Open sta9 opened 5 years ago
As described in the paper, I also used annealed version of the "L0 loss". The parameter gamma is reduced to 0 while training proceeds.
https://github.com/yu4u/noise2noise/blob/master/model.py#L10-L33
Thank you for answering. But why does not training progress just by changing the image? Even though loss = l0 before the change of the image training was going well.
Sorry but I did not understand you question. Did mean that there is no need for annealing?
Sorry for the incomprehensible question. I need annealing. What I wanted to tell you is that training will not proceed just by changing the image regardless of L0loss.
Thank you for your answer. I got it. Would you pull the repository to use the latest version of this project. I'm not sure whether it solves the problem, but there were a bug that images except ".jpg" images cannot be used.
Thanks a lot!! After changing from '.bmp' to '.jpg', training advanced.
For training, could I use the existing weight to initialize the network weight, how to do it?
I added a new argument to resume training.
https://github.com/yu4u/noise2noise/blob/master/train.py#L45-L46
Thanks for adding them. Should we add anything else?
Nothing.
But it doesn't have any effect on the whole training process. It is not included in main function. For example, if I use the weight file you trained by srresnet, it still can run with unet. There won't be any error in any situation. It seems like this code doesn't exist or get involved.
It is not included in main function
I did not understand what you intended. Weights would be loaded.
https://github.com/yu4u/noise2noise/blob/master/train.py#L75-L76
Loading resnet weights to unet model causes the following error.
ValueError: You are trying to load a weight file containing 85 layers into a model with 23 layers.
I want to change the input to one-dimensional data instead of pictures. What should I do?
Hi, I am trying to train with loss = l0 using my own image instead of image_dir: 291 and test_dir: Set14. Then the following code will appear and training will not proceed.
Epoch 00001: UpdateAnnealingParameter reducing gamma to 2.0.
I don't know what this is going on. How can I help my training to succeed?