-
Notifications
You must be signed in to change notification settings - Fork 0
/
MAIN_Prediction_Pensions.py
307 lines (253 loc) · 15.1 KB
/
MAIN_Prediction_Pensions.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
#!/usr/bin/env python
# coding: utf-8
# # Pensions
#################################### Imports Packages ###############################################
import pandas as pd
import numpy as np
import seaborn as sns
from scipy.stats import gamma, truncnorm, describe
from sklearn.preprocessing import minmax_scale
from sklearn.model_selection import KFold, StratifiedKFold, GridSearchCV
import matplotlib.pyplot as plt
import time, json, pickle, os
import tensorflow as tf
from tensorflow.keras.models import Sequential, Model, load_model, clone_model
from tensorflow.keras.optimizers import Adam, SGD, Adadelta
from tensorflow.keras.layers import Dense, Dropout, Activation, RepeatVector, Average, LSTM, Lambda, Input
from tensorflow.keras.callbacks import ModelCheckpoint, TensorBoard, EarlyStopping
from data_prep_General import data_re_transform_features
from rnn_functions import create_multiple_rnn_models, train_individual_ensembles
from statistical_analysis_functions import create_df_model_comparison, model_examine_indivual_fit
from clustering import analyze_agglomeration_test, kmeans_counts
from visualization_functions import visualize_representatives_km_ann
### Paths
cd = os.getcwd() + '/Pensions' #r"C:\Users\mark.kiermayer\Documents\Python Scripts\NEW Paper (Grouping) - Code - V1\Pensions"
path_data = cd + '/Data/' # import data
wd_rnn = cd + '/ipynb_Checkpoints/Prediction' # save/ load rnns
# dummy if saved models should be loaded (TRUE) or the all models should be recalculated (False)
dummy_load_saved_models = True
bool_fine_tune = True
bool_latex = True
BATCH_replica = 64
val_share = 0.25
tf_strategy = tf.distribute.MirroredStrategy()#['/gpu:0']) # select specific GPUs
BATCH = BATCH_replica*tf_strategy.num_replicas_in_sync
print('---------------------------------------------------------')
print("Num GPUs Available: ", tf_strategy.num_replicas_in_sync)
print('---------------------------------------------------------')
### load data
X_train = pd.read_csv(path_data+'X_train.csv', index_col= 0).values
X_test = pd.read_csv(path_data+'X_test.csv', index_col= 0).values
y_train = pd.read_csv(path_data+'y_train.csv', index_col= 0).values
y_test = pd.read_csv(path_data+'y_test.csv', index_col= 0).values
X_train_raw = pd.read_csv(path_data+'X_train_raw.csv', index_col= 0).values
X_test_raw = pd.read_csv(path_data+'X_test_raw.csv', index_col= 0).values
# Load general assumptions
with open(path_data+'Pension_params.pkl', 'rb') as f:
params = pickle.load(f)
with open(path_data+'Pension_explan_vars_range.pkl', 'rb') as f:
explan_vars_range = pickle.load(f)
#################################### Assumptions ###############################################
print('Parameters:' )
print(params)
print('Explanatory variables: ')
print(explan_vars_range)
# Dataframe representation
pd.set_option('precision', 2)
pd.set_option('display.max_colwidth', 40)
# Range of Variables
# Matrix Version of previous upper/ lower bounds on features
Min_Max = np.array([explan_vars_range['fund'][0],explan_vars_range['fund'][1],
explan_vars_range['age'][0], explan_vars_range['age'][1],
explan_vars_range['salary'][0], explan_vars_range['salary'][1],
explan_vars_range['salary_scale'][0], explan_vars_range['salary_scale'][1],
explan_vars_range['contribution'][0], explan_vars_range['contribution'][1]]).reshape(-1,2)
#################################### Build Prediction Models ###############################################
# General settings
n_output = params['pension_age_max']-explan_vars_range['age'][0]+1
n_in = X_train.shape[1]
es = EarlyStopping(monitor= 'val_loss', patience= 100 )
es_patience = 100
N_epochs = 1500
### Single Model Configurations
### MSE Training
N_ensembles = 10
# Create Multiple RNNs with identical configuration
models_mse_hist = {}
with tf_strategy.scope():
INPUT = Input(shape = (n_in,))
models_mse = create_multiple_rnn_models(number=N_ensembles, model_input=INPUT,widths_rnn =[n_output],
widths_ffn = [n_output],
dense_act_fct = 'tanh', optimizer_type='adam', loss_type='mse',
metric_type='mae', dropout_share=0,
lambda_layer = True, lambda_scale =params['V_max'], log_scale=True,
model_compile = True, return_option = 'model', branch_name = '')
# Either load model(s) or train them
if os.path.isfile(wd_rnn+r'/mse/model_0.h5') & dummy_load_saved_models:
# load model weights
for i in range(N_ensembles):
models_mse[i].load_weights(wd_rnn+r'/mse/model_{}.h5'.format(i))
with open(wd_rnn+r'/mse/model_{}_hist.json'.format(i), 'rb') as f:
models_mse_hist[i] = pickle.load(f)
print('MSE prediction models loaded!')
