Why do we need to provide A and B when testing?

[samhains]

I find it an inconvenience to have to stitch together a fake 'B' image, when testing. I have noticed that if i provide the model with a 256x256 image (instead of a 512x256 with a fake 'B' image) the output is stretched. Is there any way to disable this, so that we don't have to create the fake 'B' image? Is there something I am missing here?

[janyasir]

I also had to stitch A and B as well, for testing. If we dont stitch it takes half of the image as input. I think this was done so that the target output B can be (internally) compared with resultant output B', and can calculate errors etc. It does waste space, for separate A and B and as well as AB stitched input images. Also in the results folder it generates input A, target B, and output B'. So lot of space wastage.

[yangxingyu588]

hello,if I just have A and don't have B,can I use test.lua to get the prediction of A? thanks.

[junyanz]

The current Lua test code only supports (A, B) pairs as input. You can modify the test code by changing this line. Our PyTorch version supports this function.