文章目录
- 1.参数的更新
- 1)SGD
- 2)Momentum
- 3)AdaGrad
- 4)Adam
- 5)最优化方法的比较
- 6)基于MNIST数据集的更新方法的比较
- 2.权重的初始值
- 1)权重初始值不能为0
- 2)隐藏层的激活值的分布
- 3)ReLU的权重初始值
- 4)基于MNIST数据集的权重初始值的比较
- 3.Batch Normalization
- 1)Batch Normalization 的算法
- 2)Batch Normalization的评估
- 4.正则化
- 1)过拟合
- 2)权值衰减
- 3)Dropout
- 4)集成学习
- 5.超参数的验证
- 1)验证数据
- 2)超参数的最优化
- 3)步骤
- 4)超参数最优化的实现
1.参数的更新
神经网络的学习的目的是找到使损失函数的值尽可能小的参数。这是寻找最优参数的问题,解决这个问题的过程称为最优化(optimization)。
使用参数的梯度,沿梯度方向更新参数,并重复这个步骤多次,从而逐渐靠近最优参数,这个过程称为随机梯度下降法(stochastic gradient descent),简称SGD。
1)SGD
探险家虽然看不到周围的情况,但是能够知道当前所在位置的坡度(通过脚底感受地面的倾斜状况)。于是,朝着当前所在位置的坡度最大的方向前进,就是SGD的策略。
SGD实现:
class SGD:"""随机梯度下降法(Stochastic Gradient Descent)"""
#进行初始化时的参数lr表示learning rate(学习率)def __init__(self, lr=0.01):self.lr = lr#该方法在SGD中会被反复调用。def update(self, params, grads):for key in params.keys():params[key] -= self.lr * grads[key]
SGD呈“之”字形移动。这是一个相当低效的路径。也就是说,SGD 的缺点是,如果函数的形状非均向(anisotropic),比如呈延伸状,搜索的路径就会非常低效。因此,我们需要比单纯朝梯度方向前进的SGD更聪明的方法。SGD低效的根本原因是,梯度的方向并没有指向最小值的方向。
2)Momentum
Momentum 是“动量”的意思,和物理有关。Momentum方法给人的感觉就像是小球在地面上滚动。在物体不受任何力时,担使物体逐渐减
速的任务
实现
class Momentum:"""Momentum SGD"""def __init__(self, lr=0.01, momentum=0.9):self.lr = lrself.momentum = momentum#实例变量v会保存物体的速度。self.v = Nonedef update(self, params, grads):if self.v is None:self.v = {}for key, val in params.items(): #v会以字典型变量的形式保存与参数结构相同的数据self.v[key] = np.zeros_like(val)for key in params.keys():self.v[key] = self.momentum*self.v[key] - self.lr*grads[key] params[key] += self.v[key]
和SGD相比,我们发现“之”字形的“程度”减轻了。这是因为虽然x轴方向上受到的力非常小,但是一直在同一方向上受力,所以朝同一个方向会有一定的加速。反过来,虽然y轴方向上受到的力很大,但是因为交互地受到正方向和反方向的力,它们会互相抵消,所以y轴方向上的速度不稳定。因此,和SGD时的情形相比,
可以更快地朝x轴方向靠近,减弱“之”字形的变动程度。
3)AdaGrad
在关于学习率的有效技巧中,有一种被称为学习率衰减(learning rate decay)的方法,即随着学习的进行,使学习率逐渐减小。实际上,一开始“多”学,然后逐渐“少”学的方法,在神经网络的学习中经常被使用。
逐渐减小学习率的想法,相当于将“全体”参数的学习率值一起降低。而AdaGrad进一步发展了这个想法,针对“一个一个”的参数,赋予其“定制”的值。表达式如下:
可以按参数的元素进行学习率衰减,使变动大的参数的学习率逐渐减小。
4)Adam
融合了Momentum和AdaGrad的方法。
5)最优化方法的比较
完整代码如下:
创建文件optimizer.py,实现各种优化器 添加代码如下:
# coding: utf-8
import numpy as npclass SGD:"""随机梯度下降法(Stochastic Gradient Descent)"""def __init__(self, lr=0.01):self.lr = lrdef update(self, params, grads):for key in params.keys():params[key] -= self.lr * grads[key] class Momentum:"""Momentum SGD"""def __init__(self, lr=0.01, momentum=0.9):self.lr = lrself.momentum = momentumself.v = Nonedef update(self, params, grads):if self.v is None:self.v = {}for key, val in params.items(): self.v[key] = np.zeros_like(val)for key in params.keys():self.v[key] = self.momentum*self.v[key] - self.lr*grads[key] params[key] += self.v[key]class Nesterov:"""Nesterov's Accelerated Gradient (http://arxiv.org/abs/1212.0901)"""def __init__(self, lr=0.01, momentum=0.9):self.lr = lrself.momentum = momentumself.v = Nonedef update(self, params, grads):if self.v is None:self.v = {}for key, val in params.items():self.v[key] = np.zeros_like(val)for key in params.keys():self.v[key] *= self.momentumself.v[key] -= self.lr * grads[key]params[key] += self.momentum * self.momentum * self.v[key]params[key] -= (1 + self.momentum) * self.lr * grads[key]class AdaGrad:"""AdaGrad"""def __init__(self, lr=0.01):self.lr = lrself.h = Nonedef update(self, params, grads):if self.h is None:self.h = {}for key, val in params.items():self.h[key] = np.zeros_like(val)for key in params.keys():self.h[key] += grads[key] * grads[key]params[key] -= self.lr * grads[key] / (np.sqrt(self.h[key]) + 1e-7)class RMSprop:"""RMSprop"""def __init__(self, lr=0.01, decay_rate = 0.99):self.lr = lrself.decay_rate = decay_rateself.h = Nonedef update(self, params, grads):if self.h is None:self.h = {}for key, val in params.items():self.h[key] = np.zeros_like(val)for key in params.keys():self.h[key] *= self.decay_rateself.h[key] += (1 - self.decay_rate) * grads[key] * grads[key]params[key] -= self.lr * grads[key] / (np.sqrt(self.h[key]) + 1e-7)class Adam:"""Adam (http://arxiv.org/abs/1412.6980v8)"""def __init__(self, lr=0.001, beta1=0.9, beta2=0.999):self.lr = lrself.beta1 = beta1self.beta2 = beta2self.iter = 0self.m = Noneself.v = Nonedef update(self, params, grads):if self.m is None:self.m, self.v = {}, {}for key, val in params.items():self.m[key] = np.zeros_like(val)self.v[key] = np.zeros_like(val)self.iter += 1lr_t = self.lr * np.sqrt(1.0 - self.beta2**self.iter) / (1.0 - self.beta1**self.iter) for key in params.keys():#self.m[key] = self.beta1*self.m[key] + (1-self.beta1)*grads[key]#self.v[key] = self.beta2*self.v[key] + (1-self.beta2)*(grads[key]**2)self.m[key] += (1 - self.beta1) * (grads[key] - self.m[key])self.v[key] += (1 - self.beta2) * (grads[key]**2 - self.v[key])params[key] -= lr_t * self.m[key] / (np.sqrt(self.v[key]) + 1e-7)#unbias_m += (1 - self.beta1) * (grads[key] - self.m[key]) # correct bias#unbisa_b += (1 - self.beta2) * (grads[key]*grads[key] - self.v[key]) # correct bias#params[key] += self.lr * unbias_m / (np.sqrt(unbisa_b) + 1e-7)
创建文件optimizer_compare_naive.py,引入优化器
添加代码如下:
# coding: utf-8
import sys, os
sys.path.append(os.pardir) # 为了导入父目录的文件而进行的设定
import numpy as np
import matplotlib.pyplot as plt
from collections import OrderedDict
