CycleGAN

最新の20件

私的AI研究会 > CycleGAN

画像スタイル変換：CycleGAN †

　AI技術「GAN（敵対的生成ネットワーク）」を試してみる（その２）。「CycleGAN」で画像のスタイル変換をおこなう。

動作環境の準備と設定 †

事前準備 †

• オフィシャルサイト CycleGAN and pix2pix in PyTorch をもとに進める
  
  

• プロジェクトファイルをダウンロードしておく
  update_20231103.zip (165MB) <アップデート・データ>
  

• 以下、Windows 環境での表記とする
  Linux 環境の場合は「/anaconda_win/work」→「~/work」と読み替えること
  

Python3.8 仮想環境「py38_gan」を使用する †

• 未構築の場合は StarGan-V2 仮想環境「py38_gan」を構築 必要なパッケージをインストール を実行
  

• 「py38_gan」 仮想環境をアクティベートする
  

  conda activate py38_gan

「CycleGan」の導入 †

1. オフィシャルサイト から pytorch-CycleGAN-and-pix2pix一式をダウンロード
   

   (py38_gan) PS > cd /anaconda_win/work/
   (py38_gan) PS > git clone https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix.git

   ▼「git clone https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix.git」ログ
   ▲「git clone https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix.git」ログ

   (py38_gan) PS C:\anaconda_win> cd /anaconda_win/work/
   (py38_gan) PS C:\anaconda_win\work> git clone https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix.git
   Cloning into 'pytorch-CycleGAN-and-pix2pix'...
   remote: Enumerating objects: 2513, done.
   Receiving objects:  95% (2388/2513), 2.45 MiB | 4.50 MiB/sremote: Total 2513 (deReceiving objects:  97% (2438/2513), 2.45 MiB | 4.50 MiB/s
   Receiving objects: 100% (2513/2513), 8.20 MiB | 9.10 MiB/s, done.
   Resolving deltas: 100% (1575/1575), done.

   
   
2. 学習済みモデルのダウンロード
   
   Linux 環境の場合
   

   bash ./scripts/download_cyclegan_model.sh style_ukiyoe
   bash ./scripts/download_cyclegan_model.sh style_monet
   bash ./scripts/download_cyclegan_model.sh style_cezanne
   bash ./scripts/download_cyclegan_model.sh style_vangogh

