[R-50] Is unfreezing the first two layers for training on grayscale fingerprints a bad idea ?

[ExtReMLapin]

Hello,

I've been training a custom dataset which only contains black&white fingerprints scan, the more augmentation I do, the slower it converge, I have around 2 000 images, 10 000 annotations (for detection).

I used to have around 350 images but as i'm doing tiling, i split those ~350 images in 2k images to not lose any quality.

Is it a bad idea to "not trust" the first two layers to fit my needs and unfreeze them ?

Regards,