How to train Network Using Custom Training Loop for Semantic segmentation?

Hi, because I want to use two outputs for Semantic segmentation, so I have to train network using custom training loop. My question is how to achieve this?

Answer 1

[Aneela]

Hi zhou,

Training a neural network with two outputs for semantic segmentation using a custom training loop involves the following key steps:

• Define a network architecture that has two outputs. Here’s a sample network architecture with two output layers.

    layers = [
        imageInputLayer([256 256 3],'Name','input')
        convolution2dLayer(3,64,'Padding','same','Name','conv1')
        reluLayer('Name','relu1')
        %numClasses1, numClasses2-Replace them with the number of classes in your datasets
        convolution2dLayer(1, numClasses1, 'Name', 'convOut1')
        softmaxLayer('Name','softmax1')
        pixelClassificationLayer('Name','output1')
        convolution2dLayer(1, numClasses2, 'Name', 'convOut2')
        softmaxLayer('Name','softmax2')
        pixelClassificationLayer('Name','output2')];

• Define a custom loss function that computes the loss for both outputs.
• Use a custom training loop to train the model and compute loss. Refer to the following link for training network using a custom training loop: https://www.mathworks.com/help/deeplearning/ug/train-network-using-custom-training-loop.html
• Periodically evaluate the network on a validation set to monitor performance.

Hope this helps!