Awesome

Variation Predictability Metric

This repository contains the independent code for VP-metric in [Learning Disentangled Representations with Latent Variation Predictability].

Requirements

• Numpy.
• PyTorch >= 1.3.1

Training

Once you have a dataset of [(x1, x2) --> \delta z], you can use this code to train a simple ConvNet to do the evaluation. Remember to modify the --out_dim to fit the latent dimensions of your model.

CUDA_VISIBLE_DEVICES=0 \
    python main_vp.py \
    --result_dir /path/to/result-dir \
    --data_dir /path/to/image-pair/dir \
    --in_channels 3 \
    --out_dim 30 \
    --lr 0.01 \
    --batch_size 32 \
    --epochs 200 \
    --input_mode diff \
    --test_ratio 0.9

Then use:

python get_best_score.py --target_dir /path/to/result-dir

to obtain the VP score.

Citation

@inproceedings{VPdis_eccv20,
author={Xinqi Zhu and Chang Xu and Dacheng Tao},
title={Learning Disentangled Representations with Latent Variation Predictability},
booktitle={ECCV},
year={2020}
}