terryoo / aindnet Goto Github PK

Your paper describes that the noise estimator outputs two noise-level maps of sigma_4 (?.H/4,W/4,,3) and sigma_1 (?,H,W,3) for computing the weighted average for feeding into AIN-ResBlocks as well as for computing the asymmetric loss.

However, in your code, the noise estimator appears to output two noise-level maps of sigma_8 (?,H/8,W/8,3) and sigma_1 (?,H,W,3) for computing the weighted average and I guess also for computing the asymmetric loss (couldn't confirm this because your training code is not available).

Would appreciate if you could give a comment on this discrepancy (sigma_4 vs sigma_8).

Your work is very good, can you provide the training code

Thank you for your awesome code!

I am hoping you might open-source the log files you have from training. Maybe the training and validation loss as a function of epoch
(and/or batch) with an estimate of the runtime?

I just run the code to test the result on the DnD dataset. Here are the results. I just want to know what is the problem.
0001_01_denoised.png

Hi, thanks for your amazing work.

I want to compare my work with yours. Can you provide the number of Parameters and FLOPs of your model?

Nice work.

Could you release your training code?

In your paper, you mentioned that the weight term of the noise-level estimator is empirically determined to 0.05 (i.e., lambda_ms-asymm = 0.05).

Isn't this typo for lambda_ms-asymm = 0.5 ???

In your following code, def ain():

noisemap_beta = tf.layers.conv2d(tmp, channels, [3, 3],                                          kernel_initializer=tf.contrib.layers.variance_scaling_initializer(0.02),
                                          padding='same',
                                          name='conv_beta')

'tmp' needs to be changed into 'noisemap_gamma' ??
Otherwise, 'noisemap_gamma' and 'noisemap_beta' are the same...

Good idea and works, but I can't run this code, the checkpoint file is too small and restore error! Is the file named "checkpoint" in SIDDtransfer directory the right file?

Could you disclose the training script for your AINDNet for both gaussian and rn?