Jeremy created the Imagenette/Imagewoof datasets [1] so we could try out our models without wasting too much time and money. For both datasets, we have access to full size images, as well as 320px and 160px versions.
@LessW2020 stumbled upon an interesting effect: if we train with 128px images resized from the full sized images, we get better results than when resized from the 160px version.
I ran some tests on Imagewoof, it’s quite significant. I wonder if it’s a common occurrence in DL projects.
| Epochs | Imagewoof-160 (a) | Imagewoof-full_size (b) | # of runs | p-value |
|---|---|---|---|---|
| 5 | 61.25% | 63.89% | 20 | 0.0001 |
| 20 | 80.87% | 82.40% | 3 | 0.043 |