fmin_cg基本神经网络的精度损失

时间:2019-07-11 16:17:27

标签: neural-network cross-entropy

我正在用python实现一个简单的神经网络,将1个隐藏层(25个节点),400像素图像作为输入,并将属于10类(每位数1个)中的每一个的概率作为输出。经典。

在我看来,我理解算法并且我的实现是正确的,因为当使用我自己的(非常基本的)函数来更新权重时,它似乎可以正常工作(准确度逐渐提高到94%+)。我的问题是我无法让fmin_cg正常工作。

我将成本和梯度函数的输出与在github(相同的练习)上发现的函数的输出进行了比较,这是相同的,似乎有所不同的是,我使用了向量化的实现(出于练习的目的),而我在github上仅找到迭代版本。 而且我分开了权重和偏见(练习是吴安德(Andrew ng)的mooc,他对两者都使用一个矩阵)

我将两个权重矩阵和两个偏差向量合并在一个矩阵中,然后根据fmin_cg的要求将其展平。

我尝试跟踪权重和渐变的值,看起来还不错。在将fmin_cg用于逻辑回归之前,我遇到了一些麻烦,因为我的权重最终将变为+ inf。

我想仅提供成本和梯度函数是错误的,这是整个NeuralNetwork类,它只是期望有效的示例集和相应的标签,以及用于正则化的lambda参数。然后应调用my_learn和learn_param中的一个来开始优化。其他功能不值得一看。

import numpy as np
import matplotlib.pyplot as plt
import scipy.io as sp
import pandas as pd
import scipy.optimize as opt


class NeuralNetwork:

    def __init__(self, train_ex, labels, reg, weights: list = None, num_iters = 1500):
        self.input_layer_size = 400
        self.hidden_layer_size = 25
        self.num_labels = 10  # output layer size
        self.weights_bias = self.init_weights_biases(weights)  # weights = [np.array(),... ,np.array()]

        self.train_ex = train_ex
        self.labels = labels

        self.reg = reg  # regularization coefficient for cost and gradient
        self.num_iters = 50

    def is_input_valid(self, w):

        w1, w2, b1, b2 = self.unpack_weights_bias(w, (self.hidden_layer_size, self.input_layer_size),
                                            (self.num_labels, self.hidden_layer_size))
        try:  # check user provided weight_bias list
            error_flag = False
            if w1.shape != (self.hidden_layer_size, self.input_layer_size):
                error_flag = True
            if w2.shape != (self.num_labels, self.hidden_layer_size):
                error_flag = True
            if b1.shape != (self.hidden_layer_size,):
                error_flag = True
            if b2.shape != (self.num_labels,):
                error_flag = True

            if error_flag:
                print("Incorrect params shape")
                return False
        except Exception:
            print('Incorrect parameters')
            return False
        return True

    def init_weights_biases(self, weights = None) -> np.array:
        """randomly initialize weights if no (correct) parameters were provided"""
        if weights is None:  # if no weights_bias list were provided, initialize it randomly
            init_1 = np.sqrt(6) / np.sqrt(self.input_layer_size + self.hidden_layer_size)  # defines min/max
            init_2 = np.sqrt(6) / np.sqrt(self.hidden_layer_size + self.num_labels)  # for each parameter's values

            w1 = np.random.uniform(-init_1, init_1, (self.hidden_layer_size, self.input_layer_size))
            w2 = np.random.uniform(-init_2, init_2, (self.num_labels, self.hidden_layer_size))

            b1 = np.random.uniform(-init_2, init_2, (self.hidden_layer_size,))
            b2 = np.random.uniform(-init_2, init_2, (self.num_labels,))

            return self.pack_weights_bias(w1, w2, b1, b2)

        if not self.is_input_valid(weights):
            print('Initializing weights randomly..')
            return self.init_weights_biases()

        return weights  # if no errors were raised, returns the same list

    @staticmethod
    def pack_weights_bias(w1, w2, b1, b2):
        param_2 = np.hstack((w2, np.reshape(b2, (b2.shape[0], 1))))

        param_1 = np.hstack((w1, np.reshape(b1, (b1.shape[0], 1))))
        param_1 = np.vstack((param_1, np.ones((1, param_1.shape[1]))))

        return np.hstack((param_1, param_2.T)).flatten()

    @staticmethod
    def unpack_weights_bias(packed_params, w_1_size: tuple, w_2_size: tuple):
        packed_params = np.reshape(packed_params, (w_1_size[0] + 1, w_1_size[1] + w_2_size[0] + 1))
        param_1 = packed_params[:-1, :(w_1_size[1] + 1)]
        param_2 = packed_params[:, (w_1_size[1] + 1):].T

        w1 = param_1[:, :w_1_size[1]]
        b1 = param_1[:, -1]

        w2 = param_2[:, :w_2_size[1]]
        b2 = param_2[:, -1]

        return w1, w2, b1, b2

    @staticmethod
    def forward_prop(weights_bias, train_ex, w_size):
        w1, w2, b1, b2 = NeuralNetwork.unpack_weights_bias(weights_bias, w_size[0], w_size[1])

