如何清除MATLAB中Lambda函数内的Persistent变量以重置训练绘图

问题场景 #

在使用MATLAB的trainNetwork训练神经网络时，通过Lambda函数将plot_train_loss作为OutputFcn绘制训练进度，首次训练效果正常，但再次训练时，函数内的persistent变量（train_iteration、train_loss等）会保留上一次训练的数据，导致绘图初始显示历史训练记录。

训练配置代码 #

options = trainingOptions('sgdm', ...
    'MaxEpochs',num_epochs,...
    'InitialLearnRate',learning_rate, ...
    'ValidationData',imdsTest, ...
    'ValidationFrequency',validationFrequency, ...
    'Verbose',false, ...
    'Plots','none',...
    'OutputFcn',@(info)plot_train_loss(info,3), ...
    'MiniBatchSize', mini_batch_size);


[net,info] = trainNetwork(imdsTrain,my_net,options);

plot_train_loss核心代码片段 #

function stop  = plot_train_loss(info,N)

% initialize persitent variable in order to save the past values
persistent train_iteration
persistent train_loss
persistent train_accuracy
% Keep track of the best validation accuracy and the number of validations for which
% there has not been an improvement of the accuracy.
persistent bestValAccuracy
persistent valLag

global training_figure;
global num_iterations;

stop = false;

% check in which state the training is
% if start: do nothing
if info.State == "start"
    info.Iteration  = {};
    info.TrainingLoss = {};
    info.TrainingAccuracy = {};
    bestValAccuracy = 0;
    valLag = 0;
    return
end

% assign values of training to variable
train_iteration(info.Iteration) = info.Iteration;
train_loss(info.Iteration) = info.TrainingLoss;
train_accuracy(info.Iteration) = info.TrainingAccuracy;

% plot training progress
t = tiledlayout(2,1);

nexttile;
plot(train_iteration,train_loss,'g');
xlim([0 num_iterations]);
max_train_loss = max(train_loss(:)) + 0.2;
ylim([0 max_train_loss]);
grid on;
title('Training loss');

nexttile;
plot(train_iteration,train_accuracy,'b');
xlim([0 num_iterations]);
ylim([0 100]);
grid on;
title('Training accuracy');

解决方法 #

核心思路是利用训练启动时的info.State == "start"状态，在此阶段重置所有persistent变量，保证每次新训练开始时变量都是空的，同一次训练周期内变量持续保留数据。

修改后的plot_train_loss关键代码 #

function stop  = plot_train_loss(info,N)

% 声明持久化变量
persistent train_iteration
persistent train_loss
persistent train_accuracy
persistent bestValAccuracy
persistent valLag

global training_figure;
global num_iterations;

stop = false;

% 训练开始时重置所有持久化变量
if info.State == "start"
    % 清空所有persistent变量，为新训练初始化
    train_iteration = [];
    train_loss = [];
    train_accuracy = [];
    bestValAccuracy = 0;
    valLag = 0;
    return
end

% 赋值训练数据到持久化变量
train_iteration(info.Iteration) = info.Iteration;
train_loss(info.Iteration) = info.TrainingLoss;
train_accuracy(info.Iteration) = info.TrainingAccuracy;

% 绘图逻辑保持不变
t = tiledlayout(2,1);

nexttile;
plot(train_iteration,train_loss,'g');
xlim([0 num_iterations]);
max_train_loss = max(train_loss(:)) + 0.2;
ylim([0 max_train_loss]);
grid on;
title('Training loss');

nexttile;
plot(train_iteration,train_accuracy,'b');
xlim([0 num_iterations]);
ylim([0 100]);
grid on;
title('Training accuracy');

额外说明 #

• 原代码中修改info.Iteration等输入参数属性是无效的，MATLAB函数参数默认按值传递，修改不会影响训练过程的原始info对象，因此直接删除这部分冗余代码即可。
• 若需要在训练前强制清除函数的持久化变量，可在调用trainNetwork前执行clear plot_train_loss，但这种方法灵活性不足，优先推荐在函数内部利用start状态处理。

内容的提问来源于stack exchange，提问作者beinando