NaN performance and gradient in RNN

Hello to everyone! I'm trying to run a recurrent neural network, performing something slightly different from usual, that is time series classification, instead of prediction. By doing this, I present to the rnn same length series. If this last is not as long as the input size, I add NaN values, as usual. Then, as I care about the outcome of the rnn only at certain time step, I put NaN values in error weights related to all the others (I forgot to tell I'm running a concurrent input analysis). The learning process of the rnn starts, but during the process you can see performance and gradients bars show a NaN, intead of numerical values. At the end of learning, you're consequently not able to visualize the performance plot. Do someone know how to manage these situations?

Answer 1

[Greg Heath]

Your configuration doesn't make sense to me. If you want to classify a series

1. Find the significant correlation lengths of the series. (Search greg nncorr) 2. Use overlapping correlation length windows of the series as input vectors. 3. Use trial and error to determine the smallest effective overlap.

Hope this helps

Thank you for formally accepting my answer

Greg