403 lines
16 KiB
Python
Executable file
403 lines
16 KiB
Python
Executable file
#!/usr/bin/env python3
|
|
# -*- coding: utf-8 -*-
|
|
"""
|
|
Created on Mon May 23 23:25:26 2022
|
|
|
|
@author: tanu
|
|
"""
|
|
#%%
|
|
import os, sys
|
|
import pandas as pd
|
|
import numpy as np
|
|
import pprint as pp
|
|
from copy import deepcopy
|
|
from sklearn import linear_model
|
|
from sklearn import datasets
|
|
from collections import Counter
|
|
|
|
from sklearn.linear_model import LogisticRegression, LogisticRegressionCV
|
|
from sklearn.linear_model import RidgeClassifier, RidgeClassifierCV, SGDClassifier, PassiveAggressiveClassifier
|
|
|
|
from sklearn.naive_bayes import BernoulliNB
|
|
from sklearn.neighbors import KNeighborsClassifier
|
|
from sklearn.svm import SVC
|
|
from sklearn.tree import DecisionTreeClassifier, ExtraTreeClassifier
|
|
from sklearn.ensemble import RandomForestClassifier, ExtraTreesClassifier, AdaBoostClassifier, GradientBoostingClassifier, BaggingClassifier
|
|
from sklearn.naive_bayes import GaussianNB
|
|
from sklearn.gaussian_process import GaussianProcessClassifier, kernels
|
|
from sklearn.gaussian_process.kernels import RBF, DotProduct, Matern, RationalQuadratic, WhiteKernel
|
|
|
|
from sklearn.discriminant_analysis import LinearDiscriminantAnalysis, QuadraticDiscriminantAnalysis
|
|
from sklearn.neural_network import MLPClassifier
|
|
|
|
from sklearn.svm import SVC
|
|
from xgboost import XGBClassifier
|
|
from sklearn.naive_bayes import MultinomialNB
|
|
from sklearn.preprocessing import StandardScaler, MinMaxScaler, OneHotEncoder
|
|
|
|
from sklearn.compose import ColumnTransformer
|
|
from sklearn.compose import make_column_transformer
|
|
|
|
from sklearn.metrics import make_scorer, confusion_matrix, accuracy_score, balanced_accuracy_score, precision_score, average_precision_score, recall_score
|
|
from sklearn.metrics import roc_auc_score, roc_curve, f1_score, matthews_corrcoef, jaccard_score, classification_report
|
|
|
|
# added
|
|
from sklearn.model_selection import train_test_split, cross_validate, cross_val_score, LeaveOneOut, KFold, RepeatedKFold, cross_val_predict
|
|
|
|
from sklearn.model_selection import train_test_split, cross_validate, cross_val_score
|
|
from sklearn.model_selection import StratifiedKFold,RepeatedStratifiedKFold, RepeatedKFold
|
|
|
|
from sklearn.pipeline import Pipeline, make_pipeline
|
|
|
|
from sklearn.feature_selection import RFE, RFECV
|
|
|
|
import itertools
|
|
import seaborn as sns
|
|
import matplotlib.pyplot as plt
|
|
|
|
from statistics import mean, stdev, median, mode
|
|
|
|
from imblearn.over_sampling import RandomOverSampler
|
|
from imblearn.under_sampling import RandomUnderSampler
|
|
from imblearn.over_sampling import SMOTE
|
|
from sklearn.datasets import make_classification
|
|
from imblearn.combine import SMOTEENN
|
|
from imblearn.combine import SMOTETomek
|
|
|
|
from imblearn.over_sampling import SMOTENC
|
|
from imblearn.under_sampling import EditedNearestNeighbours
|
|
from imblearn.under_sampling import RepeatedEditedNearestNeighbours
|
|
|
|
from sklearn.model_selection import GridSearchCV
|
|
from sklearn.base import BaseEstimator
|
|
from sklearn.impute import KNNImputer as KNN
|
|
import json
|
|
import argparse
|
|
import re
|
|
#####################################
|
|
rs = {'random_state': 42}
|
|
njobs = {'n_jobs': os.cpu_count() } # the number of jobs should equal the number of CPU cores
|
|
|
|
scoring_fn = ({ 'mcc' : make_scorer(matthews_corrcoef)
|
|
, 'fscore' : make_scorer(f1_score)
|
|
, 'precision' : make_scorer(precision_score)
|
|
, 'recall' : make_scorer(recall_score)
|
|
, 'accuracy' : make_scorer(accuracy_score)
|
|
, 'roc_auc' : make_scorer(roc_auc_score)
|
|
, 'jcc' : make_scorer(jaccard_score)
|
|
})
|
|
|
|
|
|
mcc_score_fn = {'mcc': make_scorer(matthews_corrcoef)}
|
|
jacc_score_fn = {'jcc': make_scorer(jaccard_score)}
