Encoder API

MultiChannelEncoder

Bases: Module

forward(channel_inputs)

Forward pass through multi-channel encoder.

Parameters:

Name Type Description Default
channel_inputs Dict[str, Dict[str, Tensor]]

Dict mapping channel names to tokenized inputs Each channel has 'input_ids' and 'attention_mask'

required

Returns:

Type Description
Dict[str, Optional[Tensor]]

Dict with: - 'embedding': Hyperbolic embedding in Lorentz model (batch_size, embedding_dim+1) - 'embedding_euc': Euclidean embedding before projection (batch_size, embedding_dim) [optional] - 'gate_probs': MoE gating probabilities - 'top_k_indices': MoE top-k expert indices