from optimizer import *def f(x, y):return x**2 / 20.0 + y**2def df(x, y):return x / 10.0, 2.0*yinit_pos = (-7.0, 2.0)
params = {}
params['x'], params['y'] = init_pos[0], init_pos[1]
grads = {}
grads['x'], grads['y'] = 0, 0optimizers = OrderedDict()
optimizers["SGD"] = SGD(lr=0.95)
optimizers["Momentum"] = Momentum(lr=0.1)
optimizers["AdaGrad"] = AdaGrad(lr=1.5)
optimizers["Adam"] = Adam(lr=0.3)idx = 1for key in optimizers:optimizer = optimizers[key]x_history = []y_history = []params['x'], params['y'] = init_pos[0], init_pos[1]for i in range(30):x_history.append(params['x'])y_history.append(params['y'])grads['x'], grads['y'] = df(params['x'], params['y'])optimizer.update(params, grads)x = np.arange(-10, 10, 0.01)y = np.arange(-5, 5, 0.01)X, Y = np.meshgrid(x, y) Z = f(X, Y)# for simple contour line mask = Z > 7Z[mask] = 0# plot plt.subplot(2, 2, idx)idx += 1plt.plot(x_history, y_history, 'o-', color="red")plt.contour(X, Y, Z)plt.ylim(-10, 10)plt.xlim(-10, 10)plt.plot(0, 0, '+')#colorbar()#spring()plt.title(key)plt.xlabel("x")plt.ylabel("y")plt.show()
运行结果:
根据使用的方法不同,参数更新的路径也不同。只看这个图的话,AdaGrad似乎是最好的,不过也要注意,结果会根据要解决的问题而变。并且,很显然,超参数(学习率等)的设定值不同,结果也会发生变化。
并不存在能在所有问题中都表现良好的方法。这4种方法各有各的特点,都有各自擅长解决的问题和不擅长解决的问题。
很多研究中至今仍在使用SGD。Momentum和AdaGrad也是值得一试的方法。最近,很多研究人员和技术人员都喜欢用Adam。
6)基于MNIST数据集的更新方法的比较
完整代码如下:
创建文件夹common,移入optimizer.py,
在文件夹common下创建文件util.py
添加代码如下:
# coding: utf-8
import numpy as npdef smooth_curve(x):"""用于使损失函数的图形变圆滑参考:http://glowingpython.blogspot.jp/2012/02/convolution-with-numpy.html"""window_len = 11s = np.r_[x[window_len-1:0:-1], x, x[-1:-window_len:-1]]w = np.kaiser(window_len, 2)y = np.convolve(w/w.sum(), s, mode='valid')return y[5:len(y)-5]def shuffle_dataset(x, t):"""打乱数据集Parameters----------x : 训练数据t : 监督数据Returns-------x, t : 打乱的训练数据和监督数据"""permutation = np.random.permutation(x.shape[0])x = x[permutation,:] if x.ndim == 2 else x[permutation,:,:,:]t = t[permutation]return x, tdef conv_output_size(input_size, filter_size, stride=1, pad=0):return (input_size + 2*pad - filter_size) / stride + 1def im2col(input_data, filter_h, filter_w, stride=1, pad=0):"""Parameters----------input_data : 由(数据量, 通道, 高, 长)的4维数组构成的输入数据filter_h : 滤波器的高filter_w : 滤波器的长stride : 步幅pad : 填充Returns-------col : 2维数组"""N, C, H, W = input_data.shapeout_h = (H + 2*pad - filter_h)//stride + 1out_w = (W + 2*pad - filter_w)//stride + 1img = np.pad(input_data, [(0,0), (0,0), (pad, pad), (pad, pad)], 'constant')col = np.zeros((N, C, filter_h, filter_w, out_h, out_w))for y in range(filter_h):y_max = y + stride*out_hfor x in range(filter_w):x_max = x + stride*out_wcol[:, :, y, x, :, :] = img[:, :, y:y_max:stride, x:x_max:stride]col = col.transpose(0, 4, 5, 1, 2, 3).reshape(N*out_h*out_w, -1)return coldef col2im(col, input_shape, filter_h, filter_w, stride=1, pad=0):"""Parameters----------col :input_shape : 输入数据的形状(例:(10, 1, 28, 28))filter_h :filter_wstridepadReturns-------"""N, C, H, W = input_shapeout_h = (H + 2*pad - filter_h)//stride + 1out_w = (W + 2*pad - filter_w)//stride + 1col = col.reshape(N, out_h, out_w, C, filter_h, filter_w).transpose(0, 3, 4, 5, 1, 2)img = np.zeros((N, C, H + 2*pad + stride - 1, W + 2*pad + stride - 1))for y in range(filter_h):y_max = y + stride*out_hfor x in range(filter_w):x_max = x + stride*out_wimg[:, :, y:y_max:stride, x:x_max:stride] += col[:, :, y, x, :, :]return img[:, :, pad:H + pad, pad:W + pad]
在文件夹common下创建文件multi_layer_net.py
添加代码如下:
# coding: utf-8
import sys, os
sys.path.append(os.pardir) # 为了导入父目录的文件而进行的设定
import numpy as np
from collections import OrderedDict
from common.layers import *
from common.gradient import numerical_gradientclass MultiLayerNet:"""全连接的多层神经网络Parameters----------input_size : 输入大小(MNIST的情况下为784)hidden_size_list : 隐藏层的神经元数量的列表(e.g. [100, 100, 100])output_size : 输出大小(MNIST的情况下为10)activation : 'relu' or 'sigmoid'weight_init_std : 指定权重的标准差(e.g. 0.01)指定'relu'或'he'的情况下设定“He的初始值”指定'sigmoid'或'xavier'的情况下设定“Xavier的初始值”weight_decay_lambda : Weight Decay(L2范数)的强度"""def __init__(self, input_size, hidden_size_list, output_size,activation='relu', weight_init_std='relu', weight_decay_lambda=0):self.input_size = input_sizeself.output_size = output_sizeself.hidden_size_list = hidden_size_listself.hidden_layer_num = len(hidden_size_list)self.weight_decay_lambda = weight_decay_lambdaself.params = {}# 初始化权重self.__init_weight(weight_init_std)# 生成层activation_layer = {'sigmoid': Sigmoid, 'relu': Relu}self.layers = OrderedDict()for idx in range(1, self.hidden_layer_num+1):self.layers['Affine' + str(idx)] = Affine(self.params['W' + str(idx)],self.params['b' + str(idx)])self.layers['Activation_function' + str(idx)] = activation_layer[activation]()idx = self.hidden_layer_num + 1self.layers['Affine' + str(idx)] = Affine(self.params['W' + str(idx)],self.params['b' + str(idx)])self.last_layer = SoftmaxWithLoss()def __init_weight(self, weight_init_std):"""设定权重的初始值Parameters----------weight_init_std : 指定权重的标准差(e.g. 0.01)指定'relu'或'he'的情况下设定“He的初始值”指定'sigmoid'或'xavier'的情况下设定“Xavier的初始值”"""all_size_list = [self.input_size] + self.hidden_size_list + [self.output_size]for idx in range(1, len(all_size_list)):scale = weight_init_stdif str(weight_init_std).lower() in ('relu', 'he'):scale = np.sqrt(2.0 / all_size_list[idx - 1]) # 使用ReLU的情况下推荐的初始值elif str(weight_init_std).lower() in ('sigmoid', 'xavier'):scale = np.sqrt(1.0 / all_size_list[idx - 1]) # 使用sigmoid的情况下推荐的初始值self.params['W' + str(idx)] = scale * np.random.randn(all_size_list[idx-1], all_size_list[idx])self.params['b' + str(idx)] = np.zeros(all_size_list[idx])def predict(self, x):for layer in self.layers.values():x = layer.forward(x)return xdef loss(self, x, t):"""求损失函数Parameters----------x : 输入数据t : 教师标签Returns-------损失函数的值"""y = self.predict(x)weight_decay = 0for idx in range(1, self.hidden_layer_num + 2):W = self.params['W' + str(idx)]weight_decay += 0.5 * self.weight_decay_lambda * np.sum(W ** 2)return self.last_layer.forward(y, t) + weight_decaydef accuracy(self, x, t):y = self.predict(x)y = np.argmax(y, axis=1)if t.ndim != 1 : t = np.argmax(t, axis=1)accuracy = np.sum(y == t) / float(x.shape[0])return accuracydef numerical_gradient(self, x, t):"""求梯度(数值微分)Parameters----------x : 输入数据t : 教师标签Returns-------具有各层的梯度的字典变量grads['W1']、grads['W2']、...是各层的权重grads['b1']、grads['b2']、...是各层的偏置"""loss_W = lambda W: self.loss(x, t)grads = {}for idx in range(1, self.hidden_layer_num+2):grads['W' + str(idx)] = numerical_gradient(loss_W, self.params['W' + str(idx)])grads['b' + str(idx)] = numerical_gradient(loss_W, self.params['b' + str(idx)])return gradsdef gradient(self, x, t):"""求梯度(误差反向传播法)Parameters----------x : 输入数据t : 教师标签Returns-------具有各层的梯度的字典变量grads['W1']、grads['W2']、...是各层的权重grads['b1']、grads['b2']、...是各层的偏置"""# forwardself.loss(x, t)# backwarddout = 1dout = self.last_layer.backward(dout)layers = list(self.layers.values())layers.reverse()for layer in layers:dout = layer.backward(dout)# 设定grads = {}for idx in range(1, self.hidden_layer_num+2):grads['W' + str(idx)] = self.layers['Affine' + str(idx)].dW + self.weight_decay_lambda * self.layers['Affine' + str(idx)].Wgrads['b' + str(idx)] = self.layers['Affine' + str(idx)].dbreturn grads
在文件夹common下创建文件optimizer_compare_mnist.py
添加代码如下:
# coding: utf-8
import os
import sys
sys.path.append(os.pardir) # 为了导入父目录的文件而进行的设定
import matplotlib.pyplot as plt
from dataset.mnist import load_mnist
from common.util import smooth_curve
from common.multi_layer_net import MultiLayerNet
from common.optimizer import *# 0:读入MNIST数据==========
(x_train, t_train), (x_test, t_test) = load_mnist(normalize=True)train_size = x_train.shape[0]
batch_size = 128
max_iterations = 2000# 1:进行实验的设置==========
optimizers = {}
optimizers['SGD'] = SGD()
optimizers['Momentum'] = Momentum()
optimizers['AdaGrad'] = AdaGrad()
optimizers['Adam'] = Adam()
#optimizers['RMSprop'] = RMSprop()networks = {}
train_loss = {}
for key in optimizers.keys():networks[key] = MultiLayerNet(input_size=784, hidden_size_list=[100, 100, 100, 100],output_size=10)train_loss[key] = [] # 2:开始训练==========
for i in range(max_iterations):batch_mask = np.random.choice(train_size, batch_size)x_batch = x_train[batch_mask]t_batch = t_train[batch_mask]for key in optimizers.keys():grads = networks[key].gradient(x_batch, t_batch)optimizers[key].update(networks[key].params, grads)loss = networks[key].loss(x_batch, t_batch)train_loss[key].append(loss)if i % 100 == 0:print( "===========" + "iteration:" + str(i) + "===========")for key in optimizers.keys():loss = networks[key].loss(x_batch, t_batch)print(key + ":" + str(loss))# 3.绘制图形==========
markers = {"SGD": "o", "Momentum": "x", "AdaGrad": "s", "Adam": "D"}
x = np.arange(max_iterations)
for key in optimizers.keys():plt.plot(x, smooth_curve(train_loss[key]), marker=markers[key], markevery=100, label=key)
plt.xlabel("iterations")
plt.ylabel("loss")
plt.ylim(0, 1)
plt.legend()
plt.show()
与SGD相比,其他3种方法学习得更快,而且速度基本相同,仔细看的话,AdaGrad的学习进行得稍微快一点。这个实验需要注意的地方是,实验结果会随学习率等超参数、神经网络的结构(几层深等)的不同而发生变化。不过,一般而言,与SGD相比,其他3种方法可以学习得更快,有时最终的识别精度也更高。
2.权重的初始值
权值衰减就是一种提高泛化能力的技巧,以减小权重参数的值为目的进行学习的方法。通过减小权重参数的值来抑制过拟合的发生。
如果想减小权重的值,一开始就将初始值设为较小的值才是正途。
1)权重初始值不能为0
将权重初始值设为0不是一个好主意。事实上,将权重初始值设为0 的话,将无法正确进行学习。
这是因为在误差反向传播法中,所有的权重值都会进行相同的更新。比如,在2层神经网络中,假设第1层和第2层的权重为0。这样一来,正向传播时,因为输入层的权重为0,所以第2层的神经元全部会被传递相同的值。
2)隐藏层的激活值的分布
将激活函数的输出数据称为“激活值”,但是有的文献中会将在层之间流动的数据也称为“激活值”。
实现:
# coding: utf-8
import numpy as np
import matplotlib.pyplot as pltdef sigmoid(x):return 1 / (1 + np.exp(-x))def ReLU(x):return np.maximum(0, x)def tanh(x):return np.tanh(x)#高斯分布随机生成1000个数据作为输入数据,并把它们传给5层神经网络。
input_data = np.random.randn(1000, 100) # 1000个数据
node_num = 100 # 各隐藏层的节点(神经元)数
hidden_layer_size = 5 # 隐藏层有5层
activations = {} # 激活值的结果保存在这里x = input_datafor i in range(hidden_layer_size):if i != 0:x = activations[i-1]# 改变初始值进行实验!w = np.random.randn(node_num, node_num) * 1# w = np.random.randn(node_num, node_num) * 0.01# w = np.random.randn(node_num, node_num) * np.sqrt(1.0 / node_num)# w = np.random.randn(node_num, node_num) * np.sqrt(2.0 / node_num)a = np.dot(x, w)# 将激活函数的种类也改变,来进行实验!z = sigmoid(a)# z = ReLU(a)# z = tanh(a)activations[i] = z# 绘制直方图
for i, a in activations.items():plt.subplot(1, len(activations), i+1)plt.title(str(i+1) + "-layer")if i != 0: plt.yticks([], [])# plt.xlim(0.1, 1)# plt.ylim(0, 7000)plt.hist(a.flatten(), 30, range=(0,1))
