data

DDPMOutput

extendsDiffusionModelOutput

DDPMOutput(sample: Tensor, loss: Tensor | None = None, loss_simple: Tensor | None = None, loss_vlb: Tensor | None = None)

source edit

Training output of DDPMForImageGeneration.

Adds the two halves of Improved-DDPM's hybrid objective alongside the inherited total, so a training loop can watch them separately — $L_{\text{vlb}}$ is far noisier than $L_{\text{simple}}$ and a single blended number hides which one is moving.

Attributes

loss_simple(Tensor or None, optional)

The Ho 2020 Eq. 14 MSE term on its own.

loss_vlb(Tensor or None, optional)

The variational term

L_{t-1}

at the sampled timesteps, in bits, before the

\lambda T

scaling folded into loss. None unless learn_sigma=True.

Examples

>>> out = model(x_t, t, target=noise)
>>> float(out.loss_simple.item()) > 0
True

Used by 1

lucid.models.generative.ddpm

Constructors

dunder

init

→None

__init__(sample: Tensor, loss: Tensor | None = None, loss_simple: Tensor | None = None, loss_vlb: Tensor | None = None)

source edit

data

DDPMOutput

extendsDiffusionModelOutput

DDPMOutput(sample: Tensor, loss: Tensor | None = None, loss_simple: Tensor | None = None, loss_vlb: Tensor | None = None)

source edit

Training output of DDPMForImageGeneration.

Attributes

loss_simple(Tensor or None, optional)

The Ho 2020 Eq. 14 MSE term on its own.

loss_vlb(Tensor or None, optional)

The variational term

L_{t-1}

at the sampled timesteps, in bits, before the

\lambda T

scaling folded into loss. None unless learn_sigma=True.

Examples

>>> out = model(x_t, t, target=noise)
>>> float(out.loss_simple.item()) > 0
True

Used by 1

lucid.models.generative.ddpm

Constructors

dunder

init

→None

__init__(sample: Tensor, loss: Tensor | None = None, loss_simple: Tensor | None = None, loss_vlb: Tensor | None = None)

source edit