def train(self,x,y,params=None,randomize=True,
split_sizes=None, batch_size=None):
if params is None:
self.params = NeuralNet.default_params
else:
for param in NeuralNet.default_params:
if(param not in params):
params[param] = NeuralNet.default_params[param]
self.params = params
#prepare the data
(x,y,
val_in,val_out,
test_in,test_out,
do_validation,do_test) =\
self.__init_data(x, y, split_sizes, randomize)
self.y_ranges, self.y_mins = dp.get_ranges(y)
self.x_ranges, self.x_mins = dp.get_ranges(x)
x_c = dp.map_to_ranges(x,self.x_ranges,self.x_mins)
y_c = dp.map_to_ranges(y,self.y_ranges,self.y_mins)
#input size is number of features
num_features = x.shape[1]
num_output = y.shape[1]
#default batch size is the whole train dataset
#self.ranges_y = d
n_samples = len(x)
if(batch_size is None):
batch_size = n_samples
self.__build_architecture(num_features,num_output)
self.sigma = self.params["sigma"]
self.initial_lambda = self.lambda_ = self.params["lambda"]
self.lambda_bar = 0.0
self.max_epochs = self.params["max_epochs"]
self.goal = self.params["goal"]
output = self.__predict(x_c, self.layers)
self.error = self.get_mse(output, y_c) / len(y_c)
self.reciprocals = -self.compute_gradients(x_c,y_c,self.layers)
self.search_dir = np.copy(self.reciprocals)
self.pSqNorm = 0.0
self.success = True
self.delta = 0.0
self.k = self.num_params
i_it = 0
while(i_it < self.max_epochs and self.error > self.goal):
self.train_iteration(x_c,y_c)
print self.error
i_it +=1
def predict(self,x):
x = dp.map_to_ranges(x,self.x_ranges,self.x_mins)
return dp.unmap_y(self.__predict(x,self.layers),self.y_ranges)