I trained a few-shot-gan model with default parameters (trained to 30 kimg). After that I decided this needs more training so changed the resume-pkl to my latest few-shot-gan model snapshot and set the resume-kimg to 30 and total-kimg to 60
Here's the log that it says it loaded my few-shot-gan model
Loading networks from "/content/drive/My Drive/FewShotGAN/Test/00003-stylegan2-Test-Test-config-ada-sv-flat-aug/network-snapshot-000030.pkl"
But the problem is the generated fake-init and subsequent generated image ticks look like it is training completely from scratch why? It wasn't like this when I set the resume-pkl to a vanilla StyleGAN2 model.
How could I possibly resume training with more kimg without training it from scratch (i.e. resuming to an original StyleGAN2 model)?
eps696
commented
3 years ago
I trained a few-shot-gan model with default parameters (trained to 30 kimg). After that I decided this needs more training so changed the
resume-pklto my latest few-shot-gan model snapshot and set theresume-kimgto30andtotal-kimgto60Here's the log that it says it loaded my few-shot-gan modelLoading networks from "/content/drive/My Drive/FewShotGAN/Test/00003-stylegan2-Test-Test-config-ada-sv-flat-aug/network-snapshot-000030.pkl"But the problem is the generated fake-init and subsequent generated image ticks look like it is training completely from scratch why? It wasn't like this when I set the
resume-pklto a vanilla StyleGAN2 model.How could I possibly resume training with more kimg without training it from scratch (i.e. resuming to an original StyleGAN2 model)?