-
Notifications
You must be signed in to change notification settings - Fork 3
/
create_dataset.py
318 lines (282 loc) · 11.1 KB
/
create_dataset.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
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
import numpy as np
from torch.utils.data import Dataset, DataLoader
import xarray as xr
import pandas as pd
import datetime
def strTodate(x):
result=''
for i in x.split('-'):
for j in i:
try:
result+=str(int(j))
except:
continue
result+='-'
return result[:-1]
def load_enso_indices(instrument_data):
"""
Reads in the txt data file to output a pandas Series of ENSO vals
outputs
-------
pd.Series : monthly ENSO values starting from 1870-01-01
"""
with open(instrument_data) as f:
line = f.readline()
enso_vals = []
while line:
yearly_enso_vals = map(float, line.split()[1:])
enso_vals.extend(yearly_enso_vals)
line = f.readline()
enso_vals = pd.Series(enso_vals)
enso_vals.index = pd.date_range('1870-01-01',freq='MS',
periods=len(enso_vals))
enso_vals.index = pd.to_datetime(enso_vals.index)
return enso_vals
def assemble_basic_predictors_predictands(opt, train=False):
"""
inputs
------
start_date str : the start date from which to extract sst
end_date str : the end date
lead_time str : the number of months between each sst
value and the target Nino3.4 Index
use_pca bool : whether or not to apply principal components
analysis to the sst field
n_components int : the number of components to use for PCA
outputs
-------
Returns a tuple of the predictors (np array of sst temperature anomalies)
and the predictands (np array the ENSO index at the specified lead time).
"""
start_date = opt.startdate
end_date = opt.enddate
lead_time=opt.leadtime
use_pca = opt.pca
num_input_time_steps = opt.num_input_time_steps
lat_slice=opt.lat_slice
lon_slice=opt.lon_slice
n_components = opt.n_components
fname = opt.dataroot
variable_name = opt.variable_name
data_format = opt.data_format
dataset = opt.dataset
if opt.variable_name:
variable_name = opt.variable_name
else:
variable_name = {'observations' : 'sst',
'observations2': 't2m',
'CNRM' : 'tas',
'MPI' : 'tas'}[dataset]
ds = xr.open_dataset(fname)
sst = ds[variable_name].sel(time=slice(start_date, end_date))
num_time_steps = sst.shape[0]
sst = sst.values.reshape(num_time_steps, -1)
sst[np.isnan(sst)] = 0
if use_pca:
pca = sklearn.decomposition.PCA(n_components=n_components)
pca.fit(sst)
X = pca.transform(sst)
else:
X = sst
if train:
sst1 = ds[variable_name].sel(time=slice(start_date, end_date))
if lat_slice is not None:
try:
sst1=sst1.sel(lat=lat_slice)
except:
raise NotImplementedError("Implement slicing!")
if lon_slice is not None:
try:
sst1=sst1.sel(lon=lon_slice)
except:
raise NotImplementedError("Implement slicing!")
num_samples = sst1.shape[0]
sst1 = np.stack([sst1.values[n-num_input_time_steps:n] for n in range(num_input_time_steps,
num_samples+1)])
sst1[np.isnan(sst1)] = 0
if data_format=='flatten':
sst1 = sst1.reshape(num_samples, -1)
if use_pca:
pca = sklearn.decomposition.PCA(n_components=n_components)
pca.fit(sst1)
X1 = pca.transform(sst1)
else:
X1 = sst1
else: # data_format=='spatial'
X1 = sst1
start_date_plus_lead1 = pd.to_datetime(start_date) + \
pd.DateOffset(months=lead_time+num_input_time_steps-1)
end_date_plus_lead1 = pd.to_datetime(end_date) + \
pd.DateOffset(months=lead_time)
X1 = X1.astype(np.float32)
target_start_date_with_2_month1 = start_date_plus_lead1 - pd.DateOffset(months=2)
subsetted_ds = ds[variable_name].sel(time=slice(target_start_date_with_2_month1,
end_date_plus_lead1))
#Calculate the Nino3.4 index
y = subsetted_ds.sel(lat=slice(5,-5), lon=slice(360-170,360-120)).mean(dim=('lat','lon'))
y = pd.Series(y.values).rolling(window=3).mean()[2:].values
y = y.astype(np.float32)
#X = X[1:]
else:
start_date_plus_lead = pd.to_datetime(start_date) + \
pd.DateOffset(months=lead_time)
end_date_plus_lead = pd.to_datetime(end_date) + \
pd.DateOffset(months=lead_time)
if opt.compare_ground_truth:
y = load_enso_indices(opt.instrument_data)[slice(start_date_plus_lead,
end_date_plus_lead)]
else:
y = np.array([0]*X.shape[0])
## ds = xr.open_dataset(opt.dataroot)
## sst = ds['sst'].sel(time=slice(start_date, end_date))
## num_time_steps = sst.shape[0]
##
## #sst is a 3D array: (time_steps, lat, lon)
## #in this tutorial, we will not be using ML models that take
## #advantage of the spatial nature of global temperature
## #therefore, we reshape sst into a 2D array: (time_steps, lat*lon)
## #(At each time step, there are lat*lon predictors)
## sst = sst.values.reshape(num_time_steps, -1)
## sst[np.isnan(sst)] = 0
##
## #Use Principal Components Analysis, also called
## #Empirical Orthogonal Functions, to reduce the
## #dimensionality of the array
## if use_pca:
## pca = sklearn.decomposition.PCA(n_components=n_components)
## pca.fit(sst)
## X = pca.transform(sst)
## else:
## X = sst
##
## start_date_plus_lead = pd.to_datetime(start_date) + \
## pd.DateOffset(months=lead_time)
## end_date_plus_lead = pd.to_datetime(end_date) + \
## pd.DateOffset(months=lead_time)
## y = load_enso_indices(opt.instrument_data)[slice(start_date_plus_lead,
## end_date_plus_lead)]
## #print(type(X))
## #print(type(y))
##
##
## ds.close()
return X, y
def assemble_predictors_predictands(opt,train=False):
"""
inputs
------
start_date str : the start date from which to extract sst
end_date str : the end date
lead_time str : the number of months between each sst
value and the target Nino3.4 Index
dataset str : 'observations' 'CNRM' or 'MPI'
data_format str : 'spatial' or 'flatten'. 'spatial' preserves
the lat/lon dimensions and returns an
array of shape (num_samples, num_input_time_steps,
lat, lon). 'flatten' returns an array of shape
(num_samples, num_input_time_steps*lat*lon)
num_input_time_steps int : the number of time steps to use for each
predictor sample
use_pca bool : whether or not to apply principal components
analysis to the sst field
n_components int : the number of components to use for PCA
lat_slice slice: the slice of latitudes to use
lon_slice slice: the slice of longitudes to use
outputs
-------
Returns a tuple of the predictors (np array of sst temperature anomalies)
and the predictands (np array the ENSO index at the specified lead time).
