/InvTF

Easy to use tensorflow 2.0 implementation of invertible generative neural networks.

Primary LanguagePython

cite us using bibtex

InvTF

Train invertible generative models using the simplicity of Keras (aka normalizing flows).

Example: Invertible Residual Networks.

from invtf import Generator, InvResNet, faces

gen = Generator()

for _ in range(10): 
    gen.add(InvResNet())

gen.compile()
gen.fit(faces())

Most recent invertible generative model [1,2,3,4] have been reproduced in InvTF. Pretrained models are automatically downloaded when needed.

Example: Use pretrained model.

from invtf import Glow

glow.interpolate(faces()[0], faces()[1])

glow.generate(10)

Please see our tutorial and documentation for more information.

TLDR: Easily train reversible generative models with Keras.

Details

"invtf/" the package directory.

"articles/": notes / ipython notebook explaining previous articles reproducing their experiments in ipython notebooks.

"examples": implementation of a lot of examples, naming conforms with "dataset" x "model".py examples: "syntheticNormal_feedforward.py" "cifar10_glow.py" Besides show-casing how the framework can be used, these files are also used for test casing, just run 'run_all.py' (might take itme).

Development details

Developed using tensorflow-gpu 2.0beta, ubuntu 16.04 LTS, cuda 10, RTX 2080 8 GB