search

NVlabs / pacnet

Pixel-Adaptive Convolutional Neural Networks (CVPR '19)
https://suhangpro.github.io/pac/
Other
511 stars 79 forks source link

Batch size #12

Open neuralchen opened 5 years ago

neuralchen commented 5 years ago

Can i set a larger batch size for example 16? I learn that the default batch size is 1. Actually, i have some questiones about the kernel, if i input an batch(B) of different images and correponding guided features into the pacconv. Different guided features will produce different kernels. There are multiple different kernels in a batch that will increase the complexity of the forward calculation. How do you deal with this kind of problem?

suhangpro commented 5 years ago

Depending on the size of your image, network, gpu mem, etc., you can definitely try higher batch size.

A major advantage of our operation is exactly that -- different kernels for different images (and locations). Computation-wise there isn't really much overhead compared to a naive conv implementation relying on im2col, however admittedly we are not able to utilize as much optimization and is somewhat memory-heavy.