torch-ieof 0.2.0

PyTorch TIEOF: higher-order tensor (PARAFAC/HOOI/HOSVD) DINEOF reconstruction for missing satellite data

torch-ieof: torch-based implementation of tieof

This is a slimmed-down, PyTorch-backed rewrite of the original TIEOF package. All three of the paper's higher-order tensor decompositions — PARAFAC (CP-ALS), HOOI (iterative Tucker), and TruncHOSVD (closed-form Tucker) — are re-implemented in pure PyTorch with no tensorly, no oct2py, and no ray. A classical 2-D DINEOF estimator (iterative truncated SVD) is also included for comparison. Runs on numpy>=2 / scikit-learn>=1.5 / torch>=2.2 under Python 3.12+. The legacy GHER Fortran bridge, CLI scripts, and interpolator/ package were dropped (they live on the main_backup branch).

📖 Richer docs: open README.html for a tabbed reference that summarises the three decomposition methods and the DINEOF reconstruction loop as described in the paper.

Install

pip install -e .            # or:  uv pip install -e .
pip install -e ".[dev]"     # with pytest

Or with uv

uv add torch-ieof

Dependencies: numpy>=2, scipy>=1.13, scikit-learn>=1.5, torch>=2.2, tqdm, loguru.

Usage

Raw numpy tensor

import numpy as np
from torch_ieof import DINEOF3

shape = (n_lat, n_lon, n_time)
tensor = ...                # np.ndarray, NaN for missing values
model = DINEOF3(R=3, tensor_shape=shape)   # decomp_type="parafac" (default)
model._fit(tensor)
filled = model.reconstructed_tensor

Choosing a decomposition (decomp_type)

DINEOF3 exposes all three engines from the paper. Pick with decomp_type:

DINEOF3(R=3, tensor_shape=shape, decomp_type="parafac")      # CP-ALS (default)
DINEOF3(R=(4, 4, 6), tensor_shape=shape, decomp_type="hooi") # iterative Tucker
DINEOF3(R=4, tensor_shape=shape, decomp_type="trunchosvd")   # closed-form Tucker

• parafac — CP/PARAFAC via alternating least squares; R is a single integer rank shared across all modes. Most parsimonious and interpretable.
• hooi — Higher-Order Orthogonal Iteration (Tucker-ALS). R may be an int (broadcast to every mode) or a per-mode tuple (R_lat, R_lon, R_time).
• trunchosvd — truncated HOSVD: the closed-form Tucker initialiser. Cheapest, but lower quality (no ALS refinement). R as for hooi.

In the paper the three variants perform within each other's confidence intervals — the gain over classical DINEOF comes from working in the full 3-D feature space, not from the specific decomposition.

After fitting, model.predict_rank(k) reconstructs the tensor using only the first k components (an int for PARAFAC, an int or per-mode tuple for Tucker).

xarray DataArray (single variable)

from torch_ieof import reconstruct_dataarray

rec, model = reconstruct_dataarray(
    sst_da, R=3,
    lat_dim="lat", lon_dim="lon", time_dim="time",
    mask=land_mask,            # optional xr.DataArray or ndarray (lat, lon) bool
    to_center=True, nitemax=80, toliter=1e-4,
)
# rec has the same dim order, coords, and attrs as sst_da.

xarray Dataset — multivariate joint reconstruction

When one variable has heavy cloud cover but a related variable (sharing temporal dynamics) is more complete, jointly reconstructing them couples the variables through a shared temporal factor (Alvera-Azcárate-style multivariate DINEOF). Cloud gaps in the sparse variable are constrained by simultaneous observations in the others.

from torch_ieof import reconstruct_dataset

ds_recon, model = reconstruct_dataset(
    ds,                        # xr.Dataset of vars sharing (lat, lon, time)
    R=2 * 3,                   # rank — typically larger than per-var rank
    variables=["sst", "chl"],  # which vars to couple (default: all)
    masks={"sst": land_mask},  # optional per-var masks
    nitemax=120, toliter=1e-6,
)

Each variable is z-scored before stacking along the latitude axis, fit with a single CP decomposition, then split back and denormalised. to_center=False is the default for the joint path (z-scoring handles centering).

