AMP-224-50W+模型

aaaaqtqt · 发表于 2021-8-23 21:42:13

AMP模型自练58W+迭代，换点灵石买镇坛之宝。参数如下
=============== Model Summary ===============
==          Model name: z_AMP          ==
==    Current iteration: 587804          ==
==------------- Model Options -------------==
==          resolution: 224             ==
==          face_type: wf             ==
==    models_opt_on_gpu: True          ==
==             ae_dims: 256             ==
==          inter_dims: 1024          ==
==             e_dims: 64             ==
==             d_dims: 64             ==
==          d_mask_dims: 22             ==
==       morph_factor: 0.5             ==
==          uniform_yaw: True          ==
==          lr_dropout: n             ==
==          random_warp: True          ==
==             ct_mode: none          ==
==             clipgrad: False          ==
==    autobackup_hour: 0             ==
== write_preview_history: False          ==
==          target_iter: 0             ==
==    random_src_flip: False          ==
==    random_dst_flip: True          ==
==          batch_size: 14             ==
==          gan_power: 0.0             ==
==       gan_patch_size: 28             ==
==             gan_dims: 16             ==
==                                        ==

有几点需要说明：
1、该模型为从0开始自练，train AMP SRC-SRC阶段为提高速度，设置的BS8，其他好像都是默认，3天多练到50W次迭代，然后按教程已删除inter_dst.npy。
2、继续正常训练（train AMP）8W+迭代，目前实际迭代58W+，经实测win10系统2080可开BS16但有OOM风险，建议BS14或12，开GAN后BS12~8均可。
3、58W迭代时测试过一次合成效果，还行，后期关掉随机翻转和开GAN后效果应该会更好。

链接：https://pan.baidu.com/s/1PyoGFzvR18zYsog-XfD6zA

购买主题 已有 4 人购买 本主题需向作者支付 30 灵石 才能浏览