Same attribute classification code as AttGAN. The model is trained on the training set (most images of official training and validation sets) of CelebA, and achieves 94.5% mean accuracy on the 13 attributes we use on the test set (official test set), please refer to the paper for the dataset configuration.

To produce the same result (see this file given in issue #18) as given in the paper, please use the following command to generate the manipulated images since I removed the / thres_int in line 172 of test.py when merging test.py, test_multi.py and test_slide.py into one file before releasing the code.

python test.py --experiment_name 128 --test_int 2
cd att_classification
python test.py --img_dir ../output/128/sample_testing

Note that the images used in the reconstruction image evulation and the user study are also generated by the above command.

Also note that the tfrecord file is generated by LynnHo/TfrecordCreator.