"""
if train:
start_date = opt.startdate
end_date = opt.enddate
else:
start_date = opt.test_start
end_date = opt.test_end
lead_time=opt.leadtime
use_pca = opt.pca
n_components = opt.n_components
dataset = opt.dataset
data_format = opt.data_format
num_input_time_steps = opt.num_input_time_steps
lat_slice = opt.lat_slice
lon_slice = opt.lon_slice
file_name = opt.dataroot
if opt.variable_name:
variable_name = opt.variable_name
else:
variable_name = {'observations' : 'sst',
'observations2': 't2m',
'CNRM' : 'tas',
'MPI' : 'tas'}[dataset]
ds = xr.open_dataset(file_name)
sst = ds[variable_name].sel(time=slice(start_date, end_date))
if lat_slice is not None:
try:
sst=sst.sel(lat=lat_slice)
except:
raise NotImplementedError("Implement slicing!")
if lon_slice is not None:
try:
sst=sst.sel(lon=lon_slice)
except:
raise NotImplementedError("Implement slicing!")
num_samples = sst.shape[0]
#sst is a (num_samples, lat, lon) array
#the line below converts it to (num_samples, num_input_time_steps, lat, lon)
sst = np.stack([sst.values[n-num_input_time_steps:n] for n in range(num_input_time_steps,
num_samples+1)])
#CHALLENGE: CAN YOU IMPLEMENT THE ABOVE LINE WITHOUT A FOR LOOP?
num_samples = sst.shape[0]
sst[np.isnan(sst)] = 0
if data_format=='flatten':
#sst is a 3D array: (time_steps, lat, lon)
#in this tutorial, we will not be using ML models that take
#advantage of the spatial nature of global temperature
#therefore, we reshape sst into a 2D array: (time_steps, lat*lon)
#(At each time step, there are lat*lon predictors)
sst = sst.reshape(num_samples, -1)
#Use Principal Components Analysis, also called
#Empirical Orthogonal Functions, to reduce the
#dimensionality of the array
if use_pca:
pca = sklearn.decomposition.PCA(n_components=n_components)
pca.fit(sst)
X = pca.transform(sst)
else:
X = sst
else: # data_format=='spatial'
X = sst
start_date_plus_lead = pd.to_datetime(start_date) + \
pd.DateOffset(months=lead_time+num_input_time_steps-1)
end_date_plus_lead = pd.to_datetime(end_date) + \
pd.DateOffset(months=lead_time)
if dataset == 'observations' and opt.compare_ground_truth == True:
y = load_enso_indices(opt.instrument_data)[slice(start_date_plus_lead,
end_date_plus_lead)]
elif not opt.compare_ground_truth:
y=np.array([0]*X.shape[0])
else: #the data is from a GCM
X = X.astype(np.float32)
#The Nino3.4 Index is composed of three month rolling values
#Therefore, when calculating the Nino3.4 Index in a GCM
#we have to extract the two months prior to the first target start date
target_start_date_with_2_month = start_date_plus_lead - pd.DateOffset(months=2)
subsetted_ds = ds[variable_name].sel(time=slice(target_start_date_with_2_month,
end_date_plus_lead))
#Calculate the Nino3.4 index
y = subsetted_ds.sel(lat=slice(5,-5), lon=slice(360-170,360-120)).mean(dim=('lat','lon'))
y = pd.Series(y.values).rolling(window=3).mean()[2:].values
y = y.astype(np.float32)
ds.close()
return X.astype(np.float32), y.astype(np.float32)
class ENSODataset(Dataset):
def __init__(self, predictors, predictands):
self.predictors = predictors
self.predictands = predictands
assert self.predictors.shape[0] == self.predictands.shape[0], \
"The number of predictors must equal the number of predictands!"
def __len__(self):
return self.predictors.shape[0]
def __getitem__(self, idx):
return self.predictors[idx], self.predictands[idx]