Is there a good rule of thumb to choose a starting model for any image classification problems. I ask because here is what I have noticed:
- in lesson 1, dogs vs. cats – we start with resnet34 and get good accuracy off the bat,
- In lesson 2 - planet competition example in the end – we again start with resnet34 and get good accuracy.
- But in lesson 2 again, dog breed competition, we just start with resnext101_64.
I actually started with resnet34 in dog breed competition and increased the image size from 64 to 246 (similar to planet competition) and got only 83% accuracy – and as soon as I switched to resnext101_64, the starting accuracy was 91% – without any changes.
So how should we choose an initial model. Should we just cycle through different models till we get 90% or so to start. Even if we cycle, is there a good order to cycle so we can save time. Is there any correlation to the type of images (dog breed images are already part of imagenet, while planet images are new). Any insight into this will be very helpful.