Regularization images

[elismasilva]

If add a second bucket of images with repeats setted it could work like regularization images?

[bghira]

yep, similar to onetrainer in that way.

[elismasilva]

so if i have 16 images on main dataset to train and 2300 reg images i dont need do 16 / 2300 to put it value in repeats ? how many could be my repeat in this case with ignore_epochs.?

[elismasilva]

adding how do I force it to re-cache the buckets without always having to delete the json?

[bghira]

is there a reason you have to recache the buckets? they should just remain static - right? if you add more images they should be picked up automatically

for repeats it's probably enough to have a value of 2 when there's that big of a difference in the image counts.

with just 16 images on batch size 4 you'll hit 1 epoch every 4 steps unless you have repeats enabled on that dataset too.

at the end of each of those epochs you'll still have the 2300 samples being chipped-away at. so, at the end of epoch 1, you'll have pulled maybe 8-12 images from the regularisation set.

this means you have something like 500 epochs of regularisation images when ignore_epochs=True on the large regularisation images dataset.

[bghira]

you can also set probability lower on your training set, which will sample regularisation images more often.

depends on the desired outcome and how your initial results go

[elismasilva]

so 1 epoch is total_images / batch_size