Error rate not reducing

Hello guys!

Currently I am working with this dataset -

This dataset has relatively less images. I have tried transfer learning using both resnet34 and resnet50 but the error rate I am getting is 0.20. This is causing my model to predict incorrect labels while predicing a single image. I have also tried data augmentation but the error doesn’t decrease at all.

With resnet34-


With resnet50-


I have using following code for Resnet34 -

learn.fit_one_cycle(8, max_lr=slice(1e-3, 1e-2))

For Resnet50 I am getting the following learning rate finder plot:

What is the learning rate I should use with this graph?

Also, I would like to have error rate less than 0.1. Peferably close to 0.05. But With all the variations I have tried of Learning Rate and Epochs, I am getting 0.2 error rate.

What can I do to resolve this issue?

So I figured out what the problem was. Make sure to check the image size when you are creating a databunch. I had it set to 64. When I increased it to 128, the accuracy increased.