itouchz's picture
Upload methods/ (18-method baseline panel implementations)
02412f8 verified
Raw
History Blame Contribute Delete
20.7 kB
"""ITFormer authors' inference pipeline wrapped behind a Pythonic engine.
This module exposes :class:`ITFormerEngine`, a thin adapter that drives
the official ITFormer (ICML 2025 Poster) ``inference.py`` script in
batched mode. Each batch of (prompt, time-series) pairs is written to a
temp ``ts_data.h5`` + ``qa_data.jsonl`` + ``infer.yaml`` triple, the
authors' CLI is invoked once, and the resulting ``inference_results_*.json``
is read back and split into one generated string per input.
Usage from :mod:`methods.llm_ts_reason.ITFormer` (per-task predictors):
engine = ITFormerEngine.create(
upstream_dir=..., # auto-detected if None
checkpoint=..., # default Pandalin98/ITFormer-ICML25 (0.5B)
input_len=600, # ITFormer architecture default
prefix_num=10,
batch_size=8,
max_new_tokens=200,
device="cuda" or "cpu",
)
texts = engine.predict_batch(
ts_arr=np.ndarray (N, L) float32, # one signal per prompt
prompts=[str, ...] (length N),
)
Why this exists (limitation disclosure)
---------------------------------------
ITFormer's published inference is a CLI driver over the EngineMT-QA
schema (HDF5 single-channel signals + JSONL conversation pairs), not a
Python-importable function. The authors' ``main_inference(args)`` pulls
in ``Accelerator()``, parses YAML, materialises a ``DataLoader``, and
calls ``model.generate()``. Wrapping that as a per-prompt Python call
would mean re-implementing dataset / collator / generate semantics,
which the user explicitly forbids ("ALWAYS use the authors' official
repo code; never re-implement from the paper text").
The compromise here is faithful but coarse: we drive the authors' code
verbatim via subprocess, paying one process startup per
:meth:`predict_batch` call. The Method-level adapter
(:class:`methods.llm_ts_reason.ITFormer`) calls ``predict_batch`` ONCE
per task (not once per sample), so subprocess overhead is amortised.
What is faithful
----------------
- Authors' ``inference.py`` runs unmodified: same model class, same
``generate`` call, same tokenizer, same chat-template substitution of
``<ts>`` -> ``<|image_pad|> * prefix_num``.
- HDF5 + JSONL format matches the authors' ``TsQaDataset`` reader.
- YAML config fields mirror ``yaml/infer.yaml`` defaults verified against
upstream commit on 2026-05-05 via WebFetch.
What is NOT faithful (documented limitation; mirrored in DRAFT.md)
------------------------------------------------------------------
- The published pipeline expects EngineMT-QA (industrial engine sensor
signals); we feed it (close-price lookback, financial QA prompt) pairs.
This is the same domain transfer the legacy
:func:`baselines.llm_ts_reason.run_itformer` already performed; we are
not making any NEW deviation, only honouring the published format.
- Stages 2/3 in ITFormer's protocol are MULTIPLE-CHOICE; our T2..T7 are
open-ended numeric / structured. We therefore mark every prompt as
``stage='1'`` (open question) so the model emits free-form text, which
the calling parser then coerces to a number / dict / DataFrame row.
"""
from __future__ import annotations
import glob
import json
import logging
import os
import shutil
import subprocess
import sys
import tempfile
import time
from dataclasses import dataclass, field
from pathlib import Path
from typing import Any, Sequence
import numpy as np
logger = logging.getLogger(__name__)
# Upstream repo locator. We prefer to reuse the existing legacy clone at
# ``baselines/_vendor/itformer/`` (3.5 MB; already cloned by the legacy
# baselines path) over a duplicate clone under ``methods/_vendored/``.
# If neither path exists we shallow-clone into ``methods/_vendored/itformer/upstream``.
_UPSTREAM_URL = "https://github.com/Pandalin98/ITFormer-ICML25.git"
_UPSTREAM_REF = "main"
def _legacy_clone_dir() -> Path:
"""Return the legacy ``baselines/_vendor/itformer`` clone path."""
return (
Path(__file__).resolve().parents[3]
/ "baselines"
/ "_vendor"
/ "itformer"
)
def _vendored_clone_dir() -> Path:
"""Return the canonical ``methods/_vendored/itformer/upstream`` clone path."""
return Path(__file__).resolve().parent / "upstream"
def _is_repo(path: Path) -> bool:
return (path / "inference.py").is_file() and (path / "models").is_dir()
def _ensure_upstream(*, allow_clone: bool = True) -> Path:
"""Locate an upstream ITFormer clone, optionally cloning if missing.
Lookup order:
1. ``$MACROLENS_ITFORMER_REPO`` env var (explicit override).
