data

ScoreSDEOutput

extendsModelOutput

ScoreSDEOutput(score: Tensor, noise: Tensor, target: Tensor, std: Tensor, loss: Tensor | None = None)

source edit

What the model returns for a batch.

Attributes

scoreTensor

\nabla_x \log p_t(x)

at the perturbed sample, same shape as the input.

noiseTensor

What the network predicted, before it was divided by the standard deviation.

targetTensor

The noise actually added by the perturbation kernel. Returned because the objective is the error against it and recomputing it would mean redrawing.

stdTensor

The perturbation kernel's scale at the sampled times, (N,).

lossTensor or None

The denoising score-matching objective, set only by ScoreSDEForImageGeneration.

Examples

>>> import lucid
>>> from lucid.models import score_sde_vp
>>> model = score_sde_vp(sample_size=8, base_channels=8,
...     channel_mult=(1,), num_res_blocks=1, resnet_groups=4,
...     attention_resolutions=()).eval()
>>> out = model(lucid.randn((1, 3, 8, 8)))
>>> out.score.shape
(1, 3, 8, 8)

Used by 1

lucid.models.generative.score_sde

Constructors

dunder

init

→None

__init__(score: Tensor, noise: Tensor, target: Tensor, std: Tensor, loss: Tensor | None = None)

source edit

>>> import lucid >>> from lucid.models import score_sde_vp >>> model = score_sde_vp(sample_size=8, base_channels=8, ... channel_mult=(1,), num_res_blocks=1, resnet_groups=4, ... attention_resolutions=()).eval() >>> out = model(lucid.randn((1, 3, 8, 8))) >>> out.score.shape (1, 3, 8, 8)