GAN-Steerability-without-optimization
GAN-Steerability-without-optimization copied to clipboard
nsping13
GAN "Steerability" without optimization, ICLR 2021 (Spotlight)
This is an official pytorch implementation
GENERAL
Generally, our methods are coded in the file BigGAN.py. Each path can be extracted by directing the "method" flag accordingly. Full description can be found below. Also, for easier reproducing, you can use main.py.
USER-SPECIFIED GEOMETRIC TRANSFORMATIONS
| Path | Flag |
|---|---|
| Linear | 'l_shifty', 'l_zoom' and 'l_shiftx' |
| Newmann | 'nl_shifty', 'nl_zoom' and 'nl_shiftx' |
| Great circle | 'gcircle_shifty', 'gcircle_zoom' and 'gcircle_shiftx' |
UNSUPERVISED EXPLORATION OF TRANSFORMATIONS
The principal latent space directios are based on the right singular vectors of the SVD of the FC weight matrix.
Here the flags for the first layer unsupervised paths i.e., all principal latent space directions extracted from the first weight matrix:
| Path | Flag |
|---|---|
| Linear | 'svd_linear' |
| Great circle | 'svd_gcircle' |
| Small circle | 'svd_scircle' |
And here the corresponding paths of other scales:
| Path | Flag |
|---|---|
| Linear | 'linearin' |
| Great circle | 'greatin' |
| Small circle | 'smallin' |
The specific order direction is set by the index parameter.
PLUG AND PLAY
For convenience we also uploaded separate modules for directions and walks. Please see modules.py. You are invited to take the module you need and just plant in your code environment. This allows using separately methods for extracting directions and methods for constructing walks from given directions.
BigGAN model
We downloaded the G_ema.pth for the BigGAN model from here, however you can also use the TFhub and the pytorch implementation. Please see here. You are also invited to explore Imagenet classes here
For any question or issue, write to me! [email protected]
nsping13