I am processing time-series for classification using 2D CNN model (single channel) where I have converted stationary time series data into 2D image using an "imaging algorithm" known as Gramian Angular Field and Markov Transition Field as described in this paper Imaging Time-Series to Improve Classification and Imputation using pyts library.

My question is logically do I need two different CNN models and then concatenate the output before giving it to the classifier, as I have implemented two different imaging algorithms on the same data?


Can I use a single CNN model with 2 channels, one channel for each kind of imaging algorithm/technique? and then feed the output to the classifier?

In other words, does it makes sense to create different models with different kernels for each imaging technique or can we have all kernels in one model?

  • $\begingroup$ Did you check this code? github.com/cauchyturing/… $\endgroup$ Commented Nov 25, 2022 at 8:38
  • 1
    $\begingroup$ This is known as "early" vs "mid" vs "late" fusion. $\endgroup$
    – Jon Nordby
    Commented Nov 25, 2022 at 11:35
  • $\begingroup$ @NicolasMartin Yes, checked the code, however, it does not say anything about CNN's or the best architecture to convolve these images in single experiment $\endgroup$ Commented Nov 25, 2022 at 18:53


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.