Newer
Older
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
from sf_runoff_lt import create_it_matrix_lt
import pandas as pd
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import numpy as np
from sklearn.svm import SVR
from sklearn.decomposition import PCA
from sklearn.preprocessing import StandardScaler
from sklearn.pipeline import make_pipeline
from sklearn.compose import TransformedTargetRegressor
from sklearn.model_selection import GridSearchCV,TimeSeriesSplit
from sklearn.metrics import mean_squared_error
import pdb
import seaborn as sns
def SVR_nested_CV_gridsearch(daily_input, C_range,epsilon_range, t_range,t_unit,n_splits,test_size,lt):
for t_length in t_range:
it_matrix=create_it_matrix_lt(daily_input,t_length,t_unit,lt).astype('float32')
tscv = TimeSeriesSplit(gap=t_unit ,n_splits=n_splits, test_size=test_size)
sets = tscv.split(it_matrix.index)
all_models= []
for train_index, test_index in sets:
#validation set is the last 2 years of the "old_training"
val_index = train_index[-365:]
#training set is reduced by 2y and 1month"
train_index_update = train_index[:-395]
trainCvSplit = [(list(train_index_update),list(val_index))]
X = it_matrix.drop(columns='Q')
y = it_matrix['Q']
svr_estimator = SVR(kernel='rbf', gamma='scale', cache_size=1000)
svr_estimator = make_pipeline(StandardScaler(),
TransformedTargetRegressor(regressor=svr_estimator, transformer=StandardScaler()))
parameters = {'transformedtargetregressor__regressor__C': C_range,
'transformedtargetregressor__regressor__epsilon': epsilon_range}
svr_model = GridSearchCV(svr_estimator, cv=trainCvSplit, param_grid=parameters, n_jobs=-1, verbose=1, refit=True, return_train_score=True)
# execute the grid search
svr_model.fit(X, y)
all_models.append(pd.DataFrame(svr_model.cv_results_))
#PUT ALL THE TRAINED MODELS AND RESULTS IN A DATAFRAME
all_m = pd.DataFrame(data=None,columns=all_models[0].columns)
for i in all_models :
all_m=all_m.append(i);
#GROUP BY PARAMETERS AND AVERAGE OVER THE DIFFERENT VALIDATION SETS
par=(['param_transformedtargetregressor__regressor__C','param_transformedtargetregressor__regressor__epsilon'])
avg_models = all_m.groupby(par).mean()
avg_models['train_test_diff']= avg_models.mean_train_score - avg_models.mean_test_score
# SELECT THE SINGLE BEST MODEL OVERALL
best_model_overall = avg_models.loc[[avg_models.mean_test_score.idxmax()]]
best_C=best_model_overall.reset_index().param_transformedtargetregressor__regressor__C[0]
best_epsilon = best_model_overall.reset_index().param_transformedtargetregressor__regressor__epsilon[0]
#INVESTIGATE WITH HEATMAPS THE FACT THAT WE'RE NOT OVERFITTING
# get the coordinates of the "best model"
y=np.where(epsilon_range==best_epsilon)[0]+0.5
x=np.where(C_range==best_C)[0]+0.5
#get the models with a certain number of components
hm_test = avg_models.reset_index().pivot( columns='param_transformedtargetregressor__regressor__C', index='param_transformedtargetregressor__regressor__epsilon', values='mean_test_score')
hm_train= avg_models.reset_index().pivot(columns='param_transformedtargetregressor__regressor__C',index='param_transformedtargetregressor__regressor__epsilon',values='mean_train_score')
plt.figure(figsize=(15,7))
plt.subplot(1,2,1)
sns.heatmap(hm_test,vmin=hm_test.min().min(),vmax=hm_train.max().max())
plt.title('VALIDATION')
plt.plot(x,y,marker='o')
plt.subplot(1,2,2)
sns.heatmap(hm_train,vmin=hm_test.min().min(),vmax=hm_train.max().max())
plt.title('TRAIN')
plt.plot(x,y,marker='o')
plt.tight_layout()
# Check if the best C (or epsion) is in the border of the grid
if best_C == max(C_range) or best_C == min(C_range):
print(f'Warning: best C found on the grid limit: C = {best_C}')
