Hello everyone !
I have a quick question regarding the input size of vgg16 for example. As we know the recommended size is (224,224) but when using global average pooling, any size is accepted, why that ?
thank you
Adaptive pooling layers return a defined output size and accept variable input shapes. This allows you to use a fixed number of input features to the first nn.Linear layer in the classifier section of the model while the input (images) might have different spatial sizes.
But if the model is trained on size of (224x224) and then we use transfer learning on other images of size (500,500) doesn’t that infect the weights ? will the mode behave the same even if we change the input size?
I’m unsure what “infect the weights” means, but yes the input resolution will have an effect on the model performance and I would expect to see a drop if the spatial size differs by a lot (but I haven’t seen any experiments checking this behavior).