loader.py

"""PyTorch Dataset for the LANSCE surrogate training data.

Reads `index.parquet` to enumerate runs, then lazily opens each run's
`fields.h5` on `__getitem__`. Returns a dict so the caller can ignore
modalities they don't want for a given training phase:

    Phase 1 (scalar regression): only use `params` + `metrics`
    Phase 2 (+ 1D histograms):    also read the reducedfiles via
                                  `load_reduced(run_dir)`
    Phase 3 (full field surrogate): use `early` + `full`

Usage:

    from torch.utils.data import DataLoader
    from loader import LansceFieldsDataset

    ds = LansceFieldsDataset("surrogate_model_dataset", phase="scalar")
    for sample in DataLoader(ds, batch_size=8, num_workers=4):
        ...

`phase` gates which tensors are returned to keep I/O minimal:

    "scalar"     -> {"params", "metrics"} only
    "histogram"  -> adds {"hist_uz", "hist_x", "hist_y"}
    "field"      -> adds {"early", "full"} read from fields.h5
    "all"        -> everything

HDF5 opens are per-sample (one file per run), so num_workers>0 is safe.
"""

from __future__ import annotations

from pathlib import Path
from typing import Literal

import h5py
import numpy as np
import pandas as pd
import torch
from torch.utils.data import Dataset


PARAM_COLUMNS = [
    "beam_energy", "source_density", "puller_offset", "T_eV",
    "r_exit", "r_entrance", "r_puller",
]

METRIC_COLUMNS = [
    "envelope_x_m", "envelope_y_m",
    "emit_x_norm_mmmrad", "emit_y_norm_mmmrad",
    "energy_mean_kev", "energy_std_kev",
    "theta_x_rms_mrad", "theta_y_rms_mrad",
    "transmission",
]

HIST_BASENAMES = {
    "hist_uz": "reducedfilesbeam_uz_hist.txt",
    "hist_x":  "reducedfilesbeam_x_hist.txt",
    "hist_y":  "reducedfilesbeam_y_hist.txt",
}


Phase = Literal["scalar", "histogram", "field", "all"]


def _load_hist_tensor(path: Path) -> torch.Tensor:
    """Parse a WarpX ParticleHistogram reducedfile into a (T, bins) tensor."""
    if not path.exists() or path.stat().st_size == 0:
        return torch.empty(0, 0)
    arr = np.loadtxt(path)
    if arr.ndim == 1:
        arr = arr[None, :]
    # Columns 0 and 1 are step and time; the rest are bin counts.
    return torch.from_numpy(arr[:, 2:].astype(np.float32))


class LansceFieldsDataset(Dataset):
    def __init__(
        self,
        root: str | Path,
        phase: Phase = "scalar",
        index_path: str | Path | None = None,
        drop_failed: bool = True,
    ):
        self.root = Path(root)
        self.phase = phase
        # Prefer index.parquet (fast column projection); fall back to
        # manifest.csv if pyarrow/fastparquet wasn't installed when the
        # sweep ran.
        if index_path is not None:
            index_path = Path(index_path)
            df = (pd.read_parquet(index_path) if index_path.suffix == ".parquet"
                  else pd.read_csv(index_path))
        else:
            pq = self.root / "index.parquet"
            csv = self.root / "manifest.csv"
            if pq.exists():
                df = pd.read_parquet(pq)
            elif csv.exists():
                df = pd.read_csv(csv)
                df["run_dir"] = df["run_id"].map(lambda i: f"runs/run_{int(i):04d}")
                df["fields_h5"] = df["run_dir"] + "/fields.h5"
            else:
                raise FileNotFoundError(
                    f"neither index.parquet nor manifest.csv found under {self.root}"
                )
        if drop_failed:
            df = df[df["status"].fillna("") == "ok"].reset_index(drop=True)
        self.df = df

    def __len__(self) -> int:
        return len(self.df)

    def __getitem__(self, idx: int) -> dict[str, torch.Tensor]:
        row = self.df.iloc[idx]
        run_dir = self.root / row["run_dir"]

        sample: dict[str, torch.Tensor] = {
            "run_id": torch.tensor(int(row["run_id"]), dtype=torch.int64),
            "params": torch.tensor(
                [float(row[c]) for c in PARAM_COLUMNS], dtype=torch.float32
            ),
            "metrics": torch.tensor(
                [float(row[c]) for c in METRIC_COLUMNS], dtype=torch.float32
            ),
        }

        if self.phase in ("histogram", "all"):
            for key, fname in HIST_BASENAMES.items():
                sample[key] = _load_hist_tensor(run_dir / "diags" / fname)

        if self.phase in ("field", "all"):
            fields_h5 = run_dir / "fields.h5"
            if not fields_h5.exists():
                raise FileNotFoundError(f"{fields_h5} not packed yet")
            with h5py.File(fields_h5, "r") as h5:
                if "early" in h5:
                    sample["early"] = torch.from_numpy(h5["early/fields"][()])
                    sample["early_times"] = torch.from_numpy(h5["early/times"][()])
                if "full" in h5:
                    sample["full"] = torch.from_numpy(h5["full/fields"][()])
                    sample["full_times"] = torch.from_numpy(h5["full/times"][()])

        return sample


def load_reduced(run_dir: Path) -> dict[str, np.ndarray]:
    """Read scalar reduced diagnostics (beam_number, beam_jz, eb_charge).

    Returns time-series as raw numpy arrays with shape (T, n_cols).
    """
    d = Path(run_dir) / "diags"
    out = {}
    for basename in ("beam_number", "beam_jz", "eb_charge"):
        p = d / f"reducedfiles{basename}.txt"
        if p.exists() and p.stat().st_size > 0:
            out[basename] = np.loadtxt(p)
    return out