-
Notifications
You must be signed in to change notification settings - Fork 0
/
bjpmodel.py
256 lines (161 loc) · 8.56 KB
/
bjpmodel.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
"""! Andrew's file"""
import os, sys, shutil
from datetime import datetime
import numpy as np
import pytrans
import pybjp
class BjpModel:
OBSERVED_DATA_CODE = 1
CENSORED_DATA_CODE = 2
MISSING_DATA_CODE = 3
MISSING_DATA_VALUE = -9999.0
FIXED_RANDOM_SEED = 5
NO_CENS_THRESH = -99999.0
ZERO_CENS_THRESH = 0.0
GAUGE_CENS_THRESH = 0.2
def __init__(self, num_vars, groups, burn=3000, chainlength=7000, seed='random'):
self.num_vars = num_vars
self.burn = burn
self.chainlength = chainlength
assert (seed == 'random' or seed == 'fixed')
if seed == 'fixed':
np.random.seed(self.FIXED_RANDOM_SEED)
self.seed = self.FIXED_RANDOM_SEED
else:
self.seed = np.random.randint(0,100000)
self.groups = groups
self.censors = []
for i in range(num_vars):
if groups[i] == 10 or groups[i] == 20:
censor = self.ZERO_CENS_THRESH
elif groups[i] == 30 or groups[i] == 100:
censor = self.NO_CENS_THRESH
elif groups[i] == 50:
censor = self.GAUGE_CENS_THRESH
self.censors.append(censor)
self.censors = np.array(self.censors)
self.bjp_wrapper = None
self.bjp_fitting_data = None
def prepare_bjp_data(self, fit_data, group, censor, trformer=None):
fit_data = np.array(fit_data, copy=True)
# Set the censored and missing data flags and adjust data accordingly
flags = np.ones(fit_data.shape, dtype='intc', order='C')*self.OBSERVED_DATA_CODE
censor_idx = fit_data <= censor
flags[censor_idx] = self.CENSORED_DATA_CODE
# Treat both -9999.0 and np.nan as missing values in the input data
missing_idx = np.abs(fit_data - self.MISSING_DATA_VALUE) < 1E-6
flags[missing_idx] = self.MISSING_DATA_CODE
missing_idx2 = np.isnan(fit_data)
flags[missing_idx2] = self.MISSING_DATA_CODE
# set NaNs to missing data value
fit_data[np.isnan(fit_data)] = self.MISSING_DATA_VALUE
if trformer is None:
# remove missing values before estimating transformation parameters
fit_data_for_trans = np.array(fit_data, copy=True)
fit_data_for_trans = fit_data_for_trans[flags != self.MISSING_DATA_CODE]
if group == 10 or group == 50:
#print(np.nanmax(fit_data_for_trans))
trformer = pytrans.PyLogSinh(scale=5.0/np.max(fit_data_for_trans))
trformer.optim_params(fit_data_for_trans, censor, do_rescale=True, is_map=True)
elif group == 20 or group == 30:
trformer = pytrans.PyYJT(scale=1.0/np.std(fit_data_for_trans), shift=-1.0*np.mean(fit_data_for_trans))
trformer.optim_params(fit_data_for_trans, censor, do_rescale=True, is_map=True)
else:
print('Transformation group code not recognised. Exiting.')
sys.exit()
fit_data[fit_data < censor] = censor
rs_data = trformer.rescale_many(fit_data)
tr_data = trformer.transform_many(rs_data)
rs_censor = trformer.rescale_one(censor)
tr_censor = trformer.transform_one(rs_censor)
# restore some dummy values for missing data
fit_data[np.isnan(fit_data)] = self.MISSING_DATA_VALUE
bjp_data = {}
bjp_data['trformer'] = trformer
bjp_data['tr_data'] = tr_data
bjp_data['censor'] = censor
bjp_data['tr_censor'] = tr_censor
bjp_data['flags'] = flags
return bjp_data
def prepare_fc_data(self, predictor_values, transformers):
assert len(transformers) == len(self.censors)
if np.any(np.isnan(predictor_values)) or np.any(predictor_values == self.MISSING_DATA_VALUE):
print("Warning: Predictor is NaN or missing")
