Appropriate use of Convolutional neural networks

1

Hi all,

the questions may be silly but I’ve not found a simple answer.

I’m trying to use the CNN to geophysical data inversion; I would like my net to produce as output some velocity models with velocity trends constant over the layers. This means that my ideal output is something like the blue curve in the left picture attached (the red is the net output). My net is characterized by 3 simple convolutional1d layers with LeakyRelu activation function and Kaiming weights inizialitation. What I don’t understand is:

  1. how to create a net whose output may try to generate flat results.
  2. the loss function, reported in the right picture below shows those jumps I don’t really understand since I’m using no dropout …

Schermata del 2020-02-27 13-43-57
Schermata del 2020-02-27 13-43-571920×1080 236 KB

any advise would be really helpful

2

Can you put the network code here? More specifically, I am looking for the loss function and any non linear operation in the network output like to sort of subsample (decrease the resolution) of your output. A vanilla way would be to use round, floor, ceil. Can you specify or link something on the nature of the output. I guess that would be helpful, too.

Is this related?
Fig 7, 9 and 12-c