Justin-Tan
commented
6 years ago Hello, which dataset are you training on, ADE20k? 512 epochs is just a placeholder value. From memory the authors train for 50 epochs on Cityscapes and ADE20k and their results look better than mine on the testing set.
The pre-trained model you give is trained for 50 epochs which defined in the paper. But I found the default num_epochs is defined as 512 in the config.py, why? And the result I obtained trained for 50 epochs with conditional gan and without conditional gan both perform not so good.