Minimum Hardware Requirements
hiroalchem opened this issue · 4 comments
I am new to self-supervised learning, what is the minimum hardware required to use dBOT? Is it possible to learn with my own dataset on a single GPU?
It is not possible to train dBOT with a single GPU. See table11 in our paper for detail. However, downstream research will be possible. Feel free to play with our pre-trained models.
@liuxingbin Thank you for your reply.
I wanted to apply this to a somewhat specific original dataset, so I would like to do the pre-train itself if possible, but was I correct in understanding that this is not possible because it cannot be trained in a realistic time on a single GPU? Would it be possible to work on this with a few GPUs with a bit lower performance? Or is it that the batch size would be too small for learning to work? The dataset size is quite small compared to ImageNet.
You can try to train your specific data and see how long it would take for one epoch. You can set batch_size and accum_iter to maintain the effective_batch_size (see here). We have not tested our dBOT on a few GPUs, but it should achieve the same results as long as you keep the effective_batch_size with the original setting.
I will try it once I can. Thank you very much for your kind advice.