tuvovan / NHDRRNet

Keras Implementation of the paper Deep HDR Imaging via A Non-Local Network - TIP 2020
MIT License
51 stars 6 forks source link

Shared weight #1

Closed endinblue closed 4 years ago

endinblue commented 4 years ago

Hello! I have a question that you use shared-weight convolutional layer. Is there are any reason you use share-weight layers? I can't not find any mention about this in paper. Thankyou!

tuvovan commented 4 years ago

Hi, Do you mean the encoder parts?

endinblue commented 4 years ago

Yes, share-weight layers in encoder part.

tuvovan commented 4 years ago

hi, I used different encoders for different input images. you can check it in the HDR part.

endinblue commented 4 years ago

Hi. Thank you for your answer. Maybe, I misunderstand the code because I am not used to Keras. I am so sorry.

I think this code means :

X_i_32 = self.encoder_1(X_i, 1) X_i_64 = self.encoder_1(X_i_32, 2) X_i_128 = self.encoder_1(X_i_64, 4) X_i_256 = self.encoder_1(X_i_128, 8)

using shared-weight encoder layer for each same LDR.

I think it need to be write as

X_i_32 = self.encoder_1_1(X_i, 1) X_i_64 = self.encoder_1_2(X_i_32, 2) X_i_128 = self.encoder_1_3(X_i_64, 4) X_i_256 = self.encoder_1_4(X_i_128, 8)

tuvovan commented 4 years ago

Ah,

I got your point. Yeah, I'm using same encoder (shared-weight as you say) for each input images.

One reason is that I tried to make the model to have the same number of parameters as stated in the paper. the other one is that I think sharing the weight for different levels (in one input image) makes sense because they are just feature of the input images.

Please correct me if I'm wrong or please share your progress if you try with different encoder for different level of encoding.

Thanks!

endinblue commented 4 years ago

Hello. Thank you for your answer. I try to make code with pytorch. Actually I made it And I find your code with tensorflow and keras. So I just check that is there any difference with mine. haha

I mean my question is just curious about code. I also don't know exact code of paper.

Again, thank you for your answer. Thank you.

tuvovan commented 4 years ago

have you tried to run some metrics? I mean PSNR or HDR-VDP? Have you reached the results are shown in the paper? If so, please share with me ^^

endinblue commented 4 years ago

I actually build this code for my paper. But this model is not for comparison only just for ablation study. So, I trained it with my own dataset. So I don't compare the results with paper. And Actually Paper said it's HDR's range in [0,1] but I use HDRformat (RGBE) for groundtruth. So, my network's groundtruth not in that range it also different.

tuvovan commented 4 years ago

Hope to see your paper soon.