hq-liu / pytorch-glow

PyTorch implementation of "Glow: Generative Flow with Invertible 1x1 Convolutions"

pytorch-glow

PyTorch implementation of "Glow: Generative Flow with Invertible 1x1 Convolutions"

Usage

First you need to install all requirements by

pip3 install -r requirements.txt

Training

1. Prepare dataset and corresponding profile file (like profile/celeba.json)
2. Start training:

python3 train.py [profile] 

Inference

Usage: python3 infer.py [OPTIONS] COMMAND [ARGS]...

Options:
  --profile PATH
  --snapshot PATH
  --help           Show this message and exit.

Commands:
  compute_deltaz
  interpolate
  reconstruct
  sample

Result

Reconstruction

The upper is reconstructed image, the lower is original one.

Interpolation

From left to right, the attribute offset is from -1.0 to 1.0 by step 0.25.

• Attractive

• Black hair

• Blurry

• Mouth slightly open

Acknowledgement

This project refers to:

• openai/glow (Official implementation)
• chaiyujin/glow-pytorch