search

aiff22 / DPED

Software and pre-trained models for automatic photo quality enhancement using Deep Convolutional Networks
1.68k stars 370 forks source link

MXNet Gluon version. Need help #12

Open pribadihcr opened 6 years ago

pribadihcr commented 6 years ago

Hi @aiff22,

I port your TF's version to MXNet Gluon https://github.com/pribadihcr/DPED_Gluon. Still the enhanced image is not good as yours. Please help to check the code.

BEST

qicathetine commented 6 years ago

Hi@pribadihcr, I just read your code. May I ask you some questions? Do you train your model by 100*100 image crops(random cropping) and test your model by feeding the entire image? If I am right, what is the meaning of training by image crops? Thank you !!!

pribadihcr commented 6 years ago

hi @qicathetine,

I am following the original TF code, as also mention it in the paper https://arxiv.org/abs/1704.02470, section 2, page. 5.

aiff22 commented 6 years ago

Hi @pribadihcr

Thank you for your work! I guess that the problem is that you are first rescaling the input images to 720x1280 resolution, and then passing them to a neural network. Since it was trained on larger photos, the results in this case might be worse. Thus, the two possible options are 1) to pass them in their orig res defined in utils.get_resolutions() 2) or to crop them to HD size.

qicathetine commented 6 years ago

Hi @Pribadi HCR Do you mean the section of "Generator and Discriminator CNNs"?

pribadihcr commented 6 years ago

@qicathetine Section.2 2 DSLR Photo Enhancement Dataset. Have you read it?

pribadihcr commented 6 years ago

@aiff22 ,

You talk about inference right?. I mean the result still not good in training phase, For training I am trying to save the enhanced image in line 259 im.save('./samples/enhanced_images.jpg').

qicathetine commented 6 years ago

@pribadihcr Is there any possible to add a deconvolution layer in the resnet block? There are always bugs (partial shape )when writing deconvolution layer. Thank you !