Glow

This is pytorch implementation of paper "Glow: Generative Flow with Invertible 1x1 Convolutions". Most modules are adapted from the offical TensorFlow version openai/glow.

TODO

[x] Glow model. The model is coded as described in original paper, some functions are adapted from offical TF version. Most modules are tested.
[x] Trainer, builder and hparams loaded from json.
[x] Infer after training
[ ] Test LU_decomposed 1x1 conv2d

Scripts

Train a model with

train.py <hparams> <dataset> <dataset_root>

Generate z_delta and manipulate attributes with

infer_celeba.py <hparams> <dataset_root> <z_dir>

Training result

Currently, I trained model for 45,000 batches with hparams/celeba.json using CelebA dataset. In short, I trained with follwing parameters

HParam	Value
image_shape	(64, 64, 3)
hidden_channels	512
K	32
L	3
flow_permutation	invertible 1x1 conv
flow_coupling	affine
batch_size	12 on each GPU, with 4 GPUs
learn_top	false
y_condition	false

Download pre-trained model from Dropbox

Reconstruction

Following are some samples at training phase. Row 1: reconstructed, Row 2: original.

Manipulate attribute

Use the method decribed in paper to calculate z_pos and z_neg for a given attribute. And z_delta = z_pos - z_neg is the direction to manipulate the original image.