Out of memory error

[sunwoo76]

What is the minimum capacity of memory for training with the batch_size=1?

I implemented a 2060super GPU on my computer, but this error happened.

[zhangmozhe]

Possibly you can choose pytorch 1.6 and adopt mixed-precision training for memory-efficient training.

[panzhang0212]

16G is ok, and 32G is better with a larger batch size

[sunwoo76]

@panzhang0212 thank you. I took a typo in my comment. I implemented GPU2070 super not 2060super and its memory is 8G.
Will not 8GB memory size occur out of memory error with batch size 1?

[mlxht990720]

I have a same problem which is out of memory error , I use GPU2080Ti and its memory is 10GB, though I change batch_size to 2, out of memory also happened .

[myway0101]

@mlxht990720 Did you try with batch_size = 1 using 2080Ti?

[mlxht990720]

Yes. But it's still out of memory. Maybe I need to use more gpus.🥲 发自我的iPhone

…

------------------ Original ------------------ From: myway0101 ***@***.***> Date: Fri,Apr 9,2021 3:37 PM To: microsoft/CoCosNet ***@***.***> Cc: mlxht990720 ***@***.***>, Mention ***@***.***> Subject: Re: [microsoft/CoCosNet] Out of memory error (#21)

[myway0101]

Any idea to decrease memory cost in training?

[zhangmozhe]

Any idea to decrease memory cost in training?

You can try to use mixed precision training and gradient checkpointing. Further, you can try https://github.com/facebookresearch/fairscale