def __init__(self,
training_data,
weight_decay_p,
epsilon,
label_count,
numerical=False):
self.training_data = training_data
self.n, self.m = training_data[0][0].shape
self.wd = weight_decay_p # NOTE: WEIGHT DECAY PARAMETER --> what should this be set at??
self.label_count = label_count
self.W = np.random.normal(0.0, epsilon**2,
(self.n * self.m + 1) * self.label_count)
print "training model..."
if numerical:
x, f, d = l_bfgs(self.get_J, self.W, None, (), True)
else:
x, f, d = l_bfgs(self.get_J, self.W, self.get_J_gradient)
print "minimum found, J = " + str(f)
#self.W = np.array(x[:self.n*self.m]).reshape((self.n,self.m))
# FOR DEBUG
self.x = x
self.f = f
self.d = d
def __init__(self, training_data, weight_decay_p, epsilon, label_count, numerical=False):
self.training_data = training_data
self.n, self.m = training_data[0][0].shape
self.wd = weight_decay_p # NOTE: WEIGHT DECAY PARAMETER --> what should this be set at??
self.label_count = label_count
self.W = np.random.normal(0.0,epsilon**2,(self.n*self.m+1)*self.label_count)
print "training model..."
if numerical:
x, f, d = l_bfgs(self.get_J, self.W, None, (), True)
else:
x, f, d = l_bfgs(self.get_J, self.W, self.get_J_gradient)
print "minimum found, J = " + str(f)
#self.W = np.array(x[:self.n*self.m]).reshape((self.n,self.m))
# FOR DEBUG
self.x = x
self.f = f
self.d = d
def train_params(training_trees, wd, epsilon, picloud=True, num_cores=0, bf=2, numerical=False):
training = initialize_trees(training_trees, bf)
# get the param dimensions
t = training[0]
while len(t) > 0:
t = t[0]
encoding_size = len(t.node)
decoded_size = bf * encoding_size
# train the model with backprop + L-BFGS-B
x0 = np.random.normal(0.0, epsilon ** 2, 2 * encoding_size * decoded_size + encoding_size + decoded_size)
print "training model..."
if picloud:
x, f, d = l_bfgs(J, x0, Jgrad_picloud, [encoding_size, decoded_size, training, wd, num_cores])
elif numerical:
x, f, d = l_bfgs(J, x0, None, [encoding_size, decoded_size, training, wd], True)
else:
x, f, d = l_bfgs(J, x0, Jgrad, [encoding_size, decoded_size, training, wd])
print "minimum found, J = " + str(f)
# save_params(x, '_final')
return x
