Progessive resizing = got the pets data in lesson 1 to .008 error(!)

I was very intrigued with Jeremy’s comment in lesson 3 (I believe) about how progressive resizing was underutilized and how it wasn’t clear if there was optimal size progression.
Thus, tonight, I spent about 4 hours testing out some various algorithms for the progressive resizing (ala resizing based on quartiles of the final size, etc).
Anyway, I’m not done but I was able to see pretty massive progress - I used the pets setup from notebook 1 since it had a control of what was already done with the standard approach.
By starting at 50% sizing, then jumping to 100% sizing directly, that outperformed the quartile progession method pretty handily and also beat out the notebook results of .043 best, vs. multiple .008 and .009 (basically only one image wrong) with the progressive resizing (this is with resnet50).
I used the same resnet50, etc. that was used in the notebook as I was trying to control the other variables except for the progressive resizing.
I want to try thirds and also 50/75/100% tomorrow, but at least for now I can suggest that half and then full resizing seems to produce much better results in about the same amount of training. The quartile progression did get slightly better vs standard (.041) but progress in the accuracy progression was slower as the lr_find suggested much slower learning rates (1e-6 and 1e-5)…the 50/100 approach just blazed right in by contrast using 1e-3, and ultimately had far better accuracy.

1 Like

have you tried progressive resizing 50/75/100%?

Those are pretty impressive results. Almost seems too good to be true! Would you mind posting a notebook or code of how you did the resizing? I’d like to give it a try on some of my own datasets.

Were you sure to reset your random seed so that your validation set was the same each time you resized?

My thoughts too. I’m wondering how the model told between the american dog and UK kennel board definitions of the same dog type? Maybe the USA ones were sunny or had bigger back gardens. I’d guess training set having duplicates of the validation set or as you say the validation set leaking into the training set on the two training loops and data bunch creations.