plt.show()
将权重的标准差设为1时,结果显示各层的激活值呈偏向0和1的分布。偏向0和1的数据分布会造成反向传播中梯度的值不断变小,最后消失。这个问题称为梯度消失(gradient vanishing)。层次加深的深度学习中,梯度消失的问题可能会更加严重。
将权重的标准差设为0.01时,这次呈集中在0.5附近的分布。因为不像刚才的例子那样偏向0和1,所以不会发生梯度消失的问题。
因此,激活值在分布上有所偏向会出现“表现力受限”的问题。
尝试使用Xavier Glorot等人的论文中推荐的权重初始值(俗称“Xavier初始值”)。Xavier 的论文中,为了使各层的激活值呈现出具有相同广度的分布,推导了合适的权重尺度。推导出的结论是,如果前一层的节点数为n,则初始值使用标准差为
1 / n 1/ \sqrt{n} 1/n
的分布。结果表明后面的层的分布呈稍微歪斜的形状。如果用tanh函数(双曲线函数)代替sigmoid函数,这个稍微歪斜的问题就能得到改善。实际上,使用tanh函数后,会呈漂亮的吊钟型分布。tanh函数和sigmoid函数同是S型曲线函数,但tanh函数是关于原点(0, 0)对称的S型曲线,而sigmoid函数是关于(x, y)=(0, 0.5)对称的S型曲线。众所周知,用作激活函数的函数最好具有关于原点对称的性质。
3)ReLU的权重初始值
当激活函数使用ReLU时,一般推荐使用ReLU专用的初始值,也就是Kaiming He等人推荐的初始值,也称为He初始值。
在这种情况下,随着层的加深,偏向一点点变大。实际上,层加深后,激活值的偏向变大,学习时会出现梯度消失的问题。而当初始值为He初始值时,各层中分布的广度相同。由于即便层加深,数据的广度也能保持不变,因此逆向传播时,也会传递合适的值。
总结一下,当激活函数使用ReLU时,权重初始值使用He初始值,当激活函数为sigmoid或tanh等S型曲线函数时,初始值使用Xavier初始值。这是目前的最佳实践。
4)基于MNIST数据集的权重初始值的比较
创建文件weight_init_compare.py
添加代码如下:
# coding: utf-8
import os
import syssys.path.append(os.pardir) # 为了导入父目录的文件而进行的设定
import numpy as np
import matplotlib.pyplot as plt
from dataset.mnist import load_mnist
from common.util import smooth_curve
from common.multi_layer_net import MultiLayerNet
from common.optimizer import SGD# 0:读入MNIST数据==========
(x_train, t_train), (x_test, t_test) = load_mnist(normalize=True)train_size = x_train.shape[0]
batch_size = 128
max_iterations = 2000# 1:进行实验的设置==========
weight_init_types = {'std=0.01': 0.01, 'Xavier': 'sigmoid', 'He': 'relu'}
optimizer = SGD(lr=0.01)networks = {}
train_loss = {}
for key, weight_type in weight_init_types.items():networks[key] = MultiLayerNet(input_size=784, hidden_size_list=[100, 100, 100, 100],output_size=10, weight_init_std=weight_type)train_loss[key] = []# 2:开始训练==========
for i in range(max_iterations):batch_mask = np.random.choice(train_size, batch_size)x_batch = x_train[batch_mask]t_batch = t_train[batch_mask]for key in weight_init_types.keys():grads = networks[key].gradient(x_batch, t_batch)optimizer.update(networks[key].params, grads)loss = networks[key].loss(x_batch, t_batch)train_loss[key].append(loss)if i % 100 == 0:print("===========" + "iteration:" + str(i) + "===========")for key in weight_init_types.keys():loss = networks[key].loss(x_batch, t_batch)print(key + ":" + str(loss))# 3.绘制图形==========
markers = {'std=0.01': 'o', 'Xavier': 's', 'He': 'D'}
x = np.arange(max_iterations)
for key in weight_init_types.keys():plt.plot(x, smooth_curve(train_loss[key]), marker=markers[key], markevery=100, label=key)
plt.xlabel("iterations")
plt.ylabel("loss")
plt.ylim(0, 2.5)
plt.legend()
plt.show()
运行结果:
std = 0.01时完全无法进行学习。这和刚才观察到的激活值的分布一样,是因为正向传播中传递的值很小(集中在0附近的数据)。因此,逆向传播时求到的梯度也很小,权重几乎不进行更新。相反,当权重初始值为Xavier初始值和He初始值时,学习进行得很顺利。
并且,我们发现He初始值时的学习进度更快一些。
3.Batch Normalization
Batch Normalization可以“强制性”地调整激活值的分布
1)Batch Normalization 的算法
优点:
• 可以使学习快速进行(可以增大学习率)。
• 不那么依赖初始值(对于初始值不用那么神经质)。
• 抑制过拟合(降低Dropout等的必要性)。
Batch Norm的思路是调整各层的激活值分布使其拥有适当的广度。为此,要向神经网络中插入对数据分布进行正规化的层,即Batch Normalization 层
2)Batch Normalization的评估
使用 MNIST 数据集,观察使用Batch Norm层和不使用Batch Norm层时学习的过程
实现:
# coding: utf-8
import sys, os
sys.path.append(os.pardir) # 为了导入父目录的文件而进行的设定
import numpy as np
import matplotlib.pyplot as plt
from dataset.mnist import load_mnist
from common.multi_layer_net_extend import MultiLayerNetExtend
from common.optimizer import SGD, Adam(x_train, t_train), (x_test, t_test) = load_mnist(normalize=True)# 减少学习数据
x_train = x_train[:1000]
t_train = t_train[:1000]max_epochs = 20
train_size = x_train.shape[0]
batch_size = 100
learning_rate = 0.01def __train(weight_init_std):bn_network = MultiLayerNetExtend(input_size=784, hidden_size_list=[100, 100, 100, 100, 100], output_size=10, weight_init_std=weight_init_std, use_batchnorm=True)network = MultiLayerNetExtend(input_size=784, hidden_size_list=[100, 100, 100, 100, 100], output_size=10,weight_init_std=weight_init_std)optimizer = SGD(lr=learning_rate)train_acc_list = []bn_train_acc_list = []iter_per_epoch = max(train_size / batch_size, 1)epoch_cnt = 0for i in range(1000000000):batch_mask = np.random.choice(train_size, batch_size)x_batch = x_train[batch_mask]t_batch = t_train[batch_mask]for _network in (bn_network, network):grads = _network.gradient(x_batch, t_batch)optimizer.update(_network.params, grads)if i % iter_per_epoch == 0:train_acc = network.accuracy(x_train, t_train)bn_train_acc = bn_network.accuracy(x_train, t_train)train_acc_list.append(train_acc)bn_train_acc_list.append(bn_train_acc)print("epoch:" + str(epoch_cnt) + " | " + str(train_acc) + " - " + str(bn_train_acc))epoch_cnt += 1if epoch_cnt >= max_epochs:breakreturn train_acc_list, bn_train_acc_list# 3.绘制图形==========
weight_scale_list = np.logspace(0, -4, num=16)
x = np.arange(max_epochs)for i, w in enumerate(weight_scale_list):print( "============== " + str(i+1) + "/16" + " ==============")train_acc_list, bn_train_acc_list = __train(w)plt.subplot(4,4,i+1)plt.title("W:" + str(w))if i == 15:plt.plot(x, bn_train_acc_list, label='Batch Normalization', markevery=2)plt.plot(x, train_acc_list, linestyle = "--", label='Normal(without BatchNorm)', markevery=2)else:plt.plot(x, bn_train_acc_list, markevery=2)plt.plot(x, train_acc_list, linestyle="--", markevery=2)plt.ylim(0, 1.0)if i % 4:plt.yticks([])else:plt.ylabel("accuracy")if i < 12:plt.xticks([])else:plt.xlabel("epochs")plt.legend(loc='lower right')plt.show()
