Minimum recommended GPUs

[nickk124]

Minimum recommended GPUs

[nickk124]

Excellent, this is exactly what I need. Thank you for the quick and thorough responses!

[nickk124]

Let me be specific: for working with 256x256 greyscale imgs.

[raywzy]

Could you provide some statistics of your dataset? BTW, you could also scale the parameters of transformer (layers, embedding dimensions etc.) to decrease the computational cost.

[nickk124]

Hey, thanks for getting back to me!

My data are 256x256 patches sampled randomly from much larger (about 2000x3000) single-class one-channel images. At test time, I only need to do completions on the same type of patches, for the same 128x128 mask at the center on all test imgs.

[raywzy]

Hey, thanks for getting back to me!

My data are 256x256 patches sampled randomly from much larger (about 2000x3000) single-class one-channel images. At test time, I only need to do completions on the same type of patches, for the same 128x128 mask at the center on all test imgs.

Great~ How many such 2000x3000 images do you have?

[nickk124]

Around 8000, but I will have around 90,000 total soon.

[raywzy]

Around 8000, but I will have around 90,000 total soon.

Got it. I think the training (90,000 images) will require about 2 days using 8xV100 (32G). Of course you can also reduce the batch size to make it work well on single GPU, which will require longer training time.