Every time, I am so sorry for my bother you. What is different from between 4 times learning and 15 times learning?

[star4s]

Thank you for your time consumption. I checked your new uploaded program. I have a question, one more. Every time, I am so sorry for my bother you. Last time, you made different two type learning data for image.

1.

What is different from between 4 times learning and 15 times learning? In my case, when I apply 15 times learning in the folder, SRnet-v1-color-128-rmsprop, it does not work, well for Super resolution. However, when I apply 4 times learning in the folder, SRnet-v1-color-128, it works, well for Super resolution. Could you tell me, what is the difference?

2.

In addition, when I try doing your new train_SRnet, it takes 15 times learning for training. I think that the training is for SRnet-v1-color-128-rmsprop. Can I take the train for SRnet-v1-color-128, 4 times learning? How can I control for the option? I would like to change only for SRnet-v1-color-128, 4 times learning. Thank you for your attention, every time.

[layumi]

@star4s

1. The network is hard to interpret. In fact, I just found the loss converged in 4 epochs so I stop it. If there is a reason for better result, I early stop the training which may avoid overfitting. But in most case, more training epoch do not harm the result.

2. Just change the learning rate. For example, opts.train.learningRate = [1e-5*ones(1,4) 1e-6*ones(1,1)];