memory error when compiling fitcensemble model using coder
显示 更早的评论
Hello,
I am trying to increase the speed of a random forest machine learning model I have trained in Matlab using the "fitcensemble" method by compiling it to C code. However, when I use the Coder app, the memory during the "Generating Trial Code" step continues to grow until Matlab crashes. I have 64GB of RAM while the model file size is 92MB so hardware is unlikely the problem. (Model could not be included due to file size restrictions).
I have attached a zip file with the project which recapitulates the error.
Briefly,
I first fit the model
t = templateTree('MinLeafSize',2);
Mdl = fitcensemble(train,train_lab,'Learners',t,'Method','Bag','Learners',t,'CrossVal','on','NumLearningCycles',50);
Where train is a 2D array with 18 feature columns and train_lab is a 1D array with labels (3 classes).
Next I save the best model from the cross-validation for the coder
saveLearnerForCoder(Mdl.Trained{7},'072920_bagged50Tree')
Then I generate two scripts (both in zip file)
myBagEnsemblePredict.m
function [label,score] = myBagEnsemblePredict(X,fileName) %#codegen
CompactMdl = loadLearnerForCoder(fileName);
[label,score] = predict(CompactMdl,X);
end
and
test_myBagEnsemblePredict.m
load('test_data_coder.mat')
[label,score] = myBagEnsemblePredict(test,'072920_bagged50Tree');
These are used as input for the Coder app.
When I start the coder App, I specify "myBagEnsemblePredict" as the entry-point function
My input types are autodefined from test_myBagEnsemblePredict.m
and edited as above
When I generate the code, the memory increases over time until matlab crashes.
Here are my memory settings in the coder
The complete configuration is in the attached zip file ("coder_settings.mat")
Other potentially relevant information:
Running Matlab 2020a on a 64-bit Linux machine (Ubuntu). I have tried changing the X variable to have the size (double(:100X18)) and observed the same memory error. I have tried shrinking the model by using "MaxNumSplit" option in the fitcensemble step. I have tried reducing the Dynamic memory allocation threshold to 1000.
Thank you in advance for your time and help.
1 个评论
Raunak Gupta
2020-8-8
Hi Christian,
I don't see the trained model file 072920_bagged50Tree.mat in the code_reproduce_error folder. It will be really helpful if you attach that too, for reproducing the error at my end.
回答（0 个）
类别
在 帮助中心 和 File Exchange 中查找有关 Classification Ensembles 的更多信息
2020-8-3
2020-8-8
Community Treasure Hunt
Find the treasures in MATLAB Central and discover how the community can help you!
Start Hunting!
Translated by 
