Skip to content

Latest commit

 

History

History
82 lines (63 loc) · 3.64 KB

Evaluation.md

File metadata and controls

82 lines (63 loc) · 3.64 KB
Raw

If you are using FTransGAN Dataset, see Evaluating section of here.

Train

1. Modify the configuration file ("cfgs/evaluator/train.yaml")

You can set these values by giving command-line arguments like argparse, not modifying this configuration file directly. For the detailed description, please refer here.

  • trainer: (leave blank)

    • resume: Path to the checkpoint to resume from.
    • work_dir: Path to save the checkpoints, the validation images, and log.
    • max_epoch: Epochs to train the model.

  • dset: (leave blank)

    • train: (leave blank)
      • data_dir: Path to the data to use for the training.
        • List format is allowed: the training data will be collected from all the paths in this list.
      • chars: The character list to train the model to classify.
      • extension: The extesion of training data.
      • save_list: Whether to save the list of fonts and chars.
        • The list of fonts and chars will be saved to trainer.work_dir.
        • You may need the list of fonts and characters for the evaluation.
    • val: (leave blank)
      • n_val_example: The number of data to validate.
      • data_dir: Path to the data to use for the validation.
      • extension: The extesion of validation data.

2. Run training

python train_evaluator.py cfgs/evaluator/train.yaml -g(optional) 2 -n(optional) 2 -nr(optional) -p(optional) 12241 0 --work_dir(optional) path/to/save/outputs

-g, -n, -nr, -p are arguments for the DistributedDataParallel training. You do not need to give these arguments if you are using a single GPU.

  • arguments
    • path/to/config (first argument): path to configration file.
      • Multiple values are allowed but the first one should locate in cfgs/evaluator.
    • -g : number of gpus to use for the training.
    • -n : number of nodes to use for the training.
    • -nr : the ranking of current node within the nodes.
    • -p : the port to use for the DistributedDataParallel training.
    • --work_dir : path to save outputs. The trainer.work_dir in the configuration file will be overwrited to this value.

Evaluate

1. Modify the configuration file ("cfgs/evaluator/eval.yaml")

You can set these values by giving command-line arguments like argparse, not modifying this configuration file directly. For the detailed description, please refer here.

  • style_model_path: The checkpoint file which contains the weight of style classifier.

  • content_model_path: The checkpoint file which contains the weight of content(character) classifier.

  • dset: (leave blank)

    • test: (leave blank)
      • data_dir: Path to the generated images.
      • gt_dir: Path to the ground truth data (to calculate SSIM and LPIPS).
      • gt_extension: The extension of ground truth data.
      • keylist: The list of fonts which used to train the evaluator. You can obtain this by setting dset.train.save_list to True when running the training.
      • charlist: The list of characters which used to train the evaluator. You can obtain this by setting dset.train.save_list to True when running the training.

2. Run evaluation

python eval.py cfgs/evaluator/eval.yaml \
--result_dir path/to/save/result/file \
--result_name eval
  • arguments
    • path/to/config (first argument, multiple values are allowed): path to configration file.
    • --result_dir: Path to save result json file.
    • --n_ref: Name of the result json file. (Not need to contain ".json" - it will be added automatically.)