        #   b1 is a row vector, added to every row of a_1
        z_1 = np.dot(train_ex, w1.T) + b1
        a_1 = NeuralNetwork.sigmoid(z_1)  # 5000 * 25

        z_2 = np.dot(a_1, w2.T) + b2
        return NeuralNetwork.sigmoid(z_2), a_1, z_2, z_1  # 5000 * 10

    @staticmethod
    def sigmoid(x, derivative=False):
        if derivative:
            # return np.exp(-x) / ((1 + np.exp(-x)) ** 2)
            return NeuralNetwork.sigmoid(x) * (1 - NeuralNetwork.sigmoid(x))
        return 1 / (1 + np.exp(-x))

    @staticmethod
    def compute_cost(weights_bias, train_ex, labels, reg, w_size):  # cost function of neural network

        num_train_ex = train_ex.shape[0]
        w1, w2, b1, b2 = NeuralNetwork.unpack_weights_bias(weights_bias, w_size[0], w_size[1])

        y = pd.get_dummies(labels.flatten())  # 5000 * 1
        h, a_1, z_2, z_1 = NeuralNetwork.forward_prop(weights_bias, train_ex, w_size)  # 5000 * 10

        temp1 = np.multiply(y, np.log(h))  # 5000 * 10
        temp2 = np.multiply(1 - y, np.log(1 - h))  # 5000 * 10
        cost = np.sum(temp1 + temp2)  # 1 * 1

        reg_cost_1 = np.sum(w1 ** 2)
        reg_cost_2 = np.sum(w2 ** 2)
        cost = np.sum(cost / (-num_train_ex)) + reg / (2.0 * num_train_ex) * (reg_cost_1 + reg_cost_2)
        return cost

    @staticmethod
    def compute_gradient(weights_bias, train_ex, labels, reg, w_size):

        num_train_ex = train_ex.shape[0]

        w1, w2, b1, b2 = NeuralNetwork.unpack_weights_bias(weights_bias, w_size[0], w_size[1])

        labels = pd.get_dummies(labels.flatten())
        a_2, a_1, z_2, z_1 = NeuralNetwork.forward_prop(weights_bias, train_ex, w_size)

        d_2 = a_2 - np.array(labels)  # 5000 * 10
        d_1 = np.dot(d_2, w2)    # 5000 * 25
        d_1 = np.multiply(d_1, NeuralNetwork.sigmoid(z_1, derivative=True))

        delta_1 = d_1.T @ train_ex
        delta_2 = d_2.T @ a_1

        delta_1 /= num_train_ex
        delta_2 /= num_train_ex

        delta_1 += (w1 * reg) / num_train_ex
        delta_2 += (w2 * reg) / num_train_ex

        update_b2 = np.sum(d_2 / num_train_ex, axis=0)
        update_b1 = np.sum(d_1 / num_train_ex, axis=0)

        return NeuralNetwork.pack_weights_bias(delta_1, delta_2, update_b1, update_b2)

    def my_learn(self):
        w_size = [(self.hidden_layer_size, self.input_layer_size), (self.num_labels, self.hidden_layer_size)]
        for i in range(5000):
            # w1, w2, b1, b2 = NeuralNetwork.unpack_weights_bias(self.weights_bias, w_size[0], w_size[1])
            self.weights_bias -= 0.31 * NeuralNetwork.compute_gradient(self.weights_bias, self.train_ex, self.labels, self.reg, w_size)
            # print(NeuralNetwork.compute_cost(self.weights_bias, self.train_ex, self.labels, self.reg, w_size))
            if i % 100 == 0:
                self.get_accuracy()


    def learn_param(self):
        w_size = [(self.hidden_layer_size, self.input_layer_size), (self.num_labels, self.hidden_layer_size)]
        self.weights_bias = opt.fmin_cg(f=NeuralNetwork.compute_cost, x0=self.weights_bias,
                                        fprime=NeuralNetwork.compute_gradient,
                                        args=(self.train_ex, self.labels, self.reg, w_size), maxiter=100)
        return self.weights_bias

    def predict(self, img, theta):
        """compute the likelihood for the image to belong to each class (0-9) than returns the higher percentage"""
        num_class = [None] * theta.shape[0]
        for i in range(0, 10):
            cur_theta = np.reshape(theta[i, :], (theta.shape[1], 1))
            num_class[i] = NeuralNetwork.sigmoid(np.dot(img, cur_theta))[0][0]
        return np.argmax(np.array(num_class))

    def get_accuracy(self):
        accuracy = 0
        w_size = [(self.hidden_layer_size, self.input_layer_size), (self.num_labels, self.hidden_layer_size)]
        res, no, na, ni = NeuralNetwork.forward_prop(self.weights_bias, self.train_ex, w_size)
        for i in range(5000):
            if np.argmax(res[i]) == self.labels[i]:
                accuracy += 1
        print('accuracy :', 100 * accuracy / 5000, '%')
Warning: Desired error not necessarily achieved due to precision loss.
         Current function value: 6.751827
         Iterations: 0
         Function evaluations: 45
         Gradient evaluations: 33

不到一秒钟后

这是使用的数据集: http://s000.tinyupload.com/index.php?file_id=00036473883875179878

0 个答案:

没有答案
相关问题