probabll / dgm.pt

Pytorch code for deep generative models

dgm.pt

Pytorch code for building deep generative models

Install https://github.com/probabll/dists.pt

Then

pip install -r requirements.txt
python setup.py develop

Register parameterization

Use decorators to register a parameterization of a distribution, e.g.

@register_conditional_parameterization(Normal)
def make_normal(inputs, event_size):
    params = torch.split(inputs, event_size, -1)
    return Normal(loc=params[0], scale=F.softplus(params[1]))

Use conditioners to parameterize components. For example, here we show the prior, approximate posterior, and likelihood of the classic MNIST VAE:

p_z = PriorLayer(
    event_shape=z_size,
    dist_type=Normal,
    params=[0., 1.]
)
q_z = ConditionalLayer(
    event_size=z_size,
    dist_type=Normal,
    conditioner=FFConditioner(
        input_size=x_size,
        output_size=z_size * 2,  # Gaussians take two parameters per unit
        hidden_sizes=[x_size // 2]
    )
)
p_x_given_z = FullyFactorizedLikelihood(
    event_size=x_size, 
    dist_type=Bernoulli, 
    conditioner=FFConditioner(
        input_size=z_size, 
        output_size=x_size * 1,   # Bernoullis take one parameter per unit
        hidden_sizes=[x_size // 2]
    )            
)

Design

• Bijection: an invertible transformation (which also computes log det jacobian)
• Conditioner: maps from data to D-dimensional output
• ConditionalLayer: parameterises a certain distribution (using a conditioner)
• LikelihoodLayer: parameterises a certain distribution (using a conditioner) and also provides a dedicated sampling procedure
• NF: a type of distribution

dgm.pt's Issues

Confusing argument names

The autoregressive likelihood uses the named argument history to refer to the observation (conditioned on autoregressively) and inputs to refer to the features conditioned on freely, which causes confusion. Follow other software packages and name the former observation.