LynnHo/AttGAN-Tensorflow

Cannot get a desired result on CelebA-HQ dataset

EvaFlower opened this issue · 4 comments

EvaFlower commented

Hi there,

Your work is interesting. I have a problem. Could you help figure it out?

I applied your method on CelebA-HQ dataset for a single attribute manipulation. But I cannot get the desired result. The result (the interested attribute is "Smiling") at the 59th training epoch is shown as follows. There is no change in the third column images.
attgan

Thanks and Regards,

LynnHo commented

@EvaFlower I haven't applied this code to CelebA-HQ, so I am not sure what's the problem. Have you tried the original CelebA on "Smiling"?

EvaFlower commented

@LynnHo
I have not tried that.
For the limited computation, I applied it on the CelebA-HQ, which contains much fewer samples, to shorten the computation time.
But anyway the CelebA-HQ is a high quality version of the subset of CelebA. Would it be possible due to insufficient samples?

Best,

LynnHo commented

@EvaFlower I will try this dataset in the spare time.

LynnHo commented

@EvaFlower I have tried CelebA-HQ. The results are not good enough, but it still works on 128 images (other resolutions are not tested). Please refer to the new README file.

Since fewer images of CelebA-HQ, more training loops are needed. You can use the command below for "Smiling",

CUDA_VISIBLE_DEVICES=0 \
python train.py \
--img_dir ./data/CelebAMask-HQ/CelebA-HQ-img \
--train_label_path ./data/CelebAMask-HQ/train_label.txt \
--val_label_path ./data/CelebAMask-HQ/val_label.txt \
--load_size 128 \
--crop_size 128 \
--n_epochs 200 \
--epoch_start_decay 100 \
--model model_128 \
--experiment_name AttGAN_128_CelebA-HQ_Smiling \
--att_names Smiling

image