynamical import RK4Integrator, Lorenz63
import dyn_ddim.climatology as clim
main_logger = logging.getLogger(__name__)
def flatten_dict(dictionary, parent_key='', separator='.'):
# Based on https://stackoverflow.com/a/6027615
items = []
for key, value in dictionary.items():
new_key = parent_key + separator + key if parent_key else key
if isinstance(value, MutableMapping):
items.extend(
flatten_dict(value, new_key, separator=separator).items()
)
else:
items.append((new_key, value))
return dict(items)
@hydra.main(
version_base=None, config_path='configs/', config_name='assimilate'
)
def main_assimilate(cfg: DictConfig):
if cfg.get("seed"):
pl.seed_everything(cfg.seed, workers=True)
# Load the dataset
truth = torch.load(cfg.dataset_path)
clim_scaling = torch.tensor(clim.scale)
truth = truth * clim_scaling + torch.tensor(clim.shift)
main_logger.info("Loaded the truth")
# Generate the observations
obs_op = lambda state: state[..., cfg.obs_list]
obs = obs_op(truth)
obs = obs + torch.randn_like(obs) * cfg.obs_std
main_logger.info("Generated the observations")
# Generated initial ensemble
n_ens_initial = truth.size(0) * cfg.n_ens
state_clim = truth.view(-1, 3)
ens_initial = state_clim[
torch.randperm(state_clim.size(0))[:n_ens_initial]
]
ens_initial = ens_initial.reshape(truth.size(0), cfg.n_ens, 3)
main_logger.info("Generate initial ensemble")
assimilation = instantiate(
cfg.assimilation, obs_std=cfg.obs_std, obs_op=obs_op
)
main_logger.info(f"Instantiated {cfg.assimilation._target_}")
model = Lorenz63()
integrator = RK4Integrator(model, dt=0.01)
main_logger.info("Instantiated the model")
# Instantiate logger
_ = instantiate(cfg.logger)
wandb.config["assimilation"] = flatten_dict(cfg.assimilation)
wandb.config["obs_std"] = cfg.obs_std
wandb.config["obs_every"] = cfg.obs_every
wandb.config["obs_list"] = cfg.obs_list
# Burn in phase
curr_state = ens_initial.clone()
mse = 0
spread = 0
n_stat_steps = 0
time_pbar = tqdm(range(1, (cfg.obs_every*cfg.n_burn_in)+1))
for burn_time in time_pbar:
curr_state = integrator.integrate(curr_state)
if (burn_time % cfg.obs_every) == 0:
# Assimilate
curr_state, _, curr_ens = assimilation.assimilate(
curr_state, obs[:, [burn_time]]
)
# Estimate statistics
old_stat_steps = n_stat_steps
n_stat_steps = n_stat_steps + 1
old_weight = old_stat_steps / n_stat_steps
# Update MSE and spread
curr_mse = (
curr_state.mean(dim=-2)-truth[:, burn_time]
).pow(2).mean(dim=0)
mse = mse * old_weight + curr_mse / n_stat_steps
curr_spread = curr_ens.var(dim=-2).mean(dim=0)
spread = spread * old_weight + curr_spread / n_stat_steps
# Estimate local statistics
curr_nrmse = (curr_mse/clim_scaling.pow(2)).mean().sqrt().item()
curr_nspread = (
curr_spread/clim_scaling.pow(2)
).mean().sqrt().item()
ana_nrmse = (mse/clim_scaling.pow(2)).mean().sqrt().item()
ana_nspread = (spread/clim_scaling.pow(2)).mean().sqrt().item()
wandb.log({
"assim/curr_rmse": curr_nrmse,
"assim/curr_spread": curr_nspread,
"assim/ana_rmse": ana_nrmse,
"assim/ana_spread": ana_nspread
},)
# Update rolling statistics
time_pbar.set_postfix(
curr_rmse=curr_nrmse, curr_spread=curr_nspread,
ana_rmse=ana_nrmse, ana_spread=ana_nspread
)
main_logger.info("Burn-in phase finished")
# Estimate statistics
total_steps = cfg.obs_every * cfg.n_cycles
n_stat_steps = 0
mse = torch.zeros(curr_state.size(0), cfg.obs_every+1, 3)
spread = torch.zeros(curr_state.size(0), cfg.obs_every+1, 3)
ana_mse = 0
ana_spread = 0
bg_mse = 0
bg_spread = 0
cov_ana = torch.zeros(3, 3)
cov_bg = torch.zeros(3, 3)
curr_traj = [curr_state.clone()]
