RuntimeError:CUDA out of memory

[2000lf]

Can I remove the attention layer for high resolution img?

[explainingai-code]

Yes you can try removing the attention to see if that gets rid of the error. Whats the image size you are working with ?

Adding few other things that will reduce the compute requirement in case you are working with the default config.

1. Reduce Batch Size(config uses 64 as of now)
2. Keep attention only in midblock and remove Downblock attention(here) and Upblock attention(here)
3. By default the downsampling is disabled on last downblock(since mnist images are anyways small), so change this value to be [True, True, True]

[2000lf]

Yes you can try removing the attention to see if that gets rid of the error. Whats the image size you are working with ?

Adding few other things that will reduce the compute requirement in case you are working with the default config.

1. Reduce Batch Size(config uses 64 as of now)

2. Keep attention only in midblock and remove Downblock attention([here](https://github.com/explainingai-code/DDPM-Pytorch/blob/main/models/unet_base.py#L98-L105)) and Upblock attention([here](https://github.com/explainingai-code/DDPM-Pytorch/blob/main/models/unet_base.py#L275-L281))

3. By default the downsampling is disabled on last downblock(since mnist images are anyways small), so change [this](https://github.com/explainingai-code/DDPM-Pytorch/blob/main/config/default.yaml#L14) value to be `[True, True, True]`

Thank you for your advice,I use img with size 900*1600. That consume huge memory when encounter attention layer.

[explainingai-code]

Got it. Yeah try with a batch size of 1 first.
If it works then you can train with gradient accumulation.
But if that also fails then you would have to either remove attention layers or train with smaller sized images.

[2000lf]

Got it. Yeah try with a batch size of 1 first. If it works then you can train with gradient accumulation. But if that also fails then you would have to either remove attention layers or train with smaller sized images.

Thank you ,I will try as you told.