本帖最后由 安妮不安静 于 2023-11-14 14:44 编辑
1、简单科普
AMP在SAEHD的基础上删减和引入了新的参数,使其更加适合于实时换脸的场景。如果你需要能灵活设置参数,实现高质量的视频换脸,那么肯定用SAEHD。如果你训练模型是为了应用在DeepFaceLive中实现实时换脸,那么推荐AMP。当然,SAEHD其实也是可以用于DeepFaceLvie。
2、AMP模型训练流程
先对大量人脸进行训练,然后再训练具体的人。为了实现这个操作需要如下步骤: 1)将24.35G丰富光源和各类角度的面部RTM WF faceset数据集放到源目录的aligned下面。【OK】 2)点击“ 6) 训练AMP模型 源对源 train AMP SRC-SRC.bat ” 开始训练,训练了570万次。【OK】 3)删除模型文件夹中的_AMP_inter_dst.npy文件【OK】 4. 然后再复制数据到scr和目标dst的aligned文件夹,应用源和目标的遮罩,点击"6) 训练AMP模型 train AMP"进行训练。【待你炼制你的女神】 3、本模型的意义【好模型从不贱卖,只给有缘人】 节约三周以上时间(此模型用Tesla A10 连续跑了两周,显卡都要跑冒烟。。。。哎,好心疼!),拿来就可以站在570万的肩膀上,稍微训练一下就直达600万迭代,获得更好的直播DFM模型
============= Model Summary ============== == == == Model name: AMPSRCSRC_AMP == == == == Current iteration: 5729109 == == == ==----------- Model Options ------------== == == == resolution: 224 == == face_type: wf == == models_opt_on_gpu: True == == ae_dims: 512 == == inter_dims: 1024 == == e_dims: 64 == == d_dims: 64 == == d_mask_dims: 22 == == morph_factor: 0.5 == == uniform_yaw: False == == blur_out_mask: False == == 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: False == == batch_size: 8 == == gan_power: 0.0 == == gan_patch_size: 28 == == gan_dims: 16 == == == ==------------- Running On -------------== == == == Device index: 0 == == Name: NVIDIA A10 == == VRAM: 20.52GB == == == ========================================== |