def fit_logistic_or_with_crossvalidation(X, y):
"""An ordinal model of dataset with hyperparameter
cross-validation. Ordinal Ridge (regression) variant.
Parameters & returns as per other training functions.
"""
basemod = mord.OrdinalRidge()
cv = 5
param_grid = {
'fit_intercept': [True, False],
'alpha': [0.2, 0.4, 0.6, 0.8, 1.0, 2.0, 3.0],
'normalize': [True, False]
}
return fit_classifier_with_crossvalidation(X,
y,
basemod,
cv,
param_grid,
verbose=False)
def train_ordinal_logistic(train_features, train_labels, skip_grid_search,
evaluation, num_jobs, loss, alpha, cost,
ordinal_algorithm):
"""
returns the trained ordinal logistic model. loss, alpha and cost are ignored if grid
search is requested.
alpha: used only for se, it, at, and ridge and if grid search is not requested
cost: used only for lad and if grid search is not requested
loss: used only for lad and if grid search is not requested
"""
# requested grid search. find best parameters, to achieve highest average score
if not skip_grid_search:
penalty_weights = 'dummy'
clf = grid_search.grid_search(evaluation, train_features, train_labels,
penalty_weights, ordinal_algorithm,
num_jobs)
params = clf.best_params_
if 'penalty' in params:
loss = params['loss']
if 'alpha' in params:
alpha = params['alpha']
if 'cost' in params:
cost = params['cost']
# Now perform the training on full train data.
if ordinal_algorithm == 'logisticse':
model = mord.LogisticSE(alpha=alpha, max_iter=20000)
elif ordinal_algorithm == 'logisticit':
model = mord.LogisticIT(alpha=alpha, max_iter=20000)
elif ordinal_algorithm == 'logisticat':
model = mord.LogisticAT(alpha=alpha, max_iter=20000)
elif ordinal_algorithm == 'ordinalridge':
model = mord.OrdinalRidge(alpha=alpha)
elif ordinal_algorithm == 'lad':
model = mord.LAD(C=cost, loss=loss, max_iter=10000)
model = model.fit(train_features, train_labels)
return model
def run_classification(X_train,
X_test,
y_train,
y_test,
how='rfc',
random_state=0,
n_jobs=2,
cv=False,
stand=False,
verbose=True,
full_output=False,
**classpar):
"""
"""
if stand:
X_train = StandardScaler().fit_transform(X_train)
X_test = StandardScaler().fit_transform(X_test)
if how == 'or1':
pars = {'alpha': 1e0, 'verbose': 1, 'max_iter': 1e5}
for par in pars:
if par not in classpar: classpar.update({par: pars.get(par)})
clasif = mord.LogisticAT(**classpar)
elif how == 'or2':
pars = {'alpha': 1e0, 'verbose': 1, 'max_iter': 1e5}
for par in pars:
if par not in classpar: classpar.update({par: pars.get(par)})
clasif = mord.LogisticIT(**classpar)
elif how == 'or3':
pars = {
'alpha': 1e0,
'fit_intercept': True,
'normalize': False,
'copy_X': True,
'max_iter': None,
'tol': 0.001,
'solver': 'auto'
}
for par in pars:
if par not in classpar: classpar.update({par: pars.get(par)})
clasif = mord.OrdinalRidge(random_state=random_state, **classpar)
elif how == 'or4':
pars = {
'epsilon': 0.0,
'tol': 0.0001,
'C': 1.0,
'loss': 'l1',
'fit_intercept': True,
'intercept_scaling': 1.0,
'dual': True,
'verbose': 0,
'max_iter': 10000
}
for par in pars:
if par not in classpar: classpar.update({par: pars.get(par)})
clasif = mord.LAD(random_state=random_state, **classpar)
elif how == 'prank':
pars = {'n_iter': 1000, 'shuffle': True}
for par in pars:
if par not in classpar: classpar.update({par: pars.get(par)})
clasif = ranking.PRank(random_state=random_state, **classpar)
elif how == 'kprank':
pars = {
'n_iter': 200,
'shuffle': True,
'kernel': 'rbf',
'gamma': 1e2,
'degree': 3,
'coef0': 1
}
for par in pars:
if par not in classpar: classpar.update({par: pars.get(par)})
clasif = ranking.KernelPRank(random_state=random_state, **classpar)
elif how == 'rfc':
pars = {
'n_estimators': 1000,
'criterion': 'gini',
'max_depth': None,
'min_samples_split': 2,
'min_samples_leaf': 1,
'min_weight_fraction_leaf': 0.0,
'max_features': 'auto',
'max_leaf_nodes': None,
'min_impurity_split': 1e-07,