   Windows 環境の場合
   ・上記の『事前準備』でダウンロードしたファイルを解凍し「work/」フォルダ以降を上書き配置する
   
   学習済みモデルの場所 → 「checkpoints/style_XXXXXXX_pretrained/latest_net_G.pth」

1. 「CycleGan」実行に必要な不足しているパッケージをインストール
   

   (py38_gan) PS > pip install dominate visdom

   ▼「pip install dominate visdom」ログ
   ▲「pip install dominate visdom」ログ

   (py38_gan) PS C:\anaconda_win\work> pip install dominate visdom
   Collecting dominate
     Obtaining dependency information for dominate from https://files.pythonhosted.org/packages/a4/aa/d202166ecad0349fb57abf21dd3ad90d446454710e245ab3636c2cb8f5a0/dominate-2.8.0-py2.py3-none-any.whl.metadata
     Downloading dominate-2.8.0-py2.py3-none-any.whl.metadata (13 kB)
   Collecting visdom
     Downloading visdom-0.2.4.tar.gz (1.4 MB)
        ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 1.4/1.4 MB 7.5 MB/s eta 0:00:00
     Preparing metadata (setup.py) ... done
   Requirement already satisfied: numpy>=1.8 in c:\users\izuts\anaconda3\envs\py38_gan\lib\site-packages (from visdom) (1.24.4)
   Requirement already satisfied: scipy in c:\users\izuts\anaconda3\envs\py38_gan\lib\site-packages (from visdom) (1.9.0)
   Requirement already satisfied: requests in c:\users\izuts\anaconda3\envs\py38_gan\lib\site-packages (from visdom) (2.31.0)
   Requirement already satisfied: tornado in c:\users\izuts\anaconda3\envs\py38_gan\lib\site-packages (from visdom) (6.3.3)
   Requirement already satisfied: six in c:\users\izuts\appdata\roaming\python\python38\site-packages (from visdom) (1.14.0)
   Collecting jsonpatch (from visdom)
     Obtaining dependency information for jsonpatch from https://files.pythonhosted.org/packages/73/07/02e16ed01e04a374e644b575638ec7987ae846d25ad97bcc9945a3ee4b0e/jsonpatch-1.33-py2.py3-none-any.whl.metadata
     Downloading jsonpatch-1.33-py2.py3-none-any.whl.metadata (3.0 kB)
   Collecting websocket-client (from visdom)
     Obtaining dependency information for websocket-client from https://files.pythonhosted.org/packages/c4/3c/1892ce394828c43d4f65248ebdee3854114266b75d1f5915cb211155ad7b/websocket_client-1.6.4-py3-none-any.whl.metadata
     Downloading websocket_client-1.6.4-py3-none-any.whl.metadata (7.7 kB)
   Requirement already satisfied: networkx in c:\users\izuts\anaconda3\envs\py38_gan\lib\site-packages (from visdom) (3.1)
   Requirement already satisfied: pillow in c:\users\izuts\anaconda3\envs\py38_gan\lib\site-packages (from visdom) (10.0.1)
   Collecting jsonpointer>=1.9 (from jsonpatch->visdom)
     Obtaining dependency information for jsonpointer>=1.9 from https://files.pythonhosted.org/packages/12/f6/0232cc0c617e195f06f810534d00b74d2f348fe71b2118009ad8ad31f878/jsonpointer-2.4-py2.py3-none-any.whl.metadata
     Downloading jsonpointer-2.4-py2.py3-none-any.whl.metadata (2.5 kB)
   Requirement already satisfied: charset-normalizer<4,>=2 in c:\users\izuts\anaconda3\envs\py38_gan\lib\site-packages (from requests->visdom) (3.3.0)
   Requirement already satisfied: idna<4,>=2.5 in c:\users\izuts\anaconda3\envs\py38_gan\lib\site-packages (from requests->visdom) (2.10)
   Requirement already satisfied: urllib3<3,>=1.21.1 in c:\users\izuts\anaconda3\envs\py38_gan\lib\site-packages (from requests->visdom) (2.0.6)
   Requirement already satisfied: certifi>=2017.4.17 in c:\users\izuts\anaconda3\envs\py38_gan\lib\site-packages (from requests->visdom) (2023.7.22)
   Downloading dominate-2.8.0-py2.py3-none-any.whl (29 kB)
   Downloading jsonpatch-1.33-py2.py3-none-any.whl (12 kB)
   Downloading websocket_client-1.6.4-py3-none-any.whl (57 kB)
      ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 57.3/57.3 kB ? eta 0:00:00
   Downloading jsonpointer-2.4-py2.py3-none-any.whl (7.8 kB)
   Building wheels for collected packages: visdom
     Building wheel for visdom (setup.py) ... done
     Created wheel for visdom: filename=visdom-0.2.4-py3-none-any.whl size=1408308 sha256=6e4a53fb04322c40c514bfca45ff93a990c84bde26ba719608603c4b40d13ad7
     Stored in directory: c:\users\izuts\appdata\local\pip\cache\wheels\fb\b1\fc\b05c2c1930a412f01bd07dacaeb5fd8cc4bcccf71c835b0281
   Successfully built visdom
   Installing collected packages: websocket-client, jsonpointer, dominate, jsonpatch, visdom
   Successfully installed dominate-2.8.0 jsonpatch-1.33 jsonpointer-2.4 visdom-0.2.4 websocket-client-1.6.4

CycleGan の実行 †

準備 †

1. 「py38_gan」 仮想環境下で実行する（異なる環境の場合、下記コマンドでアクティベートする）
   

   conda activate py38_gan

   
   
2. カレントディレクトリを移動する
   

   (py38_gan) PS > cd /anaconda_win/work/pytorch-CycleGAN-and-pix2pix/

   
   
3. 入力画像「datasets/images/」
   
   
   
   
   
   
   
   
   
4. 出力（結果）画像「results/style_XXXXXXXX_pretrained/test_latest/images/」（コマンド実行後に作成される）
   「results/style_XXXXXXXX_pretrained/test_latest/index.html」ファイルにより Webブラウザ上で入出力画像の一覧を見ることができる
   

浮世絵風 †

• コマンド
  

  (py38_gan) PS > python test.py --dataroot datasets/images --name style_ukiyoe_pretrained --model test --no_dropout --preprocess scale_width --load_size 1024 --gpu_ids -1

  
  
• 実行結果
  
  
  
  
  
  
  
  
  
