By model architecture, I'm interested in knowing the following:

  • Number of nodes in input layer
  • Number of nodes in subsequent layers
  • Number of layers in the architecture
  • Number of filters and kernels in each layer

For more context, I'm creating a ConvWTA that previously took in 32x32 images but now I'm modifying the model to take in 16x16 images and I'm not exactly sure about how I should go about modifying the architecture/whether this needs to be done in the first place other than the input layer. I also want to include a bottleneck in the model so I'm planning out the architecture but would appreciate some guidance or resources.

  • $\begingroup$ This is very helpful. I suspect the question is now clear enough that people will be able to help. I took the liberty of changing "preceding" to "subsequent" in your 2nd bullet point. "Preceding" means the layers before the input, which doesn't make much sense. I think you meant "subsequent", which means the layers after the input. If I misinterpreted that, roll the edit back with my apologies. $\endgroup$ – gung - Reinstate Monica Aug 15 '19 at 18:43

Strictly speaking, it doesn't have to -- most convolutional networks are technically capable of processing variable sized input images. In practice, yes, the size of the input image together with a limited computational budget does impact the choice of model.

Being very general, typically larger input images means less layers can be used due to memory constraints. The first set of filters must typically be made larger (so that they capture the "same" spatial extents). You could also lower the number of filters in each layer to lower computational costs.

However since both 16x16 and 32x32 images are very small, I doubt any of these changes need to be made in your case.

| cite | improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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