search

jjonak09 / DRAGAN-keras

1 stars 0 forks source link

DRAGAN + EDSR parameter #1

Open duanzhiihao opened 5 years ago

duanzhiihao commented 5 years ago

Hi, I'm currently trying to reproduce your result. While using your DRAGAN + EDSR generator model, it often collapse while training. Your code is great, and your result is realy cool. Could I know the training parameters you use in training? parser.add_argument("--batchsize", default=128, type=int, help="batch size") parser.add_argument("--lam", default=0.5, type=float, help="the weight of regularizer") z_dim = 128 img_shape = (64, 64, 3) image_size = 64 channels = 3 lr_D = 2e-4 lr_G = 2e-4 b1 = 0.5 b2 = 0.99 I'm using these parameters running on 64*64 anime character images, dataset of 21,000 images. Thank you so much!

jjonak09 commented 5 years ago

Hello, It is very important to create hight quality dataset for good results. This paper is very useful. (https://makegirlsmoe.github.io/assets/pdf/technical_report.pdf)

If you use 64x64 images, z_dim = 100, __lam = 10.0 is better. And , if you use EDSR model , you should increase the image size(128x128 is better ?)Because of the depth of the layer of model.

My best wishes.

duanzhiihao commented 5 years ago

I resize the input images to 128*128, then it works quite well, similar to your posted results. Thank you for your tips!