Open robotzheng opened 4 years ago
As has been discussed in the paper, the non-local operation can require a huge GPU memory, which is exactly why we make improvements to suit it for video SR. According to your description, the best way is to use a GPU with a larger memory. If that's not applicable, you might as well choose a larger number of rescaling factor r as described in Section 3.2 in the paper.
OOM at "f = tf.matmul(theta_x, phi_x)" in the "NonLocalBlock" function. train with 32*32 's NonLocalBlock, test with a whole image, maybe not right. how to edit the code?