/Evaluation-for-Motion-Puzzle

Primary LanguagePython

Evaluation for Motion Puzzle

We use the ST-GCN based recognition model of [Yan et al. 2018] to evaluate generative models in our paper.

Prepare data

First, prepare the datasets needed for training and evaluation as following:

  • Make datasets directory
  • Download the datasets from the link
  • Put them in the datasets directory

The structure of datasets directory will look like:

datasets/
├─ preprocess_styletransfer_classify.npz
├─ styletransfer_classify.npz
├─ styletransfer_generate.npz
├─ styletransfer_stylized_aberman_0.npz
├─ styletransfer_stylized_holden_0.npz
├─ styletransfer_stylized_ours_0.npz
├─ ...

Test pretrained model

Prepare pretrained models as following:

  • Make output directory
  • Download the pretrained models from the link
  • Put them in the output directory

The structure of output directory will look like:

output/
├─ SRA/
│  ├─ latest_checkpoint.pth
├─ CRA/
│  ├─ latest_checkpoint.pth

Finally, you can evaluate the generated results according to each criteria. Please refer to base_options.py or eval_options.py under options directory for model and evaluation specifications.

python evaluate.py --experiment_name CRA --criteria content --load_latest --mode eval
# or
python evaluate.py --experiment_name SRA --criteria style --load_latest --mode eval

Train from scratch

You can train your own classifier from scratch by specifing the classifying criteria, e.g., content or style.

  • For training a content classifier 
    python recognition.py --experiment_name [EXPERIMENT_NAME] --criteria content
  • For training a style classifier 
    python recognition.py --experiment_name [EXPERIMENT_NAME] --criteria style