2. ``baselines/_vendor/itformer/`` (legacy clone — preferred to avoid
duplicating 3.5 MB on disk).
3. ``methods/_vendored/itformer/upstream/`` (this module's local
clone slot).
4. Shallow-clone into slot (3) if ``allow_clone`` is True.
"""
explicit = os.environ.get("MACROLENS_ITFORMER_REPO", "").strip()
if explicit:
path = Path(explicit).expanduser().resolve()
if not _is_repo(path):
raise FileNotFoundError(
f"MACROLENS_ITFORMER_REPO={path} is not an ITFormer repo "
"(expected inference.py + models/ at the path root)."
)
return path
legacy = _legacy_clone_dir()
if _is_repo(legacy):
return legacy
vendored = _vendored_clone_dir()
if _is_repo(vendored):
return vendored
if not allow_clone:
raise FileNotFoundError(
"ITFormer upstream clone not found. Set "
"MACROLENS_ITFORMER_REPO=<path> or run with allow_clone=True."
)
vendored.parent.mkdir(parents=True, exist_ok=True)
logger.info("Cloning %s into %s ...", _UPSTREAM_URL, vendored)
subprocess.run(
[
"git", "clone", "--depth", "1", "--branch", _UPSTREAM_REF,
_UPSTREAM_URL, str(vendored),
],
check=True,
capture_output=True,
)
return vendored
# ── ITFormer config defaults (verified against upstream yaml/infer.yaml,
# fetched 2026-05-05) ─────────────────────────────────────────────────
@dataclass
class _ITFormerCfg:
# Architecture (must match the checkpoint that ``model_checkpoint`` loads).
d_model: int = 512
n_heads: int = 8
e_layers: int = 4
patch_len: int = 60
stride: int = 60
input_len: int = 600
dropout: float = 0.1
it_d_model: int = 896
it_n_heads: int = 16
it_layers: int = 2
it_dropout: float = 0.1
prefix_num: int = 25
# I/O.
fp16: bool = True
dataloader_pin_memory: bool = True
dataloader_num_workers: int = 4
# Per-batch override fields (filled at run time).
ts_path_test: str = ""
qa_path_test: str = ""
def to_dict(self) -> dict[str, Any]:
d = {
"d_model": self.d_model,
"n_heads": self.n_heads,
"e_layers": self.e_layers,
"patch_len": self.patch_len,
"stride": self.stride,
"input_len": self.input_len,
"dropout": self.dropout,
"it_d_model": self.it_d_model,
"it_n_heads": self.it_n_heads,
"it_layers": self.it_layers,
"it_dropout": self.it_dropout,
"prefix_num": self.prefix_num,
"fp16": self.fp16,
"dataloader_pin_memory": self.dataloader_pin_memory,
"dataloader_num_workers": self.dataloader_num_workers,
"ts_path_test": self.ts_path_test,
"qa_path_test": self.qa_path_test,
}
return d
# ── Engine ─────────────────────────────────────────────────────────────────
class ITFormerEngine:
"""Thin Python wrapper over the authors' ``inference.py`` CLI.
Public surface:
- :meth:`create` (classmethod) — locate / clone the upstream repo and
validate it; returns a configured engine.
- :meth:`predict_batch` — run one CLI invocation over N samples.
Construction is cheap (validates paths only); the actual model load
happens inside the subprocess on every :meth:`predict_batch` call.
"""
def __init__(
self,
*,
upstream_dir: Path,
checkpoint: str = "pandalin98/ITFormer-7B",
cfg: _ITFormerCfg | None = None,
batch_size: int = 8,
max_new_tokens: int = 200,
seed: int = 42,
timeout_sec: float = 3600.0,
python_exe: str | None = None,
) -> None:
if not _is_repo(upstream_dir):
raise FileNotFoundError(
f"upstream_dir={upstream_dir} is not a valid ITFormer repo."
)
self.upstream_dir = upstream_dir
self.checkpoint = str(checkpoint)
self.cfg = cfg or _ITFormerCfg()
self.batch_size = int(batch_size)
self.max_new_tokens = int(max_new_tokens)
self.seed = int(seed)
self.timeout_sec = float(timeout_sec)
# The authors' ``inference.py`` imports torch / transformers /
# accelerate / h5py from the active interpreter's site-packages,
# so we default to the same Python that imported this module.
self.python_exe = python_exe or sys.executable
@classmethod
def create(
cls,
*,
upstream_dir: Path | str | None = None,
checkpoint: str = "pandalin98/ITFormer-7B",
input_len: int = 600,
prefix_num: int = 25,
batch_size: int = 8,
max_new_tokens: int = 200,
seed: int = 42,
allow_clone: bool = True,
) -> "ITFormerEngine":
"""Locate / clone the upstream repo and return a configured engine."""
if upstream_dir is None:
up = _ensure_upstream(allow_clone=allow_clone)
else:
up = Path(upstream_dir).expanduser().resolve()
if not _is_repo(up):
raise FileNotFoundError(
f"upstream_dir={up} is not a valid ITFormer repo."
