Does the images re-sizing trick shown in Lesson3 with Planet dataset always work?

I am working on the Plant Patology kaggle problem and I tried to use the same image resizing trick that was used in Lesson 3 for “Planet: Understanding the Amazon from Space” problem.
Specifically, what I did was:

  1. Freeze and Train with image size = 256 (Also you can see the transformation I did):

    Screen Shot 2020-05-09 at 11.29.08 AM
  2. Unfreeze and train with same image size = 256:
    Screen Shot 2020-05-09 at 11.29.49 AM
  3. Freeze and train with image size = 512:
  4. Unfreeze and train with image size = 512:
    Screen Shot 2020-05-09 at 11.34.30 AM

As you can see, clearly resizing the training images to 512 from 256 doesn’t help at all. But, it did help in the Planet problem shown in Lesson 3. Why is that? Can we not consider resizing and training as a rule of thumb to improving results?

Did you not do a lr_find after the unfreeze? unfreeze allows many more parameters to be trainable, which can affect what learning rate your would want to use.

1 Like

I agree with @marii.
I worked briefly on that competition too and resizing helped. Not significantly, admittedly. But it helped.

1 Like