Spaces:
Sleeping
Sleeping
Commit
•
5fbf3c7
1
Parent(s):
5f32fa0
added modeling module
Browse files- modeling/__init__.py +1 -0
- modeling/data_utils.py +16 -0
- modeling/eda.py +73 -0
- modeling/ml_model_dev.py +238 -0
- modeling/ml_model_test.py +38 -0
modeling/__init__.py
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
import os, sys; sys.path.append(os.path.dirname(os.path.realpath(__file__)))
|
modeling/data_utils.py
ADDED
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import numpy as np
|
2 |
+
import pandas as pd
|
3 |
+
|
4 |
+
def read_csv_file(file_csv):
|
5 |
+
df_csv = pd.read_csv(file_csv)
|
6 |
+
return df_csv
|
7 |
+
|
8 |
+
def get_dict_nan_counts_per_col(data_frame):
|
9 |
+
dict_nan_counts_per_col = data_frame.isna().sum().to_dict()
|
10 |
+
dict_nan_counts_per_col = dict(sorted(dict_nan_counts_per_col.items(), key=lambda kv: kv[1], reverse=True))
|
11 |
+
return dict_nan_counts_per_col
|
12 |
+
|
13 |
+
def get_data_from_data_frame(data_frame):
|
14 |
+
arr = data_frame.to_numpy()
|
15 |
+
X_arr, Y_arr = arr[:, :-1], arr[:, -1:].reshape(-1)
|
16 |
+
return X_arr, Y_arr
|
modeling/eda.py
ADDED
@@ -0,0 +1,73 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import sys
|
3 |
+
import argparse
|
4 |
+
|
5 |
+
import numpy as np
|
6 |
+
import seaborn as sns
|
7 |
+
import matplotlib.pyplot as plt
|
8 |
+
|
9 |
+
from data_utils import read_csv_file, get_data_from_data_frame
|
10 |
+
|
11 |
+
def do_eda(ARGS):
|
12 |
+
data_frame = read_csv_file(ARGS.file_csv)
|
13 |
+
label_counts = dict(data_frame[ARGS.target_column].value_counts())
|
14 |
+
# print(label_counts)
|
15 |
+
|
16 |
+
# plot a histogram
|
17 |
+
plt.figure(figsize=(12, 12))
|
18 |
+
plt.bar([str(l) for l in label_counts.keys()], label_counts.values(), width=0.5)
|
19 |
+
plt.xlabel(f"{ARGS.target_column}", fontsize=20)
|
20 |
+
plt.ylabel("Number of samples", fontsize=20)
|
21 |
+
plt.title("Distribution of samples in the dataset", fontsize=20)
|
22 |
+
plt.grid()
|
23 |
+
plt.xticks(fontsize=20)
|
24 |
+
plt.yticks(fontsize=20)
|
25 |
+
plt.show()
|
26 |
+
|
27 |
+
"""
|
28 |
+
feat_cols = data_frame.columns[:-1]
|
29 |
+
num_feat_cols = len(feat_cols)
|
30 |
+
|
31 |
+
fig, axs = plt.subplots(num_feat_cols)
|
32 |
+
fig.suptitle("Distribution of features")
|
33 |
+
#axs.set_xlabel(ARGS.target_column)
|
34 |
+
|
35 |
+
for col_index in range(num_feat_cols):
|
36 |
+
column = feat_cols[col_index]
|
37 |
+
not_nan_indices = list(data_frame[column].notna())
|
38 |
+
lbl_with_not_nans = data_frame[ARGS.target_column][not_nan_indices]
|
39 |
+
col_with_not_nans = data_frame[column][not_nan_indices]
|
40 |
+
print(column, len(lbl_with_not_nans), len(col_with_not_nans))
|
41 |
+
|
42 |
+
axs[col_index].scatter(lbl_with_not_nans, col_with_not_nans)
|
43 |
+
axs[col_index].set(ylabel=column)
|
44 |
+
plt.show()
|
45 |
+
"""
|
46 |
+
|
47 |
+
plt.figure()
|
48 |
+
corr_mat = data_frame.corr()
|
49 |
+
sns.heatmap(corr_mat)
|
50 |
+
plt.title("Feature correlation matrix", fontsize=20)
|
51 |
+
plt.xticks(fontsize=20)
|
52 |
+
plt.yticks(fontsize=20)
|
53 |
+
plt.show()
|
54 |
+
|