)
cfg = _ITFormerCfg(
input_len=int(input_len),
prefix_num=int(prefix_num),
patch_len=min(60, int(input_len)),
stride=min(60, int(input_len)),
)
return cls(
upstream_dir=up,
checkpoint=checkpoint,
cfg=cfg,
batch_size=batch_size,
max_new_tokens=max_new_tokens,
seed=seed,
)
# ── HF checkpoint resolver ──
def _resolve_checkpoint_path(self) -> str:
"""Return a local-filesystem path to the checkpoint directory.
``self.checkpoint`` may be either a local path (passed straight
through after existence check) or a HuggingFace repo id like
``pandalin98/ITFormer-7B``. The authors' ``inference.py`` does
``os.path.exists(model_checkpoint)`` rather than HF auto-download,
so we resolve repo ids via ``huggingface_hub.snapshot_download``
(cached) and pass the resulting local directory.
Cached on the instance so multi-task runs hit the resolver once.
"""
if hasattr(self, "_resolved_checkpoint") and self._resolved_checkpoint:
return self._resolved_checkpoint
ckpt = str(self.checkpoint)
if os.path.isdir(ckpt) or os.path.isfile(ckpt):
self._resolved_checkpoint = ckpt
return ckpt
try:
from huggingface_hub import snapshot_download # lazy
except ImportError as exc:
raise RuntimeError(
"ITFormer engine: huggingface_hub is required to resolve "
f"checkpoint repo id {ckpt!r}; install or pass a local "
"directory path instead."
) from exc
local_path = snapshot_download(repo_id=ckpt)
self._resolved_checkpoint = str(local_path)
return self._resolved_checkpoint
# ── Batch driver ──
def predict_batch(
self,
*,
ts_arr: np.ndarray,
prompts: Sequence[str],
scratch_dir: Path | None = None,
) -> list[str]:
"""Run one CLI invocation over N samples; return N decoded strings.
Args:
ts_arr: ``(N, L)`` float32 single-channel signals. ``L`` should
match ``self.cfg.input_len``; if shorter, the array is
left-padded with zeros (the authors' loader has no notion
of attention masks for the time series so zero-padding is
the standard convention).
prompts: ``N`` natural-language prompts. ``"<ts>"`` is
inserted at position 0 if not already present (the
authors' chat-template substitutes it for prefix_num
``<|image_pad|>`` tokens).
scratch_dir: optional override; defaults to a fresh tempdir.
Returns:
A length-N list of generated text strings, in the same order
as ``prompts``. Missing entries (parse failure / index gap)
are returned as empty strings — the caller decides how to
handle parse errors.
"""
prompts = list(prompts)
n = len(prompts)
if ts_arr.ndim != 2:
raise ValueError(
f"ts_arr must be 2-D (N, L); got shape {ts_arr.shape}"
)
if ts_arr.shape[0] != n:
raise ValueError(
f"ts_arr.shape[0]={ts_arr.shape[0]} != len(prompts)={n}"
)
if n == 0:
return []
# Pad / truncate to input_len.
L_target = int(self.cfg.input_len)
L_in = ts_arr.shape[1]
if L_in == L_target:
ts_padded = ts_arr.astype(np.float32, copy=False)
elif L_in > L_target:
ts_padded = ts_arr[:, -L_target:].astype(np.float32, copy=False)
else:
ts_padded = np.zeros((n, L_target), dtype=np.float32)
ts_padded[:, -L_in:] = ts_arr.astype(np.float32, copy=False)
# Materialise scratch directory.
owns_scratch = scratch_dir is None
scratch = (
Path(tempfile.mkdtemp(prefix="itformer_engine_"))
if owns_scratch
else Path(scratch_dir)
)
scratch.mkdir(parents=True, exist_ok=True)
try:
return self._invoke(scratch, ts_padded, prompts)
finally:
if owns_scratch:
shutil.rmtree(scratch, ignore_errors=True)
# ── Internals ──
def _invoke(
self,
scratch: Path,
ts_padded: np.ndarray,
prompts: list[str],
) -> list[str]:
# Lazy h5py import so this module is importable on minimal envs.
import h5py
import yaml
n = len(prompts)
h5_path = scratch / "ts_data.h5"