time_pbar = tqdm(range(burn_time+1, burn_time+total_steps+1))
for t in time_pbar:
curr_traj.append(integrator.integrate(curr_traj[-1]))
if (t % cfg.obs_every) == 0:
# Estimate statistics
old_stat_steps = n_stat_steps
n_stat_steps = n_stat_steps + 1
old_weight = old_stat_steps / n_stat_steps
# Concatenate trajectory
curr_traj = torch.stack(curr_traj, dim=1)
# Update MSE and spread
curr_mse = (
curr_traj.mean(dim=-2)-truth[:, t-cfg.obs_every:t+1]
).pow(2)
mse = mse * old_weight + curr_mse / n_stat_steps
if cfg.n_ens > 1:
spread = spread * old_weight \
+ curr_traj.var(dim=-2) / n_stat_steps
# Update bg cov
bg_mean = curr_traj[:, -1].mean(dim=-2, keepdims=True)
bg_perts = curr_traj[:, -1]-bg_mean
curr_cov = torch.einsum("big,bih->bgh", bg_perts, bg_perts)
curr_cov = curr_cov.mean(dim=0) / (bg_perts.size(-2)-1)
cov_bg = cov_bg * old_weight + curr_cov / n_stat_steps
# Assimilate
analysis, bg_ens, ana_ens = assimilation.assimilate(
curr_traj[:, -1], obs[:, [t]]
)
# Update background scores
curr_bg_mse = (bg_ens.mean(dim=-2)-truth[:, t]).pow(2).mean(dim=0)
bg_mse = bg_mse * old_weight + curr_bg_mse / n_stat_steps
curr_bg_spread = bg_ens.var(dim=-2).mean(dim=0)
bg_spread = bg_spread * old_weight + curr_bg_spread / n_stat_steps
# Update analysis scores
curr_ana_mse = (analysis.mean(dim=-2)-truth[:, t]).pow(2).mean(dim=0)
ana_mse = ana_mse * old_weight + curr_ana_mse / n_stat_steps
curr_ana_spread = ana_ens.var(dim=-2).mean(dim=0)
ana_spread = ana_spread * old_weight + curr_ana_spread / n_stat_steps
# Update ana cov
ana_perts = ana_ens-ana_ens.mean(dim=-2, keepdims=True)
curr_cov = torch.einsum("big,bih->bgh", ana_perts, ana_perts)
curr_cov = curr_cov.mean(dim=0) / (ana_perts.size(-2)-1)
cov_ana = cov_ana * old_weight + curr_cov / n_stat_steps
# Reset curr_traj
curr_traj = [analysis]
# Estimate local statistics
curr_nrmse = (curr_ana_mse/clim_scaling.pow(2)).mean().sqrt().item()
curr_nspread = (
curr_ana_spread/clim_scaling.pow(2)
).mean().sqrt().item()
bg_nrmse = (bg_mse/clim_scaling.pow(2)).mean().sqrt().item()
bg_nspread = (bg_spread/clim_scaling.pow(2)).mean().sqrt().item()
ana_nrmse = (ana_mse/clim_scaling.pow(2)).mean().sqrt().item()
ana_nspread = (ana_spread/clim_scaling.pow(2)).mean().sqrt().item()
wandb.log({
"assim/curr_rmse": curr_nrmse,
"assim/curr_spread": curr_nspread,
"assim/bg_rmse": bg_nrmse,
"assim/bg_spread": bg_nspread,
"assim/ana_rmse": ana_nrmse,
"assim/ana_spread": ana_nspread,
},)
# Update rolling statistics
time_pbar.set_postfix(
curr_rmse=curr_nrmse, curr_spread=curr_nspread,
ana_rmse=ana_nrmse, ana_spread=ana_nspread
)
wandb.define_metric("lead_time")
wandb.define_metric("assim/rmse_mean", step_metric="lead_time")
wandb.define_metric("assim/rmse_std", step_metric="lead_time")
wandb.define_metric("assim/spread_mean", step_metric="lead_time")
rmse = (mse / clim_scaling.pow(2)).mean(dim=-1).sqrt()
spread = (spread / clim_scaling.pow(2)).mean(dim=-1).sqrt()
rmse_mean = rmse.mean(dim=0)
rmse_std = rmse.std(dim=0)
spread_mean = spread.mean(dim=0)
all_scores = zip(rmse_mean, rmse_std, spread_mean)
for ld, scores in enumerate(all_scores):
score_dict = {
"assim/rmse_mean": scores[0],
"assim/rmse_std": scores[1],
"assim/spread_mean": scores[2],
"lead_time": ld
}
wandb.log(score_dict)
wandb.run.summary["cov_bg"] = cov_bg
wandb.run.summary["cov_ana"] = cov_ana
wandb.finish()
rmse_norm = (mse[0]/clim_scaling.pow(2)).mean().sqrt()
main_logger.info("RMSE: {0:.2f}".format(rmse_norm))
return rmse_norm
if __name__ == "__main__":
main_assimilate()