yhjo09 / adatarget Goto Github PK

In your implementation, a 7x7 sampling grid is output by LocalNet and then applied on SR results instead of HR images (Ground Truth). It is slightly different from your paper. Moreover, sampling 7x7 grid from 9x9 HR piece seems to be better than from 7x7 SR pieces?

When I use torch.load() to load the pre-trained weights of LocNet, I encountered the following error:

lm = torch.load('experiments/pretrained_models/LocNet_TR.pth')
Traceback (most recent call last):
File "", line 1, in
File "/root/miniconda3/envs/troch13/lib/python3.7/site-packages/torch/serialization.py", line 426, in load
return _load(f, map_location, pickle_module, **pickle_load_args)
File "/root/miniconda3/envs/troch13/lib/python3.7/site-packages/torch/serialization.py", line 613, in _load
result = unpickler.load()
AttributeError: Can't get attribute 'LocNet' on <module 'main' (built-in)>

Hello! Thanks for the great paper and code for AdaTarget, it is very interesting.

I've integrated in a model training framework here:
https://github.com/victorca25/traiNNer/tree/master/codes/models/modules/adatarget , because I was interested in evaluating it simultaneously with other strategies and it appears to work pretty well when paired together with a complex augmentation pipeline, like the ones used in BSRGAN and Real-ESRGAN, as well as other network architectures besides ESRGAN.

From my tests I can also confirm:

• it works with AMP enabled, so training models can be done consuming less VRAM (at least)
• it can potentially also work with other cases besides SR
• it works well when paired with losses that work with unaligned images (not only GAN)

I'll have to do more and longer tests, but it's very promissing. I left the initial training of the localization network for later, since I didn't find a good way to integrate it for now.

If I find something that is useful, I can let you know. Cheers!

I can't find this path in the code. Could you help me?

Dear authors,

I would like to thank you for sharing the code. Considering that for real images, kernels are different, your method should perfectly fit to that task.

Thus, I have tried your code on real images from AIM 2020 real super-resolution challenge and got worse results compared to original GT pre-trained network. I have used yours ATG pretrained ckpt & haven`t changed anything except network (also of-the-shelf SR model) and dataset.

As far as I understand there are no experiments with real dataset, aren`t they? If it was conducted, can you (dis)confirm that adatarget does not work for real scenarios?

Thank you!