55 |
+
return
|
56 |
+
|
57 |
+
def main():
|
58 |
+
file_csv = "dataset/water_potability.csv"
|
59 |
+
target_column = "Potability"
|
60 |
+
|
61 |
+
parser = argparse.ArgumentParser(
|
62 |
+
formatter_class=argparse.ArgumentDefaultsHelpFormatter
|
63 |
+
)
|
64 |
+
parser.add_argument("--file_csv", default=file_csv,
|
65 |
+
type=str, help="full path to dataset csv file")
|
66 |
+
parser.add_argument("--target_column", default=target_column,
|
67 |
+
type=str, help="target label for which the EDA needs to be done")
|
68 |
+
ARGS, unparsed = parser.parse_known_args()
|
69 |
+
do_eda(ARGS)
|
70 |
+
return
|
71 |
+
|
72 |
+
if __name__ == "__main__":
|
73 |
+
main()
|
modeling/ml_model_dev.py
ADDED
@@ -0,0 +1,238 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import sys
|
3 |
+
import joblib
|
4 |
+
import argparse
|
5 |
+
import collections
|
6 |
+
|
7 |
+
import mlflow
|
8 |
+
import numpy as np
|
9 |
+
|
10 |
+
import lightgbm as lgbm
|
11 |
+
|
12 |
+
from sklearn.svm import SVC
|
13 |
+
from sklearn.decomposition import PCA
|
14 |
+
from sklearn.pipeline import Pipeline
|
15 |
+
from sklearn.preprocessing import StandardScaler
|
16 |
+
from sklearn.linear_model import LogisticRegression
|
17 |
+
from sklearn.experimental import enable_iterative_imputer
|
18 |
+
from sklearn.metrics import accuracy_score, f1_score, make_scorer
|
19 |
+
from sklearn.impute import KNNImputer, SimpleImputer, IterativeImputer
|
20 |
+
from sklearn.ensemble import RandomForestClassifier, AdaBoostClassifier
|
21 |
+
from sklearn.model_selection import cross_validate, train_test_split, GridSearchCV, KFold
|
22 |
+
|
23 |
+
from data_utils import read_csv_file, get_data_from_data_frame
|
24 |
+
|
25 |
+
|
26 |
+
def load_ml_model(pkl_file_name):
|
27 |
+
model_pipeline = mlflow.sklearn.load_model(pkl_file_name)
|
28 |
+
return model_pipeline
|
29 |
+
|
30 |
+
def get_imputer(imputer_type):
|
31 |
+
# setup parameter search space for different imputers
|
32 |
+
imputer, imputer_params = None, None
|
33 |
+
if imputer_type == "simple":
|
34 |
+
imputer = SimpleImputer()
|
35 |
+
imputer_params = {
|
36 |
+
"imputer__strategy": ["mean", "median", "most_frequent"],
|
37 |
+
}
|
38 |
+
elif imputer_type == "knn":
|
39 |
+
imputer = KNNImputer()
|
40 |
+
imputer_params = {
|
41 |
+
"imputer__n_neighbors": [5, 7],
|
42 |
+
"imputer__weights": ["uniform", "distance"],
|
43 |
+
}
|
44 |
+
elif imputer_type == "iterative":
|
45 |
+
imputer = IterativeImputer()
|
46 |
+
imputer_params = {
|
47 |
+
"imputer__initial_strategy": ["mean", "median", "most_frequent"],
|
48 |
+
"imputer__imputation_order": ["ascending", "descending"],
|
49 |
+
}
|
50 |
+
else:
|
51 |
+
print(f"unidentified option for arg, imputer_type: {imputer_type}")
|
52 |
+
sys.exit(0)
|
53 |
+
return imputer, imputer_params
|
54 |
+
|
55 |
+
def get_scaler():
|
56 |
+
scaler = StandardScaler()
|
57 |
+
return scaler
|
58 |
+
|
59 |
+
def get_pca(max_num_feats):
|
60 |
+
pca = PCA()
|
61 |
+
pca_params = {
|
62 |
+
"pca__n_components": np.arange(2, max_num_feats+1),
|
63 |
+
}
|
64 |
+
return pca, pca_params
|
65 |
+
|
66 |
+
