0. 摘要

因为本机安装过anaconda，python，tensorflow等环境，所以直接按照版本安装部分依赖即可。
后来发现谷歌的colab有免费的gpu可以白嫖，并且colab内置了tensorflow和图形计算框架，所以省去了搭建环境的烦恼。可以直接使用，也很方便。弊端就是不氪金每天能分配的gpu资源不定并且容易断联，session断开需要重新配置，不过图一乐还是可以的。

1. 本地手动安装

全程安装按照官方指南，见 https://github.com/deepfakes/faceswap/blob/master/INSTALL.md。

2. 问题记录

2.1 问题1

问题：import win32console # pylint: disable=import-errorImportError: DLL load failed while importing win32console: 找不到指定的模块
解决：进入“Anaconda3\Scripts”下找到“pywin32_postinstall.py”文件，并执行“python3 pywin32_postinstall.py -install”

2.2 问题2

问题：在使用pip一键安装requirements_base.txt及requirements_nvidia.txt中的依赖的时候速度慢，并且切换到anaconda的新建虚拟环境中windows下无法执行其中的git命令
解决：速度慢修改anaconda的镜像。git无法访问：手动下载git包并将“requirements_base”中的git相关命令注释掉（git慢的话修改git的代理端口）

2.3 问题3

问题：File “D:\faceswap\lib\gui\menu.py”, line 302, in _get_branchesretcode, stdout.decode(‘utf-8’).strip().replace(“\n”, “ - “)) UnicodeDecodeError: ‘utf-8’ codec can’t decode byte 0xb2 in position 6: invalid start byte
解决：查看源码发现该句只是写入日志的内容，直接注掉，成功运行

2.4 问题4

问题：一直无法启动gpu进行运算，猜想是缺少插件CUDA toolkit（英伟达计算架构框架）和cudnn（英伟达架构算法）

证明猜想：

#测试tfGPU是否能够使用：
import tensorflow as tf
import os
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2' 
print('GPU', tf.test.is_gpu_available())
a = tf.constant(2.0)
b = tf.constant(4.0)
print(a + b)

#结果为如下表示可以使用gpu
GPU True
tf.Tensor(6.0, shape=(), dtype=float32)

#但是报错
“Could not load dynamic library 'cusparse64_10.dll'; dlerror: cusparse64_10.dll not found”等dll文件不存在

#需要安装CUDA和cudnn
CUDA：https://developer.nvidia.com/cuda-downloads 
CUDNN: https://developer.nvidia.com/cudnn

2.4.1 安装cuda和cudnn后查看gpu是否启动报错

问题：Could not load dynamic library 'cudnn64_7.dll'; dlerror: cudnn64_7.dll not found

解决方法：Copy <installpath>\cuda\bin\cudnn64_7.dll to C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v9.0\bin.
Copy <installpath>\cuda\ include\cudnn.h to C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v9.0\include.
Copy <installpath>\cuda\lib\x64\cudnn.lib to C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v9.0\lib\x64

3. colab安装

参考：https://github.com/RakaMaru/Faceswap_Google_Colab_Master/blob/master/Faceswap_Google_Colab_Master.ipynb，因为colab的python版本问题又不想修改python版本，进行了部分修改。以下均在colab的笔记本中按顺序执行(前置需要配置一些文件的目录参考上述链接)

#@title Use this to check the assigned GPU
!cat /etc/os-release

def install_dependencies():
  !ln -sf /opt/bin/nvidia-smi /usr/bin/nvidia-smi;
  !pip install gputil;
  !pip install psutil;
  !pip install humanize;%%capture

def printm():
 GPUs = GPU.getGPUs()

 if len(GPUs) == 0:
  print("No GPU available.")
  return

 gpu = GPUs[0]
 process = psutil.Process(os.getpid())
 print("Gen RAM Free: " + humanize.naturalsize( psutil.virtual_memory().available ), " | Proc size: " + humanize.naturalsize( process.memory_info().rss))
 print("GPU RAM Free: {0:.0f}MB | Used: {1:.0f}MB | Util {2:3.0f}% | Total {3:.0f}MB".format(gpu.memoryFree, gpu.memoryUsed, gpu.memoryUtil*100, gpu.memoryTotal))

from IPython.utils import io
from google.colab import drive
import psutil
import humanize
import os
 
with io.capture_output() as captured:
  install_dependencies()
print("Dependencies installed.")

import GPUtil as GPU
printm()

#@title Set Time Zone
!rm /etc/localtime
!ln -s /usr/share/zoneinfo/HST /etc/localtime
!date

#above is for HST, you can find yours in
#/usr/share/zoneinfo

1	! rm -rf ./drive

#@title Mount Google Drive

from google.colab import drive
drive.mount('/content/drive', force_remount=True)

#@title Cleanup folders if needed

!rm -r face_a
!rm -r face_b

#@title Download training data
!cp "/content/drive/My Drive/colab_files/faceswap/faces/face_a.zip" .
!cp "/content/drive/My Drive/colab_files/faceswap/faces/face_b.zip" .

