大数据比赛（3）- 模型选择II

常用模型概述

神经网络与深度学习初步

基础

老规矩，先推文章：
手把手入门神经网络系列(1)_从初等数学的角度初探神经网络
http://blog.csdn.net/han_xiaoyang/article/details/50100367
深度学习概述：从感知机到深度网络
http://www.cnblogs.com/xiaowanyer/p/3701944.html
Deep Learning（深度学习）学习笔记整理系列
http://blog.csdn.net/zouxy09

相信深度学习大热的今天，大家对神经网络多少都有一些了解。在传统的“教科书”之外，寒小阳同学所写的博客从另一个容易理解的角度阐释了nn的结构，叫做“逻辑回归的逻辑回归的逻辑回归的逻辑……”

class NeuralNetworkForTianchi(Regressor):
    def __init__(self, nodes, active_function='sigmoid', lr=.1, decay=1., off=1e-5,
                 batch=128, regular='l2', penalty=.5):
        self.nodes, self.active_function = nodes, active_function
        Regressor.__init__(self, None, lr, decay, off, batch, regular, penalty)

    def _define_train_inputs(self):
        return [T.matrix('x'), T.matrix('a'), T.matrix('b'), T.matrix('y')]

    def _predict_outputs(self, *inputs):
        y = inputs[0]
        for i in range(len(self.nodes)-2):
            y = self.layers['active_layer'].outs(self.layers['liner_layer'+str(i)].outs(y))
        i = len(self.nodes)-2
        y = self.layers['ReLU_layer'].outs(self.layers['liner_layer'+str(i)].outs(y))
        return y

    def _define_layers(self):
        layers = {'active_layer': ActiveLayer(self.active_function), 'ReLU_layer': ActiveLayer('ReLU')}
        for i in range(len(self.nodes)-1):
            liner_layer = LinerLayer(self.nodes[i], self.nodes[i+1])
            layers['liner_layer'+str(i)] = liner_layer
        return layers

    def _define_predict_inputs(self):
        return [T.matrix('x')]

    def _loss_function(self):
        """
        loss function: mean_log_error, mean_squared_error, mean_absolute_error
        :return: loss function
        """
        y_regular = self._predict_outputs(self.train_inputs[0])
        a, b, y = self.train_inputs[1:]
        penalty = self._regular()
        diff = y - y_regular
        return T.mean((diff > 0) * diff * a - (diff < 0) * diff * b + penalty)