深度学习中的优化算法之AdaMax

Posted fengbingchun

tags:

篇首语:本文由小常识网(cha138.com)小编为大家整理,主要介绍了深度学习中的优化算法之AdaMax相关的知识,希望对你有一定的参考价值。

      之前在https://blog.csdn.net/fengbingchun/article/details/125018001 介绍过深度学习中的优化算法Adam,这里介绍下深度学习的另一种优化算法AdaMax。AdaMax与Adam来自于同一篇论文。论文名字为《ADAM: A METHOD FOR STOCHASTIC OPTIMIZATION》,论文地址:https://arxiv.org/pdf/1412.6980.pdf

      AdaMax:是梯度优化算法的扩展,基于无穷范数的Adam的变体(a variant of Adam based on the infinity norm)。此算法对学习率的上限提供了一个更简单的范围,并可能对某些问题进行更有效的优化。如下图所示,截图来自:https://arxiv.org/pdf/1609.04747.pdf

      AdaMax与Adam区别:本质上前者是将L2范数推广到L-infinity范数。AdaMax与Adam最终公式中仅分母的计算方式不同,AdaMax使用公式24,Adam使用公式20。

以下是与Adam不同的代码片段:

      1.在原有枚举类Optimizaiton的基础上新增AdaMax

enum class Optimization 
	BGD, // Batch Gradient Descent
	SGD, // Stochastic Gradient Descent
	MBGD, // Mini-batch Gradient Descent
	SGD_Momentum, // SGD with Momentum
	AdaGrad, // Adaptive Gradient
	RMSProp, // Root Mean Square Propagation
	Adadelta, // an adaptive learning rate method
	Adam, // Adaptive Moment Estimation
	AdaMax // a variant of Adam based on the infinity norm
;

      2.calculate_gradient_descent函数:

void LogisticRegression2::calculate_gradient_descent(int start, int end)

	switch (optim_) 
		case Optimization::AdaMax: 
			int len = end - start;
			std::vector<float> m(feature_length_, 0.), u(feature_length_, 1e-8), mhat(feature_length_, 0.);
			std::vector<float> z(len, 0.), dz(len, 0.);
			float beta1t = 1.;
			for (int i = start, x = 0; i < end; ++i, ++x) 
				z[x] = calculate_z(data_->samples[random_shuffle_[i]]);
				dz[x] = calculate_loss_function_derivative(calculate_activation_function(z[x]), data_->labels[random_shuffle_[i]]);

				beta1t *= beta1_;

				for (int j = 0; j < feature_length_; ++j) 
					float dw = data_->samples[random_shuffle_[i]][j] * dz[x];
					m[j] = beta1_ * m[j] + (1. - beta1_) * dw; // formula 19
					u[j] = std::max(beta2_ * u[j], std::fabs(dw)); // formula 24

					mhat[j] = m[j] / (1. - beta1t); // formula 20

					// Note: need to ensure than u[j] cannot be 0.
					// (1). u[j] is initialized to 1e-8, or
					// (2). if u[j] is initialized to 0., then u[j] adjusts to (u[j] + 1e-8)
					w_[j] = w_[j] - alpha_ * mhat[j] / u[j]; // formula 25
				

				b_ -= (alpha_ * dz[x]);
			
		
			break;
		case Optimization::Adam: 
			int len = end - start;
			std::vector<float> m(feature_length_, 0.), v(feature_length_, 0.), mhat(feature_length_, 0.), vhat(feature_length_, 0.);
			std::vector<float> z(len, 0.), dz(len, 0.);
			float beta1t = 1., beta2t = 1.;
			for (int i = start, x = 0; i < end; ++i, ++x) 
				z[x] = calculate_z(data_->samples[random_shuffle_[i]]);
				dz[x] = calculate_loss_function_derivative(calculate_activation_function(z[x]), data_->labels[random_shuffle_[i]]);

				beta1t *= beta1_;
				beta2t *= beta2_;

				for (int j = 0; j < feature_length_; ++j) 
					float dw = data_->samples[random_shuffle_[i]][j] * dz[x];
					m[j] = beta1_ * m[j] + (1. - beta1_) * dw; // formula 19
					v[j] = beta2_ * v[j] + (1. - beta2_) * (dw * dw); // formula 19

					mhat[j] = m[j] / (1. - beta1t); // formula 20
					vhat[j] = v[j] / (1. - beta2t); // formula 20

					w_[j] = w_[j] - alpha_ * mhat[j] / (std::sqrt(vhat[j]) + eps_); // formula 21
				

				b_ -= (alpha_ * dz[x]);
			
		
			break;
		case Optimization::Adadelta: 
			int len = end - start;
			std::vector<float> g(feature_length_, 0.), p(feature_length_, 0.);
			std::vector<float> z(len, 0.), dz(len, 0.);
			for (int i = start, x = 0; i < end; ++i, ++x) 
				z[x] = calculate_z(data_->samples[random_shuffle_[i]]);
				dz[x] = calculate_loss_function_derivative(calculate_activation_function(z[x]), data_->labels[random_shuffle_[i]]);

				for (int j = 0; j < feature_length_; ++j) 
					float dw = data_->samples[random_shuffle_[i]][j] * dz[x];
					g[j] = mu_ * g[j] + (1. - mu_) * (dw * dw); // formula 10

