Details on Training RefineNet

In the paper's Implementation Details section it is stated that the RefineNet is trained with exposure estimation and FusionNet together but freezing the latter two. I already trained the FusionNet. But I didn't see in the code that RefineNet training process used my pre-trained FusionNet, The RefineNet just defined new netG and netM models.

I'm a bit confused about this training process. If the trained FusionNet is never used, then what's the point of training a FusionNet?

There is some bugs for loading I think, previously, refineNet is bulid in a second branch, in the merging into current branch, the loading have some bugs.

if opt.load_dir and opt.load_dir != 'None':
print('load fusion net from:', opt.load_dir)
model.load_networks('latest', opt.load_dir)

here for loading fusion network weights, please have a check the if statement.

This problem confuses me a lot too! Have you solved it yet? Could you provide the solution? Thanks.

This problem confuses me a lot too! Have you solved it yet? Could you provide the solution? Thanks.

In this line, add a check after the for loop

exposure-fusion-shadow-removal/models/base_model.py

Line 175 in 531c79b

for name in self.model_names:

if name == "R":
    continue

if name == "R":
    continue

thanks bruh