Multiple inputs network about bicyclegan HOT 5 CLOSED

It's possible. You can write your own dataloader to read multiple inputs from the disk (you can modify this line for example), concatenate them and feed it to the network. (also change --input_nc accordingly)

from bicyclegan.

Thanks for your suggestion.
My idea is to make the data loader load input_1, input_2 and groundtruth_output at the same time(side by side). Then concatenate two input images. The modified code is like this:

When I trying to run this code, I got an error at the transforms.Normalize() process said: TypeError: tensor is not a torch image.

Could you give me some suggestion about how to fix this issue and make this modified code works.

Thanks!

from bicyclegan.

I have worked out the last issue. By successfully applying two corresponding images as input, the generated output does not have better performance by comparing with original BicycleGAN.

I have another question:
The generator used in bicycleGAN is U-net.As far as I know, U-net is widely used in biomedical image segmentation. What is the advantage to use U-net as generator in BicycleGAN other than other generators?

from bicyclegan.

We adopt it from pix2pix [Isola et al. 2017]. Please refer to the pix2pix paper. See Sec 3.2.1 Generator with skips and Sec 4.3 Analysis of the generator architecture for more details.

from bicyclegan.

Thanks for your suggestion!

from bicyclegan.