运行结果:
使用Batch Norm后,学习进行得更快了。
4.正则化
1)过拟合
发生过拟合的原因,主要有以下两个。
• 模型拥有大量参数、表现力强。
• 训练数据少。
过拟合代码实现,
创建文件overfit_weight_decay.py
添加代码如下:
# coding: utf-8
import os
import syssys.path.append(os.pardir) # 为了导入父目录的文件而进行的设定
import numpy as np
import matplotlib.pyplot as plt
from dataset.mnist import load_mnist
from common.multi_layer_net import MultiLayerNet
from common.optimizer import SGD(x_train, t_train), (x_test, t_test) = load_mnist(normalize=True)# 为了再现过拟合,减少学习数据
x_train = x_train[:300]
t_train = t_train[:300]# weight decay(权值衰减)的设定 =======================
#weight_decay_lambda = 0 # 不使用权值衰减的情况
weight_decay_lambda = 0.1
# ====================================================network = MultiLayerNet(input_size=784, hidden_size_list=[100, 100, 100, 100, 100, 100], output_size=10,weight_decay_lambda=weight_decay_lambda)
optimizer = SGD(lr=0.01)max_epochs = 201
train_size = x_train.shape[0]
batch_size = 100train_loss_list = []
train_acc_list = []
test_acc_list = []iter_per_epoch = max(train_size / batch_size, 1)
epoch_cnt = 0for i in range(1000000000):batch_mask = np.random.choice(train_size, batch_size)x_batch = x_train[batch_mask]t_batch = t_train[batch_mask]grads = network.gradient(x_batch, t_batch)optimizer.update(network.params, grads)if i % iter_per_epoch == 0:train_acc = network.accuracy(x_train, t_train)test_acc = network.accuracy(x_test, t_test)train_acc_list.append(train_acc)test_acc_list.append(test_acc)print("epoch:" + str(epoch_cnt) + ", train acc:" + str(train_acc) + ", test acc:" + str(test_acc))epoch_cnt += 1if epoch_cnt >= max_epochs:break# 3.绘制图形==========
markers = {'train': 'o', 'test': 's'}
x = np.arange(max_epochs)
plt.plot(x, train_acc_list, marker='o', label='train', markevery=10)
plt.plot(x, test_acc_list, marker='s', label='test', markevery=10)
plt.xlabel("epochs")
plt.ylabel("accuracy")
plt.ylim(0, 1.0)
plt.legend(loc='lower right')
plt.show()
运行结果:
过了 100 个 epoch 左右后,用训练数据测量到的识别精度几乎都为100%。但是,对于测试数据,离 100% 的识别精度还有较大的差距。如此大的识别精度差距,是只拟合了训练数据的结果。从图中可知,模型对训练时没有使用的一般数据(测试数据)拟合得不是很好。
2)权值衰减
权值衰减是一直以来经常被使用的一种抑制过拟合的方法。该方法通过在学习的过程中对大的权重进行惩罚,来抑制过拟合。很多过拟合原本就是因为权重参数取值过大才发生的。
对于刚刚进行的实验,应用λ = 0.1的权值衰减。
虽然训练数据的识别精度和测试数据的识别精度之间有差距,但是与没有使用权值衰减相比,差距变小了。
3)Dropout
Dropout 是一种在学习的过程中随机删除神经元的方法。训练时,随机选出隐藏层的神经元,然后将其删除。
在common文件夹下创建文件multi_layer_net_extend.py
添加代码如下:
# coding: utf-8
import sys, os
sys.path.append(os.pardir) # 为了导入父目录的文件而进行的设定
import numpy as np
from collections import OrderedDict
from common.layers import *
from common.gradient import numerical_gradientclass MultiLayerNetExtend:"""扩展版的全连接的多层神经网络具有Weiht Decay、Dropout、Batch Normalization的功能Parameters----------input_size : 输入大小(MNIST的情况下为784)hidden_size_list : 隐藏层的神经元数量的列表(e.g. [100, 100, 100])output_size : 输出大小(MNIST的情况下为10)activation : 'relu' or 'sigmoid'weight_init_std : 指定权重的标准差(e.g. 0.01)指定'relu'或'he'的情况下设定“He的初始值”指定'sigmoid'或'xavier'的情况下设定“Xavier的初始值”weight_decay_lambda : Weight Decay(L2范数)的强度use_dropout: 是否使用Dropoutdropout_ration : Dropout的比例use_batchNorm: 是否使用Batch Normalization"""def __init__(self, input_size, hidden_size_list, output_size,activation='relu', weight_init_std='relu', weight_decay_lambda=0, use_dropout = False, dropout_ration = 0.5, use_batchnorm=False):self.input_size = input_sizeself.output_size = output_sizeself.hidden_size_list = hidden_size_listself.hidden_layer_num = len(hidden_size_list)self.use_dropout = use_dropoutself.weight_decay_lambda = weight_decay_lambdaself.use_batchnorm = use_batchnormself.params = {}# 初始化权重self.__init_weight(weight_init_std)# 生成层activation_layer = {'sigmoid': Sigmoid, 'relu': Relu}self.layers = OrderedDict()for idx in range(1, self.hidden_layer_num+1):self.layers['Affine' + str(idx)] = Affine(self.params['W' + str(idx)],self.params['b' + str(idx)])if self.use_batchnorm:self.params['gamma' + str(idx)] = np.ones(hidden_size_list[idx-1])self.params['beta' + str(idx)] = np.zeros(hidden_size_list[idx-1])self.layers['BatchNorm' + str(idx)] = BatchNormalization(self.params['gamma' + str(idx)], self.params['beta' + str(idx)])self.layers['Activation_function' + str(idx)] = activation_layer[activation]()if self.use_dropout:self.layers['Dropout' + str(idx)] = Dropout(dropout_ration)idx = self.hidden_layer_num + 1self.layers['Affine' + str(idx)] = Affine(self.params['W' + str(idx)], self.params['b' + str(idx)])self.last_layer = SoftmaxWithLoss()def __init_weight(self, weight_init_std):"""设定权重的初始值Parameters----------weight_init_std : 指定权重的标准差(e.g. 0.01)指定'relu'或'he'的情况下设定“He的初始值”指定'sigmoid'或'xavier'的情况下设定“Xavier的初始值”"""all_size_list = [self.input_size] + self.hidden_size_list + [self.output_size]for idx in range(1, len(all_size_list)):scale = weight_init_stdif str(weight_init_std).lower() in ('relu', 'he'):scale = np.sqrt(2.0 / all_size_list[idx - 1]) # 使用ReLU的情况下推荐的初始值elif str(weight_init_std).lower() in ('sigmoid', 'xavier'):scale = np.sqrt(1.0 / all_size_list[idx - 1]) # 使用sigmoid的情况下推荐的初始值self.params['W' + str(idx)] = scale * np.random.randn(all_size_list[idx-1], all_size_list[idx])self.params['b' + str(idx)] = np.zeros(all_size_list[idx])def predict(self, x, train_flg=False):for key, layer in self.layers.items():if "Dropout" in key or "BatchNorm" in key:x = layer.forward(x, train_flg)else:x = layer.forward(x)return xdef loss(self, x, t, train_flg=False):"""求损失函数参数x是输入数据,t是教师标签"""y = self.predict(x, train_flg)weight_decay = 0for idx in range(1, self.hidden_layer_num + 2):W = self.params['W' + str(idx)]weight_decay += 0.5 * self.weight_decay_lambda * np.sum(W**2)return self.last_layer.forward(y, t) + weight_decaydef accuracy(self, X, T):Y = self.predict(X, train_flg=False)Y = np.argmax(Y, axis=1)if T.ndim != 1 : T = np.argmax(T, axis=1)accuracy = np.sum(Y == T) / float(X.shape[0])return accuracydef numerical_gradient(self, X, T):"""求梯度(数值微分)Parameters----------X : 输入数据T : 教师标签Returns-------具有各层的梯度的字典变量grads['W1']、grads['W2']、...是各层的权重grads['b1']、grads['b2']、...是各层的偏置"""loss_W = lambda W: self.loss(X, T, train_flg=True)grads = {}for idx in range(1, self.hidden_layer_num+2):grads['W' + str(idx)] = numerical_gradient(loss_W, self.params['W' + str(idx)])grads['b' + str(idx)] = numerical_gradient(loss_W, self.params['b' + str(idx)])if self.use_batchnorm and idx != self.hidden_layer_num+1:grads['gamma' + str(idx)] = numerical_gradient(loss_W, self.params['gamma' + str(idx)])grads['beta' + str(idx)] = numerical_gradient(loss_W, self.params['beta' + str(idx)])return gradsdef gradient(self, x, t):# forwardself.loss(x, t, train_flg=True)# backwarddout = 1dout = self.last_layer.backward(dout)layers = list(self.layers.values())layers.reverse()for layer in layers:dout = layer.backward(dout)# 设定grads = {}for idx in range(1, self.hidden_layer_num+2):grads['W' + str(idx)] = self.layers['Affine' + str(idx)].dW + self.weight_decay_lambda * self.params['W' + str(idx)]grads['b' + str(idx)] = self.layers['Affine' + str(idx)].dbif self.use_batchnorm and idx != self.hidden_layer_num+1:grads['gamma' + str(idx)] = self.layers['BatchNorm' + str(idx)].dgammagrads['beta' + str(idx)] = self.layers['BatchNorm' + str(idx)].dbetareturn grads
在common文件夹下创建文件layers.py
添加Dropout类
完整layers.py如下:
# coding: utf-8
import numpy as np
from common.functions import *
from common.util import im2col, col2imclass Relu:def __init__(self):self.mask = Nonedef forward(self, x):self.mask = (x <= 0)out = x.copy()out[self.mask] = 0return outdef backward(self, dout):dout[self.mask] = 0dx = doutreturn dxclass Sigmoid:def __init__(self):self.out = Nonedef forward(self, x):out = sigmoid(x)self.out = outreturn outdef backward(self, dout):dx = dout * (1.0 - self.out) * self.outreturn dxclass Affine:def __init__(self, W, b):self.W =Wself.b = bself.x = Noneself.original_x_shape = None# 权重和偏置参数的导数self.dW = Noneself.db = Nonedef forward(self, x):# 对应张量self.original_x_shape = x.shapex = x.reshape(x.shape[0], -1)self.x = xout = np.dot(self.x, self.W) + self.breturn outdef backward(self, dout):dx = np.dot(dout, self.W.T)self.dW = np.dot(self.x.T, dout)self.db = np.sum(dout, axis=0)dx = dx.reshape(*self.original_x_shape) # 还原输入数据的形状(对应张量)return dxclass SoftmaxWithLoss:def __init__(self):self.loss = Noneself.y = None # softmax的输出self.t = None # 监督数据def forward(self, x, t):self.t = tself.y = softmax(x)self.loss = cross_entropy_error(self.y, self.t)return self.lossdef backward(self, dout=1):batch_size = self.t.shape[0]if self.t.size == self.y.size: # 监督数据是one-hot-vector的情况dx = (self.y - self.t) / batch_sizeelse:dx = self.y.copy()dx[np.arange(batch_size), self.t] -= 1dx = dx / batch_sizereturn dxclass Dropout:"""http://arxiv.org/abs/1207.0580"""def __init__(self, dropout_ratio=0.5):self.dropout_ratio = dropout_ratioself.mask = Nonedef forward(self, x, train_flg=True):if train_flg:self.mask = np.random.rand(*x.shape) > self.dropout_ratioreturn x * self.maskelse:return x * (1.0 - self.dropout_ratio)def backward(self, dout):return dout * self.maskclass BatchNormalization:"""http://arxiv.org/abs/1502.03167"""def __init__(self, gamma, beta, momentum=0.9, running_mean=None, running_var=None):self.gamma = gammaself.beta = betaself.momentum = momentumself.input_shape = None # Conv层的情况下为4维,全连接层的情况下为2维 # 测试时使用的平均值和方差self.running_mean = running_meanself.running_var = running_var # backward时使用的中间数据self.batch_size = Noneself.xc = Noneself.std = Noneself.dgamma = Noneself.dbeta = Nonedef forward(self, x, train_flg=True):self.input_shape = x.shapeif x.ndim != 2:N, C, H, W = x.shapex = x.reshape(N, -1)out = self.__forward(x, train_flg)return out.reshape(*self.input_shape)def __forward(self, x, train_flg):if self.running_mean is None:N, D = x.shapeself.running_mean = np.zeros(D)self.running_var = np.zeros(D)if train_flg:mu = x.mean(axis=0)xc = x - muvar = np.mean(xc**2, axis=0)std = np.sqrt(var + 10e-7)xn = xc / stdself.batch_size = x.shape[0]self.xc = xcself.xn = xnself.std = stdself.running_mean = self.momentum * self.running_mean + (1-self.momentum) * muself.running_var = self.momentum * self.running_var + (1-self.momentum) * var else:xc = x - self.running_meanxn = xc / ((np.sqrt(self.running_var + 10e-7)))out = self.gamma * xn + self.beta return outdef backward(self, dout):if dout.ndim != 2:N, C, H, W = dout.shapedout = dout.reshape(N, -1)dx = self.