  ▼「python test.py --dataroot datasets/images --name style_ukiyoe_pretrained ...」ログ
  ▲「python test.py --dataroot datasets/images --name style_ukiyoe_pretrained ...」ログ

  (py38_gan) PS C:\anaconda_win\work\pytorch-CycleGAN-and-pix2pix> python test.py --dataroot datasets/images --name style_ukiyoe_pretrained --model test --no_dropout --preprocess scale_width --load_size 1024 --gpu_ids -1
  Warning: wandb package cannot be found. The option "--use_wandb" will result in error.
  Warning: wandb package cannot be found. The option "--use_wandb" will result in error.
  ----------------- Options ---------------
               aspect_ratio: 1.0
                 batch_size: 1
            checkpoints_dir: ./checkpoints
                  crop_size: 256
                   dataroot: datasets/images                      [default: None]
               dataset_mode: single
                  direction: AtoB
            display_winsize: 256
                      epoch: latest
                       eval: False
                    gpu_ids: -1                                   [default: 0]
                  init_gain: 0.02
                  init_type: normal
                   input_nc: 3
                    isTrain: False                                [default: None]
                  load_iter: 0                                    [default: 0]
                  load_size: 1024                                 [default: 256]
           max_dataset_size: inf
                      model: test
               model_suffix:
                 n_layers_D: 3
                       name: style_ukiyoe_pretrained              [default: experiment_name]
                        ndf: 64
                       netD: basic
                       netG: resnet_9blocks
                        ngf: 64
                 no_dropout: True                                 [default: False]
                    no_flip: False
                       norm: instance
                   num_test: 50
                num_threads: 4
                  output_nc: 3
                      phase: test
                 preprocess: scale_width                          [default: resize_and_crop]
                results_dir: ./results/
             serial_batches: False
                     suffix:
                  use_wandb: False
                    verbose: False
         wandb_project_name: CycleGAN-and-pix2pix
  ----------------- End -------------------
  dataset [SingleDataset] was created
  initialize network with normal
  model [TestModel] was created
  loading the model from ./checkpoints\style_ukiyoe_pretrained\latest_net_G.pth
  ---------- Networks initialized -------------
  [Network G] Total number of parameters : 11.378 M
  -----------------------------------------------
  creating web directory ./results/style_ukiyoe_pretrained\test_latest
  processing (0000)-th image... ['datasets/images\\20210117_091324_001m.jpg']

モネ風 †

• コマンド
  

  (py38_gan) PS > python test.py --dataroot datasets/images --name style_monet_pretrained --model test --no_dropout --preprocess scale_width --load_size 1024 --gpu_ids -1

  
  
• 実行結果
  
  
  
  
  
  
  
  
  
  ▼「python test.py --dataroot datasets/images --name style_monet_pretrained ...」ログ
  ▲「python test.py --dataroot datasets/images --name style_monet_pretrained ...」ログ

  (py38_gan) PS C:\anaconda_win\work\pytorch-CycleGAN-and-pix2pix> python test.py --dataroot datasets/images --name style_monet_pretrained --model test --no_dropout --preprocess scale_width --load_size 1024 --gpu_ids -1
  Warning: wandb package cannot be found. The option "--use_wandb" will result in error.
  Warning: wandb package cannot be found. The option "--use_wandb" will result in error.
  ----------------- Options ---------------
               aspect_ratio: 1.0
                 batch_size: 1
            checkpoints_dir: ./checkpoints
                  crop_size: 256
                   dataroot: datasets/images                      [default: None]
               dataset_mode: single
                  direction: AtoB
            display_winsize: 256
                      epoch: latest
                       eval: False
                    gpu_ids: -1                                   [default: 0]
                  init_gain: 0.02
                  init_type: normal
                   input_nc: 3
                    isTrain: False                                [default: None]
                  load_iter: 0                                    [default: 0]
                  load_size: 1024                                 [default: 256]
           max_dataset_size: inf
                      model: test
               model_suffix:
                 n_layers_D: 3
                       name: style_monet_pretrained               [default: experiment_name]
                        ndf: 64
                       netD: basic
                       netG: resnet_9blocks
                        ngf: 64
                 no_dropout: True                                 [default: False]
                    no_flip: False
                       norm: instance
                   num_test: 50
                num_threads: 4
                  output_nc: 3
                      phase: test
                 preprocess: scale_width                          [default: resize_and_crop]
                results_dir: ./results/
             serial_batches: False
                     suffix:
                  use_wandb: False
                    verbose: False
         wandb_project_name: CycleGAN-and-pix2pix
  ----------------- End -------------------
  dataset [SingleDataset] was created
  initialize network with normal
  model [TestModel] was created
  loading the model from ./checkpoints\style_monet_pretrained\latest_net_G.pth
  ---------- Networks initialized -------------
  [Network G] Total number of parameters : 11.378 M
  -----------------------------------------------
  creating web directory ./results/style_monet_pretrained\test_latest
  processing (0000)-th image... ['datasets/images\\20210117_091324_001m.jpg']