if best_epsilon == max(epsilon_range) or best_epsilon == min(epsilon_range):
print(f'Warning: best epsilon found on the grid limit: epsilon = {best_epsilon}')
print()
return best_C, best_epsilon
def SVR_PCA_nested_CV_gridsearch(daily_input, C_range, epsilon_range, components_range, t_range,t_unit,n_splits,test_size,lt):
for t_length in t_range:
it_matrix=create_it_matrix_lt(daily_input,t_length,t_unit,lt).astype('float32')
tscv = TimeSeriesSplit(gap=t_unit ,n_splits=n_splits, test_size=test_size)
sets = tscv.split(it_matrix.index)
all_models= []
for train_index, test_index in sets:
#validation set is the last 2 years of the "old_training"
val_index = train_index[-365:]
#training set is reduced by 2y and 1month"
train_index_update = train_index[:-395]
trainCvSplit = [(list(train_index_update),list(val_index))]
X = it_matrix.drop(columns='Q')
y = it_matrix['Q']
svr_estimator = SVR(kernel='rbf', gamma='scale', cache_size=1000)
svr_estimator = make_pipeline(StandardScaler(),
PCA(),
TransformedTargetRegressor(regressor=svr_estimator, transformer=StandardScaler()))
parameters = {'pca__n_components': components_range,
'transformedtargetregressor__regressor__C': C_range,
'transformedtargetregressor__regressor__epsilon': epsilon_range
}
svr_model = GridSearchCV(svr_estimator, cv=trainCvSplit, param_grid=parameters, n_jobs=-1, verbose=1, refit=True, return_train_score=True)
# execute the grid search
svr_model.fit(X, y)
all_models.append(pd.DataFrame(svr_model.cv_results_))
#PUT ALL THE TRAINED MODELS AND RESULTS IN A DATAFRAME
all_m = pd.DataFrame(data=None,columns=all_models[0].columns)
for i in all_models :
all_m=all_m.append(i);
#GROUP BY PARAMETERS AND AVERAGE OVER THE DIFFERENT VALIDATION SETS
par=(['param_pca__n_components', 'param_transformedtargetregressor__regressor__C', 'param_transformedtargetregressor__regressor__epsilon'])
avg_models = all_m.groupby(par).mean()
avg_models['train_test_diff']= avg_models.mean_train_score - avg_models.mean_test_score
# SELECT THE SINGLE BEST MODEL OVERALL
best_model_overall = avg_models.loc[[avg_models.mean_test_score.idxmax()]]
best_C=best_model_overall.reset_index().param_transformedtargetregressor__regressor__C[0]
best_epsilon = best_model_overall.reset_index().param_transformedtargetregressor__regressor__epsilon[0]
best_n = best_model_overall.reset_index().param_pca__n_components[0]
#INVESTIGATE WITH HEATMAPS THE FACT THAT WE'RE NOT OVERFITTING
# get the coordinates of the "best model"
y=np.where(epsilon_range==best_epsilon)[0]+0.5
x=np.where(C_range==best_C)[0]+0.5
#get the models with a certain number of components
query=f'param_pca__n_components=={best_n}'
nc=avg_models.query(query)
hm_test = nc.reset_index().pivot( columns='param_transformedtargetregressor__regressor__C', index='param_transformedtargetregressor__regressor__epsilon', values='mean_test_score')
hm_train = nc.reset_index().pivot( columns='param_transformedtargetregressor__regressor__C', index='param_transformedtargetregressor__regressor__epsilon', values='mean_train_score')
plt.figure(figsize=(15,7))
plt.subplot(1,2,1)
sns.heatmap(hm_test,vmin=hm_test.min().min(),vmax=hm_train.max().max())
plt.title('TEST')
plt.plot(x,y,marker='o')
plt.subplot(1,2,2)
sns.heatmap(hm_train,vmin=hm_test.min().min(),vmax=hm_train.max().max())
plt.title('TRAIN')
plt.plot(x,y,marker='o')
plt.tight_layout()
# Check if the best C (or epsion) is in the border of the grid
if best_C == max(C_range) or best_C == min(C_range):
print(f'Warning: best C found on the grid limit: C = {best_C}')
if best_epsilon == max(epsilon_range) or best_epsilon == min(epsilon_range):
print(f'Warning: best epsilon found on the grid limit: epsilon = {best_epsilon}')
print()
return best_C, best_epsilon, best_n