X = df_bands.values
y = df_meta['Megaclasse'].values
# sample data
X, _, y, _ = train_test_split(X,
y,
train_size=.3,
shuffle=True,
stratify=y,
random_state=random_state)
cv = StratifiedKFold(n_splits=5, shuffle=True, random_state=0)
## Final model - Parameter tuning
pipelines, param_grids = check_pipelines([anomally_detection, classifiers_1],
random_state=0,
n_runs=1)
fit_params = check_fit_params(pre_fit_params)
model_search = ModelSearchCV(pipelines,
param_grids,
scoring=scorers,
refit='accuracy',
n_jobs=-1,
cv=cv,
verbose=1)
model_search.fit(X, y, **fit_params)
pickle.dump(model_search,
open(RESULTS_PATH + 'final_pipeline_parameter_tuning.pkl', 'wb'))
# 'selection_strategy': ['combined', 'minority', 'majority'],
# 'truncation_factor': [-.5,0,.5],
# 'deformation_factor': [0,.5,1],
# 'k_neighbors_filter': [3,5]
# })
#]
classifiers = [
('randomforestclassifier', RandomForestClassifier(n_estimators=100, random_state=random_state), {})
]
objects_list = [noise_objs, data_filters, #oversamplers,
classifiers]
pipelines, param_grid = check_pipelines(objects_list, random_state, 1)
fit_params = {}
for clf_name in list(dict(pipelines).keys()):
clf_name_split = clf_name.split('|')
if clf_name_split[1]=='DenoisedGeometricSMOTE':
pass
elif clf_name_split[1]=='singlefilter':
fit_params[f'{clf_name_split[1]}__filters'] = [single_filter]
elif clf_name_split[1]!='no_filter':
fit_params[f'{clf_name_split[1]}__filters'] = filts
cv = StratifiedKFold(n_splits=5, shuffle=True, random_state=0)
model_search = ModelSearchCV(pipelines, param_grid, n_jobs=-1, cv=cv, verbose=1)
model_search.fit(X,y,**fit_params)
df_meta = df.drop(df.columns[df.columns.str.startswith('X201')|df.columns.str.startswith('ND')], axis=1)
df_bands = df.drop(columns=df_meta.columns)
# normalize
znorm = StandardScaler()
df_bands = pd.DataFrame(znorm.fit_transform(df_bands.values), columns=df_bands.columns, index=df_bands.index)
X = df_bands.values
y = df_meta['Label'].values
ids = df_meta['Object'].values
cv = StratifiedKFold(n_splits=5, shuffle=True, random_state=0)
## Experiment 1 (feature selection)
pipelines_feature, param_grid_feature = check_pipelines(
[feature_selection, classifiers_1],
random_state=0,
n_runs=1
)
model_search_feature = ModelSearchCV(pipelines_feature, param_grid_feature, n_jobs=-1, cv=cv, verbose=1)
model_search_feature.fit(X,y)
df_results_feature = report_model_search_results(model_search_feature)\
.sort_values('mean_test_score', ascending=False)
#df_results_feature.to_csv('results_feature_selection.csv')
pickle.dump(model_search_feature, open('gini_feature_selection.pkl','wb'))
rfc = RandomForestClassifier(
'multi_class': ['ovr', 'multinomial'],
'penalty': ['l2', 'none']
}),
]
# setup scorers
def geometric_mean_macro(X, y):
return geometric_mean_score(X, y, average='macro')
SCORERS['geometric_mean_macro'] = make_scorer(geometric_mean_macro)
scorers = ['accuracy', 'f1_macro', 'geometric_mean_macro']
pipelines, param_grid = check_pipelines([classifiers],
random_state=random_state,
n_runs=1)
cv = StratifiedKFold(n_splits=5, shuffle=True, random_state=random_state)
# run experiment
model_search = ModelSearchCV(pipelines,
param_grid,
scoring=scorers,
refit='accuracy',
n_jobs=-1,
cv=cv,
verbose=1)
model_search.fit(X, y)
pickle.dump(model_search,
X = df_bands.values
y = df_meta['Megaclasse'].values
# sample data
X, _, y, _ = train_test_split(X,
y,
train_size=.1,
shuffle=True,
stratify=y,
random_state=random_state)
cv = StratifiedKFold(n_splits=5, shuffle=True, random_state=0)
## Experiment 1 (feature selection)
pipelines_feature, param_grid_feature = check_pipelines(
[feature_selection, classifiers], random_state=0, n_runs=1)
model_search_feature = ModelSearchCV(pipelines_feature,
param_grid_feature,
scoring=scorers,
refit='accuracy',
n_jobs=-1,
cv=cv,
verbose=1)
model_search_feature.fit(X, y)
df_results_feature = report_model_search_results(model_search_feature)\
.sort_values('mean_test_accuracy', ascending=False)
df_results_feature.to_csv('results_feature_selection.csv')
pickle.dump(model_search_feature,
open('model_search_feature_selection.pkl', 'wb'))
axis=1)
df_bands = df.drop(columns=df_meta.columns)
# normalize
znorm = StandardScaler()
df_bands = pd.DataFrame(znorm.fit_transform(df_bands.values),
columns=df_bands.columns,
index=df_bands.index)
X = df_bands.values
y = df_meta['Label'].values
ids = df_meta['Object'].values
cv = StratifiedKFold(n_splits=5, shuffle=True, random_state=0)
## Experiment 1 (feature selection)
pipelines_feature, param_grid_feature = check_pipelines(
[feature_selection, classifiers_1], random_state=0, n_runs=1)
model_search_feature = ModelSearchCV(pipelines_feature,
param_grid_feature,
n_jobs=-1,
cv=cv,
verbose=1)
model_search_feature.fit(X, y)
df_results_feature = report_model_search_results(model_search_feature)\
.sort_values('mean_test_score', ascending=False)
df_results_feature.to_csv('results_feature_selection.csv')
pickle.dump(model_search_feature,
open('model_search_feature_selection.pkl', 'wb'))
## Experiment 2 (anomally detection)