From 7d697ad0a5c315fe4a7191648608b744e0299d1d Mon Sep 17 00:00:00 2001 From: iperov Date: Fri, 20 Aug 2021 18:49:11 +0400 Subject: [PATCH] upd FAQ, temporary remove AMP training instruction --- doc/faq/faq.md | 14 +------------- 1 file changed, 1 insertion(+), 13 deletions(-) diff --git a/doc/faq/faq.md b/doc/faq/faq.md index 5a0120d..c7b95ba 100644 --- a/doc/faq/faq.md +++ b/doc/faq/faq.md @@ -60,20 +60,8 @@ You can reuse this model to train new src faceset. In this case you should to de > Using AMP model. +-- coming soon -- -Do not mix different ages of people in the src dataset, otherwise the model will approximate age and make the celebrity older/younger depending on the input person. - -res:224, WF, ae_dims:256, inter_dims:1024, e_dims:64, d_dims:64, d_mask_dims:22, morph factor:0.5, batch more is better. Others by default. - -Make a backup before every stage ! - -Assuming 1kk iters with batch 8. If the batch is higher, iters is possible less. - -1) lrd:Y, train src-src for 1kk iters -2) delete inter_dst, lrd:N, uniform_yaw:True, color_transfer:lct, train +500..1kk -3) color_transfer:none, lrd:Y, train +500..1kk -4) random_warp:False, train +500..1kk -5) enable gan 0.1 gan_dims:32, train +100..300k iters