Environment: TFv1.3.0-rc1-1302-g593dc8e. Tesla P100.
Code: DCGAN-tensorflow at commit b13830.
- DCGAN-tensorflow:
python main.py --dataset celebA --train --crop
This command takes 0.36s per iteration, where each iteration is 1 update to D and 2 updates to G.
Modify the code to use SeparateGANTrainer(..., d_period=2), and run with:
python DCGAN.py --data /path/to/img_align_celebA --crop-size 108 --batch 64
This script runs at 15.5 it/s, where every two iterations is equivalent to 1 iteration in DCGAN-tensorflow. Therefore this script is roughly 2.8x faster.