Hi, I started training an rf-detr nano model. It said that the estimated cost was around 5 credits and would take 5 hours. I checked again and it has already used over 5 credits and now that it will take 3 days to train…
Hey, @mz0g , thanks for the post. training time estimates are tough to get right - factors like image size, number of classes, annotations, types of objects all play a role. RF-DETR has had fewer overall training jobs, so estimating times will only get more accurate from here with use.
Would you be open to inviting Support to your account so we can investigate the attributes of your dataset specifically? This will help us validate why the training estimate is as it is. if so, please email me (joseph.nelson at roboflow dot com) with a link to your Workspace once you’ve invited us. i’ll also include another team member to dive in.
again, thanks for flagging this for us.
Hi @mz0g what resolution was your dataset exported at? A lot of what makes the Nano model fast is its low resolution – when ran at high resolution it can be unexpectedly slow!
That being said, I am pushing a change that will make train time estimates a bit more accurate. Won’t take into account resolution yet, though.
Hey Peter, I trained using 640 by 640 resolution slide!