Random filter after learning

[chapaev28]

Hello, guys.

I hope you help me. I'm newbie in deep learning. I solve a problem of super resolution and some days ago i found super resolution using cnn.

repository this solution https://github.com/SolessChong/srcnn-caffe.

I want repeat theirs results. I creating datasets using theirs python and shell scripts. but in this repository dont has solvers configuration.

I create solver conguration like this:

solver_type:ADAGRAD

weight_decay: 0.00005

base_lr: 0.001   

lr_policy: "step" 

gamma: 0.1        

stepsize: 200000 

also i try using nesterov with lr = 0.00001

but result is not improved.

in theirs repository there is script python/vizualize_loss.py. In this script there is a log from caffe train(they have 0.09 loss after 400 k iteration, i have 0.4 - 0.3 loss after some million iteration).

a train this network more than 400 000 iteration and more than 10 millions iteration(about 10 hours of training). and also result is not improved.

after visualization of first layer of convolution i have a random filter like this:

what i can do wrong(i dont know anything about substraction of mean or may be i have bad solver configuration)?