MolecularDiffusion.modules.models.ligandiff.noise¶

Diffusion noise schedules. Verbatim port of LigandDiff’s src/noise.py.

Both branches matter: polynomial_2 is upstream’s config.yml default for fresh training, while the released model/pretrained.ckpt was trained with the learned GammaNetwork (its edm.gamma.{gamma_0,gamma_1,l1,l2,l3} tensors are in the checkpoint). Selecting the wrong one makes those eight tensors arrive as unexpected keys, which cli/generate.py’s strict=False load silently drops.

Classes¶

`GammaNetwork`	Monotonic learned gamma schedule (VDM paper construction).
`PositiveLinear`	Linear layer with weights forced to be positive.
`PredefinedNoiseSchedule`	Lookup table for a fixed (non-learned) noise schedule.

Functions¶

`clip_noise_schedule`(alphas2[, clip_value])	Clip alpha_t / alpha_{t-1} for sampling stability.
`cosine_beta_schedule`(timesteps[, s, raise_to_power])	Cosine schedule (Nichol & Dhariwal).
`polynomial_schedule`(timesteps[, s, power])	Noise schedule based on `1 - x^power`.

Module Contents¶

class MolecularDiffusion.modules.models.ligandiff.noise.GammaNetwork¶

Bases: torch.nn.Module

Monotonic learned gamma schedule (VDM paper construction).

forward(t)¶

gamma_tilde(t)¶

gamma_0¶

gamma_1¶

l1¶

l2¶

l3¶

class MolecularDiffusion.modules.models.ligandiff.noise.PositiveLinear(in_features: int, out_features: int, bias: bool = True, weight_init_offset: int = -2)¶

Bases: torch.nn.Module

Linear layer with weights forced to be positive.

forward(x)¶

reset_parameters() → None¶

in_features¶

out_features¶

weight¶

weight_init_offset = -2¶

class MolecularDiffusion.modules.models.ligandiff.noise.PredefinedNoiseSchedule(noise_schedule, timesteps, precision)¶

Bases: torch.nn.Module

Lookup table for a fixed (non-learned) noise schedule.

forward(t)¶

gamma¶

timesteps¶

MolecularDiffusion.modules.models.ligandiff.noise.clip_noise_schedule(alphas2, clip_value=0.001)¶: Clip alpha_t / alpha_{t-1} for sampling stability.

MolecularDiffusion.modules.models.ligandiff.noise.cosine_beta_schedule(timesteps, s=0.008, raise_to_power: float = 1)¶: Cosine schedule (Nichol & Dhariwal).

MolecularDiffusion.modules.models.ligandiff.noise.polynomial_schedule(timesteps: int, s=0.0001, power=3.0)¶: Noise schedule based on 1 - x^power.