__backward(dout)dx = dx.reshape(*self.input_shape)return dxdef __backward(self, dout):dbeta = dout.sum(axis=0)dgamma = np.sum(self.xn * dout, axis=0)dxn = self.gamma * doutdxc = dxn / self.stddstd = -np.sum((dxn * self.xc) / (self.std * self.std), axis=0)dvar = 0.5 * dstd / self.stddxc += (2.0 / self.batch_size) * self.xc * dvardmu = np.sum(dxc, axis=0)dx = dxc - dmu / self.batch_sizeself.dgamma = dgammaself.dbeta = dbetareturn dxclass Convolution:def __init__(self, W, b, stride=1, pad=0):self.W = Wself.b = bself.stride = strideself.pad = pad# 中间数据(backward时使用)self.x = None self.col = Noneself.col_W = None# 权重和偏置参数的梯度self.dW = Noneself.db = Nonedef forward(self, x):FN, C, FH, FW = self.W.shapeN, C, H, W = x.shapeout_h = 1 + int((H + 2*self.pad - FH) / self.stride)out_w = 1 + int((W + 2*self.pad - FW) / self.stride)col = im2col(x, FH, FW, self.stride, self.pad)col_W = self.W.reshape(FN, -1).Tout = np.dot(col, col_W) + self.bout = out.reshape(N, out_h, out_w, -1).transpose(0, 3, 1, 2)self.x = xself.col = colself.col_W = col_Wreturn outdef backward(self, dout):FN, C, FH, FW = self.W.shapedout = dout.transpose(0,2,3,1).reshape(-1, FN)self.db = np.sum(dout, axis=0)self.dW = np.dot(self.col.T, dout)self.dW = self.dW.transpose(1, 0).reshape(FN, C, FH, FW)dcol = np.dot(dout, self.col_W.T)dx = col2im(dcol, self.x.shape, FH, FW, self.stride, self.pad)return dxclass Pooling:def __init__(self, pool_h, pool_w, stride=1, pad=0):self.pool_h = pool_hself.pool_w = pool_wself.stride = strideself.pad = padself.x = Noneself.arg_max = Nonedef forward(self, x):N, C, H, W = x.shapeout_h = int(1 + (H - self.pool_h) / self.stride)out_w = int(1 + (W - self.pool_w) / self.stride)col = im2col(x, self.pool_h, self.pool_w, self.stride, self.pad)col = col.reshape(-1, self.pool_h*self.pool_w)arg_max = np.argmax(col, axis=1)out = np.max(col, axis=1)out = out.reshape(N, out_h, out_w, C).transpose(0, 3, 1, 2)self.x = xself.arg_max = arg_maxreturn outdef backward(self, dout):dout = dout.transpose(0, 2, 3, 1)pool_size = self.pool_h * self.pool_wdmax = np.zeros((dout.size, pool_size))dmax[np.arange(self.arg_max.size), self.arg_max.flatten()] = dout.flatten()dmax = dmax.reshape(dout.shape + (pool_size,)) dcol = dmax.reshape(dmax.shape[0] * dmax.shape[1] * dmax.shape[2], -1)dx = col2im(dcol, self.x.shape, self.pool_h, self.pool_w, self.stride, self.pad)return dx
common下创建文件trainer.py
添加代码如下:
# coding: utf-8
import sys, os
sys.path.append(os.pardir) # 为了导入父目录的文件而进行的设定
import numpy as np
from common.optimizer import *class Trainer:"""进行神经网络的训练的类"""def __init__(self, network, x_train, t_train, x_test, t_test,epochs=20, mini_batch_size=100,optimizer='SGD', optimizer_param={'lr':0.01}, evaluate_sample_num_per_epoch=None, verbose=True):self.network = networkself.verbose = verboseself.x_train = x_trainself.t_train = t_trainself.x_test = x_testself.t_test = t_testself.epochs = epochsself.batch_size = mini_batch_sizeself.evaluate_sample_num_per_epoch = evaluate_sample_num_per_epoch# optimzeroptimizer_class_dict = {'sgd':SGD, 'momentum':Momentum, 'nesterov':Nesterov,'adagrad':AdaGrad, 'rmsprpo':RMSprop, 'adam':Adam}self.optimizer = optimizer_class_dict[optimizer.lower()](**optimizer_param)self.train_size = x_train.shape[0]self.iter_per_epoch = max(self.train_size / mini_batch_size, 1)self.max_iter = int(epochs * self.iter_per_epoch)self.current_iter = 0self.current_epoch = 0self.train_loss_list = []self.train_acc_list = []self.test_acc_list = []def train_step(self):batch_mask = np.random.choice(self.train_size, self.batch_size)x_batch = self.x_train[batch_mask]t_batch = self.t_train[batch_mask]grads = self.network.gradient(x_batch, t_batch)self.optimizer.update(self.network.params, grads)loss = self.network.loss(x_batch, t_batch)self.train_loss_list.append(loss)if self.verbose: print("train loss:" + str(loss))if self.current_iter % self.iter_per_epoch == 0:self.current_epoch += 1x_train_sample, t_train_sample = self.x_train, self.t_trainx_test_sample, t_test_sample = self.x_test, self.t_testif not self.evaluate_sample_num_per_epoch is None:t = self.evaluate_sample_num_per_epochx_train_sample, t_train_sample = self.x_train[:t], self.t_train[:t]x_test_sample, t_test_sample = self.x_test[:t], self.t_test[:t]train_acc = self.network.accuracy(x_train_sample, t_train_sample)test_acc = self.network.accuracy(x_test_sample, t_test_sample)self.train_acc_list.append(train_acc)self.test_acc_list.append(test_acc)if self.verbose: print("=== epoch:" + str(self.current_epoch) + ", train acc:" + str(train_acc) + ", test acc:" + str(test_acc) + " ===")self.current_iter += 1def train(self):for i in range(self.max_iter):self.train_step()test_acc = self.network.accuracy(self.x_test, self.t_test)if self.verbose:print("=============== Final Test Accuracy ===============")print("test acc:" + str(test_acc))
