Use only the schema without using pre-trained weights

I want to use RESnet-50 without using pre-trained weights by deep learning Toolbox,how can i do?

Answer 1

[Sameer]

Hi Wenlin

From my understanding, you want to use the "ResNet-50" architecture in MATLAB's "Deep Learning Toolbox" without utilizing pre-trained weights.

This involves defining the "ResNet-50" model architecture from scratch, initializing it with random weights, and then training it on your dataset. Here's how you can achieve this:

1. Define the ResNet-50 Architecture: You need to manually define the "ResNet-50" architecture, which involves specifying the layers and how they connect. MATLAB allows you to define custom layers and networks.

2. Initialize Weights: By default, MATLAB initializes the weights randomly when you define the layers from scratch, so you won't be using any pre-trained weights in this setup.

3. Train the Network: Once the architecture is defined, you can train the network using your dataset with the "trainNetwork" function.

Here’s a basic outline:

% Define the layers of ResNet-50
layers = [
    imageInputLayer([224 224 3],"Name","input")
    convolution2dLayer(7,64,'Stride',2,'Padding','same','Name','conv1')
    batchNormalizationLayer('Name','bn_conv1')
    reluLayer('Name','conv1_relu')
    maxPooling2dLayer(3,'Stride',2,'Padding','same','Name','pool1')
    % Add more layers according to the ResNet-50 architecture
    % This includes multiple residual blocks
    fullyConnectedLayer(numClasses,'Name','fc')
    softmaxLayer('Name','softmax')
    classificationLayer('Name','output')
    ];
% Create a layer graph from the layers
lgraph = layerGraph(layers);
% Add connections for ResNet-50 (skip connections)
% Use the addLayers and connectLayers functions to create the full graph
% Define training options
options = trainingOptions('sgdm', ...
    'MaxEpochs',30, ...
    'InitialLearnRate',0.01, ...
    'Verbose',true, ...
    'Plots','training-progress');
% Load your dataset
% [trainImages, trainLabels] = loadYourDataFunction();
% Train the network
% net = trainNetwork(trainImages, trainLabels, lgraph, options);

Hope this helps!