Trained on over 15000 images and still getting lackluster results

carpedm20 / DCGAN-tensorflow

A tensorflow implementation of "Deep Convolutional Generative Adversarial Networks"

http://carpedm20.github.io/faces/

MIT License

7.16k stars 2.62k forks source link

Trained on over 15000 images and still getting lackluster results #361

Open avaseghi opened 5 years ago

avaseghi commented 5 years ago

I'm training the model on over 15000 images but it's still outputting final test images like the one below:

train_00006000

I'm running the following command:

python main.py --data_dir dataset --dataset food --input_height 128 --output_height 128 --train

What can I do to get better results??

WasabiThumb commented 5 years ago

First, ensure the uniformity of your dataset. Dataset "food" likely contains many different images, different kinds of food, etc. If that were the case, I would expect results like this. DCgan figured out background separation, a plate, and a subject.