Merge pull request #65 from TensorBFS/feature/tropical-einsum-omeco

feat: add tropical einsum module with OMEinsum-style design and tropical-gemm acceleration

Author: ChanceSiyuan

tropical_in_new/src/__init__.py

@@ -1,9 +1,22 @@
 """Tropical tensor network tools for MPE (independent package)."""
 
-from .contraction import build_contraction_tree, choose_order, contract_tree
+from .contraction import (
+    build_contraction_tree,
+    choose_order,
+    contract_omeco_tree,
+    contract_tree,
+    get_omeco_tree,
+)
 from .mpe import mpe_tropical, recover_mpe_assignment
 from .network import TensorNode, build_network
-from .primitives import argmax_trace, safe_log, tropical_einsum
+from .primitives import safe_log
+from .tropical_einsum import (
+    Backpointer,
+    argmax_trace,
+    match_rule,
+    tropical_einsum,
+    tropical_reduce_max,
+)
 from .utils import (
     Factor,
     UAIModel,
@@ -14,6 +27,7 @@
 )
 
 __all__ = [
+    "Backpointer",
     "Factor",
     "TensorNode",
     "UAIModel",
@@ -22,12 +36,16 @@
     "build_network",
     "build_tropical_factors",
     "choose_order",
+    "contract_omeco_tree",
     "contract_tree",
+    "get_omeco_tree",
+    "match_rule",
     "mpe_tropical",
     "read_evidence_file",
     "read_model_file",
     "read_model_from_string",
     "recover_mpe_assignment",
     "safe_log",
     "tropical_einsum",
+    "tropical_reduce_max",
 ]

tropical_in_new/src/contraction.py

@@ -10,8 +10,7 @@
 import omeco
 
 from .network import TensorNode
-from .primitives import Backpointer, tropical_reduce_max
-from .utils import build_index_map
+from .tropical_einsum import tropical_einsum, tropical_reduce_max, Backpointer
 
 
 @dataclass
@@ -36,12 +35,6 @@ class ReduceNode:
 TreeNode = TensorNode | ContractNode | ReduceNode
 
 
-@dataclass(frozen=True)
-class ContractionTree:
-    order: Tuple[int, ...]
-    nodes: Tuple[TensorNode, ...]
-
-
 def _infer_var_sizes(nodes: Iterable[TensorNode]) -> dict[int, int]:
     sizes: dict[int, int] = {}
     for node in nodes:
@@ -54,16 +47,128 @@ def _infer_var_sizes(nodes: Iterable[TensorNode]) -> dict[int, int]:
     return sizes
 
 
-def _extract_leaf_index(node_dict: dict) -> int | None:
-    for key in ("leaf", "leaf_index", "index", "tensor"):
-        if key in node_dict:
-            value = node_dict[key]
-            if isinstance(value, int):
-                return value
-    return None
+def get_omeco_tree(nodes: list[TensorNode]) -> dict:
+    """Get the optimized contraction tree from omeco.
+
+    Args:
+        nodes: List of tensor nodes to contract.
+
+    Returns:
+        The omeco tree as a dictionary with structure:
+        - Leaf: {"tensor_index": int}
+        - Node: {"args": [...], "eins": {"ixs": [[...], ...], "iy": [...]}}
+    """
+    ixs = [list(node.vars) for node in nodes]
+    sizes = _infer_var_sizes(nodes)
+    method = omeco.GreedyMethod()
+    tree = omeco.optimize_code(ixs, [], sizes, method)
+    return tree.to_dict()
+
+
+def contract_omeco_tree(
+    tree_dict: dict,
+    nodes: list[TensorNode],
+    track_argmax: bool = True,
+) -> TreeNode:
+    """Contract tensors following omeco's optimized tree structure.
+
+    Uses tropical-gemm for accelerated binary contractions when available.
+
+    Args:
+        tree_dict: The omeco tree dictionary from get_omeco_tree().
+        nodes: List of input tensor nodes.
+        track_argmax: Whether to track argmax for MPE backtracing.
+
+    Returns:
+        Root TreeNode with contracted result and backpointers.
+    """
+
+    def recurse(node: dict) -> TreeNode:
+        # Leaf node - return the input tensor
+        if "tensor_index" in node:
+            return nodes[node["tensor_index"]]
+
+        # Internal node - contract children
+        args = node["args"]
+        eins = node["eins"]
+        out_vars = tuple(eins["iy"])
+
+        # Recursively contract children
+        children = [recurse(arg) for arg in args]
+
+        # Use tropical_einsum for the contraction
+        tensors = [c.values for c in children]
+        child_ixs = [c.vars for c in children]
+
+        values, backpointer = tropical_einsum(
+            tensors, list(child_ixs), out_vars, track_argmax=track_argmax
+        )
+
+        # Build result node (for binary, use ContractNode)
+        if len(children) == 2:
+            all_input = set(children[0].vars) | set(children[1].vars)
+            elim_vars = tuple(v for v in all_input if v not in out_vars)
+
+            return ContractNode(
+                vars=out_vars,
+                values=values,
+                left=children[0],
+                right=children[1],
+                elim_vars=elim_vars,
+                backpointer=backpointer,
+            )
+        else:
+            # For n-ary, chain as binary
+            result = children[0]
+            for i, child in enumerate(children[1:], 1):
+                is_final = (i == len(children) - 1)
+                target_out = out_vars if is_final else tuple(dict.fromkeys(result.vars + child.vars))
+
+                step_tensors = [result.values, child.values]
+                step_ixs = [result.vars, child.vars]
+
+                step_values, step_bp = tropical_einsum(
+                    step_tensors, list(step_ixs), target_out, track_argmax=track_argmax
+                )
+
+                all_input = set(result.vars) | set(child.vars)
+                elim_vars = tuple(v for v in all_input if v not in target_out)
+
+                result = ContractNode(
+                    vars=target_out,
+                    values=step_values,
+                    left=result,
+                    right=child,
+                    elim_vars=elim_vars,
+                    backpointer=step_bp,
+                )
+            return result
+
+    return recurse(tree_dict)
+
+
+# =============================================================================
+# Legacy API for backward compatibility
+# =============================================================================
+
+@dataclass(frozen=True)
+class ContractionTree:
+    """Legacy contraction tree structure."""
+    order: Tuple[int, ...]
+    nodes: Tuple[TensorNode, ...]
+
+
+def choose_order(nodes: list[TensorNode], heuristic: str = "omeco") -> list[int]:
+    """Legacy: Select elimination order. Use get_omeco_tree() instead."""
+    if heuristic != "omeco":
+        raise ValueError("Only the 'omeco' heuristic is supported.")
+    tree_dict = get_omeco_tree(nodes)
+    ixs = [list(node.vars) for node in nodes]
+    return _elim_order_from_tree_dict(tree_dict, ixs)
 
