from typing import Dict, Any from .network.nn import DenoisingGraphTransformer from .diffusion import PermutableGaussianDiffusion from ..data import Kinematic def num_graph_nodes(skeleton) -> int: """Number of graph nodes of a skeleton: bones for the bone direction representation, joints otherwise.""" return (skeleton.bone_adj_matrix() if skeleton.node_type == "bone" else skeleton.adj_matrix).shape[-1] class DiffusionManager(): """Builds the denoiser and the diffusion process for the skeletons (datasets) seen during training.""" def __init__(self, skeletons: Dict[str, Kinematic], num_nodes: int, diffusion_type: str = 'PermutableGaussianDiffusion', **kwargs): assert diffusion_type == 'PermutableGaussianDiffusion', diffusion_type model = self.get_network(num_nodes=num_nodes, **kwargs) self.diffusion = PermutableGaussianDiffusion( num_nodes_per_dataset={name: num_graph_nodes(skel) for name, skel in skeletons.items()}, num_nodes=num_nodes, model=model, **kwargs) def get_diffusion(self): return self.diffusion def get_network(self, latent_size=96, diffusion_arch: Dict[str, Any] = {}, diffusion_conditioning=True, **kwargs): assert diffusion_conditioning, "the denoiser is always conditioned on the observed motion" return DenoisingGraphTransformer(dim=latent_size, cond_dim=latent_size, out_dim=latent_size, **diffusion_arch, **kwargs)