isonetpp_loader.py

# isonetpp_loader.py
from __future__ import annotations
import os
from typing import Optional, Dict
from huggingface_hub import hf_hub_download

from subiso_dataset import (
    SubgraphIsomorphismDataset,
    # TRAIN_MODE, VAL_MODE, TEST_MODE, BROAD_TEST_MODE
    TRAIN_MODE, VAL_MODE, TEST_MODE
)

# ----------------------------
# Helpers
# ----------------------------

def _pairs_for_size(dataset_size: str) -> str:
    return "80k" if dataset_size == "small" else "240k"

def _folder_for_size(dataset_size: str) -> str:
    return "small_dataset" if dataset_size == "small" else "large_dataset"

def _normalize_name(base_name: str, dataset_size: str) -> str:
    """
    Accepts 'aids' or 'aids240k' (and similarly for other sets).
    If bare name -> append pairs; if already has 80k/240k -> keep as-is.
    """
    pairs = _pairs_for_size(dataset_size)
    if base_name.endswith(("80k", "240k")):
        return base_name
    return f"{base_name}{pairs}"

def _mode_prefix_and_dir(mode: str) -> tuple[str, str]:
    """
    File prefix uses 'test' when mode contains 'test' (repo convention).
    Directory has train/val/test. Map Extra_test_300 => 'test'.
    """
    prefix = "test" if "test" in mode.lower() else mode
    mode_dir = "test" if "test" in mode.lower() else mode
    return prefix, mode_dir

# ----------------------------
# Path resolution + downloads
# ----------------------------

def _ensure_paths(
    repo_id: str,
    mode: str,
    dataset_name: str,      # 'aids' or 'aids240k'
    dataset_size: str,      # 'small' | 'large'
    local_root: Optional[str] = None,
) -> Dict[str, str]:
    """
    Download the three files needed into cache (or local_root if set):
      - large_dataset/splits/<mode_dir>/<prefix>_<base>_query_subgraphs.pkl
      - large_dataset/splits/<mode_dir>/<prefix>_<base>_rel_nx_is_subgraph_iso.pkl
      - large_dataset/corpus/<base>_corpus_subgraphs.pkl
    where <base> is normalized (contains 80k/240k exactly once).
    """
    folder = _folder_for_size(dataset_size)                # "large_dataset" or "small_dataset"
    base   = _normalize_name(dataset_name, dataset_size)   # e.g., "aids240k"
    # prefix, mode_dir = _mode_prefix_and_dir(mode)

    query_fname  = f"{mode}_{base}_query_subgraphs.pkl"
    rel_fname    = f"{mode}_{base}_rel_nx_is_subgraph_iso.pkl"
    corpus_fname = f"{base}_corpus_subgraphs.pkl"

    repo_query_path  = f"{folder}/splits/{mode}/{query_fname}"
    repo_rel_path    = f"{folder}/splits/{mode}/{rel_fname}"
    repo_corpus_path = f"{folder}/corpus/{corpus_fname}"

    kwargs = dict(
        repo_id=repo_id,
        repo_type="dataset",
        local_dir=local_root,
        local_dir_use_symlinks=False,
    )

    query_path  = hf_hub_download(filename=repo_query_path,  **kwargs)
    rel_path    = hf_hub_download(filename=repo_rel_path,    **kwargs)
    corpus_path = hf_hub_download(filename=repo_corpus_path, **kwargs)

    return {"query": query_path, "rel": rel_path, "corpus": corpus_path}

# ----------------------------
# Public entrypoint
# ----------------------------

def load_isonetpp_benchmark(
    repo_id: str = "structlearning/isonetpp-benchmark",
    mode: str = "train",            # "train" | "val" | "test" | "Extra_test_300"
    dataset_name: str = "aids",     # "aids" or "aids240k" (same for mutag/ptc_*)
    dataset_size: str = "large",    # "small" | "large"
    batch_size: int = 128,
    data_type: str = "gmn",         # "pyg" or "gmn"
    device: Optional[str] = None,
    download_root: Optional[str] = None,
):
    # Map to class constantss
    mode_map = {
        "train": TRAIN_MODE,
        "val": VAL_MODE,
        "test": TEST_MODE,
        # "extra_test_300": BROAD_TEST_MODE,
        # "Extra_test_300": BROAD_TEST_MODE,
    }
    mode_norm = mode_map.get(mode, mode)

    paths = _ensure_paths(
        repo_id=repo_id,
        mode=mode_norm,
        dataset_name=dataset_name,
        dataset_size=dataset_size,
        local_root=download_root,
    )

    # paths["query"] = .../<folder>/splits/<mode_dir>/<file>
    # We want dataset_base_path to be the **parent of <folder>** so that:
    #   dataset_base_path / dataset_path_override / splits/<mode>/... exists
    # Compute levels carefully:
    #   file_dir   = .../<folder>/splits/<mode_dir>
    #   splits_dir = .../<folder>/splits
    #   folder_dir = .../<folder>
    #   parent_dir = parent of <folder>
    file_dir   = os.path.dirname(paths["query"])
    splits_dir = os.path.dirname(file_dir)
    folder_dir = os.path.dirname(splits_dir)
    parent_dir = os.path.dirname(folder_dir)   # <-- this is the correct dataset_base_path

    dataset_config = dict(
        mode=mode_norm,
        dataset_name=dataset_name,
        dataset_size=dataset_size,
        batch_size=batch_size,
        data_type=data_type,
        dataset_base_path=parent_dir,                    # parent of <folder>
        dataset_path_override=None,#  _folder_for_size(dataset_size),  # "large_dataset"/"small_dataset"
        experiment=None,
        device=device,
    )

    return SubgraphIsomorphismDataset(**dataset_config)