File size: 7,892 Bytes
ae73c7f | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 | """
Loaders for The Well data.
CONTRACT (see provenance.py for shared error types):
- get_dataset() loads REAL data only. It NEVER silently substitutes
synthetic data. On any failure it raises DataLoadError or
SchemaValidationError. This is a deliberate design decision:
retraining on the wrong data (or on synthetic data believed to be
real) is worse than a loud failure.
- Synthetic data is only available via the separate, explicitly-named
get_synthetic_dataset() — callers must opt in on purpose.
- Tensor layout is never guessed from shape magnitude. Callers must
declare the expected channel_layout; a mismatch is a hard error,
not a silent transpose.
Setup for real data:
1. Run: the-well-download --base-path ./data/well --dataset active_matter --split train
2. Or place .hdf5 files under data/real/ with a known, declared schema.
"""
from __future__ import annotations
import os
from pathlib import Path
from typing import Optional, List, Literal
import torch
from torch.utils.data import Dataset
import numpy as np
from .provenance import DataLoadError, SchemaValidationError
ChannelLayout = Literal["channels_first", "channels_last"]
class LocalWellHDF5(Dataset):
KNOWN_KEYS = ("t0_fields", "fields", "data", "x", "trajectory")
def __init__(
self,
root: str,
max_samples: int = 256,
n_steps: Optional[int] = None,
expected_channels: int = 2,
channel_layout: ChannelLayout = "channels_first",
strict: bool = True,
):
self.root = Path(root)
self.files = sorted(self.root.rglob("*.hdf5")) + sorted(self.root.rglob("*.h5"))
if not self.files:
raise DataLoadError(
f"no .hdf5/.h5 files found under {root}",
outcome_code="NO_FILES_FOUND",
)
self.max_samples = max_samples
self.n_steps = n_steps
self.expected_channels = expected_channels
self.channel_layout = channel_layout
self.strict = strict
self.samples: List[torch.Tensor] = []
self._load()
def _load(self):
try:
import h5py
except ImportError as e:
raise DataLoadError(
"h5py is required to load local Well HDF5 files but is not installed",
outcome_code="MISSING_DEPENDENCY",
) from e
count = 0
rejected = []
for fp in self.files:
if count >= self.max_samples:
break
with h5py.File(fp, "r") as f:
arr = None
for k in self.KNOWN_KEYS:
if k in f:
arr = f[k]
break
if arr is None and list(f.keys()):
arr = f[list(f.keys())[0]]
if arr is None:
rejected.append((str(fp), "no recognizable dataset key"))
continue
data = np.array(arr)
try:
if data.ndim == 5: # N, T, C/H, H/C, W or similar
for i in range(min(data.shape[0], self.max_samples - count)):
traj = self._validate_and_orient(data[i], fp)
self.samples.append(torch.from_numpy(traj).float())
count += 1
elif data.ndim == 4:
traj = self._validate_and_orient(data, fp)
self.samples.append(torch.from_numpy(traj).float())
count += 1
else:
rejected.append((str(fp), f"unsupported ndim={data.ndim}"))
except SchemaValidationError as e:
rejected.append((str(fp), e.detail))
if self.strict:
raise
if rejected and not self.strict:
print(f"[LocalWellHDF5] WARNING: {len(rejected)} file(s) rejected: {rejected}")
if not self.samples:
raise SchemaValidationError(
f"no valid trajectories loaded from {self.root}; "
f"rejected files: {rejected}",
outcome_code="NO_VALID_TRAJECTORIES",
)
print(f"[LocalWellHDF5] loaded {len(self.samples)} trajectories "
f"from {len(self.files)} files (schema={self.channel_layout}, "
f"channels={self.expected_channels})")
def _validate_and_orient(self, data: np.ndarray, fp: Path) -> np.ndarray:
if data.ndim != 4:
raise SchemaValidationError(
f"{fp.name}: expected 4D (T,C,H,W)-like array, got ndim={data.ndim}",
outcome_code="WRONG_NDIM",
)
T, A, B, C_ = data.shape
if self.channel_layout == "channels_first":
channel_axis_size = A
oriented = data
elif self.channel_layout == "channels_last":
channel_axis_size = C_
oriented = np.transpose(data, (0, 3, 1, 2))
else:
raise SchemaValidationError(
f"unknown channel_layout '{self.channel_layout}'",
outcome_code="INVALID_LAYOUT_SPEC",
)
if channel_axis_size != self.expected_channels:
raise SchemaValidationError(
f"{fp.name}: declared channel_layout='{self.channel_layout}' "
f"implies {channel_axis_size} channels, but "
f"expected_channels={self.expected_channels}. "
f"Refusing to guess a different layout for this file — "
f"pass the correct channel_layout/expected_channels explicitly.",
outcome_code="CHANNEL_COUNT_MISMATCH",
)
return oriented
def __len__(self):
return len(self.samples)
def __getitem__(self, idx):
traj = self.samples[idx]
if self.n_steps is not None and traj.size(0) > self.n_steps:
traj = traj[: self.n_steps]
return {"fields": traj, "idx": idx}
def get_dataset(
max_samples: int = 128,
n_steps: int = 14,
expected_channels: int = 2,
channel_layout: ChannelLayout = "channels_first",
search_roots: Optional[List[str]] = None,
):
roots = search_roots or ["./data/real", "./data/well"]
existing_dirs = [r for r in roots if os.path.isdir(r)]
if not existing_dirs:
raise DataLoadError(
f"no real-data directory found among candidates: {roots}. "
f"Run `the-well-download ...` or place .hdf5 files under one "
f"of these paths before training.",
outcome_code="NO_DATA_DIRECTORY",
)
last_error = None
for root in existing_dirs:
try:
ds = LocalWellHDF5(
root,
max_samples=max_samples,
n_steps=n_steps,
expected_channels=expected_channels,
channel_layout=channel_layout,
strict=True,
)
print(f"[data] using REAL local Well data from {root} ({len(ds)} trajs)")
return ds, "REAL_LOCAL"
except (DataLoadError, SchemaValidationError) as e:
last_error = e
continue
raise last_error or DataLoadError(
f"no usable real data found in {existing_dirs}",
outcome_code="NO_DATA_DIRECTORY",
)
def get_synthetic_dataset(max_samples: int = 128, n_steps: int = 14):
"""
Explicit, opt-in synthetic data. Never called implicitly by get_dataset().
Returns: (dataset, provenance) where provenance == "SYNTHETIC".
"""
from .synthetic_fields import SyntheticWellLike
print("[data] EXPLICIT synthetic mode requested — not real Well data")
ds = SyntheticWellLike(n_samples=max_samples, n_steps=n_steps, height=32, width=32)
return ds, "SYNTHETIC"
|