セザンヌ風 †

• コマンド
  

  (py38_gan) PS > python test.py --dataroot datasets/images --name style_cezanne_pretrained --model test --no_dropout --preprocess scale_width --load_size 1024 --gpu_ids -1

  
  
• 実行結果
  
  
  
  
  
  
  
  
  
  ▼「python test.py --dataroot datasets/images --name style_cezanne_pretrained ...」ログ
  ▲「python test.py --dataroot datasets/images --name style_cezanne_pretrained ...」ログ

  (py38_gan) PS C:\anaconda_win\work\pytorch-CycleGAN-and-pix2pix> python test.py --dataroot datasets/images --name style_cezanne_pretrained --model test --no_dropout --preprocess scale_width --load_size 1024 --gpu_ids -1
  Warning: wandb package cannot be found. The option "--use_wandb" will result in error.
  Warning: wandb package cannot be found. The option "--use_wandb" will result in error.
  ----------------- Options ---------------
               aspect_ratio: 1.0
                 batch_size: 1
            checkpoints_dir: ./checkpoints
                  crop_size: 256
                   dataroot: datasets/images                      [default: None]
               dataset_mode: single
                  direction: AtoB
            display_winsize: 256
                      epoch: latest
                       eval: False
                    gpu_ids: -1                                   [default: 0]
                  init_gain: 0.02
                  init_type: normal
                   input_nc: 3
                    isTrain: False                                [default: None]
                  load_iter: 0                                    [default: 0]
                  load_size: 1024                                 [default: 256]
           max_dataset_size: inf
                      model: test
               model_suffix:
                 n_layers_D: 3
                       name: style_cezanne_pretrained             [default: experiment_name]
                        ndf: 64
                       netD: basic
                       netG: resnet_9blocks
                        ngf: 64
                 no_dropout: True                                 [default: False]
                    no_flip: False
                       norm: instance
                   num_test: 50
                num_threads: 4
                  output_nc: 3
                      phase: test
                 preprocess: scale_width                          [default: resize_and_crop]
                results_dir: ./results/
             serial_batches: False
                     suffix:
                  use_wandb: False
                    verbose: False
         wandb_project_name: CycleGAN-and-pix2pix
  ----------------- End -------------------
  dataset [SingleDataset] was created
  initialize network with normal
  model [TestModel] was created
  loading the model from ./checkpoints\style_cezanne_pretrained\latest_net_G.pth
  ---------- Networks initialized -------------
  [Network G] Total number of parameters : 11.378 M
  -----------------------------------------------
  creating web directory ./results/style_cezanne_pretrained\test_latest
  processing (0000)-th image... ['datasets/images\\20210117_091324_001m.jpg']

ゴッホ風 †

• コマンド
  

  (py38_gan) PS > python test.py --dataroot datasets/images --name style_vangogh_pretrained --model test --no_dropout --preprocess scale_width --load_size 1024 --gpu_ids -1

  
  
• 実行結果
  
  
  
  
  
  
  
  
  
  ▼「python test.py --dataroot datasets/images --name style_vangogh_pretrained ...」ログ
  ▲「python test.py --dataroot datasets/images --name style_vangogh_pretrained ...」ログ