predictor_values[np.isnan(predictor_values)] = -9999.0
# should we limit extreme new predictor values ???
bjp_data_new = {}
bjp_data_new['tr_data'] = np.array([self.MISSING_DATA_VALUE]*self.num_vars)
bjp_data_new['flags'] = np.array([self.MISSING_DATA_CODE]*self.num_vars, dtype='intc')
bjp_data_new['censor'] = np.array([self.censors[i] for i in range(self.num_vars)])
bjp_data_new['tr_censor'] = np.array([transformers[i].transform_one(transformers[i].rescale_one(self.censors[i])) for i in range(self.num_vars)])
for i in range(len(predictor_values)):
if np.abs(predictor_values[i] - self.MISSING_DATA_VALUE) < 1E-6:
bjp_data_new['flags'][i] = self.MISSING_DATA_CODE
elif predictor_values[i] <= bjp_data_new['censor'][i]:
bjp_data_new['flags'][i] = self.CENSORED_DATA_CODE
predictor_values[i] = self.censors[i]
else:
bjp_data_new['flags'][i] = self.OBSERVED_DATA_CODE
trformer = transformers[i]
bjp_data_new['trformer'] = trformer
rs_pred = trformer.rescale_one(predictor_values[i])
tr_pred = trformer.transform_one(rs_pred)
#print(predictor_values[i], tr_pred, rs_pred, trformer.get_params())
bjp_data_new['tr_data'][i] = tr_pred
return bjp_data_new
def join_ptor_ptand_data(self, bjp_fitting_data):
joined_data = {'trformer': [], 'tr_data': [], 'censor': [], 'tr_censor': [], 'flags':[]}
for i in range(len(bjp_fitting_data)):
joined_data['trformer'].append(bjp_fitting_data[i]['trformer'])
joined_data['tr_data'].append(bjp_fitting_data[i]['tr_data'])
joined_data['censor'].append(bjp_fitting_data[i]['censor'])
joined_data['tr_censor'].append(bjp_fitting_data[i]['tr_censor'])
joined_data['flags'].append(bjp_fitting_data[i]['flags'])
joined_data['tr_data'] = np.array(joined_data['tr_data'], order='C')
joined_data['censor'] = np.array(joined_data['censor'], order='C')
joined_data['tr_censor'] = np.array(joined_data['tr_censor'], order='C')
joined_data['flags'] = np.array(joined_data['flags'], order='C')
return joined_data
def inv_transform(self, data, trformer):
inv_tr_data = trformer.inv_transform_many(data)
inv_rs_data = trformer.inv_rescale_many(inv_tr_data)
return inv_rs_data
def sample(self, obs):
# obs has dimensions num_vars x num_time_periods
bjp_fitting_data = []
for i in range(self.num_vars):
bjp_fitting_data.append(self.prepare_bjp_data(obs[i], self.groups[i], self.censors[i]))
bjp_fitting_data = self.join_ptor_ptand_data(bjp_fitting_data)
bjp_wrapper = pybjp.PyBJP(self.num_vars, self.burn, self.chainlength, self.seed)
mu, cov = bjp_wrapper.sample(bjp_fitting_data['tr_data'], bjp_fitting_data['flags'], bjp_fitting_data['tr_censor'])
self.bjp_wrapper = bjp_wrapper
self.bjp_fitting_data = bjp_fitting_data
tparams = []
for trformer in bjp_fitting_data['trformer']:
tparams.append(trformer.get_params())
tparams = np.array(tparams)
return mu, cov, tparams
def forecast(self, predictor_values, transformers, mu, cov, gen_climatology=False, convert_cens=True):
self.bjp_wrapper = pybjp.PyBJP(self.num_vars, self.burn, self.chainlength, self.seed)
bjp_fc_data = self.prepare_fc_data(predictor_values, transformers)
#print(bjp_fc_data['tr_data'].dtype, bjp_fc_data['flags'].dtype, bjp_fc_data['tr_censor'].dtype, mu.dtype, cov.dtype, (cov.shape[0] / 2))
forecasts = self.bjp_wrapper.forecast2(bjp_fc_data['tr_data'], bjp_fc_data['flags'], bjp_fc_data['tr_censor'], mu.astype(np.float64), cov.astype(np.float64), int(cov.shape[0] / 2))
for i in range(self.num_vars):
trformer = transformers[i]
forecasts[:, i] = self.inv_transform(forecasts[:, i], trformer)
if convert_cens:
cens = self.censors[i]
forecasts[:, i][forecasts[:, i] < cens] = cens
res = {}
res['forecast'] = forecasts
if gen_climatology:
bjp_clims = self.bjp_wrapper.gen_climatology()
clim = np.empty(forecasts.shape)*np.nan
for i in range(self.num_vars):
trformer = self.bjp_fitting_data['trformer'][i]
clim[:,i] = self.inv_transform(bjp_clims[:, i], trformer)
if convert_cens:
cens = bjp_fc_data['censor'][i]
clim[:, i][clim[:, i] < cens] = cens
res['clim'] = clim
return res