data

FlowMatchingConfig

extendsGenerativeModelConfig

FlowMatchingConfig(sample_size: int | tuple[int, int] = 32, in_channels: int = 3, out_channels: int = 3, act_fn: GenerativeActivation = 'silu', base_channels: int = 128, channel_mult: tuple[int, ...] = (lambda: (1, 2, 2, 2))(), num_res_blocks: int = 2, attention_resolutions: tuple[int, ...] = (lambda: (32, 16, 8))(), num_head_channels: int = 64, dropout: float = 0.0, resnet_groups: int = 32, path: ProbabilityPath = 'ot', sigma_min: float = 0.0001, beta_min: float = 0.1, beta_max: float = 20.0, solver: str = 'dopri5', rtol: float = 1e-05, atol: float = 1e-05, trace_method: TraceMethod | None = None, trace_noise: TraceNoise = 'rademacher', exact_trace_max_dim: int = 32)

source edit

Frozen configuration for every Flow Matching variant.

Parameters

sample_sizeint or tuple of int= 32

Spatial resolution of a sample.

in_channelsint= 3

Channels per sample.

out_channelsint= 3

Kept equal to in_channels: the field is a velocity in the same space as the sample, so it cannot change the dimension.

act_fn(silu, swish, relu, gelu)= "silu"

Activation inside the velocity field.

base_channelsint= 128

Width of the first U-Net stage.

channel_multtuple of int= (1, 2, 2, 2)

Per-stage width multipliers; the length sets the depth of the encoder/decoder ladder.

num_res_blocksint= 2

Residual blocks per resolution.

attention_resolutionstuple of int= (32, 16, 8)

Spatial sizes at which self-attention is applied.

num_head_channelsint= 64

Channels per attention head. The head count is derived per block, so a stage twice as wide gets twice as many heads — the multi-resolution attention the architecture this borrows from specifies.

dropoutfloat= 0.0

Dropout inside residual blocks.

resnet_groupsint= 32

GroupNorm groups; clamped down when a stage has fewer channels.

path(ot, diffusion)= "ot"

Which conditional probability path the regression target comes from. Both are trained by the same objective — that equivalence is the paper's point — and "ot" is the one it recommends.

sigma_minfloat= 1e-4

Width of the Gaussian left around a data point at t = 1. The paper requires only that it be "sufficiently small" and publishes no value, so this is a choice rather than a citation; anything small and positive behaves the same. At exactly zero the path degenerates to a pure straight line between the two samples.

beta_minfloat= 0.1, 20.0

Endpoints of the linear

\beta(t)

used by the diffusion path. Ignored when path="ot".

beta_maxfloat= 0.1, 20.0

Endpoints of the linear

\beta(t)

used by the diffusion path. Ignored when path="ot".

solverstr= "dopri5"

Method handed to lucid.diffeq.odeint for sampling and likelihood — the paper's own choice.

rtolfloat= 1e-5, 1e-5

Solver tolerances; the values the paper evaluates with.

atolfloat= 1e-5, 1e-5

Solver tolerances; the values the paper evaluates with.

trace_method(exact, hutchinson)= "exact"

How the divergence is obtained when scoring likelihood. None picks by dimension. Irrelevant to training.

trace_noise(rademacher, gaussian)= "rademacher"

Probe distribution for the Hutchinson estimate.

exact_trace_max_dimint= 32

Dimension below which trace_method=None resolves to "exact".

Notes

in_channels and out_channels must agree: the field lives in the sample's own space.

Examples

>>> from lucid.models.generative.flow_matching import FlowMatchingConfig
>>> FlowMatchingConfig().path
'ot'
>>> FlowMatchingConfig(sample_size=32, in_channels=3).data_dim
3072

Used by 3

Constructors

dunder

init

→None

__init__(sample_size: int | tuple[int, int] = 32, in_channels: int = 3, out_channels: int = 3, act_fn: GenerativeActivation = 'silu', base_channels: int = 128, channel_mult: tuple[int, ...] = (lambda: (1, 2, 2, 2))(), num_res_blocks: int = 2, attention_resolutions: tuple[int, ...] = (lambda: (32, 16, 8))(), num_head_channels: int = 64, dropout: float = 0.0, resnet_groups: int = 32, path: ProbabilityPath = 'ot', sigma_min: float = 0.0001, beta_min: float = 0.1, beta_max: float = 20.0, solver: str = 'dopri5', rtol: float = 1e-05, atol: float = 1e-05, trace_method: TraceMethod | None = None, trace_noise: TraceNoise = 'rademacher', exact_trace_max_dim: int = 32)

source edit

Properties

prop

data_dim

→int

data_dim: int

source edit

int: Flattened width of one sample.

prop

resolved_trace_method

→TraceMethod

resolved_trace_method: TraceMethod

source edit

TraceMethod: trace_method, or the one implied by the dimension.

FlowMatchingConfig(sample_size: int | tuple[int, int] = 32, in_channels: int = 3, out_channels: int = 3, act_fn: GenerativeActivation = 'silu', base_channels: int = 128, channel_mult: tuple[int, ...] = (lambda: (1, 2, 2, 2))(), num_res_blocks: int = 2, attention_resolutions: tuple[int, ...] = (lambda: (32, 16, 8))(), num_head_channels: int = 64, dropout: float = 0.0, resnet_groups: int = 32, path: ProbabilityPath = 'ot', sigma_min: float = 0.0001, beta_min: float = 0.1, beta_max: float = 20.0, solver: str = 'dopri5', rtol: float = 1e-05, atol: float = 1e-05, trace_method: TraceMethod | None = None, trace_noise: TraceNoise = 'rademacher', exact_trace_max_dim: int = 32)

__init__(sample_size: int | tuple[int, int] = 32, in_channels: int = 3, out_channels: int = 3, act_fn: GenerativeActivation = 'silu', base_channels: int = 128, channel_mult: tuple[int, ...] = (lambda: (1, 2, 2, 2))(), num_res_blocks: int = 2, attention_resolutions: tuple[int, ...] = (lambda: (32, 16, 8))(), num_head_channels: int = 64, dropout: float = 0.0, resnet_groups: int = 32, path: ProbabilityPath = 'ot', sigma_min: float = 0.0001, beta_min: float = 0.1, beta_max: float = 20.0, solver: str = 'dopri5', rtol: float = 1e-05, atol: float = 1e-05, trace_method: TraceMethod | None = None, trace_noise: TraceNoise = 'rademacher', exact_trace_max_dim: int = 32)