Just a general question, why the model to go is resnet34 if we know that there are better archs, that are faster to train and with better predictive capabilities?
I can see that even in this years course Jeremy is using resnet34 all over.
For dawnbench wasn’t better wrn22 than resnets? Yolo v3 and wide resenet paper didn’t prove this?
Am I missing something? From my very noob point of view, training wrn22 and darknet is harder than resnet34 in general.
Its a good question - i have been thinking about that too .
I think that its because resnet34 is relatively fast to train compared to many more accurate models with hugh number of parameters. Hugh models are not practical in a course where many have trouble with memory capacity and overfitting
Just read this paper that compares many models: https://arxiv.org/abs/1605.07678
1 Like
But even Jeremy says that resnets are slow, it is the conclusion of the wideresnet paper and yolo