|
|
skf_cv = StratifiedKFold(n_splits = 10
|
|
#, shuffle = False, random_state= None)
|
|
, shuffle = True,**rs)
|
|
|
|
rskf_cv = RepeatedStratifiedKFold(n_splits = 10
|
|
, n_repeats = 3
|
|
, **rs)
|
|
|
|
###############################################################################
|
|
def fsgs_rfecv(input_df
|
|
, target
|
|
, param_gridLd = [{'fs__min_features_to_select' : [1]}]
|
|
, blind_test_df = pd.DataFrame()
|
|
, blind_test_target = pd.Series(dtype = 'int64')
|
|
, estimator = LogisticRegression(**rs) # placeholder
|
|
, use_fs = False # uses estimator as the RFECV parameter for fs. Set to TRUE if you want to supply custom_fs as shown below
|
|
, custom_fs = RFECV(DecisionTreeClassifier(**rs) , cv = skf_cv, scoring = 'matthews_corrcoef')
|
|
, cv_method = skf_cv
|
|
, var_type = ['numerical', 'categorical' , 'mixed']
|
|
, resampling_type = 'none'
|
|
, verbose = 3
|
|
, random_state = 42
|
|
, n_jobs = os.cpu_count()
|
|
):
|
|
'''
|
|
returns
|
|
Dict containing results from FS and hyperparam tuning for a given estiamtor
|
|
|
|
>>> ADD MORE <<<
|
|
|
|
optimised/selected based on mcc
|
|
|
|
'''
|
|
rs = {'random_state': random_state}
|
|
njobs = {'n_jobs': n_jobs}
|
|
|
|
|
|
###########################################################################
|
|
#================================================
|
|
# Determine categorical and numerical features
|
|
#================================================
|
|
numerical_ix = input_df.select_dtypes(include=['int64', 'float64']).columns
|
|
numerical_ix
|
|
categorical_ix = input_df.select_dtypes(include=['object', 'bool']).columns
|
|
categorical_ix
|
|
|
|
#================================================
|
|
# Determine preprocessing steps ~ var_type
|
|
#================================================
|
|
if var_type == 'numerical':
|
|
t = [('num', MinMaxScaler(), numerical_ix)]
|
|
|
|
if var_type == 'categorical':
|
|
t = [('cat', OneHotEncoder(), categorical_ix)]
|
|
|
|
if var_type == 'mixed':
|
|
t = [('cat', OneHotEncoder(), categorical_ix)
|
|
, ('num', MinMaxScaler(), numerical_ix)]
|
|
|
|
col_transform = ColumnTransformer(transformers = t
|
|
, remainder='passthrough')
|
|
|
|
###########################################################################
|
|
#==================================================
|
|
# Create var_type ~ column names
|
|
# using one hot encoder with RFECV means
|
|
# the names internally are lost. Hence
|
|
# fit col_transformeer to my input_df and get
|
|
# all the column names out and stored in a var
|
|
# to allow the 'selected features' to be subsetted
|
|
# from the numpy boolean array
|
|
#=================================================
|
|
col_transform.fit(input_df)
|
|
col_transform.get_feature_names_out()
|
|
|
|
var_type_colnames = col_transform.get_feature_names_out()
|
|
var_type_colnames = pd.Index(var_type_colnames)
|
|
|
|
if var_type == 'mixed':
|
|
print('\nVariable type is:', var_type
|
|
, '\nNo. of columns in input_df:', len(input_df.columns)
|
|
, '\nNo. of columns post one hot encoder:', len(var_type_colnames))
|
|
else:
|
|
print('\nNo. of columns in input_df:', len(input_df.columns))
|
|
|
|
#==================================
|
|
# Build FS with supplied estimator
|
|
#==================================
|
|
if use_fs:
|
|
fs = custom_fs
|
|
else:
|
|
fs = RFECV(estimator, cv = skf_cv, scoring = 'matthews_corrcoef')
|
|
|
|
#==================================
|
|
# Build basic param grid
|
|
#==================================
|
|
# param_gridD = [
|
|
# {'fs__min_features_to_select' : [1]
|
|
# }]
|
|
|
|
############################################################################
|
|
# Create Pipeline object
|
|
pipe = Pipeline([
|
|
('pre', col_transform),
|
|
('fs', fs),
|
|
('clf', estimator)])
|
|
############################################################################
|
|
