No error when transfer learning with wrong shape?

vision
1

Hi,
I have a funny “non-error”: In my training setup I changed from Inception V3 to Resnet50 but did not change the Resizing (was still at 299 image size), but when I then tried to retrain on these Images, I noticed, that even though I had the wrong size, the training would work without any problem. Why is that?

Best,
Nico

2

The torchvision models use an adaptive pooling layer as shown here, which makes the input size more flexible.

3

Ah this is good to know thank you! I really appreciate your help here, it’s a great community :+1:

1 Like