src/flowfm/checkpointing.py

"""Checkpoint loading helpers shared across train/inference/submission."""

from pathlib import Path

import re
import torch


def load_state_dict(path: Path, device: torch.device) -> dict:
    """Load a state_dict with compatibility for multiple PyTorch versions."""
    try:
        checkpoint = torch.load(path, map_location=device, weights_only=True)
    except TypeError:
        checkpoint = torch.load(path, map_location=device)

    if isinstance(checkpoint, dict) and "state_dict" in checkpoint:
        state_dict = checkpoint["state_dict"]
        if isinstance(state_dict, dict):
            return state_dict

    if not isinstance(checkpoint, dict):
        raise TypeError(f"Expected checkpoint dict at {path}, got {type(checkpoint)}")

    return checkpoint


def resolve_stage2_checkpoint(ckpt_dir: Path, stage2_ckpt_name: str | None) -> Path:
    """Resolve a Stage 2 checkpoint path, defaulting to the latest epoch checkpoint."""
    if stage2_ckpt_name:
        candidate = ckpt_dir / stage2_ckpt_name
        if not candidate.exists():
            raise FileNotFoundError(f"Stage 2 checkpoint not found: {candidate}")
        return candidate

    stage2_paths = list(ckpt_dir.glob("stage2_epoch_*.pt"))
    if not stage2_paths:
        raise FileNotFoundError(f"No stage2 checkpoints found in {ckpt_dir}")

    def _epoch_num(path: Path) -> int:
        match = re.search(r"stage2_epoch_(\d+)\.pt$", path.name)
        return int(match.group(1)) if match else -1

    return max(stage2_paths, key=_epoch_num)