					float alpha = (eps_ + std::sqrt(p[j])) / (eps_ + std::sqrt(g[j]));
					float change = alpha * dw;
					p[j] = mu_ * p[j] +  (1. - mu_) * (change * change); // formula 15

					w_[j] = w_[j] - change;
				

				b_ -= (eps_ * dz[x]);
			
		
			break;
		case Optimization::RMSProp: 
			int len = end - start;
			std::vector<float> g(feature_length_, 0.);
			std::vector<float> z(len, 0), dz(len, 0);
			for (int i = start, x = 0; i < end; ++i, ++x) 
				z[x] = calculate_z(data_->samples[random_shuffle_[i]]);
				dz[x] = calculate_loss_function_derivative(calculate_activation_function(z[x]), data_->labels[random_shuffle_[i]]);

				for (int j = 0; j < feature_length_; ++j) 
					float dw = data_->samples[random_shuffle_[i]][j] * dz[x];
					g[j] = mu_ * g[j] + (1. - mu_) * (dw * dw); // formula 18
					w_[j] = w_[j] - alpha_ * dw / (std::sqrt(g[j]) + eps_);
				

				b_ -= (alpha_ * dz[x]);
			
		
			break;
		case Optimization::AdaGrad: 
			int len = end - start;
			std::vector<float> g(feature_length_, 0.);
			std::vector<float> z(len, 0), dz(len, 0);
			for (int i = start, x = 0; i < end; ++i, ++x) 
				z[x] = calculate_z(data_->samples[random_shuffle_[i]]);
				dz[x] = calculate_loss_function_derivative(calculate_activation_function(z[x]), data_->labels[random_shuffle_[i]]);

				for (int j = 0; j < feature_length_; ++j) 
					float dw = data_->samples[random_shuffle_[i]][j] * dz[x];
					g[j] += dw * dw;
					w_[j] = w_[j] - alpha_ * dw / (std::sqrt(g[j]) + eps_);
				

				b_ -= (alpha_ * dz[x]);
			
		
			break;
		case Optimization::SGD_Momentum: 
			int len = end - start;
			std::vector<float> change(feature_length_, 0.);
			std::vector<float> z(len, 0), dz(len, 0);
			for (int i = start, x = 0; i < end; ++i, ++x) 
				z[x] = calculate_z(data_->samples[random_shuffle_[i]]);
				dz[x] = calculate_loss_function_derivative(calculate_activation_function(z[x]), data_->labels[random_shuffle_[i]]);

				for (int j = 0; j < feature_length_; ++j) 
					float new_change = mu_ * change[j] - alpha_ * (data_->samples[random_shuffle_[i]][j] * dz[x]);
					w_[j] += new_change;
					change[j] = new_change;
				

				b_ -= (alpha_ * dz[x]);
			
		
			break;
		case Optimization::SGD:
		case Optimization::MBGD: 
			int len = end - start;
			std::vector<float> z(len, 0), dz(len, 0);
			for (int i = start, x = 0; i < end; ++i, ++x) 
				z[x] = calculate_z(data_->samples[random_shuffle_[i]]);
				dz[x] = calculate_loss_function_derivative(calculate_activation_function(z[x]), data_->labels[random_shuffle_[i]]);

				for (int j = 0; j < feature_length_; ++j) 
					w_[j] = w_[j] - alpha_ * (data_->samples[random_shuffle_[i]][j] * dz[x]);
				

				b_ -= (alpha_ * dz[x]);
			
		
			break;
		case Optimization::BGD:
		default: // BGD
			std::vector<float> z(m_, 0), dz(m_, 0);
			float db = 0.;
			std::vector<float> dw(feature_length_, 0.);
			for (int i = 0; i < m_; ++i) 
				z[i] = calculate_z(data_->samples[i]);
				o_[i] = calculate_activation_function(z[i]);
				dz[i] = calculate_loss_function_derivative(o_[i], data_->labels[i]);

				for (int j = 0; j < feature_length_; ++j) 
					dw[j] += data_->samples[i][j] * dz[i]; // dw(i)+=x(i)(j)*dz(i)
				
				db += dz[i]; // db+=dz(i)
			

			for (int j = 0; j < feature_length_; ++j) 
				dw[j] /= m_;
				w_[j] -= alpha_ * dw[j];
			

			b_ -= alpha_*(db/m_);
	

      执行结果如下图所示:测试函数为test_logistic_regression2_gradient_descent,多次执行每种配置,最终结果都相同。图像集使用MNIST,其中训练图像总共10000张,0和1各5000张,均来自于训练集;预测图像总共1800张,0和1各900张,均来自于测试集。Adam和AdaMax配置参数相同的情况下,即eps为1e-8,学习率为0.002,beta1为0.9,beta2为0.999的情况下,Adam耗时30秒,AdaMax耗时为25秒;它们的识别率均为100%

      GitHub: https://github.com/fengbingchun/NN_Test

以上是关于深度学习中的优化算法之AdaMax的主要内容,如果未能解决你的问题,请参考以下文章

深度学习最全优化方法总结比较(SGD,Adagrad,Adadelta,Adam,Adamax,Nadam)

深度学习中的优化算法之Adadelta

深度学习中的优化算法之Nadam

深度学习中的优化算法之RMSProp

深度学习中的优化算法之Adam

深度学习中的优化算法之NAG