else:
# Train multiple RNNs with identical configuration
for i in range(len(models_mse)):
models_mse[i].fit(x=X_train, y=y_train, validation_split = val_share, batch_size= BATCH, epochs=N_epochs, callbacks=[es], verbose = 2)
models_mse[i].save_weights(wd_rnn+r'/mse/model_{}.h5'.format(i))
models_mse_hist[i] = models_mse[i].history
with open(wd_rnn+r'/mse/model_{}_hist.json'.format(i), 'wb') as f: # alternative: option #'w'
pickle.dump(models_mse_hist[i], f, pickle.HIGHEST_PROTOCOL)
#models_mse, models_mse_hist = train_individual_ensembles(models_mse, X_train, y_train,
# n_epochs= N_epochs,
# n_batch= BATCH, es_patience= es_patience,
# path = wd_rnn+r'/mse')
# Note: Save Model (and History) is integrated in function 'train_individual_ensembles'
#### MAE Training
N_ensembles = 10
# Create Multiple RNNs with identical configuration
models_mae_hist = {}
with tf_strategy.scope():
INPUT = Input(shape = (n_in,))
models_mae = create_multiple_rnn_models(number=N_ensembles, model_input=INPUT,widths_rnn =[n_output],
widths_ffn = [n_output],
dense_act_fct = 'tanh', optimizer_type='adam', loss_type='mae',
metric_type='mae', dropout_share=0,
lambda_layer = True, lambda_scale =params['V_max'], log_scale=True,
model_compile = True, return_option = 'model', branch_name = '')
# Either load model(s) or train them
if os.path.isfile(wd_rnn+r'/mae/model_0.h5') & dummy_load_saved_models:
# load model weights
for i in range(N_ensembles):
models_mae[i].load_weights(wd_rnn+r'/mae/model_{}.h5'.format(i))
with open(wd_rnn+r'/mae/model_{}_hist.json'.format(i), 'rb') as f:
models_mae_hist[i] = pickle.load(f)
print('MAE prediction models loaded!')
