def load_feature_matrix(args):
if args.feature_format % 3 == 0:
X1 = HierarchicalMLModel.load_feature_matrix(args.input_inst_feat1)
X2 = HierarchicalMLModel.load_feature_matrix(args.input_inst_feat2)
X = smat.hstack([sk_normalize(X1, axis=1),
sk_normalize(X2, axis=1)]).tocsr()
elif args.feature_format % 3 == 1 and args.input_inst_feat1:
X = HierarchicalMLModel.load_feature_matrix(args.input_inst_feat1)
elif args.feature_format % 3 == 2 and args.input_inst_feat2:
X = HierarchicalMLModel.load_feature_matrix(args.input_inst_feat2)
else:
raise NotImplementedError(
f"args.feature_format = {args.feature_format} is not supported.")
if args.feature_format // 3 == 0:
X = sk_normalize(X, axis=1, copy=False)
return X
def train(cls,
X,
Y,
C,
mode='full-model',
shallow=False,
solver_type=solver_dict['L2R_L2LOSS_SVC_DUAL'],
Cp=1.0,
Cn=1.0,
threshold=0.1,
max_iter=100,
threads=-1,
bias=-1.0,
Z_pred=None):
if mode in ['full-model', 'matcher']:
if mode == 'full-model':
prob = MLProblem(X, Y, C, bias=bias, Z_pred=Z_pred)
elif mode == 'matcher':
assert C is not None
Y = Y.dot(C)
prob = MLProblem(X, Y, bias=bias)
hierarchical = True
min_labels = 2
if shallow:
if prob.C is None:
min_labels = prob.Y.shape[1]
else:
min_labels = prob.C.shape[1]
elif mode == 'ranker':
assert C is not None
prob = MLProblem(X, Y, C, bias=bias)
hierarchical = False
min_labels = 2
model = HierarchicalMLModel.train(prob,
hierarchical=hierarchical,
min_labels=min_labels,
solver_type=solver_type,
Cp=Cp,
Cn=Cn,
threshold=threshold,
threads=threads,
max_iter=max_iter)
return cls(model)
def load(cls, model_folder):
return cls(HierarchicalMLModel.load(model_folder))
def train(
cls,
X,
Y,
C,
mode="full-model",
shallow=False,
solver_type=solver_dict["L2R_L2LOSS_SVC_DUAL"],
Cp=1.0,
Cn=1.0,
threshold=0.1,
max_iter=100,
threads=-1,
bias=-1.0,
Z_pred=None,
negative_sampling_scheme=None,
):
if mode in ["full-model", "matcher"]:
if mode == "full-model":
prob = MLProblem(
X,
Y,
C,
Z_pred=Z_pred,
negative_sampling_scheme=negative_sampling_scheme,
)
elif mode == "matcher":
assert C is not None
Y = Y.dot(C)
prob = MLProblem(X, Y, C=None)
hierarchical = True
min_labels = 2
if shallow:
if prob.C is None:
min_labels = prob.Y.shape[1]
else:
min_labels = prob.C.shape[1]
elif mode == "ranker":
assert C is not None
prob = MLProblem(
X,
Y,
C,
Z_pred=Z_pred,
negative_sampling_scheme=negative_sampling_scheme,
)
hierarchical = False
min_labels = 2
model = HierarchicalMLModel.train(
prob,
hierarchical=hierarchical,
min_labels=min_labels,
solver_type=solver_type,
Cp=Cp,
Cn=Cn,
threshold=threshold,
threads=threads,
bias=bias,
max_iter=max_iter,
)
return cls(model)