Generative AI
Follow


James Prestegard

Training a deep neural net

James Prestegard on 3 Jan 2024
Latest activity Reply by James Prestegard on 12 Jan 2024

I have been developing a neural net to extract a set of generative parameters from an image of a 2-D NMR spectrum. I use a pair of convolution layers each followed by a fullyconnected layer; the pair are joined by an addtion layer and that fed to a regression layer. This trains fine, but answers are sub-optimal. I woudl like to add a fully connected layer between the addtion layer and regression, but training using default training scripts simply won't converge. Any suggestions? Maybe I can start with the pre-trained weights for the convolution layers, but I don't know how to do this.
JHP

2 Comments

Hans Scharler
Hans Scharler on 12 Jan 2024
Are you using the Deep Learning Toolbox? I use the Deep Network Designer App for these situations.
James Prestegard
James Prestegard on 12 Jan 2024
Thanks for the response.
I am using the DeepLearning Toolbox, but have not been using the designer app. I build the network as a layer graph and then train using the command:
net=trainNetwork(input,target,lgraph,options)
The training cycle runs; the RMSE just doesn't converge as it does without the additional fully connected layer.
I'll look into the designer app to see if it makes any difference.

See Also

Tags

No tags entered yet.