else:
# Train multiple RNNs with identical configuration
for i in range(len(models_mse)):
models_mae[i].fit(x=X_train, y=y_train, validation_split = val_share, batch_size= BATCH, epochs=N_epochs, callbacks=[es], verbose = 2)
models_mae[i].save_weights(wd_rnn+r'/mae/model_{}.h5'.format(i))
models_mae_hist[i] = models_mae[i].history
with open(wd_rnn+r'/mae/model_{}_hist.json'.format(i), 'wb') as f: # alternative: option #'w'
pickle.dump(models_mae_hist[i], f, pickle.HIGHEST_PROTOCOL)
#models_mae, models_mae_hist = train_individual_ensembles(models_mae, X_train, y_train,
# n_epochs= N_epochs,
# n_batch= BATCH, es_patience= es_patience,
# path = wd_rnn+r'/mae')
# Save Model (and History) is integrated in function 'train_individual_ensembles'
# Insight in early stopping behaviour
print('Observe early stopping times of mse/mae-trained models:')
for i in range(10):
print('\t'+str(len(models_mse_hist[i]['val_loss']))+ " / " + str(len(models_mae_hist[i]['val_loss'])))
################ 4.2. Ensemble(s) ###############################
# Fix Number of Ensembles used
N_ensembles = 5
N_epochs = 4000
dummy_load_saved_models_ensembles = True
batchsize = 64
es_patience = 50
# Create Ensembles, using pre-trained weak learners
with tf_strategy.scope():
#----------------------------------------------------
N_ensembles = 5
# Note: cloning of models in order to perform fine-tuning independent of weak learners
ensemble_mse_5 = clone_model(Model(INPUT, Average()([models_mse[i](INPUT) for i in range(N_ensembles)])))
ensemble_mse_5.set_weights(Model(INPUT, Average()([models_mse[i](INPUT) for i in range(N_ensembles)])).get_weights())
ensemble_mse_5.compile(loss = 'mse', metrics=['mae'], optimizer = Adam(0.001))
ensemble_mae_5 = clone_model(Model(INPUT, Average()([models_mae[i](INPUT) for i in range(N_ensembles)])))
ensemble_mae_5.set_weights(Model(INPUT, Average()([models_mae[i](INPUT) for i in range(N_ensembles)])).get_weights())
ensemble_mae_5.compile(loss = 'mae', optimizer = Adam(0.001))
#----------------------------------------------------
N_ensembles = 10
ensemble_mse_10 = clone_model(Model(INPUT, Average()([models_mse[i](INPUT) for i in range(N_ensembles)])))
ensemble_mse_10.set_weights(Model(INPUT, Average()([models_mse[i](INPUT) for i in range(N_ensembles)])).get_weights())
ensemble_mse_10.compile(loss = 'mse', metrics=['mae'], optimizer = Adam(0.001))
ensemble_mae_10 = clone_model(Model(INPUT, Average()([models_mae[i](INPUT) for i in range(N_ensembles)])))
ensemble_mae_10.set_weights(Model(INPUT, Average()([models_mae[i](INPUT) for i in range(N_ensembles)])).get_weights())
ensemble_mae_10.compile(loss = 'mae', optimizer = Adam(0.001))
#----------------------------------------------------
#results_statistic = create_df_model_comparison(model_single_lst=models_mse[0:2]+models_mae[0:2],
# x_test = X_test, y_test= y_test,
# model_ens_lst = [ensemble_mse_5, ensemble_mse_10,
# ensemble_mae_5, ensemble_mae_10], #, ensemble_mse_mae_5, ensemble_mse_mae_10],
# names_number= ['5', '10','5', '10'], #,'5','10'],
# names_loss= ['MSE', 'MSE','MAE','MAE'], # 'Mixed','Mixed'],
# names_loss_single = ['MSE']*2+['MAE']*2)
#print('Statistics before fine-tuning:')
#print(results_statistic[0])
#print('\n')
# fine tuning
if bool_fine_tune:
if os.path.isfile(wd_rnn+r'/ensemble_mse_5.h5'):
ensemble_mse_5.load_weights(wd_rnn+r'/ensemble_mse_5.h5')
print('Fine-tuned mse-5 ensemble loaded.')
else:
print('Fine tuning mse-5 ensemble ...')
ensemble_mse_5.fit(x=X_train, y=y_train, validation_split = val_share, batch_size= BATCH, epochs=N_epochs, callbacks=[es], verbose = 2)
print('Ensemble mse-5 fine-tuned for {} epochs.'.format(len(ensemble_mse_5.history.history['loss'])))
ensemble_mse_5.save_weights(wd_rnn+r'/ensemble_mse_5.h5')
if os.path.isfile(wd_rnn+r'/ensemble_mse_10.h5'):
ensemble_mse_10.load_weights(wd_rnn+r'/ensemble_mse_10.h5')
print('Fine-tuned mse-10 ensemble loaded.')
else:
print('Fine tuning mse-10 ensemble ...')
