You are here

Home » Neural Network Forums » General » General Discussion

Is there a way to estimate the time of each epoch in a training?

Clifte's picture
published by Clifte on Thu, 08/02/2012 - 08:21

In the feed-Forward with back propagation network with:

example1:
10n input
5n hiden layer 1
7n output

example2:
100n input
50n hiden layer 1
70n output

Neural Network Forums: 
General Discussion
jeffheaton's picture

Submitted by jeffheaton on Wed, 08/08/2012 - 07:25.

In theory you could. The longest part of training is typically the activation functions and calculation of each connection. The time to train is going to be some forumla like:

(numWeights * weightTime) + ((numNeurons - inputNeurons) * activationTime)

numWeights = number of weights in the network
numNeurons = total number of neurons in the network
inputNeurons = input neuron count

activationTime & weightTime will depend on your hardware and will need to be determined experimentally.

I subtract out input neurons since they are not processed by an activation function.

cws4173's picture

Submitted by cws4173 on Sat, 09/15/2012 - 18:03.

Couldn't you take the total training time and divide by number of epochs? This gives you the average training time per epoch. If you want an even finer grained recording of time per epoch you can try wrapping each train.iteration() statement in a timing block.

Theme by Danetsoft and Danang Probo Sayekti inspired by Maksimer