def get_classifier(classifier_type):
|
67 |
+
# setup parameter search space for different classifiers
|
68 |
+
|
69 |
+
classifier, classifier_params = None, None
|
70 |
+
if classifier_type == "ada_boost":
|
71 |
+
classifier = AdaBoostClassifier()
|
72 |
+
classifier_params = {
|
73 |
+
"classifier__learning_rate": [0.5, 1, 1.5, 2, 2.5, 3],
|
74 |
+
"classifier__n_estimators": [100, 200, 500],
|
75 |
+
}
|
76 |
+
elif classifier_type == "log_reg":
|
77 |
+
classifier = LogisticRegression(max_iter=200, solver="saga")
|
78 |
+
classifier_params = {
|
79 |
+
"classifier__penalty": ["l1", "l2", "elasticnet"],
|
80 |
+
"classifier__class_weight": [None, "balanced"],
|
81 |
+
"classifier__C": [0.1, 0.5, 1, 2],
|
82 |
+
"classifier__l1_ratio": np.arange(0.1, 1, 0.1),
|
83 |
+
}
|
84 |
+
elif classifier_type == "random_forest":
|
85 |
+
classifier = RandomForestClassifier()
|
86 |
+
classifier_params = {
|
87 |
+
"classifier__n_estimators": [100, 250],
|
88 |
+
"classifier__criterion": ["gini", "entropy"],
|
89 |
+
"classifier__max_depth": [None, 10, 25, 50, 75],
|
90 |
+
"classifier__min_samples_leaf": [1, 5, 10, 20],
|
91 |
+
"classifier__min_samples_split": [2, 3, 4, 5],
|
92 |
+
}
|
93 |
+
elif classifier_type == "svc":
|
94 |
+
classifier = SVC()
|
95 |
+
classifier_params = {
|
96 |
+
"classifier__C": [0.5, 1, 1.5, 2, 2.5],
|
97 |
+
"classifier__kernel": ["linear", "poly", "rbf", "sigmoid"],
|
98 |
+
"classifier__degree": [2, 3, 4],
|
99 |
+
}
|
100 |
+
elif classifier_type == "light_gbm":
|
101 |
+
classifier = lgbm.LGBMClassifier(
|
102 |
+
boosting_type="gbdt", objective="binary", metric="auc", verbosity=-1)
|
103 |
+
classifier_params = {
|
104 |
+
"classifier__num_leaves": [15, 31, 63, 127, 255],
|
105 |
+
"classifier__learning_rate": [0.1, 0.5, 1, 2],
|
106 |
+
"classifier__n_estimators": [100, 500, 1000],
|
107 |
+
"classifier__reg_lambda": [0.1, 0.5, 1],
|
108 |
+
"classifier__min_data_in_leaf": [10, 20, 30, 50],
|
109 |
+
}
|
110 |
+
else:
|
111 |
+
print(f"unidentified option for arg, classifier_type: {classifier_type}")
|
112 |
+
sys.exit(0)
|
113 |
+
|
114 |
+
return classifier, classifier_params
|
115 |
+
|
116 |
+
def get_pipeline_params(imputer_params, classifier_params):
|
117 |
+
pipeline_params = {**imputer_params, **classifier_params}
|
118 |
+
return pipeline_params
|
119 |
+
|
120 |
+
def train_model(df_train, df_test, imputer_type, classifier_type):
|
121 |
+
# get data arrays from the data frame for train and test sets
|
122 |
+
X_train, Y_train = get_data_from_data_frame(df_train)
|
123 |
+
X_test, Y_test = get_data_from_data_frame(df_test)
|
124 |
+
|
125 |
+
# get imputer and its params
|
126 |
+
imputer, imputer_params = get_imputer(imputer_type)
|
127 |
+
|
128 |
+
# get classifier and its params
|
129 |
+
classifier, classifier_params = get_classifier(classifier_type)
|
130 |
+
|
131 |
+
# get the pipeline params
|
132 |
+
pipeline_params = get_pipeline_params(imputer_params, classifier_params)
|
133 |
+
|
134 |
+
print("\n" + "-"*100)
|
135 |
+
# build the model pipeline
|
136 |
+
if classifier_type == "svc" or classifier_type == "log_reg":
