search

carpedm20 / DCGAN-tensorflow

A tensorflow implementation of "Deep Convolutional Generative Adversarial Networks"
http://carpedm20.github.io/faces/
MIT License
7.15k stars 2.62k forks source link

Train the DCGAN using 102 flowers, but it doesn't work well #283

Open jononny opened 6 years ago

jononny commented 6 years ago

Hello, I used the 102 flower database to train the model, after 600 epoch, it even cannot generate something like flowers. How could I do to make it generate more realistic one? Can change the parameters make it better? the parameter I put in python main.py --input_height 96 --input_width 96 --output_height 48 --output_width 48 --dataset jpg --crop --train --epoch 600 --input_fname_pattern "*.jpg"

Please help me with that Thanks!

neelindap commented 6 years ago

Hi,

It would be difficult to train the model with such a small number of images. I am trying out a similar scenario with 100 custom images, and so far nothing looks promising.

I have added another layer though to both the networks and played around with the hyper-parameters. Also epochs need to be really high for a small dataset

jononny commented 6 years ago

Hi, Thank you for replying I had used this database: http://www.robots.ox.ac.uk/~vgg/data/flowers/102/ with about 8191 pictures, is this database too small? if change some parameters could help?

neelindap commented 6 years ago

My bad. I thought you mean 102 flower database as in there were 102 images of flowers. :)

8191 is still less though, as compared to CelebA. You could try getting the epochs up to see if that has any impact.

tuoniaoren commented 6 years ago

my data is 130k,but it generate bad

lzzlxxlsz commented 5 years ago

my resulted picture are so blurred,and seems like the same color.

enochkan commented 5 years ago

@jononny what is the size of your flower images? I am confused about input_height/ input_width vs. output_height/ output_width. If I have 512x512 images should I set them all to 512?