IFT6266-project
Sunday 26 February 2017
Simple convolution network
I've been experiencing with a convolution network. The network is a series of convolutions and max poolings followed by a series of deconvolutions and upscalings using the mean squared error as a cost function. Here is an example of result with 32 filters for every convolutions.
I then tried to increase the number of filters in the 1x1 layer. I tried with 1000 filters. The results were very similar.
Adding 2 fully connected layers in the middle of the network improved the quality of the blur a little.
Finally, I tried to get rid of the grid by removing the upscaling layers and adding more deconvolutions instead. Here is the result
There are a few more experiences I will do, but I won't spend much more times on this model since the results are always gonna be a blur. It is possible to improve the quality of the blur, but it won't show clear shapes.
Subscribe to:
Posts (Atom)