  (py38_gan) PS C:\anaconda_win\work\pytorch-CycleGAN-and-pix2pix> python test.py --dataroot datasets/images --name style_vangogh_pretrained --model test --no_dropout --preprocess scale_width --load_size 1024 --gpu_ids -1
  Warning: wandb package cannot be found. The option "--use_wandb" will result in error.
  Warning: wandb package cannot be found. The option "--use_wandb" will result in error.
  ----------------- Options ---------------
               aspect_ratio: 1.0
                 batch_size: 1
            checkpoints_dir: ./checkpoints
                  crop_size: 256
                   dataroot: datasets/images                      [default: None]
               dataset_mode: single
                  direction: AtoB
            display_winsize: 256
                      epoch: latest
                       eval: False
                    gpu_ids: -1                                   [default: 0]
                  init_gain: 0.02
                  init_type: normal
                   input_nc: 3
                    isTrain: False                                [default: None]
                  load_iter: 0                                    [default: 0]
                  load_size: 1024                                 [default: 256]
           max_dataset_size: inf
                      model: test
               model_suffix:
                 n_layers_D: 3
                       name: style_vangogh_pretrained             [default: experiment_name]
                        ndf: 64
                       netD: basic
                       netG: resnet_9blocks
                        ngf: 64
                 no_dropout: True                                 [default: False]
                    no_flip: False
                       norm: instance
                   num_test: 50
                num_threads: 4
                  output_nc: 3
                      phase: test
                 preprocess: scale_width                          [default: resize_and_crop]
                results_dir: ./results/
             serial_batches: False
                     suffix:
                  use_wandb: False
                    verbose: False
         wandb_project_name: CycleGAN-and-pix2pix
  ----------------- End -------------------
  dataset [SingleDataset] was created
  initialize network with normal
  model [TestModel] was created
  loading the model from ./checkpoints\style_vangogh_pretrained\latest_net_G.pth
  ---------- Networks initialized -------------
  [Network G] Total number of parameters : 11.378 M
  -----------------------------------------------
  creating web directory ./results/style_vangogh_pretrained\test_latest
  processing (0000)-th image... ['datasets/images\\20210117_091324_001m.jpg']

発生したエラーと対処法法 †

AttributeError: module 'torch._C' has no attribute '_cuda_setDevice' †

• エラー内容
  
  ▼　AttributeError: ログ
  ▲　AttributeError: ログ

  (py38_gan) PS C:\anaconda_win\work\pytorch-CycleGAN-and-pix2pix>  python test.py --dataroot datasets/ukiyoe2photo/testB --name style_ukiyoe_pretrained --model test --no_dropout
  Warning: wandb package cannot be found. The option "--use_wandb" will result in error.
  Warning: wandb package cannot be found. The option "--use_wandb" will result in error.
  ----------------- Options ---------------
               aspect_ratio: 1.0
                 batch_size: 1
            checkpoints_dir: ./checkpoints
                  crop_size: 256
                   dataroot: datasets/ukiyoe2photo/testB          [default: None]
               dataset_mode: single
                  direction: AtoB
            display_winsize: 256
                      epoch: latest
                       eval: False
                    gpu_ids: 0
                  init_gain: 0.02
                  init_type: normal
                   input_nc: 3
                    isTrain: False                                [default: None]
                  load_iter: 0                                    [default: 0]
                  load_size: 256
           max_dataset_size: inf
                      model: test
               model_suffix:
                 n_layers_D: 3
                       name: style_ukiyoe_pretrained              [default: experiment_name]
                        ndf: 64
                       netD: basic
                       netG: resnet_9blocks
                        ngf: 64
                 no_dropout: True                                 [default: False]
                    no_flip: False
                       norm: instance
                   num_test: 50
                num_threads: 4
                  output_nc: 3
                      phase: test
                 preprocess: resize_and_crop
                results_dir: ./results/
             serial_batches: False
                     suffix:
                  use_wandb: False
                    verbose: False
         wandb_project_name: CycleGAN-and-pix2pix
  ----------------- End -------------------
  Traceback (most recent call last):
    File "test.py", line 43, in <module>
      opt = TestOptions().parse()  # get test options
    File "N:\anaconda_win\work\pytorch-CycleGAN-and-pix2pix\options\base_options.py", line 136, in parse
      torch.cuda.set_device(opt.gpu_ids[0])
    File "C:\Users\izuts\anaconda3\envs\py38_gan\lib\site-packages\torch\cuda\__init__.py", line 404, in set_device
      torch._C._cuda_setDevice(device)
  AttributeError: module 'torch._C' has no attribute '_cuda_setDevice'

• 参考サイト
  How do I fix this error? '_cuda_setDevice' #360
  

• 対処法：コマンドにオプションを付加する
  

  --gpu_ids -1

更新履歴 †

• 2023/10/30 初版
  
   

参考資料 †

• CycleGAN
  
  • CycleGAN and pix2pix in PyTorch
    
  • CycleGANで浮世絵風の画像を生成してみる
    
  • AI(学習済みCycleGAN)による画像スタイル変換