Download Latest Version
HyperGAN 1.0 RC 2.zip (53.3 MB)
Get an email when there's a new version of HyperGAN
| Name | Modified | Size | Downloads / Week |
|---|---|---|---|
| Parent folder | |||
| HyperGAN 1.0 RC 1.tar.gz | 2020-08-02 | 34.2 MB | |
| HyperGAN 1.0 RC 1.zip | 2020-08-02 | 34.4 MB | |
| README.md | 2020-08-02 | 2.3 kB | |
| Totals: 3 Items | 68.7 MB | 0 | |
HyperGAN 1.0 RC 1
HyperGAN 1.0 is the result of roughly one year of directed effort to stabilize and speed up training.
When we started the project we decided that 1.0 was going to be released when we have a stable convergent default. Now that's here we're excited to bring you HyperGAN 1.0!
With this release we are also announcing the github sponsorship program. Help us develop HyperGAN into the GAN framework we know it can be!
https://github.com/sponsors/hypergan
If you have tried HyperGAN before, please give it another shot.
If you are new to HyperGAN, welcome, let us know more about your use case in the discord.
Changes
- Hypergan 1.0 now uses Pytorch instead of Tensorflow 1.0!
- New gitbook documentation! https://hypergan.gitbook.io/hypergan/
- New formalized HyperGAN layer definition language for defining your own network layouts
- New stabilization technique independently discovered and not available in other frameworks. See the default configuration for more.
- UI improvements
- Faster training
- WAY faster startup time
- More robust input handling - 0 byte files are not a problem!
--cropnow does a resize THEN a crop to capture as much of the target image as possible.- Fix samplers and enable more of them
- Switch samplers while training through menu
- New --width for specifying number sampler columns
- Updated configurations
- New working default configuration
- New AlignedGAN for aligning one distribution to another
- New AlignedInterpolatedGAN for discovering novel distributions
- Added TOML support ( [#240] )
- Lots of new layers, including exciting layers such as multi_head_attention. See the (auto-generated) documentation for more.
- Unique layers not available in other frameworks, such as
segment_softmaxwhich increases generator capacity significantly - Much much more...
We've tried to keep the good things about hypergan and remove the gnarly parts. We're still documenting and improving things but it's working and ready to be tested out.
To be continued...
We still have a lot of things planned for this release. Consider this an open invitation to start seeing experimenting with 1.0 and let us know what you think in the discord!