# Authors' TimeSeriesEncoder.forward expects (B, L, F) tensors (it
# transposes dims (1, 2) before patching). The TsQaDataset returns
# ``ts_values`` straight from this HDF5 row, so we must write the
# dataset as (N, L, 1) — single-channel signals — to get
# batched (B, L, 1) tensors at training time.
if ts_padded.ndim == 2:
ts_padded_3d = ts_padded[..., None]
else:
ts_padded_3d = ts_padded
with h5py.File(str(h5_path), "w") as f:
f.create_dataset("seq_data", data=ts_padded_3d)
# Build JSONL — one entry per prompt; id is 1-indexed (the authors
# subtract 1 in TsQaDataset before HDF5 lookup). All prompts are
# tagged stage='1' (open question) since our parsers consume
# free-form text, not multiple-choice answers; the upstream
# build_index loop drops any entry whose stage is outside
# {'1','2','3','4'}, so '1' is the only safe open-text choice.
jsonl_path = scratch / "qa_data.jsonl"
with open(jsonl_path, "w", encoding="utf-8") as f:
for i, prompt in enumerate(prompts):
# Insert <ts> marker once if absent so chat-template
# substitutes the prefix_num image_pad tokens.
q = prompt if "<ts>" in prompt else f"<ts> {prompt}"
entry = {
"id": i + 1,
"conversations": [
{"stage": "1", "attribute": "open",
"value": q},
# The upstream loader requires an answer slot for
# eval mode (it tokenises labels for free-form
# comparison). A blank suffices since we only
# consume ``prediction``, never ``is_correct``.
{"stage": "1", "attribute": "open", "value": ""},
],
}
f.write(json.dumps(entry) + "\n")
# Write YAML.
yaml_path = scratch / "infer.yaml"
cfg_dict = self.cfg.to_dict()
cfg_dict["ts_path_test"] = str(h5_path)
cfg_dict["qa_path_test"] = str(jsonl_path)
with open(yaml_path, "w", encoding="utf-8") as f:
yaml.safe_dump(cfg_dict, f)
# Output dir for inference_results_*.json.
output_dir = scratch / "results"
output_dir.mkdir(parents=True, exist_ok=True)
# Compose the subprocess command. We invoke ``inference.py``
# directly (no ``accelerate launch``) since single-GPU /
# single-process is sufficient for our batch sizes; ``Accelerator()``
# gracefully degrades to a no-op single-process when launched
# without ``accelerate``.
# Authors' inference.py does ``os.path.exists(model_checkpoint)``
# rather than HF auto-download, so an HF repo id like
# ``pandalin98/ITFormer-7B`` is rejected. Resolve to a local
# snapshot path (cached after first hit) so the subprocess gets
# an actual directory.
resolved_ckpt = self._resolve_checkpoint_path()
cmd = [
self.python_exe,
str(self.upstream_dir / "inference.py"),
"--config", str(yaml_path),
"--model_checkpoint", resolved_ckpt,
"--output_dir", str(output_dir),
"--seed", str(self.seed),
"--batch_size", str(self.batch_size),
"--max_new_tokens", str(self.max_new_tokens),
]
env = dict(os.environ)
# Harden against the upstream's relative imports (``from dataset...``
# ``from models...``) by pinning PYTHONPATH and cwd to the repo root.
env["PYTHONPATH"] = (
f"{self.upstream_dir}{os.pathsep}{env.get('PYTHONPATH','')}"
)
t0 = time.time()
try:
rc = subprocess.run(
cmd,
cwd=str(self.upstream_dir),
env=env,
capture_output=True,
text=True,
timeout=self.timeout_sec,
)
except subprocess.TimeoutExpired as exc:
logger.error(
"ITFormer inference timed out after %.0fs (n=%d): %s",
self.timeout_sec, n, exc,
)
return [""] * n
elapsed = time.time() - t0
if rc.returncode != 0:
tail = (rc.stderr or rc.stdout or "")[-1000:]
logger.error(
"ITFormer inference failed (rc=%d, n=%d, %.1fs): %s",
rc.returncode, n, elapsed, tail,
)
return [""] * n
return self._read_results(output_dir, n)
@staticmethod
def _read_results(output_dir: Path, n: int) -> list[str]:
files = sorted(glob.glob(str(output_dir / "inference_results_*.json")))
if not files:
logger.error("ITFormer: no result files in %s", output_dir)
return [""] * n
with open(files[-1], "r", encoding="utf-8") as f:
results = json.load(f)
if not isinstance(results, list):
logger.error("ITFormer: result file is not a list: %s", files[-1])
return [""] * n
# Result entries carry ``index`` matching JSONL line_num (0-based).
# Build a dense list aligned to input order.
out: list[str] = [""] * n
for entry in results:
try:
idx = int(entry.get("index"))
pred = str(entry.get("prediction", ""))
except (TypeError, ValueError):
continue
if 0 <= idx < n:
out[idx] = pred
return out
__all__ = ["ITFormerEngine"]