search

heykeetae / Self-Attention-GAN

Pytorch implementation of Self-Attention Generative Adversarial Networks (SAGAN)
2.53k stars 475 forks source link

large image generation will take a lot of memory. #6

Closed watsonyanghx closed 6 years ago

watsonyanghx commented 6 years ago

If I want to generate image of size 512512, the attention module will take `int(64 / 8)(512 ** 2) = 2097152`，the parameters will take a lot of memory. Is the self-attention not suitable for large image generation? How to solve this?

heykeetae commented 6 years ago

In my view, there are two solutions. In this implementation, feature number N is equal to pixel number. However, the paper does not indicate strictly if feature should be a pixel. Thus a quick solution is 1) EntonyTang's interpretation, https://github.com/heykeetae/Self-Attention-GAN/issues/1 (where the pitfall is you have only one feature. Not sure if this is correct. In this way, are attention maps as in the paper possible?) or another possible application is to 2) group a set of pixels as one feature, such as 10x10 pixels to get one attention map, to form less number of attention maps per layer.

heykeetae commented 6 years ago

We have updated the 'sagan_models.py'! please check it out!