# Define GridSearchCV
|
|
gscv_fs = GridSearchCV(pipe
|
|
#, param_gridLd = param_gridD
|
|
, param_gridLd
|
|
, cv = cv_method
|
|
, scoring = scoring_fn
|
|
, refit = 'mcc'
|
|
, verbose = 3
|
|
, return_train_score = True
|
|
, **njobs)
|
|
|
|
gscv_fs.fit(input_df, target)
|
|
|
|
###########################################################################
|
|
# Get best param and scores out
|
|
gscv_fs.best_params_
|
|
gscv_fs.best_score_
|
|
|
|
# Training best score corresponds to the max of the mean_test<score>
|
|
train_bscore = round(gscv_fs.best_score_, 2); train_bscore
|
|
print('\nTraining best score (MCC):', train_bscore)
|
|
gscv_fs.cv_results_['mean_test_mcc']
|
|
round(gscv_fs.cv_results_['mean_test_mcc'].max(),2)
|
|
round(np.nanmax(gscv_fs.cv_results_['mean_test_mcc']),2)
|
|
|
|
check_train_score = [round(gscv_fs.cv_results_['mean_test_mcc'].max(),2)
|
|
, round(np.nanmax(gscv_fs.cv_results_['mean_test_mcc']),2)]
|
|
|
|
check_train_score = np.nanmax(check_train_score)
|
|
|
|
# Training results
|
|
gscv_tr_resD = gscv_fs.cv_results_
|
|
mod_refit_param = gscv_fs.refit
|
|
|
|
# sanity check
|
|
if train_bscore == check_train_score:
|
|
print('\nVerified training score (MCC):', train_bscore )
|
|
else:
|
|
sys.exit('\nTraining score could not be internatlly verified. Please check training results dict')
|
|
|
|
#-------------------------
|
|
# Dict of CV results
|
|
#-------------------------
|
|
cv_allD = gscv_fs.cv_results_
|
|
cvdf0 = pd.DataFrame(cv_allD)
|
|
cvdf = cvdf0.filter(regex='mean_test', axis = 1)
|
|
cvdfT = cvdf.T
|
|
cvdfT.columns = ['cv_score']
|
|
cvdfTr = cvdfT.loc[:,'cv_score'].round(decimals = 2) # round values
|
|
cvD = cvdfTr.to_dict()
|
|
print('\n CV results dict generated for:', len(scoring_fn), 'scores'
|
|
, '\nThese are:', scoring_fn.keys())
|
|
|
|
#-------------------------
|
|
# Blind test: REAL check!
|
|
#-------------------------
|
|
#tp = gscv_fs.predict(X_bts)
|
|
tp = gscv_fs.predict(blind_test_df)
|
|
|
|
print('\nMCC on Blind test:' , round(matthews_corrcoef(blind_test_target, tp),2))
|
|
print('\nAccuracy on Blind test:', round(accuracy_score(blind_test_target, tp),2))
|
|
|
|
#=================
|
|
# info extraction
|
|
#=================
|
|
# gives input vals??
|
|
gscv_fs._check_n_features
|
|
|
|
# gives gscv params used
|
|
gscv_fs._get_param_names()
|
|
|
|
# gives ??
|
|
gscv_fs.best_estimator_
|
|
gscv_fs.best_params_ # gives best estimator params as a dict
|
|
gscv_fs.best_estimator_._final_estimator # similar to above, doesn't contain max_iter
|
|
gscv_fs.best_estimator_.named_steps['fs'].get_support()
|
|
gscv_fs.best_estimator_.named_steps['fs'].ranking_ # array of ranks for the features
|
|
|
|
gscv_fs.best_estimator_.named_steps['fs'].grid_scores_.mean()
|
|
gscv_fs.best_estimator_.named_steps['fs'].grid_scores_.max()
|
|
#gscv_fs.best_estimator_.named_steps['fs'].grid_scores_
|
|
|
|
estimator_mask = gscv_fs.best_estimator_.named_steps['fs'].get_support()
|
|
|
|
|
|
############################################################################
|
|
#============
|
|
# FS results
|
|
#============
|
|
# Now get the features out
|
|
|
|
#--------------
|
|
# All features
|
|
#--------------
|
|
all_features = gscv_fs.feature_names_in_
|
|
n_all_features = gscv_fs.n_features_in_
|
|
#all_features = gsfit.feature_names_in_
|
|
|
|
#--------------
|
|
# Selected features by the classifier
|
|
# Important to have var_type_colnames here
|
|
#----------------
|
|
#sel_features = X.columns[gscv_fs.best_estimator_.named_steps['fs'].get_support()] 3 only for numerical df
|
|
sel_features = var_type_colnames[gscv_fs.best_estimator_.named_steps['fs'].get_support()]
|
|
n_sf = gscv_fs.best_estimator_.named_steps['fs'].n_features_
|
|
|
|
#--------------
|
|
# Get model name
|
|
#--------------
|
|
model_name = gscv_fs.best_estimator_.named_steps['clf']
|
|
b_model_params = gscv_fs.best_params_
