Home

Awesome

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

Scripts

Training result

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

HParamValue
image_shape(64, 64, 3)
hidden_channels512
K32
L3
flow_permutationinvertible 1x1 conv
flow_couplingaffine
batch_size12 on each GPU, with 4 GPUs
learn_topfalse
y_conditionfalse

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.

Issues

There might be some errors in my codes. Please help me to figure out.