FaceDiffuser 部署笔记

依赖项安装：

[win11 ffmpeg合并报错](#win11 ffmpeg合并报错)

修改后代码：

facebook/hubert-base-ls960报错

我的解决方法：

DiffSpeaker网络音频编码器：

头模加载

transformers

依赖项安装：

复制代码

"tokenizers": "tokenizers>=0.10.1,<0.11",

改为：

复制代码

"tokenizers": "tokenizers",

安装：

pip install tokenizers

win11 ffmpeg合并报错

python 复制代码

    ffmpeg.concat(input_video, input_audio, v=1, a=1).output(video_wA_path).run()
    del video, seq, ref_mesh

修改后代码：

python 复制代码

import os

if __name__ == '__main__':
    video_woA_path = os.path.abspath('demo/renders/tmp/0428.mp4')

    out_path=r'output.mp4'
    wav_path = os.path.abspath(r'test.wav')
    cmd = " ".join(['ffmpeg', '-i', video_woA_path,  # 输入视频文件
        '-i', wav_path,  # 输入音频文件
        '-c:v', 'copy',  # 视频编解码器为复制，不进行转码
        '-c:a', 'aac',  # 音频编解码器为AAC
        '-strict', '-2',  # 允许使用实验性AAC编解码器
        '-pix_fmt', 'yuv420p',  # 设置像素格式
        '-q:v', '0',  # 对视频使用最佳质量（无损压缩）
        out_path  # 输出文件名
    ])

    os.system(cmd)

facebook/hubert-base-ls960报错

报错代码：

python 复制代码

from hubert.modeling_hubert import HubertModel


  File "E:\project\audio\audio2face\FaceDiffuser-main\models.py", line 252, in __init__
    self.audio_encoder = HubertModel.from_pretrained("facebook/hubert-base-ls960")
  File "E:\project\audio\audio2face\FaceDiffuser-main\hubert\modeling_utils.py", line 1147, in from_pretrained
    raise EnvironmentError(msg)
OSError: Can't load weights for 'facebook/hubert-base-ls960'. Make sure that:

- 'facebook/hubert-base-ls960' is a correct model identifier listed on 'https://huggingface.co/models'

- or 'facebook/hubert-base-ls960' is the correct path to a directory containing a file named one of pytorch_model.bin, tf_model.h5, model.ckpt.

我的解决方法：

自己下载.bin模型文件

https://huggingface.co/facebook/hubert-base-ls960/tree/main

下载完直接用torch加载.bin模型

修改代码：modeling_utils.py中

python 复制代码

  if state_dict is None:
                try:
                    # state_dict = torch.load(resolved_archive_file, map_location="cpu")
                    state_dict = torch.load(r'pytorch_model.bin', map_location="cpu")
                except Exception:
                    raise OSError(
                        f"Unable to load weights from pytorch checkpoint file for '{pretrained_model_name_or_path}' "
                        f"at '{resolved_archive_file}'"
                        "If you tried to load a PyTorch model from a TF 2.0 checkpoint, please set from_tf=True. "
                    )

DiffSpeaker网络音频编码器：

python 复制代码

audio_encoder:
  train_audio_encoder: True
  model_name_or_path: 'facebook/wav2vec2-base-960h'


from transformers import Wav2Vec2Model
self.audio_encoder = Wav2Vec2Model.from_pretrained(cfg.audio_encoder.model_name_or_path)

头模加载

python 复制代码

import trimesh

if __name__ == '__main__':
    template_file=f"data/BIWI/templates/face_template.obj"
    ref_mesh = trimesh.load_mesh(template_file, process=False)
    ref_mesh.show()
    #下面也ok
    # scene = trimesh.scene.scene.Scene([ref_mesh])

    # 显示场景
    # scene.show()

安装依赖项：

pip install "pyglet<2"