François CholletVerified account

ok - i abandoned resnets in favor of xception due to the former requiring larger size of input image; will revisit

No modern CNN requires a specific input size. You can use any size you like with rn34. Be sure to use adaptive pooling to make this work (fastai does that for you)

Keras API puts a low boundary on the input image size - ex for rn50 it requires input size to be no smaller than 197

Hi Helena, all modern CNNs operate downsampling on their inputs, which means they start with large feature maps that get smaller as you do down the network. In general, convolution is a downsampling operating unless you explicit pad your inputs.

Thats means that all CNNs have a minimum input size, which is the minimum size you can call the network on without resulting in empty feature maps. At that size, the output of your network is a 1x1 feature map.

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.

Resnet34 and 50 work great on 128x128 px inputs, and quite well on 64x64, FYI.