I am a research scholar, working on transformer fault diagnosis and classification. Can anybody tell me how to calculate training time, in seconds, of transformers, using Back Propagation Neural Network, Fuzzy Logic and Adaptive Neuro-Fuzzy Inference System?
Hi Tapsi.
You cannot specify training in units of time. You want to train your NN to capture the underlying I/O mapping - you need to train the NN for sufficiently long to allow it to capture this mapping, but you want to avoid over-training (where, for example the NN is learning noise in the training data). The standard way of doing this is to partition your data into, approx. 80% training data, and 20% validation data. After each epoch of training using your training data, you test the NN on the validation data (e.g. calculate the MSE). You repeat these steps until you see an increase in the validation MSE. You now stop training and use the NN which gave you the smallest validation MSE.
If you were to test the NN on the training data, you might not see, due to over-fitting, an increase in the MSE.
Hope this helps. Sean
Its all depends on the training network and compexity of the problem. If you want to see which method takes less time (among the three you mentioned), I am not sure. You will have to try and see!
If you would like to estimate the run time of any programs in Matlab environment, you could write “tic” on the beginning of the program and “toc” at the end.
tic : start a stopwatch timer
toc: stop a stopwatch timer
enjoy
- Badges
- Science topic
- Similar topics
- Engineering
- Control Theory