Model performs worse after adjusting learning rate

Part 1 (2019)
1

Hi!

I’m playing with some simple vision CNNs and trying something similar to what someone tried between class1 and class2, categorizing similar looking relatives. After 6 epochs I’m seeing something like in the image below:

resnet50_1
resnet50_1.png1456×604 52.4 KB

What is weird is that after adjusting the learning rate the performance seems to be much worse.

resnet50_2
resnet50_2.png1568×1114 99.4 KB

This seems counterintuitive. I’m likely doing something wrong but I can’t figure out what.

Separately, which pretrained models are known to perform better for face recognition? I saw mentions of VGGFace2,. Even if it’s overkill it might be good practice to try to use it with fast.ai.

Thank you!

2

Hey,
If you have watched the Week-3 in the course Jeremy mentions that you should actually use LR of 10 magnitude less than the point where the loss starts increasing. So in your case you should replace le-2 with le-03 and see what happens
Regards,
Jayam

3

Will definitely try it out over the weekend. Have been trying to use face detection and transforms to crop the face + body. face detection is working, transforms not yet.