!unzip face_a.zip -d face_a
!unzip face_b.zip -d face_b

!rm face_a.zip
!rm face_b.zip

1 2	#@title Grab the latest Faceswap !git clone --single-branch --branch r1.0 https://github.com/deepfakes/faceswap.git

#@title Copy configuration files
!cp "/content/drive/My Drive/colab_files/faceswap/config/train.ini" faceswap/config/
!ls -lA faceswap/config/
!cat faceswap/config/train.ini

1
2
3

#@title 删除自带的tensorflow2.3安装1.15.0
! pip install tensorflow==1.15.0
! pip install gast==0.2.2

1
2
3

#@title Install Tensorflow
!pip install -r faceswap/requirements_nvidia.txt
#@ 这里会报错“ERROR: albumentations 0.1.12 has requirement imgaug<0.2.7,>=0.2.5, but you'll have imgaug 0.2.9 which is incompatible.”  但没关系

#@title 开始训练，参数根据需要修改
num_iterations = "100000"
save_every = "360"
save_model_every = "25000"
batch_num = "8"
num_gpus = "1"

trainer_type = "dlight"

model_dir = "/content/drive/My Drive/colab_files/faceswap/models/model"
alignments_file_a = "face_a/alignments.fsa"
alignments_file_b = "face_b/alignments.fsa"
timelapse_dir = "/content/drive/My Drive/colab_files/faceswap/output/timelapse"
#@title set variables end

!python3 faceswap/faceswap.py train \
  -A './face_a/faceA' -ala '{alignments_file_a}' \
  -B './face_b/faceB' -alb '{alignments_file_b}' \
  -m '{model_dir}' \
  -t '{trainer_type}' \
  -bs '{batch_num}' \
  -it '{num_iterations}' \
  -g '{num_gpus}' \
  -s '{save_every}' \
  -ss '{save_model_every}' \

#@title 查看结果：
Setting Faceswap backend to NVIDIA
09/26/2020 16:14:02 INFO     Log level set to: INFO
Using TensorFlow backend.
09/26/2020 16:14:04 INFO     Model A Directory: /content/face_a/faceA
09/26/2020 16:14:04 INFO     Model B Directory: /content/face_b/faceB
09/26/2020 16:14:04 INFO     Training data directory: /content/drive/My Drive/colab_files/faceswap/models/model
09/26/2020 16:14:04 INFO     ===================================================
09/26/2020 16:14:04 INFO       Starting
09/26/2020 16:14:04 INFO       Press 'ENTER' to save and quit
09/26/2020 16:14:04 INFO       Press 'S' to save model weights immediately
09/26/2020 16:14:04 INFO     ===================================================
09/26/2020 16:14:05 INFO     Loading data, this may take a while...
09/26/2020 16:14:05 INFO     Loading Model from Dlight plugin...
09/26/2020 16:14:05 INFO     Using configuration saved in state file
09/26/2020 16:14:10 INFO     Loaded model from disk: '/content/drive/My Drive/colab_files/faceswap/models/model'
09/26/2020 16:14:10 INFO     Loading Trainer from Original plugin...
09/26/2020 16:14:12 INFO     Enabled TensorBoard Logging
[16:14:27] [#04386] Loss A: 0.05719, Loss B: 0.05088
09/26/2020 16:14:32 INFO     [Saved models] - Average since last save: face_loss_A: 0.05719, face_loss_B: 0.05088
[16:19:42] [#04746] Loss A: 0.04279, Loss B: 0.03795
09/26/2020 16:19:46 INFO     [Saved models] - Average since last save: face_loss_A: 0.04535, face_loss_B: 0.04580
[16:25:01] [#05106] Loss A: 0.04427, Loss B: 0.04492
09/26/2020 16:25:06 INFO     [Saved models] - Average since last save: face_loss_A: 0.04406, face_loss_B: 0.04483
[16:25:19] [#05120] Loss A: 0.04438, Loss B: 0.03987

3.1 总结

根据官方指导上面的损失函数lossA和lossB分别训练到0.02和0.01就基本差不多了，最后把训练的模型下载到本地上，然后跑一下就行了。虽然是在云端训练的，但是本地似乎还是需要安装相关的环境，因为模型下载下来了本地使用还是需要环境。

云端的好处就是不用占用本机的内存什么的，还有gpu用方便很多速度也蛮快的。

Wednesday`s Blog

本地安装faceswap与一键使用google的colab搭建faceswap两种方式