MATLAB Answers

LSTM with data sequence including NaN values

13 views (last 30 days)
Mohanad Alkhodari
Mohanad Alkhodari on 31 Oct 2019
Commented: Dheeraj Singh on 4 Dec 2019
Hello,
I am using LSTM training network but the training progress is not working and a blank loss plot is coming.
The datasequence is corresponding to a signal in time and it includes NaN values, even the validation dataset.
How to overcome this issue ? is it becuse of the NaN values.
I have to use NaN values to make sure all signals are at the same length for the sequenceInputLayer(numFeatures)
regards,
Mohanad

  0 Comments

Sign in to comment.

Answers (1)

Dheeraj Singh
Dheeraj Singh on 5 Nov 2019
Too much padding can have negative impact on the network performance. To prevent the training process from adding too much padding, you can sort the training data by sequence length, and choose a mini-batch size such that the sequences in a mini-batch have similar length.
For more information refer to the following link:

  2 Comments

Mohanad Alkhodari
Mohanad Alkhodari on 6 Nov 2019
I have 1 dimension of feature using this. like 1x3400, 1x5000, 1x3200, these are my data.
The sequence length will be 1 in this case right?
I have tried this and it shows an pattern in the training with very bad validation accuracy.,
Dheeraj Singh
Dheeraj Singh on 4 Dec 2019
If you have only one feature, your sequence length would be 3400, 5000, and 3200 for the above data, respectively. It is difficult to comment on the dataset without any description, so if you could provide the dataset or some description about it, it can be more helpful in understanding the issue.

Sign in to comment.

Sign in to answer this question.


Translated by