 
 def _elim_order_from_tree_dict(tree_dict: dict, ixs: list[list[int]]) -> list[int]:
+    """Extract elimination order from omeco tree (legacy support)."""
     total_counts: dict[int, int] = {}
     for vars in ixs:
         for var in vars:
@@ -72,14 +177,13 @@ def _elim_order_from_tree_dict(tree_dict: dict, ixs: list[list[int]]) -> list[in
     eliminated: set[int] = set()
 
     def visit(node: dict) -> tuple[dict[int, int], list[int]]:
-        leaf_index = _extract_leaf_index(node)
-        if leaf_index is not None:
+        if "tensor_index" in node:
             counts: dict[int, int] = {}
-            for var in ixs[leaf_index]:
+            for var in ixs[node["tensor_index"]]:
                 counts[var] = counts.get(var, 0) + 1
             return counts, []
 
-        children = node.get("children", [])
+        children = node.get("args") or node.get("children", [])
         if not isinstance(children, list) or not children:
             return {}, []
 
@@ -106,59 +210,49 @@ def visit(node: dict) -> tuple[dict[int, int], list[int]]:
     return order + remaining
 
 
-def choose_order(nodes: list[TensorNode], heuristic: str = "omeco") -> list[int]:
-    """Select elimination order over variable indices using omeco."""
-    if heuristic != "omeco":
-        raise ValueError("Only the 'omeco' heuristic is supported.")
-    ixs = [list(node.vars) for node in nodes]
-    sizes = _infer_var_sizes(nodes)
-    method = omeco.GreedyMethod() if hasattr(omeco, "GreedyMethod") else None
-    tree = (
-        omeco.optimize_code(ixs, [], sizes, method)
-        if method is not None
-        else omeco.optimize_code(ixs, [], sizes)
-    )
-    tree_dict = tree.to_dict() if hasattr(tree, "to_dict") else tree
-    if not isinstance(tree_dict, dict):
-        raise ValueError("omeco.optimize_code did not return a usable tree.")
-    return _elim_order_from_tree_dict(tree_dict, ixs)
-
-
 def build_contraction_tree(order: Iterable[int], nodes: list[TensorNode]) -> ContractionTree:
-    """Prepare a contraction plan from order and leaf nodes."""
+    """Legacy: Prepare a contraction plan from order and leaf nodes."""
     return ContractionTree(order=tuple(order), nodes=tuple(nodes))
 