|
137 |
+
scaler = get_scaler()
|
138 |
+
pca, pca_params = get_pca(X_train.shape[1])
|
139 |
+
print(f"Started training the model with the imputer: {imputer_type}, preprocessing: std_scaler + pca, classifier: {classifier_type}")
|
140 |
+
|
141 |
+
pipeline = Pipeline([("imputer", imputer), ("scaler", scaler), ("pca", pca), ("classifier", classifier)])
|
142 |
+
pipeline_params = get_pipeline_params(pipeline_params, pca_params)
|
143 |
+
else:
|
144 |
+
print(f"Started training the model with the imputer: {imputer_type}, classifier: {classifier_type}")
|
145 |
+
pipeline = Pipeline([("imputer", imputer), ("classifier", classifier)])
|
146 |
+
print("Model pipeline params space: ")
|
147 |
+
print(pipeline_params)
|
148 |
+
print("-"*100)
|
149 |
+
|
150 |
+
# setup grid search with k-fold cross validation
|
151 |
+
k_fold_cv = KFold(n_splits=5, shuffle=True, random_state=4)
|
152 |
+
grid_cv = GridSearchCV(pipeline, pipeline_params, scoring="f1", cv=k_fold_cv)
|
153 |
+
grid_cv.fit(X_train, Y_train)
|
154 |
+
|
155 |
+
# get the cross validation score and the params for the best estimator
|
156 |
+
cv_best_estimator = grid_cv.best_estimator_
|
157 |
+
cv_best_f1 = grid_cv.best_score_
|
158 |
+
cv_best_params = grid_cv.best_params_
|
159 |
+
|
160 |
+
# predict and compute train set metrics
|
161 |
+
Y_train_pred = cv_best_estimator.predict(X_train)
|
162 |
+
train_f1 = f1_score(Y_train, Y_train_pred)
|
163 |
+
train_acc = accuracy_score(Y_train, Y_train_pred)
|
164 |
+
|
165 |
+
# predict and compute test set metrics
|
166 |
+
Y_test_pred = cv_best_estimator.predict(X_test)
|
167 |
+
test_f1 = f1_score(Y_test, Y_test_pred)
|
168 |
+
test_acc = accuracy_score(Y_test, Y_test_pred)
|
169 |
+
|
170 |
+
print("\n" + "-"*50)
|
171 |
+
# begin mlflow logging for the best estimator
|
172 |
+
mlflow.set_experiment("water_potability")
|
173 |
+
experiment = mlflow.get_experiment_by_name("water_potability")
|
174 |
+
print(f"Started mlflow logging for the best estimator")
|
175 |
+
with mlflow.start_run(experiment_id=experiment.experiment_id):
|
176 |
+
# log the model and the metrics
|
177 |
+
mlflow.sklearn.log_model(cv_best_estimator, f"{imputer_type}_{classifier_type}")
|
178 |
+
mlflow.sklearn.save_model(cv_best_estimator, f"{imputer_type}_{classifier_type}")
|
179 |
+
mlflow.log_params(cv_best_params)
|
180 |
+
mlflow.log_metric("cv_f1_score", cv_best_f1)
|
181 |
+
mlflow.log_metric("train_f1_score", train_f1)
|
182 |
+
mlflow.log_metric("train_acc_score", train_acc)
|
183 |
+
mlflow.log_metric("test_f1_score", test_f1)
|
184 |
+
mlflow.log_metric("test_acc_score", test_acc)
|
185 |
+
# end mlflow logging
|
186 |
+
mlflow.end_run()
|
187 |
+
print(f"Completed mlflow logging for the best estimator")
|
188 |
+
print("-"*50)
|
189 |
+
return
|
190 |
+
|
191 |
+
def init_and_train_model(ARGS):
|
192 |
+
df_csv = read_csv_file(ARGS.file_csv)
|
193 |
+
df_train, df_test = train_test_split(df_csv, test_size=0.1, random_state=4)
|
194 |
+
|
195 |
+
num_samples_train = df_train.shape[0]
|
196 |
+
num_samples_test = df_test.shape[0]
|
197 |
+
|
198 |
+
print("\n" + "-"*40)
