kynthesis / starsrgan Goto Github PK

Can please you share the link for the pretrained weights in order to do some testing and finetuning?
It is not found in the README file.

Thanks :D

Thanks for your interesting work :)

I just had a quick question, when I tried out training a starsrgan light model, with your training command, I somehow got a ModuleNotFoundError:

PS D:\StarSRGAN> python .\starsrgan\train.py -opt .\options\2xNomosUniStarsrgan_light.yml
Traceback (most recent call last):
File "D:\StarSRGAN\starsrgan\train.py", line 3, in
from starsrgan.utils.train import train_pipeline
ModuleNotFoundError: No module named 'starsrgan'

I am not sure why this is or what the correct solution were. It worked after I copied train.py and init.py from the starsrgan folder into the main folder aswell and just ran python train.py -opt optionsfile but not sure if that is a correct solution. Also trained a testmodel for around 19k iter but if I use it with inference_starsrgan.py it produces a black output image, probably because I broke something with copying the two files into their parent folder. (also id rather like to give a path to a specific .pth file as input argument rather than some model name where it searches for a hardcoded net_g_latest.pth file from experiments folder, also it forces a naming convention for the model names, model name needs to start with 'lite' for it to use the correct network/arch this way)

(PS in the Readme the finetune command needs to be adapted since its starsrgan/train.py not realesrgan/train.py, probably from copying)

Yea I simply wanted to try out StarSRGAN, looking forward for your pretrains to try out :)

I cannot seem to get auto_resume to work. It continues to start back at 0. Please advise

What is the purpose of degradation_params in [starsrgan_model.py]?

Hello! Thank you so much for you project but I have a couple of questions.
Do you have any pretrained models? It would be great if you could give access to both starrsnet and starsrgan! :)
Also are there any recommendations for amount of iterations depending on how many GPUs are used in training?

I'm trying to leverage google colab and locally export PYTHONPATH="$PWD" works but I cannot seem to get it to work in Google Colab. Any suggestions would be fantastic! Cheers

I am trying to train StarSRGAN by following all the instructions on the README file so that I can eventually train it using my own dataset for research. I am using images from DIV2K and have my meta_info.txt file, but I did not perform the optional steps to generate multiscale/cropped images. However, when I try to train with a single GPU in debug mode (python starsrgan/train.py -opt options/train_starsrnet.yml --debug), I get the following error about tensor size:

RuntimeError: The expanded size of the tensor (16) must match the existing size (128) at non-singleton dimension 0. Target sizes: [16, 1]. Tensor sizes: [128, 1]

Which is coming from:

self.degradation_params[:, self.road_map[0] : self.road_map[0] + 1] = (data["kernel1"]["kernel_size"].unsqueeze(1) - kernel_size_range1[0]
                ) / (kernel_size_range1[1] - kernel_size_range1[0])

since self.degradation_params[:, self.road_map[0] : self.road_map[0] + 1] is of size [16,1] and (data["kernel1"]["kernel_size"].unsqueeze(1) - kernel_size_range1[0]) / (kernel_size_range1[1] - kernel_size_range1[0]) is of size [128,1].

Has anyone had this issue and resolved it/have any idea what may be causing this? Any help is appreciated.

after training the net and moving to the gan stage I get the following error:
KeyError: "No object named 'PerceptualLoss' found in 'loss' registry!"

Please advise