2

I am working on making prediction from 2d data. Data size is 7640x200x2; for each 200x2 matrix, I want a 2x1 array predicted from it. I am a beginner, and I am confused how to bulid a useful model. I have tried a cnn+lstm model, but the result was really bad. Could anyone please give me some advice?

Sarah Leidy
  • 21
  • 2

1 Answers1

0

Doesn't seem there are any sequences as said in the question, so LSTM isn't applicable here.

(200x2) --> (2,1) can be simply done by a dense network after flattening:

inp (200x2) --> flatten (to 400) --> dense(2, activation=identity)

2D convolutional layers may be put in between inp layer and flatten layer:

inp --> conv2d --> flatten

Depends on the range of the expected output, the activation in output dense layer can be changed, for example, use 'relu' instead of 'identity' if output is always positive.

Dee
  • 7,455
  • 6
  • 36
  • 70
  • 1
    Thanks! I am now facing another problem, the training loss isn't decreasing and I think I used a wrong way to normalize the data. Each (200,2) matrix has 2 features, how should I normalize it correctly? – Sarah Leidy Apr 22 '21 at 09:34
  • the popular way to normalise is finding abs max, and divide all values by that abs max – Dee Apr 22 '21 at 14:35