|
|
|
|
print('\n========================================'
|
|
, '\nRunning model:'
|
|
, '\nModel name:', model_name
|
|
, '\n==============================================='
|
|
, '\nRunning feature selection with RFECV for model'
|
|
, '\nTotal no. of features in model:', len(all_features)
|
|
, '\nThese are:\n', all_features, '\n\n'
|
|
, '\nNo of features for best model: ', n_sf
|
|
, '\nThese are:', sel_features, '\n\n'
|
|
, '\nBest Model hyperparams:', b_model_params
|
|
)
|
|
|
|
###########################################################################
|
|
############################## OUTPUT #####################################
|
|
###########################################################################
|
|
#=========================
|
|
# Blind test: BTS results
|
|
#=========================
|
|
# Build the final results with all scores for a feature selected model
|
|
#bts_predict = gscv_fs.predict(X_bts)
|
|
bts_predict = gscv_fs.predict(blind_test_df)
|
|
|
|
print('\nMCC on Blind test:' , round(matthews_corrcoef(blind_test_target, bts_predict),2))
|
|
print('\nAccuracy on Blind test:', round(accuracy_score(blind_test_target, bts_predict),2))
|
|
bts_mcc_score = round(matthews_corrcoef(blind_test_target, bts_predict),2)
|
|
|
|
# Diff b/w train and bts test scores
|
|
train_test_diff = train_bscore - bts_mcc_score
|
|
print('\nDiff b/w train and blind test score (MCC):', train_test_diff)
|
|
|
|
lr_btsD ={}
|
|
#lr_btsD['bts_mcc'] = bts_mcc_score
|
|
lr_btsD['bts_fscore'] = round(f1_score(blind_test_target, bts_predict),2)
|
|
lr_btsD['bts_precision'] = round(precision_score(blind_test_target, bts_predict),2)
|
|
lr_btsD['bts_recall'] = round(recall_score(blind_test_target, bts_predict),2)
|
|
lr_btsD['bts_accuracy'] = round(accuracy_score(blind_test_target, bts_predict),2)
|
|
lr_btsD['bts_roc_auc'] = round(roc_auc_score(blind_test_target, bts_predict),2)
|
|
lr_btsD['bts_jcc'] = round(jaccard_score(blind_test_target, bts_predict),2)
|
|
lr_btsD
|
|
|
|
#===========================
|
|
# Add FS related model info
|
|
#===========================
|
|
model_namef = str(model_name)
|
|
# FIXME: doesn't tell you which it has chosen
|
|
fs_methodf = str(gscv_fs.best_estimator_.named_steps['fs'])
|
|
all_featuresL = list(all_features)
|
|
fs_res_arrayf = str(list( gscv_fs.best_estimator_.named_steps['fs'].get_support()))
|
|
fs_res_array_rankf = str(list( gscv_fs.best_estimator_.named_steps['fs'].ranking_))
|
|
sel_featuresf = list(sel_features)
|
|
n_sf = int(n_sf)
|
|
|
|
output_modelD = {'model_name': model_namef
|
|
, 'model_refit_param': mod_refit_param
|
|
, 'Best_model_params': b_model_params
|
|
, 'n_all_features': n_all_features
|
|
, 'fs_method': fs_methodf
|
|
, 'fs_res_array': fs_res_arrayf
|
|
, 'fs_res_array_rank': fs_res_array_rankf
|
|
, 'all_feature_names': all_featuresL
|
|
, 'n_sel_features': n_sf
|
|
, 'sel_features_names': sel_featuresf}
|
|
#output_modelD
|
|
|
|
#========================================
|
|
# Update output_modelD with bts_results
|
|
#========================================
|
|
output_modelD.update(lr_btsD)
|
|
output_modelD
|
|
|
|
output_modelD['train_score (MCC)'] = train_bscore
|
|
output_modelD['bts_mcc'] = bts_mcc_score
|
|
output_modelD['train_bts_diff'] = round(train_test_diff,2)
|
|
output_modelD['resampling'] = resampling_type
|
|
|
|
print(output_modelD)
|
|
|
|
nlen = len(output_modelD)
|
|
|
|
#========================================
|
|
# Update output_modelD with cv_results
|
|
#========================================
|
|
output_modelD.update(cvD)
|
|
|
|
if (len(output_modelD) == nlen + len(cvD)):
|
|
print('\nFS run complete for model:', estimator
|
|
, '\nFS using:', fs
|
|
, '\nOutput dict size:', len(output_modelD))
|
|
return(output_modelD)
|
|
else:
|
|
sys.exit('\nFAIL:numbers mismatch output dict length not as expected. Please check')
|
|
|