 
 def contract_tree(
     tree: ContractionTree,
-    einsum_fn,
+    einsum_fn=None,
     track_argmax: bool = True,
 ) -> TreeNode:
-    """Contract along the tree using the tropical einsum."""
+    """Legacy: Contract using elimination order. Use contract_omeco_tree() instead."""
     active_nodes: list[TreeNode] = list(tree.nodes)
+
     for var in tree.order:
         bucket = [node for node in active_nodes if var in node.vars]
         if not bucket:
             continue
         bucket_ids = {id(node) for node in bucket}
         active_nodes = [node for node in active_nodes if id(node) not in bucket_ids]
+
         combined: TreeNode = bucket[0]
         for i, other in enumerate(bucket[1:]):
             is_last = i == len(bucket) - 2
             elim_vars = (var,) if is_last else ()
-            index_map = build_index_map(combined.vars, other.vars, elim_vars=elim_vars)
-            values, backpointer = einsum_fn(
-                combined.values, other.values, index_map,
+
+            # Use tropical_einsum
+            target_out = tuple(v for v in dict.fromkeys(combined.vars + other.vars) if v not in elim_vars)
+            values, backpointer = tropical_einsum(
+                [combined.values, other.values],
+                [combined.vars, other.vars],
+                target_out,
                 track_argmax=track_argmax if is_last else False,
             )
+
             combined = ContractNode(
-                vars=index_map.out_vars,
+                vars=target_out,
                 values=values,
                 left=combined,
                 right=other,
                 elim_vars=elim_vars,
                 backpointer=backpointer,
             )
+
         if var in combined.vars:
             # Single-node bucket: eliminate via reduce
             values, backpointer = tropical_reduce_max(
@@ -172,20 +266,27 @@ def contract_tree(
                 backpointer=backpointer,
             )
         active_nodes.append(combined)
+
     while len(active_nodes) > 1:
         left = active_nodes.pop(0)
         right = active_nodes.pop(0)
-        index_map = build_index_map(left.vars, right.vars, elim_vars=())
-        values, _ = einsum_fn(left.values, right.values, index_map, track_argmax=False)
+        target_out = tuple(dict.fromkeys(left.vars + right.vars))
+        values, _ = tropical_einsum(
+            [left.values, right.values],
+            [left.vars, right.vars],
+            target_out,
+            track_argmax=False,
+        )
         combined = ContractNode(
-            vars=index_map.out_vars,
+            vars=target_out,
             values=values,
             left=left,
             right=right,
             elim_vars=(),
             backpointer=None,
         )
         active_nodes.append(combined)
+
     if not active_nodes:
         raise ValueError("Contraction produced no nodes.")
     return active_nodes[0]

tropical_in_new/src/mpe.py

@@ -4,10 +4,14 @@
 
 from typing import Dict, Iterable
 
-from .contraction import ContractNode, ReduceNode, build_contraction_tree, choose_order
-from .contraction import contract_tree as _contract_tree
+from .contraction import (
+    ContractNode,
+    ReduceNode,
+    contract_omeco_tree,
+    get_omeco_tree,
+)
 from .network import TensorNode, build_network
-from .primitives import argmax_trace, tropical_einsum, tropical_reduce_max
+from .tropical_einsum import argmax_trace, tropical_reduce_max
 from .utils import UAIModel, build_tropical_factors
 
 
@@ -70,16 +74,22 @@ def traverse(node, out_assignment: Dict[int, int]) -> None:
 def mpe_tropical(
     model: UAIModel,
     evidence: Dict[int, int] | None = None,
-    order: Iterable[int] | None = None,
 ) -> tuple[Dict[int, int], float, Dict[str, int | tuple[int, ...]]]:
-    """Return MPE assignment, score, and contraction metadata."""
+    """Return MPE assignment, score, and contraction metadata.
+
+    Uses omeco for optimized contraction order and tropical-gemm for acceleration.
+    """
     evidence = evidence or {}
     factors = build_tropical_factors(model, evidence)
     nodes = build_network(factors)
-    if order is None:
-        order = choose_order(nodes, heuristic="omeco")
-    tree = build_contraction_tree(order, nodes)
-    root = _contract_tree(tree, einsum_fn=tropical_einsum)
+
+    # Get optimized contraction tree from omeco
+    tree_dict = get_omeco_tree(nodes)
+
+    # Contract using the optimized tree
+    root = contract_omeco_tree(tree_dict, nodes, track_argmax=True)
+
+    # Final reduction if there are remaining variables
     if root.vars:
         values, backpointer = tropical_reduce_max(
             root.values, root.vars, tuple(root.vars), track_argmax=True
@@ -91,12 +101,11 @@ def mpe_tropical(
             elim_vars=tuple(root.vars),
             backpointer=backpointer,
         )
+
     assignment = recover_mpe_assignment(root)
     assignment.update({int(k): int(v) for k, v in evidence.items()})
     score = float(root.values.item())
     info = {
-        "order": tuple(order),
         "num_nodes": len(nodes),
-        "num_elims": len(tuple(order)),
     }
     return assignment, score, info