-
Notifications
You must be signed in to change notification settings - Fork 8
/
model.py
244 lines (192 loc) · 10.5 KB
/
model.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
import numpy as np
from scipy import stats
import statsmodels.api as sm
from sklearn.linear_model import ElasticNet
from sklearn.feature_selection import SelectKBest, f_regression
from sklearn import preprocessing
from sklearn.pipeline import Pipeline
from sklearn.model_selection import LeaveOneOut
from sklearn.metrics import mean_squared_error, mean_absolute_error, median_absolute_error, r2_score, explained_variance_score
from sklearn.model_selection import GridSearchCV
from sklearn.model_selection import cross_val_predict
from sklearn.model_selection import cross_validate
import PAINTeR.plot as plot
def train(X, y, model, p_grid, nested=False, model_averaging=True, ):
inner_cv = LeaveOneOut()
outer_cv = LeaveOneOut()
clf = GridSearchCV(estimator=model, param_grid=p_grid, cv=inner_cv, scoring="neg_mean_squared_error", verbose=False, return_train_score=False, n_jobs=8)
clf.fit(X, y)
print("**** Non-nested analysis ****")
print("** Best hyperparameters: " + str(clf.best_params_))
print("** Score on full data as training set:\t" + str(-mean_squared_error(y_pred=clf.best_estimator_.predict(X), y_true=y)))
print("** Score on mean as model: " + str(-mean_squared_error(np.repeat(y.mean(), len(y)), y)))
print("** Best Non-nested cross-validated score on test:\t" + str(clf.best_score_))
model=clf.best_estimator_
print("XXXXX Explained Variance: " + str(
1 - clf.best_score_ / -mean_squared_error(np.repeat(y.mean(), len(y)), y)))
avg_model = None
all_models = []
if nested:
print("**** Nested analysis ****")
#nested_scores = cross_val_score(clf, X, y, cv=outer_cv, scoring="explained_variance")
#print "** Nested Score on test:\t" + str(nested_scores.mean())
# this above has the same output as this below:
best_params = []
predicted = np.zeros(len(y))
actual = np.zeros(len(y))
nested_scores_train = np.zeros(outer_cv.get_n_splits(X))
nested_scores_test = np.zeros(outer_cv.get_n_splits(X))
nested_scores_test2 = np.zeros(outer_cv.get_n_splits(X))
i = 0
avg = []
# doing the crossval itewrations manually
print("model\tinner_cv mean score\touter vc score")
for train, test in outer_cv.split(X, y):
clf.fit(X[train], y[train])
# model avaraging
RES, mat, labels = get_full_coef(X[train], clf.best_estimator_, plot=False)
avg.append(RES)
all_models.append(clf.best_estimator_)
# plot histograms to check distributions
#bins = np.linspace(-1.5, 1.5, 6)
#pyplot.hist(y[train], bins, alpha=0.5, label='train')
#pyplot.hist(y[test], bins, alpha=0.5, label='test')
#pyplot.legend(loc='upper right')
#pyplot.show()
print(str(clf.best_params_) + " " + str(clf.best_score_) + " " + str(clf.score(X[test], y[test])))
predicted[i] = clf.predict(X[test])
actual[i] = y[test]
best_params.append(clf.best_params_)
nested_scores_train[i] = clf.best_score_
nested_scores_test[i] = clf.score(X[test], y[test])
# clf.score is the same as calculating the score to the prediced values of the test dataset:
#nested_scores_test2[i] = explained_variance_score(y_pred=clf.predict(X[test]), y_true=y[test])
i = i+1
print("*** Score on mean as model:\t" + str(-mean_squared_error(np.repeat(y.mean(), len(y)), y)))
print("** Mean score in the inner crossvaludation (inner_cv):\t" + str(nested_scores_train.mean()))
print("** Mean Nested Crossvalidation Score (outer_cv):\t" + str(nested_scores_test.mean()))
print("Explained Variance: " + str( 1- nested_scores_test.mean()/-mean_squared_error(np.repeat(y.mean(), len(y)), y) ))
print("Correlation: " + str(np.corrcoef(actual, predicted)[0,1]))
avg_model = np.mean(np.array(avg), axis=0)
#plot the prediction of the outer cv
fig, ax = plt.subplots()
ax.scatter(actual, predicted, edgecolors=(0, 0, 0))
ax.plot([y.min(), y.max()],
[y.min(), y.max()],
'k--',
lw=2)
ax.set_xlabel('Pain Sensitivity')
ax.set_ylabel('Predicted (Nested LOO)')
plt.title("Expl. Var.:" + str( 1- nested_scores_test.mean()/-mean_squared_error(np.repeat(y.mean(), len(y)), y) ) +
"\nCorrelation: " + str(np.corrcoef(actual, predicted)[0, 1]) )
plt.show()
else:
all_models = [model]
model.fit(X, y) # fit to whole data
return model, avg_model, all_models
def pipe_scale_fsel_elnet(scaler=preprocessing.RobustScaler(),
fsel=SelectKBest(f_regression),
model=ElasticNet(max_iter=100000),
p_grid = {'fsel__k': [20, 25, 30, 35, 40, 45, 50, 60, 70, 80], 'model__alpha': [.001, .005, .01, .05, .1, .5], 'model__l1_ratio': [.999] } # for fast re-calculation
