added FS to MultClfs.py and modified data for different splits for consistency
This commit is contained in:
parent
edb7aebd6a
commit
e2bc384155
12 changed files with 1585 additions and 994 deletions
|
|
@ -502,10 +502,11 @@ def ProcessMultModelsCl(inputD = {}):
|
|||
# Combine WF+Metadata: Final output
|
||||
#-------------------------------------
|
||||
# checking indices for the dfs to combine:
|
||||
c1 = list(set(combined_baseline_wf.index))
|
||||
c2 = list(metaDF.index)
|
||||
c1L = list(set(combined_baseline_wf.index))
|
||||
c2L = list(metaDF.index)
|
||||
|
||||
if c1 == c2:
|
||||
#if set(c1L) == set(c2L):
|
||||
if set(c1L) == set(c2L) and all(x in c2L for x in c1L) and all(x in c1L for x in c2L):
|
||||
print('\nPASS: proceeding to merge metadata with CV and BT dfs')
|
||||
combDF = pd.merge(combined_baseline_wf, metaDF, left_index = True, right_index = True)
|
||||
else:
|
||||
|
|
@ -531,5 +532,302 @@ def ProcessMultModelsCl(inputD = {}):
|
|||
return combDF
|
||||
|
||||
###############################################################################
|
||||
#%% Feature selection function ################################################
|
||||
############################
|
||||
# fsgs_rfecv()
|
||||
############################
|
||||
# Run FS using some classifier models
|
||||
#
|
||||
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']
|
||||
, verbose = 3
|
||||
):
|
||||
'''
|
||||
returns
|
||||
Dict containing results from FS and hyperparam tuning for a given estiamtor
|
||||
|
||||
>>> ADD MORE <<<
|
||||
|
||||
optimised/selected based on mcc
|
||||
|
||||
'''
|
||||
###########################################################################
|
||||
#================================================
|
||||
# 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)
|
||||
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')
|
||||
|
||||
Loading…
Add table
Add a link
Reference in a new issue