jgkwak95 / au-gan Goto Github PK

Firstly, thank you for sharing your wonderful work.
I was about to try your project, but seems there is an unknown issue to the Alderley dataset. It is unavailable to access to the link. Also, I've tried to reach to the website via Google but seems the website itself has issues.
If possible, could you provide a link(e.g. Google drive) to access to the Alderley dataset you have?
Thank you.

Hello, thanks for your work. I don't understand if your implementation has a multi-GPU mode or it has to be implemented from scratch. Thanks

This is a promising project, and could have some use cases, although a better and higher resolution models should be trained, and perhaps the algorithm tweaked.

I was unable to set it up nor run using the provided readme instructions, so here are the adjusted commands that I used to get it up and running on Windows, if the original author is no longer interested in this project, then someone could use this and perhaps fork a new repo and improve the project:

git clone https://github.com/jgkwak95/AU-GAN.git
cd AU-GAN
conda create -y --name augan python=3.6.7
conda activate augan
conda install tensorflow-gpu==1.15.0
pip install --no-cache-dir -r requirements.txt
conda install tensorflow-estimator==1.15.1
mkdir datasets
mkdir datasets\bdd100k
mkdir datasets\bdd100k\testA

place whatever night time images files you have in datasets\bdd100k\testA

can use this pretrained model: https://drive.google.com/file/d/1rvIF3yE9MwPWj0kD4IEstETyMQXYAHzr/view?usp=sharing
unzip to ./check/bdd_exp/bdd100k_256/

now run:
python main.py --dataset_dir bdd100k --phase test --experiment_name bdd_exp --batch_size 1 --load_size 286 --fine_size 256
output is in .\check\bdd_exp\testa2b

@jgkwak95
I am running train.py and test.py on 512x256 pix images.
Each is executed as follows.

python main.py --dataset_dir bdd100k --phase train--experiment_name bdd_exp --batch_size 1 --load_size 258 --fine_size 256

python main.py --dataset_dir bdd100k --phase test --experiment_name bdd_exp --batch_size 1 --load_size 258 --fine_size 256 --use_uncertainty True

I am facing the following problems.

① When I am running --phase train
loss function does not decrease as shown in the picture.
(It's always been nan.)

② When I am running --phase test
the fake_img is not displayed as shown in the picture.

I believe ① and ② are related.
If you have a solution, I would like to know.
I'm looking forward to hearing from you soon.

Thanks @jgkwak95 for the interesting result. However, I cannot download pre-train model. Cloud you fix it?

dear authors, thanks for your contribution· for this filed, but now we still cannot download the pre-trained model, can you reshare it?

I am currently training my own dataset for day-to-night image translation.
However, the generated images produced during training do not accurately depict the transformation from night to day in the original images.
The positions or colors of the generated objects are incorrect.
Could I ask how many epochs were set for the training, or are there other ways to improve this issue?

Hello, can I ask which kind of license does the project comply with? Can you add a license to this project?

Thankyou for your amazing work. I was interested in the pre-trained model.
Is there a pre-trained model available to convert a single image from night -> day.
Thankyou for your time. Looking forward for your response.

Thanks for your great work! I tried to reproduce your work. But I found some strange results with the pre-trained file and BDD10K dataset. @jgkwak95

When the fine_size is set to 512, I got the black results:

When the fine_size is set to 1024, I got the scary results:

Hi, thanks for your nice work :)
I was wondering if I can make a night version of the cityscapes dataset using this model.

or just in case, do u have a night-version cityscapes dataset?