Beyond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising

Train

$ python generate_patches.py
$ python main.py
(note: You can add command line arguments according to the source code, for example
    $ python main.py --batch_size 64 )

For the provided model, it took about 4 hours in GTX 1080TI.

Here is my training loss:

Note: This loss figure isn't suitable for this trained model any more, but I don't want to update the figure 🌚

Test

$ python main.py --phase test

Run

for i in `cat machine.txt`;do echo $i;ssh  -i ./key.pem root@$i "cd /mnt/jinfan/tensorflow/multiMachineDenoise/ && python3 main.py --epoch=10 &";done

j=0;for i in `cat machine.txt`;do echo $i $j;ssh -i ./key-jr-nonprod-huabei2.pem root@$i "cd /mnt/jinfan/tensorflow/multiMachineDenoise/ && (python3 main.py --task_index=$j --job_name='worker' &)";((j+=1));done


nohup

compare

sigma=25, denoise细节消失的很厉害, 细节基本都被抹平了

Restore

使用MonitoredTrainingSession方式跑出的结果，.index和.data文件都保存在ps机器上而.meta文件保存在master worker机器上，restore的时候需要将文件拷贝过去

输出的时候，需要输出self.X, 不然会报错，不知道为什么

TODO

• Write code to support multi machine to learning
• Implement server to support restful API denoise
• Compare with original DnCNN.
• Replace tf.nn with tf.layer.
• Replace PIL with OpenCV.
• Try tf.dataset API to speed up training process.
• Train a noise level blind model.