StackAutoEncoder.py 文件源码

def minibatch_update(self,x,y,lr,regularization):
        n_sample = x.shape[0]
        info = x
        hidden_cache = []
        for i in xrange(self.n_hidden + 1):
            if i == self.n_hidden:
                probs = softmax(info.dot(self.W[i]) + self.b[i])
            else:
                info = sigmoid(info.dot(self.W[i]) + self.b[i])
                hidden_cache.append(info)
        loss = neg_log_likelihood(probs,y)
        probs[np.arange(n_sample),y] -= 1.0
        errors = probs
        for i in range(self.n_hidden,-1,-1):
            if i >= 1:
                hidden_out = hidden_cache[i - 1]
                grad_hidden_out = errors.dot(self.W[i].T)
                self.W[i] -= (lr * (hidden_out.T).dot(errors) + regularization * self.W[i])
                self.b[i] -= lr * np.sum(errors,axis = 0)
                errors = hidden_out * (1 - hidden_out) * grad_hidden_out
            else:
                hidden_out = x
                self.W[i] -= (lr * (hidden_out.T).dot(errors) + regularization * self.W[i])
                self.b[i] -= lr * np.sum(errors,axis = 0)

        return loss