OOM using EfficientNet B3 #5801
Replies: 5 comments
-
Looks fine for me. Have you trained the same network without lightning successfully? Because I remember it being a quite large network, so it is likely that you run OOM here just because of that. |
Beta Was this translation helpful? Give feedback.
-
@SeanNaren could sharded help as you showed that it reduces memory usage...? |
Beta Was this translation helpful? Give feedback.
-
@justusschock No I have not. However I've read a paper which managed to train the network with 11Gb of vram in a similar setup, can't access the code though. |
Beta Was this translation helpful? Give feedback.
-
How are you defining |
Beta Was this translation helpful? Give feedback.
-
What's the image size you're using? |
Beta Was this translation helpful? Give feedback.
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
-
Hi,
I am currently doing the ISIC 2019 Challenge using pretrained EfficientNet B3 on Kaggle with a 16GB GPU. I instantly run into out of memory problems, tried reducing the batchsize and 16 bit precision, doesn't help. I don't know how to debug this and can't find anything special in my code. Here's the important parts of the model:
`
class LightModel(pl.LightningModule):
Beta Was this translation helpful? Give feedback.
All reactions