What if my learning rate plot shows no clear dip?

Part 1 2022
1

Hi there!

I’m trying to improve a classification model I’ve trained which distinguishes video game screenshots between super nintendo and sega megadrive

I’m getting okish results, but while looking for ways to speed up iteration, i tried the learning rate finder, and it doesnt have the pronounced dip i’ve seen in the examples.

Screenshot 2024-03-04 at 11.45.19 pm
Screenshot 2024-03-04 at 11.45.19 pm1166×878 61.3 KB

this is resnet18

Screenshot 2024-03-05 at 12.30.16 am
Screenshot 2024-03-05 at 12.30.16 am1182×886 60.2 KB

this is resnet34

Any guidance on what this graph suggests?

My guess is that it suggests my data isnt very good for learning?

2

this is my project so far

3

This lesson might be helpful: Practical Deep Learning for Coders - 17: Initialization/normalization

1 Like
4

Watched the vid. Thanks!

Looks like incrementing the lr smaller MAY help get a closer picture but most likely my model is not great…

I guess it literally means

“adjusting the learning rate aint gonna help you son, your problems are elsewhere”!