train neural net with prior solution

Net training finished with 10000 epochs. Need to strat where it finished.

Answer 1

[Matt J]

Your post is under-detailed and does not tell us how the network and training are implemented. If I assume you are using trainnet, e.g.,

netTrained = trainnet(images,net,lossFcn,options)

you can simply run the training again, giving as the second input argument your pre-existing, partially trained network.

[Dilip Madan]

I ran a two layer neural net with 50 neurons at each layer using

mdl=train(net,X,Y)

X and Y were 2000 by 4.

The training stopped on maximum epochs reached that was 10000.

Would now like to start where it finished. Performance of mse was 0.007 target is 0.0005. It took some 6 hours.

Thanks.

[Matt J]

Similary, train() can be given a previously trained net as its first input,

mdl=train(netPrevious,X,Y)

and the training will resume using the weights contained in netPrevious.

See also Train Network Using Checkpoint Saves,

https://www.mathworks.com/help/deeplearning/ref/network.train.html#btzb5il-1

[Dilip Madan]

When I do this the initial Performance is much larger than the final Performance of the previous net. Is this a concern or should one not worry about this. Many Thanks.

[Matt J]

This method of resuming training is not optimal. The optimal method is using checkpoint saves, as explained at the link I gave you. But since you did not set checkpoints, the training algorithm does not have everything it needs to resume gracefully.

Even though you have the network weights and biases, there is no record of prior algorithm state variables, like learning rate schedules and momentum, etc... Therefore, the algorithm will need time to reconverge. You still might end up saving iterative effort as compared to starting from scratch, but next time you should use checkpoints. Or, consider moving to the Deep Learning Toolbox, which does give you finer control of algorithm variables.