Didn't reproduce the result about tensorflow-vdsr HOT 5 CLOSED

• I am not the author of the paper.. this is just a simple implementation of the paper.

• I am using the same framework as you, since I am using the latest docker image from Tensorflow. So it should not be a problem with the environment..

• It looks like the values are very close to the bicubic result.

• Have you uses the checkpoint file provided in the README file? How is the result?

from tensorflow-vdsr.

No, i didn't use the checkpoint file provided in the README file. I just train the model and do evaluate.

from tensorflow-vdsr.

@GavinZhou2014 It seems the network is not trained at all. I need more information to figure this out. Does the loss go down? What modification did you make? How many epochs did you train?

I am kinda busy for a close deadline, so I will get back to you next week.

from tensorflow-vdsr.

Using AdamOptimizer with 1e-4 learning rate improves the results drastically. And it does not require gradient clipping either. I know it's not what they did in the paper but I'm just saying.

from tensorflow-vdsr.

@harungunaydin The author stated they used 0.9 momentum, so I thought they used MomentumOptimizer. Thank you for pointing out new optimizer. I will certainly try that out and add it into the repository.

from tensorflow-vdsr.