def __init__(self, net, training_data, validation_data): # self.net = net validation_inputs_data, validation_results_data = \ Net_Util.seperate(validation_data) training_inputs_data, training_results_data = \ Net_Util.seperate(training_data) self.evaluation_cost = np.sum(Net_Util.Cost.crossentropycost( net.feedforword(validation_inputs_data), validation_results_data)) self.evaluation_accurate = Net_Util.accurate( net.feedforword(training_inputs_data), training_results_data) self.training_cost = np.sum(Net_Util.Cost.crossentropycost( net.feedforword(training_inputs_data), training_results_data)) self.training_accurate = Net_Util.accurate( net.feedforword(training_inputs_data), training_results_data)
def __init__(self, size, pre_layer=None, next_layer=None): self.size = size self.pre_layer = pre_layer self.next_layer = next_layer self.inputs = None self.outputs = None if pre_layer is not None: # initialize the weights of layer self.weights, self.biases = Net_Util.weights_initialize(self.size, self.pre_layer.size) else: self.weights = None self.biases = None self.delta_weights = None self.delta_biases = None self.delta = None self.regular = None
def net_training(self, training_data, learning_rate, lamda): """Training the net work with stochastic gradient descent eta: learning rate lamda: regulation rate """ # seperate input and output data input_data, result_data = Net_Util.seperate(training_data) num_sample = len(input_data) # feedforword self.outputs = self.feedforword(input_data) # backpropagation self.input_layer.backpropagation(result_data) # update weights and biases for layer in self.hiden_layers: layer.weights = layer.weights - learning_rate*(np.sum(layer.delta_weights, axis=1)/num_sample + layer.regularization(lamda)) # regularization layer.biases = layer.biases - learning_rate*np.sum(layer.delta_biases, axis=1)/num_sample
