Lesson 1 - Oxford 102 Flowers dataset - low accuracy

Part 1 (2019)
1

Hi there! I’ve just watched lesson 1 so that’s my level of background.

I’m trying to train a classifier using the Oxford 102 Flowers dataset, doing pretty much what’s shown in the lecture, but the error rate is too high

This is how I created the databunch:

databunchdetailed
databunchdetailed959×1086 106 KB

And these were the results:

results
results515×794 42.5 KB

I’ve also tried resnet50 and it was the same (0.127)

Any ideas?
Thanks!

2

Hey,
I don’t think the error rate is too high. You are almost getting 90 percent accuracy with ResNet34 which is really good. If you take a look at this link you will see that you are close to the top 10 by using just fastai’s defaults for hyper-parameters. (and after just a few epochs!)
I suggest using Progressive Resizing (a technique you will learn in lesson 3 I think) which could help the accuracy. The other thing you might want to consider is other sets of hyper parameters (the most important being lr)
Regards :slight_smile:

1 Like
3

hey, thanks! good to know. I’ll keep calm then haha. Good information
Cheers :slight_smile: