style-based-gan-pytorch
style-based-gan-pytorch copied to clipboard
rosinality
About generate.py and continue the training
Sorry to disturb you again. During the training process, due to some personal reasons, when the checkpoint is 150000.model, the training is interrupted. Before the training is not interrupted, the generated image effect is very good, as shown in Figure 1.
But when the checkpoint is used to continue the training, it is found that the sample is very poor, as shown in Figure 2.
The second problem is that when using the saved checkpoint to run generate.py, the generated image is also very poor, as shown in Figure 3.
Do you know why? I am looking forward to your reply.
After each resolution phase is finished, train_step-{step}.model file will be saved. You can load that checkpoints and train from the next resolutions, for example, if you have finished 64px training, then you can load train_step-5.model and start 128px training.
And you should properly set --size parameters when using generate.py to resolutions that model has been trained.
