Question for example: Train Classification Network to Classify Object in 3-D Point Cloud, how to evaluate data?

Hello,

I am new to using neural networks and have a question about a tutorial "https://de.mathworks.com/help/vision/ug/train-classification-network-to-classify-object-in-3-d-point-cloud.html#TrainNetwork3DPointCloudDataExample-2".

I have been working through this example and would like to know how I can evaluate data with a test set well after the network has been trained?

I know I can use the command "predict(net,data)" or "classify(net,data)" but the data in this tutorial is stored in fileDatastores. But what about if I want to evaluate the network with a single test file, i.e. a pointCloud, from the workspace? Because then, as far as I know, I can't use fileDatastore because fileDatastore asks for a saved file on disk?

Thanks for your help

Answer 1

[Sai Pavan]

Hi Kr,

I understand that you are trying to evaluate a trained model on a single test file loaded into the base workspace.

If you want to evaluate a trained network with a single test point cloud from the workspace in MATLAB, you can create a custom datastore using the “matlab.io.datastore.MiniBatchable” class. This allows you to load the point cloud data directly from the workspace and feed it into the network for evaluation.

Please refer to the below code snippet to learn how to create a custom datastore and evaluate the network with a single point cloud:

customDatastore = matlab.io.Datastore(@() testPointCloud, 'ReadFcn', @identityFcn); % Create a custom datastore 
function data = identityFcn() 
    data = evalin('base', 'testPointCloud'); % Define the identity function to return the point cloud 
end
customDatastore.MiniBatchSize = 1; % Set the 'MiniBatchSize' property to 1 to process one point cloud at a time 
while hasdata(customDatastore) % Loop over the point clouds in the custom datastore and evaluate the network 
    pointCloud = read(customDatastore); % Read the next point cloud from the custom datastore 
    predictedLabels = classify(net, pointCloud); % Evaluate the network on the point cloud 
end

Please refer to the below documentation to learn more about the “matlab.io.datastore.Minibatchable” class and understand the workflow with an example:

https://www.mathworks.com/help/deeplearning/ref/matlab.io.datastore.minibatchable-class.html

Hope it helps.

Regards,

Sai Pavan