-
Notifications
You must be signed in to change notification settings - Fork 22
/
train.py
215 lines (189 loc) · 6.96 KB
/
train.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
import numpy as np
from scipy import stats
from sklearn import metrics
from sklearn.feature_selection import SelectKBest, chi2
from sklearn.model_selection import train_test_split, RandomizedSearchCV
from sklearn.linear_model import SGDClassifier
from sklearn.feature_extraction.text import TfidfVectorizer
# in order to use SMOTE, you've got to import Pipeline from imblearn
from imblearn.pipeline import Pipeline
import dill as pickle
import logging
from fbo_scraper.predict import Predict
from fbo_scraper.binaries import binary_path, Path
logger = logging.getLogger(__name__)
class log_uniform:
"""
Provides an instance of the log-uniform distribution with an .rvs() method. Meant to be used with
RandomizedSearchCV, particularly for params like alpha, C, gamma, etc.
Attributes:
a (int or float): the exponent of the beginning of the range
b (int or float): the exponent of the end of range.
base (int or float): the base of the logarithm. 10 by default.
"""
def __init__(self, a=-1, b=0, base=10):
self.loc = a
self.scale = b - a
self.base = base
def rvs(self, size=1, random_state=None):
uniform = stats.uniform(loc=self.loc, scale=self.scale)
return np.power(self.base, uniform.rvs(size=size, random_state=random_state))
def get_param_distribution():
"""
Utility function that returns the param distribution for the grid search
"""
param_dist = {
"vectorizer__ngram_range": [(1, 1), (1, 2)],
"vectorizer__min_df": stats.randint(1, 3),
"vectorizer__max_df": stats.uniform(0.60, 0.35),
"vectorizer__sublinear_tf": [True, False],
"select__k": [10, 100, 200, 500, 1000, 1500, 2000, 5000],
"clf__alpha": log_uniform(-5, 2),
"clf__penalty": ["l2", "l1", "elasticnet"],
"clf__loss": ["hinge", "log", "modified_huber", "squared_hinge", "perceptron"],
}
return param_dist
def train(
X, y, weight_classes=True, n_iter_search=500, score="roc_auc", random_state=123
):
"""
Train a binary SGD classifier using a randomized grid search with given scoring metric.
Parameters:
X (list-like): list of normalized attachment texts
y (list-like): list of validated targets (0 = red, 1 = green)
weight_classes (bool): whether or not to use the “balanced” mode to adjust class weights.
n_iter_search (int): number of parameter settings that are sampled. Trades off runtime vs quality
of the solution.
score (str): the scorer used to evaluate the predictions on the test set. `roc_auc` by
default. Available options include: accuracy, roc_auc, precision, fbeta, recall.
Note: for fbeta, beta is set to 1.5 to favor recall of the positive class.
random_state (int): sets the random seed for reproducibility.
Returns:
results (dict): a dict of scoring metrics and their values
best_score (float): mean cross-validated score of the best_estimator.
best_estimator (sklearn estimator): estimator that was chosen by the search
best_params (dict): parameter setting that gave the best results on the hold out data.
"""
if weight_classes:
clf = SGDClassifier(class_weight="balanced")
else:
clf = clf = SGDClassifier()
scoring = {
"accuracy": metrics.make_scorer(metrics.accuracy_score),
"roc_auc": metrics.make_scorer(metrics.roc_auc_score),
"precision": metrics.make_scorer(metrics.average_precision_score),
"fbeta": metrics.make_scorer(metrics.fbeta_score, beta=0.5),
"recall": metrics.make_scorer(metrics.recall_score),
}
X_train, X_test, y_train, y_test = train_test_split(
X, y, stratify=y, test_size=0.2, random_state=random_state
)
pipe = Pipeline(
[
("vectorizer", TfidfVectorizer(stop_words="english")),
("select", SelectKBest(chi2)),
("clf", clf),
]
)
param_dist = get_param_distribution()
random_search = RandomizedSearchCV(
pipe,
param_distributions=param_dist,
scoring=scoring,
refit=score,
n_iter=n_iter_search,
cv=5,
n_jobs=-1,
verbose=1,
random_state=random_state,
)
try:
random_search.fit(X_train, y_train)
except Exception as e:
logger.error(
f"Exception occurred training a new model: \
{e}",
exc_info=True,
)
y_pred = random_search.predict(X_test)
# get the col number of the positive class (i.e. green)
positive_class_col = list(random_search.classes_).index(1)
try:
y_score = random_search.predict_proba(X_test)[:, positive_class_col]
except AttributeError:
y_score = random_search.decision_function(X_test)
average_precision = metrics.average_precision_score(y_test, y_score)
acc = metrics.accuracy_score(y_test, y_pred)
try:
roc_auc = metrics.roc_auc_score(y_test, y_pred)
except ValueError:
roc_auc = None
precisions, recalls, _ = metrics.precision_recall_curve(y_test, y_score)
try:
auc = metrics.auc(recalls, precisions)
except ValueError:
auc = None
fbeta = metrics.fbeta_score(y_test, y_pred, beta=1.5)
recall = metrics.recall_score(y_test, y_pred)
best_estimator = random_search.best_estimator_
best_params = random_search.best_params_
best_score = random_search.best_score_
result_values = [
y_pred,
y_score,
precisions,
recall,
average_precision,
acc,
roc_auc,
auc,
fbeta,
recalls,
best_score,
best_estimator,
y_test,
]
result_keys = [
"y_pred",
"y_score",
"precisions",
"recall",
"average_precision",
"acc",
"roc_auc",
"auc",
"fbeta",
"recalls",
"best_score",
"best_estimator",
"y_test",
]
results = {k: v for k, v in zip(result_keys, result_values)}
return results, best_score, best_estimator, best_params
def prepare_samples(attachments):
"""
Prepares attachment data for training
Parameters:
attachments (list): list of dicts, with each dict containing an attachment's text and
validated target
Returns:
X (list): list of normalized attachment texts
y (list): list of validated targets
"""
X = []
y = []
for attachment in attachments:
text = Predict.transform_text(attachment["text"])
X.append(text)
y.append(attachment["target"])
return X, y
def pickle_model(best_estimator):
"""
Pickles an estimator
Parameters:
best_estimator (sklearn estimator): estimator that was chosen by a grid search
Returns:
None
"""
with open(Path(binary_path, "atc_estimator.pkl"), "wb") as f:
pickle.dump(best_estimator, f)