# p_grid = {'fsel__k': np.linspace(10, 200, 39), 'model__alpha': [.001, .005, .01, .05, .1, .5], 'model__l1_ratio': [1, .5, .7, .9, .95, .99, .999] } # exhaustive, takes a lot of time
):
mymodel = Pipeline(
[('scaler', scaler), ('fsel', fsel),
('model', model)])
return mymodel, p_grid
def pred_stat(observed, predicted, robust=False):
# convert to np.array
observed = np.array(observed)
predicted = np.array(predicted)
#EXCLUDE NA-s:
predicted = predicted[~np.isnan(observed)]
observed = observed[~np.isnan(observed)]
if robust:
res = sm.RLM(observed, sm.add_constant(predicted)).fit()
p_value = res.pvalues[1]
regline = res.fittedvalues
residual = res.sresid
# this is a pseudo r_squared, see: https://stackoverflow.com/questions/31655196/how-to-get-r-squared-for-robust-regression-rlm-in-statsmodels
r_2 = sm.WLS(observed, sm.add_constant(predicted), weights=res.weights).fit().rsquared
else:
slope, intercept, r_value, p_value, std_err = stats.linregress(observed, predicted)
regline = slope*observed+intercept
r_2 = r_value**2
residual = observed - regline
return p_value, r_2, residual, regline
def learning_curve(model, X, y, Ns = [15, 20, 25, 30, 35]):
from random import shuffle
train = []
test = []
for n in Ns:
print("******************")
print(n)
tr=[]
te=[]
for s in range(10):
idx = range(len(y))
shuffle(idx)
idx=idx[:n]
#model, p_grid = pipe_scale_fsel_elnet()
cv = cross_validate(model, [X[i] for i in idx], [y[i] for i in idx], scoring="neg_mean_squared_error",
cv=LeaveOneOut(), return_train_score = True)
#clf = GridSearchCV(estimator=model, param_grid=p_grid, cv=LeaveOneOut(), scoring="neg_mean_squared_error",
# verbose=False, return_train_score=True, n_jobs=8)
#clf.fit(X[:i], y[:i])
#train_score = -mean_squared_error(y_pred=clf.best_estimator_.predict(X[:i]), y_true=y[:i])
#test_score = clf.best_score_
tr.append(np.median(cv["train_score"]))
te.append(np.median(cv["test_score"]))
print(np.mean(tr), np.mean(te))
#print "******************"
#print np.mean(cv["test_score"])
#print np.mean(cv["train_score"])
train.append(np.median(tr))
test.append(np.median(te))
#fitted_model = model.fit()
#predicted = model.predict(X)
return train, test
def evaluate_prediction(model, X, y, orig_mean=None, outfile="", robust=False, covar=[]):
predicted = model.predict(X)
p_value, r_2, residual, regline = pred_stat(y, predicted, robust=robust)
if orig_mean:
y_base = orig_mean
else:
y_base = y.mean()
expl_var = (1 - (-mean_squared_error(y_pred=predicted, y_true=y)
/
-mean_squared_error(np.repeat(y_base, len(y)), y))) * 100
print("R2=" + "{:.3f}".format(r_2) + " R=" + "{:.3f}".format(np.sqrt(r_2))\
+ " p=" + "{:.6f}".format(p_value) +" Expl. Var.: " + "{:.1f}".format(expl_var) + "%" \
+ " Expl. Var.2: " + "{:.1f}".format(explained_variance_score(y_pred=predicted, y_true=y)*100) + "%" \
+ " MSE=" + "{:.3f}".format(mean_squared_error(y_pred=predicted, y_true=y))\
+ " RMSE=" + "{:.3f}".format(np.sqrt(mean_squared_error(y_pred=predicted, y_true=y))) \
+ " MAE=" + "{:.3f}".format(mean_absolute_error(y_pred=predicted, y_true=y)) \
+ " MedAE=" + "{:.3f}".format(median_absolute_error(y_pred=predicted, y_true=y)) \
+ " R^2=" + "{:.3f}".format(r2_score(y_pred=predicted, y_true=y)))
plot.plot_prediction(y, predicted, outfile, robust=robust, sd=True, covar=covar,
text="$R^2$ = " + "{:.3f}".format(r_2) +
" p = " + "{:.3f}".format(p_value)+
" Expl. Var.: " + "{:.1f}".format(expl_var)
)
return predicted
def evaluate_crossval_prediction(model, X, y, outfile="", cv=LeaveOneOut(), robust=False):
predicted = cross_val_predict(model, X, y, cv=cv)
p_value, r_2, residual, regline = pred_stat(y, predicted, robust=robust)
expl_var = ( 1- (-mean_squared_error(y_pred=predicted, y_true=y)
/
-mean_squared_error(np.repeat(y.mean(), len(y)), y) ))*100
print("R2=" + "{:.3f}".format(r_2) + " R=" + "{:.3f}".format(np.sqrt(r_2)) \
+ " p=" + "{:.6f}".format(p_value) + " Expl. Var.: " + "{:.1f}".format(expl_var) + "%" \
+ " Expl. Var.2: " + "{:.1f}".format(explained_variance_score(y_pred=predicted, y_true=y)*100) + "%" \
+ " MSE=" + "{:.3f}".format(mean_squared_error(y_pred=predicted, y_true=y)) \
+ " RMSE=" + "{:.3f}".format(np.sqrt(mean_squared_error(y_pred=predicted, y_true=y))) \
+ " MAE=" + "{:.3f}".format(mean_absolute_error(y_pred=predicted, y_true=y)) \
+ " MedAE=" + "{:.3f}".format(median_absolute_error(y_pred=predicted, y_true=y)) \
+ " R^2=" + "{:.3f}".format(r2_score(y_pred=predicted, y_true=y)))
plot.plot_prediction(y, predicted, outfile, robust=robust, sd=True,
text="$R2$=" + "{:.3f}".format(r_2) +
" p=" + "{:.3f}".format(p_value) +
" Expl. Var.: " + "{:.1f}".format(expl_var) + "%"
)
return predicted