Optimize MPNCOV.py for less memory usage and faster inference

daitao / SAN

Second-order Attention Network for Single Image Super-resolution (CVPR-2019)

557 stars 105 forks source link

Optimize MPNCOV.py for less memory usage and faster inference #29

Open csmliu opened 4 years ago

csmliu commented 4 years ago

I am going to compare your method in my paper, and am now generating your results using the released model. According to https://github.com/daitao/SAN/issues/22, I modified the code for a faster inference and less memory usage. Kindly have a check.

aryan-f commented 4 years ago

Could not train the model after this modification. Torch 0.4 did not have a diagnol method defined for Tensors, and working around it usingtorch.diagnol(tensor) causes the loss to jump to infinity.

ZFhuang commented 3 years ago

Nice commit, saves a lot of memory.

antonelse commented 9 months ago

the problem still exist ! Does anyone knows where is the problem ?