Classical 2-D DINEOF (truncated-SVD baseline)

For comparison, DINEOF implements the original 2-D method (Beckers & Rixen 2003): restrict to valid ocean pixels, reshape to a (space, time) matrix, and alternate truncated-SVD reconstruction with re-imputation of the gaps.

from torch_ieof import DINEOF

model = DINEOF(R=5, tensor_shape=shape, mask=ocean_mask)
model._fit(tensor)               # same NaN-for-missing convention as DINEOF3
filled = model.reconstructed_tensor

Same sklearn-style fit/predict/score API and predict_rank(k) helper as DINEOF3.

sklearn-style fit/predict

# X: (N, 3) integer coords (lat_idx, lon_idx, t_idx)
# y: (N,) values (NaN allowed for missing points)
model = DINEOF3(R=3, tensor_shape=(n_lat, n_lon, n_time))
model.fit(X, y)
y_pred = model.predict(X_test)
score = model.score(X_test, y_test)   # negative NRMSE

Key options:

• R — rank. Int (CP rank, or broadcast Tucker rank) or per-mode tuple for Tucker variants.
• decomp_type — "parafac" (default), "hooi", or "trunchosvd".
• mask — (n_lat, n_lon) boolean array (or .npy path). True = inside the investigated area. Cells outside are zeroed during fitting and set to NaN in the output.
• to_center / lat_lon_sep_centering — centre the tensor before fitting.
• keep_non_negative_only — clamp negative reconstructed values to 0.
• early_stopping — stop on absolute or gradient convergence (the paper's early-stopping mode); else stop on absolute error only.
• nitemax, toliter — outer reconstruction loop budget / tolerance.
• td_iter_max, tol — inner decomposition (CP-ALS / HOOI) budget / tolerance.

Performance, logging & progress

The hot loop runs in PyTorch. By default the device is cuda if available, otherwise CPU torch (MPS is skipped automatically because per-iter CPU fallbacks for SVD / pinv make it slower than CPU for typical CP ranks). Override with device="mps" for very large tensors, or dtype=torch.float64 for tighter numerics. Default dtype is float32.

Each fit logs a single INFO line at the start (shape, R, device, dtype, missing fraction) and one at the end (iters, final error). Iteration progress is shown via a live tqdm bar with err and Δerr in the postfix. Disable the bar with progress=False if you're piping output to a file or running in CI:

DINEOF3(R=4, tensor_shape=shape, progress=False)
reconstruct_dataarray(da, R=4, progress=False)

To suppress the INFO logs too:

from loguru import logger
logger.disable("torch_ieof")

Tests

pytest -q

Covers Kolda-convention unfolding, Khatri-Rao, CP reconstruction, end-to-end recovery of a noisy synthetic rank-R tensor, and a cloud-like patchy-mask case with a moving spatial feature.

Examples

Install the extra dependencies (matplotlib, xarray) and run:

pip install -e ".[examples]"
python examples/timeseries_demo.py   # SST-like field, cloud-blob gaps, saves PNG
python examples/xarray_workflow.py   # round-trip via xarray.DataArray

examples/xarray_workflow.py includes a small reconstruct_dataarray() helper showing how to wrap DINEOF3 for an xarray.DataArray with arbitrary dim order — useful as a template for plugging into an existing oceanographic workflow.

Credit, citing & license

The TIEOF method and its original implementation are the work of Kulikov, Inkova, Cherniuk, Teslyuk & Namsaraev. This package is a PyTorch repackaging; all scientific credit belongs to them. If you use this in academic work, please cite the original paper:

Kulikov, L.; Inkova, N.; Cherniuk, D.; Teslyuk, A.; Namsaraev, Z. TIEOF: Algorithm for Recovery of Missing Multidimensional Satellite Data on Water Bodies Based on Higher-Order Tensor Decompositions. Water 2021, 13(18), 2578. https://doi.org/10.3390/w13182578

Original code: https://github.com/theleokul/tieof.

Licensed under CC BY 4.0 — free to use, share and adapt with attribution to the authors above. See LICENSE.