logistic.py 文件源码

def _buildModel(self):
        self.updates_ack= True
        X             = T.matrix('X',   dtype=config.floatX)
        Y             = T.matrix('Y',   dtype=config.floatX)
        X.tag.test_value, Y.tag.test_value    = self._fakeData()
        #output_params_t= T.nnet.sigmoid(self._LinearNL(self.tWeights['W_lr'], self.tWeights['b_lr'], X, onlyLinear=True))
        output_params_t= T.nnet.sigmoid(self._BNlayer(self.tWeights['W_lr'], self.tWeights['b_lr'], X, validation=False, onlyLinear=True))
        nll_t          = T.nnet.binary_crossentropy(output_params_t, Y).sum()

        #output_params_e = T.nnet.sigmoid(self._LinearNL(self.tWeights['W_lr'], self.tWeights['b_lr'], X, onlyLinear=True))
        output_params_e= T.nnet.sigmoid(self._BNlayer(self.tWeights['W_lr'], self.tWeights['b_lr'], X, validation=True, onlyLinear=True))
        nll_e          = T.nnet.binary_crossentropy(output_params_e, Y).sum()

        if not self.params['validate_only']:
            model_params  = self._getModelParams()
            print len(self.updates),' extraneous updates'
            optimizer_up, norm_list  = self._setupOptimizer(nll_t, 
                                                        model_params, 
                                                        lr=self.params['lr'],
                                                        divide_grad = T.cast(X.shape[0],config.floatX))
            optimizer_up+=self.updates
            self.train      = theano.function([X,Y], [nll_t,self.tWeights['_lr_BN_running_mean'], self.tWeights['_lr_BN_running_var']], updates = optimizer_up)
        self.evaluate   = theano.function([X,Y],nll_e)