actually besides http://fast.ai i always credit @chollet's DL book, starting with its very early draft; so the choice of Xception and use of Keras is due to the latter :)
-
For most nets, this tends to be about ~50% of the input size the network was designed for. Downsampling is absolutely necessary when you have large inputs. To get rid of this limitation, you need to create your own network, where you would do less downsampling.
For instance, you could edit Xception and 1) remove max pooling layers, 2) use "same" padding in all convolution layers. This would create a version of Xception that would be compatible with smaller input sizes.
