I have seven measurements (Obs1-7), each measurement has the dimension of [x,y,t] where x and y are coordinates and t is time. Now I want to build a model that uses the first 6 measurements to predict the last one, saying Obs7=f(Obs1~6). I want to use CNN to distil this relationship f. I had some experience using CNN for classification, but have no idea how to deal with such 2D/3D regression. Could someone please give me some ideas? Thanks!


1 Answer 1


It should even be easier than classification: you do not need the final layer.

Your input layer should have 18 nodes $(x_i,y_i,t_i)|i=1..6$

Hidden layers as you see fit (experiment with it, depending on data and results). I would expect the best results from a fully connected layer and a 'relu' layer.

The output layer has 3 nodes $(x_7, y_7, t_7)$

Train the model.

Evaluation is just feed forward of a set of 18 values, and retrieving 3 results from the output layer nodes.

  • $\begingroup$ Thank you @Pieter21 for your comments. After several days' research, I know CNN better and have a general idea on how to handle the data. But I am still unclear about the following questions: (1) my data have 6 channels [batch, x, y, 6], if I set 32 kernels at first layer and padding as same, what is the dimension of the first feature map? Is it [batch,x,y, 32*6]? (2) I need keep padding as same, so do I need pooling step (downsampling)? (3) How to set up the last layer to convert [batch, x, y, number] map to my [batch, x, y,1] map? $\endgroup$ Jan 22, 2020 at 18:09
  • $\begingroup$ I now understand pooling has several functions. But if my input dimension is [x,y,n], after several convolution and pooling steps, I may get a pooling layer with dimension as [x/2,y/2, n]. My target dimension is [x,y,1], how can I connect [x/2,y/2, n] to [x,y,1]? $\endgroup$ Jan 22, 2020 at 18:22

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.