IP-FSRGAN

Requirements

• python3
• torch 1.0.1
• torchvision 0.2.2
• Pretrained LightCNN model: download LightCNN-29 v2 from https://github.com/AlfredXiangWu/LightCNN and place it in recognition/pre-trained

training

python3 train.py --opt [path_to_training_config]

dataset

• An directory with all the high resolution images for training
• Should be specified in config file, such as options/train/train_example.json

  ...
  ,datasets": {
  "train": {
    "name": "VISHR"
    , "mode": "LRHR"
    , "resize": 0.9
    , "dataroot_HR": "/mnt/WXRG0235/jnli/datasets/VIS/VISHR/VISHR_train"
    //, "image_lists": ["/mnt/ficuszambia/jnli/facesr/datasets/zhengjian/train.list", "/mnt/WXRG0235/jnli/datasets/VIS/VISHR/VISHR_train.list"]
    , "subset_file": null
    , "use_shuffle": true
    , "n_workers": 8
    , "batch_size": 16
    , "HR_size": 128
    , "use_flip": true
    , "use_rot": true
    , "downsample": "numpy"
  }
  ...

• subset_file the list to HR images
• name the name of the training dataset
• mode LRHR: user can only provide HR, or provide both HR & LR
• dataroot_HR path to the directory
• HR_size size of patches random cropped for training
• use_flip use_rot flip or rotate
• crop to crop or just resize the whole face
• resize resize the init HR face (to find the best performance of LR size), 1 for not resize
• downsample downsample method for generating LR image. numpy | cubic | linear

• image_lists the path to all the HR images when dataroot_HR is not given.

  path

  "path": {
  "root": "/mnt/ficuszambia/jnli/facesr"
  // , "resume_state": "../experiments/debug_002_RRDB_ESRGAN_x4_DIV2K/training_state/16.state"
  //    , "pretrain_model_G": "../experiments/pretrained_models/RRDB_PSNR_x4.pth"
  }

  • root the root path of this project for saving checkpoint (saving at root/Experiments)

  training settings

  , "train": {
  "lr_G": 1e-4
  , "weight_decay_G": 0
  , "beta1_G": 0.9
  , "lr_D": 1e-4
  , "weight_decay_D": 0
  , "beta1_D": 0.9
  , "lr_scheme": "MultiStepLR"
  , "lr_steps": [50000, 100000, 200000, 300000]
  , "lr_gamma": 0.5
  
  , "pixel_criterion": "l1"
  , "pixel_weight": 1e-2
  , "feature_criterion": "l1"
  , "feature_weight": 1
  , "gan_type": "vanilla"
  , "gan_weight": 5e-3
  , "recloss": {
    "name":"lightcnn",
    "path": "/mnt/ficuszambia/jnli/facesr/recognition/pre-trained/LightCNN_29Layers_V2_checkpoint.pth",
    "weight": 1
  }
  
  //for wgan-gp
  // , "D_update_ratio": 1
  // , "D_init_iters": 0
  // , "gp_weigth": 10
  
  , "manual_seed": 0
  , "niter": 1e6
  , "val_freq": 1000
  }

• recloss-name the name of recognition loss to be used. currently onlt lightcnn is supported. Leaving blank '' for not using recognition loss
• recloss-path path to the pre-trained model
• recloss-weight weight of recloss, when set to 1:

  
  19-04-03 17:00:12.447 - INFO: <epoch:  0, iter:     200, lr:1.000e-04> l_g_pix: 4.4571e-03 l_g_fea: 2.6769e+00 l_g_rec: 6.0616e-01 l_g_gan: 1.3311e-01 l_d_real: 0.0000e+00 l_d_fake: 1.6391e-06 D_real: 2.3803e+01 D_fake: -2.8195e+00
  ...
  19-04-03 20:50:43.945 - INFO: <epoch:  1, iter:  19,600, lr:1.000e-04> l_g_pix: 2.0350e-04 l_g_fea: 8.0707e-01 l_g_rec: 4.8295e-02 l_g_gan: 5.6367e-02 l_d_real: 7.4582e-04 l_d_fake: 7.7790e-05 D_real: -3.0652e+00 D_fake: -1.4338e+01

- `niter` how many batches to train?
- `val_freq` val image will be saved in `/valid`. Need manually deleted for each experiment
#### other settings

"name": "lightcnn_VISHR_numpy128crop" // please remove "debug" during training , "use_tb_logger": true , "model":"srragan" ,"scale": 4 , "gpu_ids": [0,1,2,3,4,5,6]

- `name` checkpoints will be saved in `root/Experiments/name/models`. Attention: new experiment with the same name will overwrite the old one

## test

python3 test.py --opt [path_to_test_config]

#### testset

"test 1": {
"name": "FaceTest_Crop_1",
"resize": 1,
"mode": "LRHR",
 "downsample": "numpy",
  "dataroot_HR": "/mnt/WXRG0235/jnli/05_RegWorkspace/image/A40P/__backups_rgbfull__collected_crop"

},

- `resize` resize the HR image before downsampling

### Reference
ESRGAN https://github.com/xinntao/BasicSR