Training time and number of GPUS about enhancing-transformers HOT 6 CLOSED

Hi @JialeTao, the training vit-vqgan small is faster than I expected and it is just released. The speed is 1.05s per iteration for a batch size of 8 and it can even go up to 16 but 8 is good enough, again this is for A100 40GB. Also if you dont have any further question, I will close this issue. Feel free to reopen

Hi @thuanz123 , thanks for sharing the checkpoint of vit-vqgan small. Then for this small-small model, how many GPUs and how many iterations you have trained?

For quick training, I use 32 gpus A100 40GB and train for 500000 iterations on ImageNet. But I think a decent GPU with 8GB VRAM is enough, just lower the batch size and train longer

from enhancing-transformers.

Hi @JialeTao, depend on the config, training can be fast or slow. For the config in this repo which is ViT-VGAN base, it takes about 1.45s per iteration with a batch size of 4 on a A100, this is quite demanding. So if you dont have much gpus, I recommend training a much smaller config than the config I have in this repo. Also, there are plans to train smaller models so you can wait if you want but it will be a long time later since I'm too busy these days 😭

from enhancing-transformers.

Thanks for the reply. Then for the vit-vqgan base, what how many iterations you have trained? And the 1.45s means stage 1 training or stage 2 training? And the last, A100 with 40G menmory or 80G?

from enhancing-transformers.

Hi @JialeTao, 1,45s per iteration is for stage 1 training and the GPU is A100 40GB. I have trained vit-vqgan base for 1000000 iterations with 32 A100s and each gpu has batch size of 4. For stage 2 training, it is currently buggy so I dont have any estimate or numbers for it 😅

from enhancing-transformers.

Hi @JialeTao, the training vit-vqgan small is faster than I expected and it is just released. The speed is 1.05s per iteration for a batch size of 8 and it can even go up to 16 but 8 is good enough, again this is for A100 40GB. Also if you dont have any further question, I will close this issue. Feel free to reopen

from enhancing-transformers.

Hi @JialeTao, the training vit-vqgan small is faster than I expected and it is just released. The speed is 1.05s per iteration for a batch size of 8 and it can even go up to 16 but 8 is good enough, again this is for A100 40GB. Also if you dont have any further question, I will close this issue. Feel free to reopen

Hi @thuanz123 , thanks for sharing the checkpoint of vit-vqgan small. Then for this small-small model, how many GPUs and how many iterations you have trained?

from enhancing-transformers.