def fit_ftdata(self, x_ft, y_ft, x_ab, y_ab):
""" Fit the model with featurized data as input
:param x_ft: x_featurized
:param y_ft: y_featurized
:param x_ab: x_inverse_featurized
:param y_ab: y_inverse_featurized
"""
self.clf0 = CLF(**self.params).fit(x_ft,
y_ft != 0) # causal or confounded?
self.clf1 = CLF(**self.params).fit(x_ab,
y_ab == 1) # causal or anticausal?
def fit(
self,
x,
y,
is_q=False
): # x,y = original data and original labels => x=(npairs,2),y=(npairs,1)
'''This function takes the entire dataframe and labels as input => x and y are original "data" and original "labels" respectively.
The output is a trained CLF.
The idea of "train" is to 1) separately go through each row, 2) featurize each row two times (for A->B and B->A), and 3) vectically stack them.
The idea of "labels" is to duplicate the given labels (i.e., the y variable) by cosidering both A->B and B->A directions.
Example:
obj.fit(data, labels)
'''
train = np.vstack((
np.array([
self.featurize_row(row.iloc[0], row.iloc[1], is_q=is_q)
for idx, row in x.iterrows()
]), #(npairs,3m) #Does this really work for quantum?!
np.array([
self.featurize_row(row.iloc[1], row.iloc[0], is_q=is_q)
for idx, row in x.iterrows()
]))) #(npairs,3m)
#train=(2*npairs,3m)
labels = np.vstack((y, -y)).ravel() #(2*npairs,)
verbose = 1 if self.verbose else 0
self.clf = CLF(verbose=verbose,
min_samples_leaf=self.L,
n_estimators=self.E,
max_depth=self.max_depth,
n_jobs=self.njobs).fit(train, labels)
def fit(self, x, y):
# CAUTION this x and y should not be dataframe, but preprocessed above
print('training CLF ..')
verbose = 1 if self.verbose else 0
# FIXME and this is very im-balanced
self.clf = CLF(verbose=verbose,
min_samples_leaf=self.L,
n_estimators=self.E,
max_depth=self.max_depth,
n_jobs=self.njobs).fit(x, y)
def fit(self, x, y):
"""Train the model.
args:
x: pandas.Dataframe of the data
y: targets
"""
train = np.vstack((np.array([self.featurize_row(row.iloc[0],
row.iloc[1]) for idx, row in x.iterrows()]),
np.array([self.featurize_row(row.iloc[1],
row.iloc[0]) for idx, row in x.iterrows()])))
labels = np.vstack((y, -y)).ravel()
verbose = 1 if self.verbose else 0
self.clf = CLF(verbose=verbose,
min_samples_leaf=self.L,
n_estimators=self.E,
max_depth=self.max_depth,
n_jobs=self.n_jobs).fit(train, labels)
def fit(self, x, y):
"""Train the model.
Args:
x_tr (pd.DataFrame): CEPC format dataframe containing the pairs
y_tr (pd.DataFrame or np.ndarray): labels associated to the pairs
"""
train = np.vstack((np.array([self.featurize_row(row.iloc[0],
row.iloc[1]) for idx, row in x.iterrows()]),
np.array([self.featurize_row(row.iloc[1],
row.iloc[0]) for idx, row in x.iterrows()])))
labels = np.vstack((y, -y)).ravel()
verbose = 1 if self.verbose else 0
self.clf = CLF(verbose=verbose,
min_samples_leaf=self.L,
n_estimators=self.E,
max_depth=self.max_depth,
n_jobs=self.njobs).fit(train, labels)
y_te = np.hstack((y_te, -y_te))
d_tr = np.hstack((d_tr, d_tr))
d_te = np.hstack((d_te, d_te))
x_ab = x_tr[(y_tr == 1) | (y_tr == -1)]
y_ab = y_tr[(y_tr == 1) | (y_tr == -1)]
params = {
'random_state': 0,
'n_estimators': E,
'max_features': None,
'max_depth': 50,
'min_samples_leaf': 10,
'verbose': 10
}
params = {
'random_state': 0,
'n_estimators': E,
'min_samples_leaf': L,
'n_jobs': 16
}
clf0 = CLF(**params).fit(x_tr, y_tr != 0) # causal or confounded?
clf1 = CLF(**params).fit(x_ab, y_ab == 1) # causal or anticausal?
clfd = CLF(**params).fit(x_tr, d_tr) # dependent or independent?
p_te = clf0.predict_proba(x_te)[:,
1] * (2 * clf1.predict_proba(x_te)[:, 1] - 1)
print([score(y_te, p_te), clf0.score(x_te, y_te != 0), clfd.score(x_te, d_te)])