Serializing a model on PC and Deserializing it on Android

[SaraKhalifa]

Part of my research work is to classify sensor data on Android using Multilayer Perceptron (MLP) classifier.

I used WekaSTRIPPED to train (build) the MLP classifier on the phone and save the trained model on the sdcard by serializing it. This works fine in a small scale problem even with large training data. I mean by small scale problem that the MLP model to be trained consists of about 10 inputs (features) and 4 outputs (classes). However, when the MLP model increases to have 25 inputs and 10 outputs the building process fails on the phone. I tried to increase the heap size to 1024 m but unfortunately this doesn’t solve the problem. I also tried to build the classifier on a virtual machine but again this doesn’t solve the problem.

Another trial was to train the model using WEKA on the PC and then serialize the model to a specific place on the machine. I copied this saved model from the pc and pasted it on the sdcard of the phone. Then I deserialized this model using Weka for Android but unfortunately, the deserialization of the model fails.

My question here is do I really need to train (build) my classifier on the phone specially that my training data is supposed to be very large and even the model to be trained itself is going to be more and more complex?

Any help is greatly appreciated.

Thanks.

[georgikoemdzhiev]

Try serialising the model on a PC using the same version of WEKA you have on the phone (I believe it is Weka 3.7.3) I was able to serialise my model on PC and deserialise it successfully in my android app. :)

[feliperpv]

@georgikoemdzhiev can you show your code? I have problems with it and i'm using the same version of weka (3.8.2)

[georgikoemdzhiev]

@feliperpv take a look at my project here => https://github.com/georgikoemdzhiev/ActiveMinutes
You want to check this file specifically => https://github.com/georgikoemdzhiev/ActiveMinutes/blob/master/app/src/main/java/georgikoemdzhiev/activeminutes/data_layer/FileManager.java

[feliperpv]

@georgikoemdzhiev Thanks for the code. I noticed you don't work with a .model file extension, you use .data, what is the difference? Because I serialise my model on my PC with .model extension and I need to deserialize it on my app. In other words I have the .arff file only on my PC. In other words, I have the .arff file only on my PC and from it I generate my .model file

[georgikoemdzhiev]

@feliperpv I can't remember why I used .data extension, unfortunately. It's been a while since my Thesis project. However, I also trained/created my model from a .arff file and then serialised it on PC.