CODEWITHSUNDEEP
machine-learningtensorflowdata-science
BioInfoBrett
BioInfoBrett

Reputation: 305

How to represent a linear data in TensorFlow

I'm trying to model some oscilloscope-like data in TensorFlow - a linear stream of energy pulses with a duration, intensity, etc. - but otherwise performing very similar classification tasks, and I'm having trouble figuring out how best to represent it in TensorFlow.

The tutorials are aimed at image classification, and the framework seems to be built around 4d tensors I'm having trouble figuring out how to represent my relatively simpler data.

Specifically, I'm trying to figure out the following:

1) If I have a string of 100 pulses with 4 tracks of information about them, is that equivalent to a 4x100 image or a 100x4 image? Or perhaps a 100x1 image with 4 "channels"? How the tensors are folded/unfolded is unclear.

2) How does max-pooling translate into this lower dimensional space? Like, how do I ensure that I'm pooling across pulses instead of just within a channel, or some other non-sensical pattern? How the "strides" are calculated and used is not well explained.

Had anyone tried to model similar data in TensorFlow?

Upvotes: 2

Views: 229

Answers (1)

mrry
mrry

Reputation: 126154

Recurrent neural networks (RNNs) are a possible representation for sequential data like a stream of energy pulses. The TensorFlow website has a tutorial on building an RNN for predicting the next word in a sentence of words, but this could possibly be adapted to predicting the next value in your scenario.

Upvotes: 1

Related Questions