def train(self,
X,
y,
learning_rate=0.1,
cost_threshold=1e-6,
diff_threshold=1e-16,
max_iter=10000,
min_iter=0,
lambda_=3):
nclass = NeuralNetClassification(self)
#self.thetas = npe.reshape_for_neuralnet(fmin_cg(
# f=nclass._flat_cost,
# x0=npe.flatten(self.thetas),
# fprime=nclass._flat_gradients,
# args=(X, y),
# epsilon=learning_rate,
# # gtol=gtol,
# disp=False,
# maxiter=50), self)
self.thetas = npe.reshape_for_neuralnet(
gradient_descent(start_thetas=npe.flatten(self.thetas),
cost_func=nclass._flat_cost,
gradient_func=nclass._flat_gradients,
args=(X, y, lambda_),
learning_rate=learning_rate,
min_iter=min_iter,
max_iter=max_iter,
cost_threshold=cost_threshold,
diff_threshold=diff_threshold), self)
def _flat_forwardpropagate(self, flat_thetas, X, *args):
thetas = npe.reshape_for_neuralnet(flat_thetas,
self.neuralnet)
r,_,_ = self._forwardpropagate(thetas, self.neuralnet.num_layers(), X)
return npe.flatten(r)
def _flat_regularized_gradients(self, flat_thetas, X, y, l, *args):
thetas = npe.reshape_for_neuralnet(flat_thetas, self.neuralnet)
gradients = self._regularized_gradients(thetas, X, y, l)
f = npe.flatten(gradients)
return f
def _flat_regularized_cost(self, flat_thetas, X, y, l, *args):
thetas = npe.reshape_for_neuralnet(flat_thetas, self.neuralnet)
return self._regularized_cost(thetas, X, y, l)
def _flat_cost(self, flat_thetas, X, y, *args):
thetas = npe.reshape_for_neuralnet(flat_thetas, self.neuralnet)
return self._cost(thetas, X, y)