CODEWITHSUNDEEP
pythontensorflowkerasconv-neural-networkmnist
rbsc
rbsc

Reputation: 53

Input numerical arrays instead of images into Keras/TF CNN

I have been building some variations of CNN's off of Keras/Tensorflow examples that use the MNIST data images (ubyte files) for feature extraction. My eventual goal is to do a similar thing but with a collection (~10000) 2D FFT arrays of signal data that I have made (n x m ~ 1000 x 50)(32 bite float data)

I have been looking for an example that uses something other than image files but can not seem to find any.

My questions are: Is this possible to do without converting them to images. Can the dataset be exported to a pickle or some other file that I could input? Whats the best way to achieve this?

Thanks!

Upvotes: 1

Views: 641

Answers (1)

Zad
Zad

Reputation: 26

Yes, you can use CNN for data other than images like sequential/time-series data(1D convolution but you can use 2D convolution as well).

CNN does its job pretty good for these types of data.

You should provide your input as an image matrix i.e a window on which CNN can perform convolution on.

And you can store those input matrices/window in a numpy array and then load those file and train your CNN on it.

Upvotes: 1

Related Questions