Hi,
I am trying to reproduce the red graph provided in the paper which shows the Average PSNR vs epochs for the Set68:
I am using the Keras implementation provided here (it includes Residual Learning and Batch Normalization, and Adam optimization as well), and I am training 50 epochs.
However, instead of a stable graph like the one in the paper, I get a graph with pronounced drops, similar to the blue and black graphs in the paper. Furthermore, even in the last epochs, where the network should be converging, the Average PSNR often shows significant drops, which are related to the presence of odd spots in some of the denoised images:
I have experimented with different settings, in particular with different BN momentum, and the situation becomes better when it is set to a high value, as 0.9. Even so, the graph I get is never something similar to the red one in the paper. I am afraid these spots in the denoised images appear quite randomly, with different shapes and locations for different trainings.
Has anyone experienced the same or something similar? I would appreciate any help, thank you!!
Hi, I am trying to reproduce the red graph provided in the paper which shows the Average PSNR vs epochs for the Set68:![psnrvsepoch](https://user-images.githubusercontent.com/56032684/71247410-0566d000-2319-11ea-993a-807eef643a0f.png)
I am using the Keras implementation provided here (it includes Residual Learning and Batch Normalization, and Adam optimization as well), and I am training 50 epochs.
However, instead of a stable graph like the one in the paper, I get a graph with pronounced drops, similar to the blue and black graphs in the paper. Furthermore, even in the last epochs, where the network should be converging, the Average PSNR often shows significant drops, which are related to the presence of odd spots in some of the denoised images:
![test001_dncnn](https://user-images.githubusercontent.com/56032684/71249945-ac9a3600-231e-11ea-8ac0-f752a8fd3ca2.png)
I have experimented with different settings, in particular with different BN momentum, and the situation becomes better when it is set to a high value, as 0.9. Even so, the graph I get is never something similar to the red one in the paper. I am afraid these spots in the denoised images appear quite randomly, with different shapes and locations for different trainings.
Has anyone experienced the same or something similar? I would appreciate any help, thank you!!