def svm_train(X, y, b, alpha, n_samples, n_features, learner, loop, eta,
max_iter=100, step_probability=0.5):
"""
Minimizes an expression of the form
Loss(X, y, b) + 0.5 * alpha * (||w|| ** 2)
where Loss is an Hinge loss defined on pairs of images
Parameters
----------
X : input data
y : target labels
b : blocks (aka query_id)
alpha: float
loop : {'rank', 'combined-ranking', 'roc', 'stochastic', 'balanced-stochastic'}
Returns
-------
coef
"""
if isinstance(X, bstring):
if n_features is None:
n_features = 2**17 # the default in sofia-ml TODO: parse file to see
w = _sofia_ml.train(X, n_features, alpha, max_iter, False,
learner.value, loop.value, eta.value, step_probability)
elif isinstance(X, np.ndarray):
if n_features is None:
n_features = X.shape[1]
if n_samples is None:
n_samples = X.shape[0]
w = _sofia_ml.train_fast(np.float64(X), np.float64(y), n_samples, n_features, alpha, max_iter, False,
learner.value, loop.value, eta.value, step_probability)
else:
if n_features is None:
n_features = X.shape[1]
with tempfile.NamedTemporaryFile() as f:
datasets.dump_svmlight_file(X, y, f.name, query_id=b)
w = _sofia_ml.train(f.name, n_features, alpha, max_iter, False,
learner.value, loop.value, eta.value, step_probability)
return w
def sgd_train(X,
y,
b,
alpha,
n_features=None,
model='rank',
max_iter=100,
step_probability=0.5):
"""
Minimizes an expression of the form
Loss(X, y, b) + 0.5 * alpha * (||w|| ** 2)
where Loss is an Hinge loss defined on pairs of images
Parameters
----------
X : input data
y : target labels
b : blocks (aka query_id)
alpha: float
model : {'rank', 'combined-ranking', 'roc'}
Returns
-------
coef
None
"""
if isinstance(X, bstring):
if n_features is None:
n_features = 2**17 # the default in sofia-ml TODO: parse file to see
w = _sofia_ml.train(X, n_features, alpha, max_iter, False, model,
step_probability)
else:
with tempfile.NamedTemporaryFile() as f:
datasets.dump_svmlight_file(X, y, f.name, query_id=b)
w = _sofia_ml.train(f.name, X.shape[1], alpha, max_iter, False,
model, step_probability)
return w, None
def train(X, y, alpha, query_id, max_iter=100, model='rank', step_probability=0.5):
"""
model : {'rank', 'combined-ranking'}
"""
if query_id is None:
query_id = np.ones(y.size)
with tempfile.NamedTemporaryFile() as f:
datasets.dump_svmlight_file(X, y, f.name, query_id=query_id)
w = _sofia_ml.train(f.name, X.shape[1], alpha, max_iter, False, model,
step_probability)
return w
def sgd_train(X, y, b, alpha, n_features=None, model='rank', max_iter=100, step_probability=0.5):
"""
Minimizes an expression of the form
Loss(X, y, b) + 0.5 * alpha * (||w|| ** 2)
where Loss is an Hinge loss defined on pairs of images
Parameters
----------
X : input data
y : target labels
b : blocks (aka query_id)
alpha: float
model : {'rank', 'combined-ranking', 'roc'}
Returns
-------
coef
None
"""
if isinstance(X, bstring):
if n_features is None:
n_features = 2 ** 17 # the default in sofia-ml TODO: parse file to see
w = _sofia_ml.train(X, n_features, alpha, max_iter, False, model,
step_probability)
else:
with tempfile.NamedTemporaryFile() as f:
datasets.dump_svmlight_file(X, y, f.name, query_id=b)
w = _sofia_ml.train(f.name, X.shape[1], alpha, max_iter, False, model,
step_probability)
return w, None
def train(X,
y,
alpha,
query_id,
max_iter=100,
model='rank',
step_probability=0.5):
"""
model : {'rank', 'combined-ranking'}
"""
if query_id is None:
query_id = np.ones(y.size)
with tempfile.NamedTemporaryFile() as f:
datasets.dump_svmlight_file(X, y, f.name, query_id=query_id)
w = _sofia_ml.train(f.name, X.shape[1], alpha, max_iter, False, model,
step_probability)
return w