pretrained network for small images about recipes HOT 1 CLOSED

I'm not sure how much pretraining on ImageNet is worth for 20x20 images, but looking at the ResNet-50 specification, there's 3x3-max-pooling of stride 2x2 in the beginning, and 7x7-pooling in the end, with a lot of convolutions in between, some of which are strided. Try adapting the input layer specification to your input size of 32x32 or 20x20 instead of 224x224 and Lasagne will warn you that at some layer, there won't be any pixels left.

You'll need a different architecture (maybe pretrained on CIFAR-10/100?), or only use the very first few layers.

from recipes.