|
199 |
+
print("Num samples after splitting the dataset")
|
200 |
+
print("-"*40)
|
201 |
+
print(f"train: {num_samples_train}, test: {num_samples_test}")
|
202 |
+
|
203 |
+
print("\n" + "-"*40)
|
204 |
+
print("A few samples from train data")
|
205 |
+
print("-"*40)
|
206 |
+
print(df_train.head())
|
207 |
+
|
208 |
+
if ARGS.is_train:
|
209 |
+
train_model(df_train, df_test, ARGS.imputer_type, ARGS.classifier_type)
|
210 |
+
return
|
211 |
+
|
212 |
+
def main():
|
213 |
+
file_csv = "dataset/water_potability.csv"
|
214 |
+
classifier_type = "ada_boost"
|
215 |
+
imputer_type = "knn"
|
216 |
+
is_train = 1
|
217 |
+
|
218 |
+
parser = argparse.ArgumentParser(
|
219 |
+
formatter_class=argparse.ArgumentDefaultsHelpFormatter
|
220 |
+
)
|
221 |
+
|
222 |
+
parser.add_argument("--file_csv", default=file_csv,
|
223 |
+
type=str, help="full path to dataset csv file")
|
224 |
+
parser.add_argument("--is_train", default=is_train,
|
225 |
+
type=int, choices=[0, 1], help="to train or not")
|
226 |
+
parser.add_argument("--classifier_type", default=classifier_type,
|
227 |
+
type=str, choices=["ada_boost", "log_reg", "random_forest", "svc", "light_gbm"],
|
228 |
+
help="classifier to be used in the training model pipeline")
|
229 |
+
parser.add_argument("--imputer_type", default=imputer_type,
|
230 |
+
type=str, choices=["simple", "knn", "iterative"],
|
231 |
+
help="imputer to be used in the training model pipeline")
|
232 |
+
|
233 |
+
ARGS, unparsed = parser.parse_known_args()
|
234 |
+
init_and_train_model(ARGS)
|
235 |
+
return
|
236 |
+
|
237 |
+
if __name__ == "__main__":
|
238 |
+
main()
|
modeling/ml_model_test.py
ADDED
@@ -0,0 +1,38 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import argparse
|
2 |
+
|
3 |
+
import mlflow
|
4 |
+
import numpy as np
|
5 |
+
from sklearn.metrics import classification_report
|
6 |
+
|
7 |
+
from ml_model_dev import load_ml_model, train_test_split, read_csv_file
|
8 |
+
|
9 |
+
def test_ml_pipeline(ARGS):
|
10 |
+
df_csv = read_csv_file(ARGS.file_csv)
|
11 |
+
df_train, df_test = train_test_split(df_csv, test_size=0.1, random_state=4)
|
12 |
+
arr_test = df_test.to_numpy()
|
13 |
+
X_test, Y_test = arr_test[:, :-1], arr_test[:, -1:].reshape(-1)
|
14 |
+
|
15 |
+
model_pipeline = load_ml_model(ARGS.pkl_file_name)
|
16 |
+
Y_pred_test = model_pipeline.predict(X_test)
|
17 |
+
print(classification_report(Y_test, Y_pred_test))
|
18 |
+
return
|
19 |
+
|
20 |
+
def main():
|
21 |
+
file_csv = "dataset/water_potability.csv"
|
22 |
+
pkl_file_name = "trained_models/knn_ada_boost"
|
23 |
+
|
24 |
+
parser = argparse.ArgumentParser(
|
25 |
+
formatter_class=argparse.ArgumentDefaultsHelpFormatter
|
26 |
+
)
|
27 |
+
|
28 |
+
parser.add_argument("--file_csv", default=file_csv,
|
29 |
+
type=str, help="full path to dataset csv file")
|
30 |
+
parser.add_argument("--pkl_file_name", default=pkl_file_name,
|
31 |
+
type=str, help="full path to ml model pkl file")
|
32 |
+
|
33 |
+
ARGS, unparsed = parser.parse_known_args()
|
34 |
+
test_ml_pipeline(ARGS)
|
35 |
+
return
|
36 |
+
|
37 |
+
if __name__ == "__main__":
|
38 |
+
main()
|