结果的异常现象

[wbhu]

结果的异常现象

[wbhu]

Hi @baizhengbiao ,

As you say, some results of this model are not as smooth as the original DnCNN model, which confused me too.

I also trained a new model whose parameters are all the same with original model, and the loss converged at about 0.92, which is lower than original model. But these 3 results were still not smooth. Surprisingly, average PSNR of this model on Set12 with 25 noise level is 30.84, while the figure of original DnCNN model is only 30.44. And the PSNR of these 3 images in our new model is 32.74, 32.23, 30.60 respectively.

I will continue to pursue the reasons.

Thanks,
Wenbo

[wbhu]

Additionally, the PSNR of these 3 images in original DnCNN model is 33.06, 32.44, 30.12 respectively.

[ccqu]

The cause would probably be that you're using tf.truncated_normal to generated noisy input images in your model.build_model(). If it is replaced with tf.random_normal as in the original MATLAB code, which is more difficult, the result on Set12 will drop to 29.xx.

[wbhu]

Hi @ccqu ,

Thanks for your reply. You're right that noise generated with tf.truncated_normal is easier than that with tf.random_normal. Now, I am trying to train the model with tf.random_normal noise. We will get the figure soon.

Thanks,
Wenbo

[ccqu]

Hi Wenbo, thanks for your confirmation! I ran the training with tf.random_normal again and got:
img0 PSNR: 29.62
img1 PSNR: 31.96
img2 PSNR: 30.23
img3 PSNR: 28.92
img4 PSNR: 29.92
img5 PSNR: 28.80
img6 PSNR: 28.99
img7 PSNR: 31.13
img8 PSNR: 29.19
img9 PSNR: 29.61
img10 PSNR: 29.46
img11 PSNR: 29.46
--- Test ---- Average PSNR 29.77 ---

[baizhengbiao]

Yeah I have try it again with tf.random_normal .
img0 PSNR: 29.33
img1 PSNR: 28.88
img2 PSNR: 29.69
img3 PSNR: 29.88
img4 PSNR: 29.67
img5 PSNR: 29.75
img6 PSNR: 29.60
img7 PSNR: 32.08
img8 PSNR: 28.98
img9 PSNR: 31.32
img10 PSNR: 29.19
img11 PSNR: 30.44
--- Test ---- Average PSNR 29.90 ---

[wbhu]

Hi @baizhengbiao @ccqu ,

Thanks for all of your contributions.

I have found out the reason of the model's bad performance in these 3 images as @baizhengbiao said. It's a bug in BN layer that causes this problem. I have fixed it, and now this model could achieve average PSNR 30.38 with noise generated with tf.random_normal in Set12, which is quite similar with original DnCNN model.

This is the result:
[] Reading checkpoint...
[] Load weights SUCCESS...
[*] noise level: 25 start testing...
img0 PSNR: 33.07
img1 PSNR: 30.84
img2 PSNR: 30.03
img3 PSNR: 30.14
img4 PSNR: 29.96
img5 PSNR: 29.07
img6 PSNR: 30.05
img7 PSNR: 29.45
img8 PSNR: 30.01
img9 PSNR: 29.26
img10 PSNR: 32.33
img11 PSNR: 30.35
--- Average PSNR 30.38 ---

For more information, plz refer to the updated README and code.

Thanks,
Wenbo

[wbhu]

[ccqu]

Hi @crisb-DUT, thanks for your code update! Just wondering without e.g. the previous clipping code in BN as in the original MATLAB code, the results are much better…

[wbhu]

Hi @ccqu ,

It's not the clipping when initial the BN weights that caused the poor performance in previous model but the not frizzed moving average of beta and gamma when testing. I think it's not that important how we initial BN weights.

[lzzlxxlsz]

你好，想问一下出现这个错误该怎么办呢File "D:\Anaconda\envs\tensorflow\lib\random.py", line 282, in shuffle
x[i], x[j] = x[j], x[i]
TypeError: 'range' object does not support item assignment

[hujiayu712]

@lzzlxxlsz 你好请问您解决这个问题了吗，我也遇到相同的问题，谢谢