Home

Awesome

Disentangled VAE

CircleCI

Replicating DeepMind's papers "β-VAE: Learning Basic Visual Concepts with a Constrained Variational Framework" and "Understanding disentangling in β-VAE"

2D shape disentaglement

Result by changing latent Z from -3.0 to 3.0 with γ=100.0 and C=20.0

Latent variables with small variances seem extracting "x", "y", "rotation" and "scale" parameters.

(This experiment is using DeepMind's dsprite data set.)

ZImageParameterVariance
z00.9216
z10.9216
z2Rotation0.0011
z3Rotation?0.0038
z4Pos X0.0002
z50.9384
z6Scale?0.0004
z70.8991
z80.9483
z9Pos Y0.0004

Reconstruction result

Left: original Right: reconstructed image