search

lucidrains / DALLE2-pytorch

Implementation of DALL-E 2, OpenAI's updated text-to-image synthesis neural network, in Pytorch
MIT License
11.03k stars 1.07k forks source link

Hardware requirements #9

Closed INF800 closed 2 years ago

INF800 commented 2 years ago

Hi,

Let's say if the project completes -- what will be the hardware specs required to run the model?

Best, Rakesh

lucidrains commented 2 years ago

@INF800 :wave: reportedly the model is smaller than DALL-E v1, so whatever the requirements of ruDALLE is, but less

INF800 commented 2 years ago

@INF800 👋 reportedly the model is smaller than DALL-E v1, so whatever the requirements of ruDALLE is, but less

Hi @lucidrains, how did you come to know DALLE2 is smaller than DALLE1? Would you kindly provide reference links? I tried to read paper but couldn't find any such information.

johnpaulbin commented 2 years ago

@INF800 👋 reportedly the model is smaller than DALL-E v1, so whatever the requirements of ruDALLE is, but less

Hi @lucidrains, how did you come to know DALLE2 is smaller than DALLE1? Would you kindly provide reference links? I tried to read paper but couldn't find any such information.

I believe this is relevant:

The Dall-E 2 paper shows that the "base" model (64) is 3.5 Billion parameters in size.

image

This is from the Dall-E 1 paper:

image

johnpaulbin commented 2 years ago

So, theoretically, it is possible to run on free colab GPUs (K80 +) Training should be possible with a T4 GPU

INF800 commented 2 years ago

So, theoretically, it is possible to run on free colab GPUs (K80 +) Training should be possible with a T4 GPU

Thankyou for this @johnpaulbin.

Needless to say I was worried that parameter count will be as unwieldy as PaLM or GPT3