'bootstrap': True,
'oob_score': True,
'verbose': 0,
'warm_start': False,
'class_weight': None
}
for par in pars:
if par not in classpar: classpar.update({par: pars.get(par)})
clasif = RFC(random_state=random_state, n_jobs=n_jobs, **classpar)
elif how == 'svc':
pars = {
'C': 1.0,
'kernel': 'rbf',
'degree': 3,
'gamma': 'auto',
'coef0': 0.0,
'shrinking': True,
'probability': False,
'tol': 0.001,
'cache_size': 200,
'class_weight': None,
'verbose': False,
'max_iter': -1,
'decision_function_shape': None
}
for par in pars:
if par not in classpar: classpar.update({par: pars.get(par)})
clasif = SVC(random_state=random_state, **classpar)
else:
print 'Classifier not yet supported'
return
if cv:
crosv = ShuffleSplit(n_splits=5,
test_size=0.3,
random_state=random_state)
# y_pred = cross_val_predict(clasif, X_train, y_train, cv=5, n_jobs=n_jobs,
# verbose=1)
# f1 = f1_score(y_test, y_pred, average='weighted')
# ck = cohen_kappa_score(y_test, y_pred)
# rec = recall_score(y_test, y_pred, average='weighted')
# if verbose:
# print '\nF1={:.2f}, Recall={:.2f}, Cohen Kappa={:.2f}'.format(f1, rec, ck)
# return f1, rec, ck
f1_cv_scores = cross_val_score(clasif,
X_train,
y_train,
cv=crosv,
scoring='f1_weighted',
verbose=1,
n_jobs=n_jobs)
mean_cv_f1 = np.mean(f1_cv_scores)
if verbose:
print f1_cv_scores
print 'Mean F1 score={:.3f}'.format(mean_cv_f1)
return mean_cv_f1, f1_cv_scores
else:
if verbose:
print clasif.fit(X_train, y_train.astype(int))
else:
clasif.fit(X_train, y_train.astype(int))
y_pred = clasif.predict(X_test)
if verbose:
print '\n', imbmet.classification_report_imbalanced(y_test, y_pred)
if verbose and hasattr(clasif, 'feature_importances_'):
print 'Feature importances:'
print clasif.feature_importances_
ck = cohen_kappa_score(y_test, y_pred)
rec = recall_score(y_test, y_pred, average='weighted')
f1 = f1_score(y_test, y_pred, average='weighted')
if verbose:
print '\nF1={:.2f}, Recall={:.2f}, Cohen Kappa={:.2f}'.format(
f1, rec, ck)
if full_output:
return clasif, f1, rec, ck
else:
return f1, rec, ck
# In[13]: #train_x = StandardScaler().fit_transform(train_x) #test_x = StandardScaler().fit_transform(test_x) # train_x=preprocessing.scale(train_x) # test_x=preprocessing.scale(test_x) # In[14]: import mord as m # In[15]: model = m.OrdinalRidge(alpha=0.1, max_iter=10000) # In[16]: train_x[:1] # In[17]: model.fit(train_x, train_y) # In[18]: model.score(test_x, test_y) # In[19]:
args = parser.parse_args()
df = pd.read_csv(args.training_csv, header=None)
# Getting features and labels
y = df[0].values
features_df = df.loc[:, 1:25].fillna(0)
X = features_df.loc[:, 1:25].values
# Generating train/test
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.3)
# Training the model
reg = mord.OrdinalRidge(alpha=1.0,
fit_intercept=True,
normalize=False,
copy_X=True,
max_iter=None,
tol=0.001,
solver='auto')
reg.fit(X_train, y_train)
# Showing some results and saving model
y_pred = reg.predict(X_test)
b = 1
print(np.bincount(y_pred.astype(np.int32) + b))
filename = args.out_model
pickle.dump(reg, open(filename, 'wb'))
X = X.drop(X.columns[1], axis=1)
X = X.drop(X.columns[1], axis=1)
clf1 = linear_model.LogisticRegression(
solver='lbfgs',
multi_class='multinomial')
clf1.fit(X, y)
print('Mean Absolute Error of LogisticRegression: %s' %
metrics.mean_absolute_error(clf1.predict(X), y))
clf2 = mord.LogisticAT(alpha=1.)
clf2.fit(X, y)
print('Mean Absolute Error of LogisticAT %s' %
metrics.mean_absolute_error(clf2.predict(X), y))
clf3 = mord.LogisticIT(alpha=1.)
clf3.fit(X, y)
print('Mean Absolute Error of LogisticIT %s' %
metrics.mean_absolute_error(clf3.predict(X), y))
clf4 = mord.OrdinalRidge(alpha=1.)
clf4.fit(X, y)
print('Mean Absolute Error of LogisticSE %s' %
metrics.mean_absolute_error(clf4.predict(X), y))
y_pred = clf4.predict(X)
plt.scatter(X, y, color='black')
plt.plot(X, y_pred, color='blue', linewidth=3)
plt.show()