ensemble_mse_10.fit(x=X_train, y=y_train, validation_split = val_share, batch_size= BATCH, epochs=N_epochs, callbacks=[es], verbose = 2)
print('Ensemble mse-10 fine-tuned for {} epochs.'.format(len(ensemble_mse_10.history.history['loss'])))
ensemble_mse_10.save_weights(wd_rnn+r'/ensemble_mse_10.h5')
if os.path.isfile(wd_rnn+r'/ensemble_mae_5.h5'):
ensemble_mae_5.load_weights(wd_rnn+r'/ensemble_mae_5.h5')
print('Fine-tuned mae-5 ensemble loaded.')
else:
print('Fine tuning mae-5 ensemble ...')
ensemble_mae_5.fit(x=X_train, y=y_train, validation_split = val_share, batch_size= BATCH, epochs=N_epochs, callbacks=[es], verbose = 2)
print('Ensemble mae-5 fine-tuned for {} epochs.'.format(len(ensemble_mae_5.history.history['loss'])))
ensemble_mae_5.save_weights(wd_rnn+r'/ensemble_mae_5.h5')
if os.path.isfile(wd_rnn+r'/ensemble_mae_10.h5'):
ensemble_mae_10.load_weights(wd_rnn+r'/ensemble_mae_10.h5')
print('Fine-tuned mae-10 ensemble loaded.')
else:
print('Fine tuning mae-10 ensemble ...')
ensemble_mae_10.fit(x=X_train, y=y_train, validation_split = val_share, batch_size= BATCH, epochs=N_epochs, callbacks=[es], verbose = 2)
print('Ensemble mae-10 fine-tuned for {} epochs.'.format(len(ensemble_mae_10.history.history['loss'])))
ensemble_mae_10.save_weights(wd_rnn+r'/ensemble_mae_10.h5')
if False:
results_statistic = create_df_model_comparison(model_single_lst=models_mse[0:]+models_mae[0:],
x_test = X_test, y_test= y_test,
model_ens_lst = [ensemble_mse_5, ensemble_mse_10,
ensemble_mae_5, ensemble_mae_10], #, ensemble_mse_mae_5, ensemble_mse_mae_10],
names_number= ['5', '10','5', '10'], #,'5','10'],
names_loss= ['mse', 'mse','mae','mae'], # 'Mixed','Mixed'],
names_loss_single = ['mse']*10+['mae']*10)
print('Statistics after fine-tuning:')
print(results_statistic[0])
if bool_latex:
with open('TeX_tables/Prediction_DC_Model_Comparison.tex','w') as tf:
tf.write(results_statistic[0].to_latex())
# Relate following relative values to absolute Policy Values
interval_lst = [0,0.005, 0.01,0.2,0.4,0.6,0.8,1]
stat_ENS_0 = model_examine_indivual_fit(model = ensemble_mse_5, data = X_test, PV_max= params['V_max'],
targets = y_test, output_option = 'statistic', interval_lst= interval_lst)
print('Statistics for 5-MSE ensemble')
print(stat_ENS_0, r'\n')
if bool_latex:
with open('TeX_tables/Prediction_DC_Model_MSE_5.tex','w') as tf:
tf.write(stat_ENS_0.to_latex())
stat_ENS_1 = model_examine_indivual_fit(model = ensemble_mse_10, data = X_test, PV_max= params['V_max'],
targets = y_test, output_option = 'statistic', interval_lst= interval_lst)
print('Statistics for 10-MSE ensemble')
print(stat_ENS_1, r'\n')
with open('Prediction_TL_Model_MSE_10.tex','w') as tf:
tf.write(stat_ENS_1.to_latex())
stat_ENS_2 = model_examine_indivual_fit(model = ensemble_mae_5, data = X_test, PV_max= params['V_max'],
targets = y_test, output_option = 'statistic', interval_lst= interval_lst)
print('Statistics for 5-MAE ensemble')
print(stat_ENS_2, r'\n')
if bool_latex:
with open('TeX_tables/Prediction_DC_Model_MAE_5.tex','w') as tf:
tf.write(stat_ENS_2.to_latex())
stat_ENS_3 = model_examine_indivual_fit(model = ensemble_mae_10, data = X_test, PV_max= params['V_max'],
targets = y_test, output_option = 'statistic', interval_lst= interval_lst)
print('Statistics for 10-MAE ensemble')
print(stat_ENS_3, r'\n')
print(r'\t Analysis of prediction models completed!')