search

daitao / SAN

Second-order Attention Network for Single Image Super-resolution (CVPR-2019)
550 stars 106 forks source link

Memory requirement. #5

Open Tord-Zhang opened 5 years ago

Tord-Zhang commented 5 years ago

Impossible to run this model on single 1080ti GPU under the default setting in the code

daitao commented 5 years ago

Impossible to run this model on single 1080ti GPU under the default setting in the code The training is not a problem, since the model is trained by patches. For testing, it may be a problem if test images are too large. In practice, we can use 1x1 Conv or add pooling function to resize the resolution of feature input of SOCA for more efficiency.

Tord-Zhang commented 5 years ago

However, the default patch size is set to 192. if the patch size is smaller than 192, the kernel size will be larger than the size of feature map in some SOCA module, which will bring some error when doing cuddn_conv. Besides, setting a smaller patch size, 48 for example (I am not sure if this patch size is small enough to run on single 1080ti), the performance would definitely be influenced.