search

yerfor / GeneFace

GeneFace: Generalized and High-Fidelity 3D Talking Face Synthesis; ICLR 2023; Official code
MIT License
2.44k stars 290 forks source link

训练得到了14个 ckpt_steps，怎么看该用哪个 ckpt_steps呀，请问看哪个维度呀？ #193

Closed tailangjun closed 9 months ago

tailangjun commented 10 months ago

我用tensorboard把日志导入，生成了这样几张图

image image image
tailangjun commented 9 months ago

又读了一遍 train_models-zh.md，发现了这段话 val/mse和 val/sync较小，并且 tr/disc_neg_conf和 tr/disc_pos_conf都约等于0.5（这意味着discriminator已经无法区分正样本和postnet产生的负样本之间的差异）

于是我选择8k步的checkpoint

zqdsdfg commented 1 month ago

请问你怎么导入的呀？