neural_ode_gen

→NeuralODEForImageGeneration

neural_ode_gen(pretrained: bool = False, overrides: object = {})

source edit

Neural ODE flow with the bits/dim objective and a sampler.

Same field as neural_ode, wrapped so forward reports a loss and generate integrates a prior draw back to t = 0. The reverse pass carries no density term, which makes sampling the cheaper of the two directions — the opposite of the usual situation for a flow, where the inverse is the awkward one to write.

Model Size

Parameters

pretrainedbool= False

Reserved for future weight registration; currently a no-op.

**overridesobject= {}

Optional NeuralODEConfig field overrides forwarded into the underlying config.

Returns

NeuralODEForImageGeneration

Generator configured with the paper's $M = 64$ field.

Notes

Reference: Chen, Rubanova, Bettencourt, and Duvenaud, "Neural Ordinary Differential Equations", NeurIPS, 2018 (arXiv:1806.07366). The paper reports no bits/dim figure for the continuous flow — its density experiments are two-dimensional and reported as KL — so there is no published number for this configuration to be compared against.

Examples

>>> from lucid.models.generative.neural_ode import neural_ode_gen
>>> model = neural_ode_gen(sample_size=(1, 2), in_channels=1,
...                        out_channels=1).eval()
>>> model.generate(n_samples=2).samples.shape
(2, 1, 1, 2)

Used by 1

lucid.models.generative.neural_ode

>>> from lucid.models.generative.neural_ode import neural_ode_gen >>> model = neural_ode_gen(sample_size=(1, 2), in_channels=1, ... out_channels=1).eval() >>> model.generate(n_samples=2).samples.shape (2, 1, 1, 2)