model2.py 文件源码

def __build_positive_phase(self):

        n_iterations = T.iscalar('n_iterations')
        alphas = [T.fscalar("alpha_W"+str(r+1)) for r in range(len(self.weights))]

        def backprop(*layers):
            layers_new = [layers[-1]]
            for k in range(len(self.layers)-2,0,-1):
                layers_new += [pi(T.dot(layers[k-1], self.weights[k-1]) + T.dot(layers_new[-1], self.weights[k].T) + self.biases[k])]
            layers_new += [layers[0]]
            layers_new.reverse()
            return layers_new

        ( layers, updates ) = theano.scan(
            backprop,
            outputs_info=self.layers[:-1]+[self.y_data_one_hot],
            n_steps=n_iterations
        )
        layers_new = [layer[-1] for layer in layers]

        Delta_layers = [(layer_new-layer) for layer_new,layer in zip(layers_new[1:],self.layers[1:])]
        biases_new  = [b + alpha * T.mean(Delta, axis=0) for b,alpha,Delta in zip(self.biases[1:],alphas,Delta_layers)]
        weights_new = [W + alpha * T.dot(layer.T, Delta) / T.cast(self.layers[0].shape[0], dtype=theano.config.floatX) for W, alpha, layer, Delta in zip(self.weights, alphas, self.layers[:-1], Delta_layers)]

        for bias, bias_new in zip(self.biases[1:],biases_new):
            updates[bias]=bias_new
        for weight, weight_new in zip(self.weights,weights_new):
            updates[weight]=weight_new

        positive_phase = theano.function(
            inputs=[n_iterations]+alphas,
            outputs=[],
            updates=updates
        )

        return positive_phase