how to initialize the neural network to a set of weights ???
33 views (last 30 days)
Show older comments
I created my NN with patternet ??
0 Comments
Accepted Answer
Greg Heath
on 20 Oct 2012
Unlike the older nets (e.g., newfit, newpr, newff,...), you cannot assign weights to the newer networks (e.g., fitnet, patternnet, feedforwardnet,...) unless the net is configured.
There are two ways to configure the net before manually assigning your own initial weights. Both will assign initial weights that you can overwrite:
1. help/doc configure.
net = configure(net, x, t );
2. Train the net for 1 epoch
net.trainParam.epochs= 1.
net = train(net,x,t);
Hope this helps.
Thank you for formally accepting my answer.
Greg
2 Comments
Samisam
on 7 Jan 2018
@Greg Heath can I do a manual weight initialization before I train the net???
I mean if I have an optimal weight from a spesific algorithm and I want to create a NN to test data using these weights is there any way to do this without training the net again??
More Answers (3)
Greg Heath
on 19 Oct 2012
Edited: Greg Heath
on 20 Oct 2012
net = patternet;
will default to H = 10 hidden nodes. For other values use
net = patternnet(H);
If
size(input) = [I N ]
size(target) = [O N ]
the node topology is I-H-O.
For a manual weight initialization, first configure the net:
net = configure(net,x,t);
For a random weight initialization, initialize the random number generator. Then generate and assign the weights:
rng(0)
IW = 0.01*randn(H,I);
b1 = 0.01*randn(H,1);
LW = 0.01*randn(O,H);
b2 = 0.01*randn(O,1);
then
net.IW{1,1} = IW;
net.b{1,1} = b1;
net.LW{2,1} = LW;
net.b{2,1} = b2;
Hope this helps.
Thank you for formally accepting my answer.
Greg
4 Comments
LukasJ
on 6 Nov 2020
Dear Greg Heath,
unfortunately configuring the net doesn't do the trick for me:
I tried setting the inital weights manually e.g.
net.iw{1,1} = zeros(...
and via
net.initFcn = 'initlay';
net.layers{1,1}.initFcn = 'initwb';
net.layers{2,1}.initFcn = 'initwb';
net.InputWeights{1,1}.initFcn = 'midpoint';
net.LayerWeights{2,1}.initFcn = 'midpoint';
initFcn to call for midpoint initialization. The first won't update any weights after training, the former won't do anything (still random weights when I check before training, training results after fixed epochs are not comparable).
Your help would be very much appreciated!
Best regards,
Lukas
renz
on 19 Oct 2012
net.IW{1} = %input weights
net.LW{2} = %layer weights
% biases:
net.b{1} =
net.b{2} =
0 Comments
Sara Perez
on 12 Sep 2019
You can specify your own function for the initialization of the weights with 'WeightsInitializer' in convolution2dLayer.
layer = convolution2dLayer(filterSize,numFilters, ...
'WeightsInitializer', @(sz) rand(sz) * 0.0001, ...
'BiasInitializer', @(sz) rand(sz) * 0.0001)
info here:
0 Comments
See Also
Categories
Find more on Sequence and Numeric Feature Data Workflows in Help Center and File Exchange
Products
Community Treasure Hunt
Find the treasures in MATLAB Central and discover how the community can help you!
Start Hunting!