Model is not being trained

Part 1 (2019)
1

output
Hi, I was training an image classifier and something weird is happening, as accuracy isn’t increasing. Can anybody explain this?

This is the learning rate graph.
lr_find

2

Hi, @bhawna. I’m new at this, too, so I might not be much help, but when I had low accuracy with my classifier, I’m pretty sure it was because my data was of low quality. What kind of images are you classifying, and where did you get your data sets?

Regards,
go_go_gadget

1 Like
3

Try the following fine turning

  1. Fit for one cycle for longer epochs… then
  2. Use just 1e-4 in the learning rate

learn.fit_one_cycle(20, max_lr=slice(1e-04))

It should improve…

1 Like
4

Hi Laura,
I was building a pretty basic classifier that could differentiate between ‘Jaguar’, ‘Cheetah’ and ‘Leopard’. I collected the images from google. The means of collecting images was exactly same to the one demonstrated by Jeremy in lesson 2.

Best Regards,
Bhawna Singh

5

Hi Abduljamiu,

I tried what you were saying. Here are the results.
Before unfreezing

p1
p1.png361×627 46.6 KB

After unfreezing
d2

7

try download more images to your dataset and train again. More data may lead to higher accuracy

8

Hi, @bhawna

Hmm. What does your confusion matrix show, and have you run the ImageCleaner?

1 Like
9

Yes definitely! Thank you :smiley:

10

Yes, I did run imagecleaner. There are 90 images for each class. I think downloading better images will lead to more accurate results. Thankyou :smiley: