jessemelpolio / non-stationary_texture_syn Goto Github PK

Hi, I really appreciate your model, it gives amazing results!

However, I would like to know how can we visualize the intermediate result of the residual blocks? Each image in figure 6 is just 1 random channel of the 256 channels?

Good job! It shocks me that GAN can generate so clear and vivid textures. Could you please release the trained model?

Hi, I really appreciate your code, but it for some unknown reason, the complier complains about the connection problem all the time.

Exception in user code:

Traceback (most recent call last):
File "D:\Anaconda\envs\Py35\lib\site-packages\urllib3\connection.py", line 171, in _new_conn
(self._dns_host, self.port), self.timeout, **extra_kw)
File "D:\Anaconda\envs\Py35\lib\site-packages\urllib3\util\connection.py", line 79, in create_connection
raise err
File "D:\Anaconda\envs\Py35\lib\site-packages\urllib3\util\connection.py", line 69, in create_connection
sock.connect(sa)
ConnectionRefusedError: [WinError 10061] No connection could be made because the target machine actively refused it

This error shows up at 13th line at train.py:

dataset = data_loader.load_data()

It's really strange, because I think this program has nothing to do with network connections, isn't it?

Hi,

In your paper you mention diversity augmentation by giving shuffled tiles of the original image to the generator ; have you implemented this anywhere in github?
Thank you.

@jessemelpolio Hello, I found that scripts/train_half_style.sh only takes one image (i.e. datasets/half/202/train/202.jpg) as your training data. In the testing phase, this trained model can only take the same image (i.e. dataset/half/202/test/202.jpg) as input and output a bigger one. What I want to know is whether this already trained model is only valid for this picture. If I want to test other image, does the model need to be retrained？I think this is a bit like ZSSR.

Lines of codes are changed to work on pytorch 3.9.0 and cuda 11.3 with python 3.7:
https://github.com/Alhasan-Abdellatif/non-stationary_texture_syn

I think this dataset for train is a bit small. Do you think so?

bethgelab is refusing the connection and I am not able to download the pre trained model. Can you please provide the file or make the link working? Or please suggest any other workaround.

Thanks a lot

Hi,

I'm running the code using Linux with a Nvidia P100 card (16GB memory) and Pytorch 0.4. It's okay to train small images that's around the size 256x256 to generate 512x512. However, when I train using an image that's 500x500, it threw me an out-of-memory error below during validation stage. I think 16GB should be sufficient for this image size, do you happen to have any idea why this happens?

Traceback (most recent call last):
File "train.py", line 66, in
test_func(opt, webpage, epoch=str(epoch))
File "/home/yuzhang/non-stationary_texture_syn/test_function.py", line 48, in test_func
model.test()
File "/home/yuzhang/non-stationary_texture_syn/models/test_model.py", line 38, in test
self.fake_B = self.netG.forward(self.real_A)
File "/home/yuzhang/non-stationary_texture_syn/models/networks.py", line 221, in forward
return self.model(input)
File "/home/yuzhang/anaconda3/envs/pytorch_3.5/lib/python3.6/site-packages/torch/nn/modules/module.py", line 477, in call
result = self.forward(*input, **kwargs)
File "/home/yuzhang/anaconda3/envs/pytorch_3.5/lib/python3.6/site-packages/torch/nn/modules/container.py", line 91, in forward
input = module(input)
File "/home/yuzhang/anaconda3/envs/pytorch_3.5/lib/python3.6/site-packages/torch/nn/modules/module.py", line 477, in call
result = self.forward(*input, **kwargs)
File "/home/yuzhang/anaconda3/envs/pytorch_3.5/lib/python3.6/site-packages/torch/nn/modules/conv.py", line 301, in forward
self.padding, self.dilation, self.groups)
RuntimeError: CUDA error: out of memory

It seems that there are only pre-trained models in “download_pretrained_models.sh”. Can you share your trained models?

Hi, I have a question that is not directly related to your algorithm. It's about the experiments in your paper.

I notice you compared your result with Self-tuning and MGANs in figure 7. As far as I know, the code for MGANs can not be used for texture synthesis, although they have these results in their paper, and there are some bugs in the code of self-tuning methods. How do you run their codes?

Is there any difference between (style_loss+perceptual_loss+gan_G).backward() and style_loss.backward(retain_graph=True) + (perceptual_loss+gan_G).backward() ?
thanks

Do you think it would be possible to train this GAN. And use the discrimator to detect the same surface with defects ?

Sry to use issues ;)

Hi,
I was trying to download you pretrained models via the download_pretrained_models.sh script, but the resource (as well as the entire domain) seems unavailable. Is it possible to download the models using another source?
Than you in advance for the answer and for the contribution!

Hi,
I'm facing this problem when trying to run "train.py". It asks me to put the dataroot parameter but whatever the directory it always gives me this problem because it doesn't find the "A" and "B" folders.
What am I missing?

I'm training the model on the following options:
python train.py --dataroot ./datasets/half/1 --name 1_half_style_14x14 --use_style --no_lsgan --padding_type replicate --model half_style --which_model_netG resnet_2x_6blocks --which_model_netD n_layers --n_layers_D 4 --which_direction AtoB --lambda_A 100 --dataset_mode half_crop --norm batch --pool_size 0 --resize_or_crop no --niter_decay 50000 --niter 50000 --save_epoch_freq 2000 --display_id 0 --gpu_ids 0
Once the model finishes training, it seems that the image is expanded as it just doubles the resolution, without adding patterns to the surroundings as it seems on your paper.
I've tried with images without patterns and images with repetitive macro patterns.

sorry

Hi, this is more of a question, than an issue.
In the definition backward_G, we take gram_matrix of the real_B, but not of the targets. So to calculate loss we compare gram_matrix of source to simple (not gram_matrix) of targets. Any reason for that?