正向传播时传递了信号的神经元,反向传播时按原样传递信号;正向传播时没有传递信号的神经元,反向传播时信号将停在那里。
使用MNIST数据集进行验证
创建文件overfit_dropout.py
添加代码如下:
# coding: utf-8
import os
import sys
sys.path.append(os.pardir) # 为了导入父目录的文件而进行的设定
import numpy as np
import matplotlib.pyplot as plt
from dataset.mnist import load_mnist
from common.multi_layer_net_extend import MultiLayerNetExtend
from common.trainer import Trainer(x_train, t_train), (x_test, t_test) = load_mnist(normalize=True)# 为了再现过拟合,减少学习数据
x_train = x_train[:300]
t_train = t_train[:300]# 设定是否使用Dropuout,以及比例 ========================
use_dropout = True # 不使用Dropout的情况下为False
dropout_ratio = 0.2
# ====================================================network = MultiLayerNetExtend(input_size=784, hidden_size_list=[100, 100, 100, 100, 100, 100],output_size=10, use_dropout=use_dropout, dropout_ration=dropout_ratio)
trainer = Trainer(network, x_train, t_train, x_test, t_test,epochs=301, mini_batch_size=100,optimizer='sgd', optimizer_param={'lr': 0.01}, verbose=True)
trainer.train()train_acc_list, test_acc_list = trainer.train_acc_list, trainer.test_acc_list# 绘制图形==========
markers = {'train': 'o', 'test': 's'}
x = np.arange(len(train_acc_list))
plt.plot(x, train_acc_list, marker='o', label='train', markevery=10)
plt.plot(x, test_acc_list, marker='s', label='test', markevery=10)
plt.xlabel("epochs")
plt.ylabel("accuracy")
plt.ylim(0, 1.0)
plt.legend(loc='lower right')
plt.show()
运行结果:
通过使用Dropout,训练数据和测试数据的识别精度的差距变小了。并且,训练数据也没有到达100%的识别精度。像这样,通过使用Dropout,即便是表现力强的网络,也可以抑制过拟合。
4)集成学习
机器学习中经常使用集成学习。所谓集成学习,就是让多个模型单独进行学习,推理时再取多个模型的输出的平均值。用神经网络的语境来说,比如,准备5个结构相同(或者类似)的网络,分别进行学习,测试时,以这5个网络的输出的平均值作为答案。
通过进行集成学习,神经网络的识别精度可以提高好几个百分点。
5.超参数的验证
超参数是指,比如各层的神经元数量、batch大小、参数更新时的学习率或权值衰减等。如果这些超参数没有设置合适的值,模型的性能就会很差。虽然超参数的取值非常重要,但是在决定超参数的过程中一般会伴随很多的试错。
1)验证数据
不能使用测试数据评估超参数的性能。
因为如果使用测试数据调整超参数,超参数的值会对测试数据发生过拟合。换句话说,用测试数据确认超参数的值的“好坏”,就会导致超参数的值被调整为只拟合测试数据。这样的话,可能就会得到不能拟合其他数据、泛化能力低的模型。
因此,调整超参数时,必须使用超参数专用的确认数据。用于调整超参数的数据,一般称为验证数据
训练数据用于参数(权重和偏置)的学习,验证数据用于超参数的性能评估。
2)超参数的最优化
进行超参数的最优化时,逐渐缩小超参数的“好值”的存在范围非常重要。所谓逐渐缩小范围,是指一开始先大致设定一个范围,从这个范围中随机选出一个超参数(采样),用这个采样到的值进行识别精度的评估;然后,多次重复该操作,观察识别精度的结果,根据这个结果缩小超参数的“好值”的范围。通过重复这一操作,就可以逐渐确定超参数的合适范围。
3)步骤
步骤0
设定超参数的范围。
步骤1
从设定的超参数范围中随机采样。
步骤2
使用步骤1中采样到的超参数的值进行学习,通过验证数据评估识别精度(但是要将epoch设置得很小)。
步骤3
重复步骤1和步骤2(100次等),根据它们的识别精度的结果,缩小超参数的范围。
在超参数的最优化中,如果需要更精炼的方法,可以使用贝叶斯最优化
4)超参数最优化的实现
创建文件hyperparameter_optimization.py
添加代码如下:
# coding: utf-8
import sys, os
sys.path.append(os.pardir) # 为了导入父目录的文件而进行的设定
import numpy as np
import matplotlib.pyplot as plt
from dataset.mnist import load_mnist
from common.multi_layer_net import MultiLayerNet
from common.util import shuffle_dataset
from common.trainer import Trainer(x_train, t_train), (x_test, t_test) = load_mnist(normalize=True)# 为了实现高速化,减少训练数据
x_train = x_train[:500]
t_train = t_train[:500]# 分割验证数据
validation_rate = 0.20
validation_num = x_train.shape[0] * validation_rate
x_train, t_train = shuffle_dataset(x_train, t_train)
# x_val = x_train[:validation_num]
# t_val = t_train[:validation_num]
# x_train = x_train[validation_num:]
# t_train = t_train[validation_num:]x_val = x_train[:int(validation_num)]
t_val = t_train[:int(validation_num)]
x_train = x_train[int(validation_num):]
t_train = t_train[int(validation_num):]def __train(lr, weight_decay, epocs=50):network = MultiLayerNet(input_size=784, hidden_size_list=[100, 100, 100, 100, 100, 100],output_size=10, weight_decay_lambda=weight_decay)trainer = Trainer(network, x_train, t_train, x_val, t_val,epochs=epocs, mini_batch_size=100,optimizer='sgd', optimizer_param={'lr': lr}, verbose=False)trainer.train()return trainer.test_acc_list, trainer.train_acc_list# 超参数的随机搜索======================================
optimization_trial = 100
results_val = {}
results_train = {}
for _ in range(optimization_trial):# 指定搜索的超参数的范围===============weight_decay = 10 ** np.random.uniform(-8, -4)lr = 10 ** np.random.uniform(-6, -2)# ================================================val_acc_list, train_acc_list = __train(lr, weight_decay)print("val acc:" + str(val_acc_list[-1]) + " | lr:" + str(lr) + ", weight decay:" + str(weight_decay))key = "lr:" + str(lr) + ", weight decay:" + str(weight_decay)results_val[key] = val_acc_listresults_train[key] = train_acc_list# 绘制图形========================================================
print("=========== Hyper-Parameter Optimization Result ===========")
graph_draw_num = 20
col_num = 5
row_num = int(np.ceil(graph_draw_num / col_num))
i = 0for key, val_acc_list in sorted(results_val.items(), key=lambda x:x[1][-1], reverse=True):print("Best-" + str(i+1) + "(val acc:" + str(val_acc_list[-1]) + ") | " + key)plt.subplot(row_num, col_num, i+1)plt.title("Best-" + str(i+1))plt.ylim(0.0, 1.0)if i % 5: plt.yticks([])plt.xticks([])x = np.arange(len(val_acc_list))plt.plot(x, val_acc_list)plt.plot(x, results_train[key], "--")i += 1if i >= graph_draw_num:breakplt.show()
运行结果:
学习率在0.001到0.01、权值衰减系数在10−8到10−6之间时,学习可以顺利进行。像这样,观察可以使学习顺利进行的超参数的范围,从而缩小值的范围。然后,在这个缩小的范围中重复相同的操作。学习率在0.001到0.01、权值衰减系数在10−8到10−6之间时,学习可以顺利进行。像这样,观察可以使学习顺利进行的超参数的范围,从而缩小值的范围。然后,在这个缩小的范围中重复相同的操作。