I have two encoder-decoder models.

*First model:

enter image description here

*Second model:

enter image description here

When I check the performance of the models I get approximately the same performance time (First model ~ 42 sec, Second model ~ 40 sec). I train my model on GPU and check performance on CPU. I test it only on one large image where the size is 12348x12348. I was expecting the larger model that has more parameters to train (second model) to give me longer run time. Anyone can help me to understand why it is not the case here? Am I doing something wrong?

  • $\begingroup$ Could you tell us a little bit more about those 2 models ? If your small network is recurrent, the computation graph of the model is not very parallelizable. It could explain the latency. $\endgroup$
    – Adrien D
    Commented May 11, 2018 at 13:30
  • 1
    $\begingroup$ side note: when you tested it on an image where the size is 12348x12348, you are not following the assumption that says, training and testing data must always come from the same distribution, and for the larger model I guess there are some memory issues because the larger model may simple used swap like memory $\endgroup$ Commented May 11, 2018 at 17:03


Your Answer

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