Ctrl-World-Graph/graphwm/models/graph_encoder_pyg.py

from __future__ import annotations

from typing import Literal

import torch
import torch.nn as nn
from torch_geometric.nn import GINEConv, GPSConv

from graphwm.models.graph_encoder_edge_transformer import EdgeTransformerGraphSpatialEncoder
from graphwm.models.graph_encoder_gatv2 import GATv2GraphSpatialEncoder
from graphwm.models.graph_encoder_hybrid import HybridGINETransformerGraphSpatialEncoder
from graphwm.models.graph_encoder_transformer import TransformerGraphSpatialEncoder


class _GINEGPSGraphSpatialEncoder(nn.Module):
    """Per-frame spatial graph encoder using GINE or GPSConv."""

    def __init__(
        self,
        node_in_dim: int,
        edge_in_dim: int,
        hidden_dim: int = 256,
        num_layers: int = 4,
        dropout: float = 0.1,
        backbone: Literal["gine", "gps"] = "gine",
        num_heads: int = 8,
    ):
        super().__init__()
        self.backbone = backbone
        self.node_proj = nn.Linear(node_in_dim, hidden_dim)
        self.edge_proj = nn.Linear(edge_in_dim, hidden_dim)
        self.dropout = nn.Dropout(dropout)

        self.layers = nn.ModuleList()
        for _ in range(num_layers):
            mlp = nn.Sequential(
                nn.Linear(hidden_dim, hidden_dim),
                nn.SiLU(),
                nn.Linear(hidden_dim, hidden_dim),
            )
            if backbone == "gine":
                conv = GINEConv(mlp, edge_dim=hidden_dim)
            elif backbone == "gps":
                conv = GPSConv(
                    channels=hidden_dim,
                    conv=GINEConv(mlp, edge_dim=hidden_dim),
                    heads=num_heads,
                    attn_type="multihead",
                    attn_kwargs={"dropout": dropout},
                )
            else:
                raise ValueError(f"Unsupported graph backbone: {backbone}")
            self.layers.append(conv)

        self.norm = nn.LayerNorm(hidden_dim)

    def forward(self, data) -> torch.Tensor:
        x = self.node_proj(data.x)
        edge_attr = self.edge_proj(data.edge_attr)

        for layer in self.layers:
            residual = x
            if self.backbone == "gps":
                x = layer(x, data.edge_index, data.batch, edge_attr=edge_attr)
            else:
                x = layer(x, data.edge_index, edge_attr=edge_attr)
            x = self.norm(x + self.dropout(x) + residual)

        return x


class GraphSpatialEncoder(nn.Module):
    """Factory wrapper for per-frame spatial graph encoders."""

    SUPPORTED_BACKBONES = (
        "gine",
        "gps",
        "gatv2",
        "transformer",
        "edge_transformer",
        "hybrid_gine_transformer",
    )

    def __init__(
        self,
        node_in_dim: int,
        edge_in_dim: int,
        hidden_dim: int = 256,
        num_layers: int = 4,
        dropout: float = 0.1,
        backbone: str = "gine",
        num_heads: int = 8,
    ):
        super().__init__()
        if backbone in {"gine", "gps"}:
            encoder_cls = _GINEGPSGraphSpatialEncoder
            kwargs = {"backbone": backbone}
        elif backbone == "gatv2":
            encoder_cls = GATv2GraphSpatialEncoder
            kwargs = {}
        elif backbone == "transformer":
            encoder_cls = TransformerGraphSpatialEncoder
            kwargs = {}
        elif backbone == "edge_transformer":
            encoder_cls = EdgeTransformerGraphSpatialEncoder
            kwargs = {}
        elif backbone == "hybrid_gine_transformer":
            encoder_cls = HybridGINETransformerGraphSpatialEncoder
            kwargs = {}
        else:
            raise ValueError(
                f"Unsupported graph backbone: {backbone}. "
                f"Supported: {', '.join(self.SUPPORTED_BACKBONES)}"
            )

        self.backbone = backbone
        self.encoder = encoder_cls(
            node_in_dim=node_in_dim,
            edge_in_dim=edge_in_dim,
            hidden_dim=hidden_dim,
            num_layers=num_layers,
            dropout=dropout,
            num_heads=num_heads,
            **kwargs,
        )

    def forward(self, data) -> torch.Tensor:
        return self.encoder(data)