Identical error_rate with the fit_one_cycle method

Hi folks, I haven’t found any posts or hints for my problem so far. Please see the picture below.

I keep getting identical error_rates when I use the fit_one_cycle method. It seems kind of wrong to me. Any ideas why this could be happening?

I have two folders with ultrasound images of the heart and liver. I would like to build a classifier that distinguishes between those two.

Thank you

Grigorij

your lr looks very small, i’d surprised if it can’t go higher. train loss higher than val loss also indicates that you haven’t trained enough yet which is probably also because lr is too low.

show us an lr_finder() plot?

also doing 3 or 4 epochs then unfreezing and doing some more will normally get you better results.

Hi @joedockrill, I am very thankful for your detailed and quick response. Please see the screenshot below.

hmmm. looking at that i think i’d probably try max_lr=3e-4 but you can also just try higher and see if you get away with it.

maybe try

learn.fit_one_cycle(3, max_lr=3e-4)
learn.unfreeze()
learn.fit_one_cycle(5, max_lr=3e-4)

and play around from there? you want your train loss to be lower than your val loss by the end though or you haven’t trained enough. if you can’t get it there then you’re doing something wrong.

How big is your dataset? Have you used ClassificationInterpretation to see what kinds of images the model is failing on?