torchflat 0.9.21

GPU-accelerated photometric preprocessing with UMI detrending for exoplanet transit searches

TorchFlat

GPU-native photometric preprocessing pipeline for exoplanet transit searches.

TorchFlat replaces the standard CPU preprocessing workflow (quality filtering, gap handling, sigma clipping, detrending, normalization, and windowing) with a GPU-accelerated pipeline. It uses UMI (Unified Median Iterative), a novel asymmetric robust location estimator implemented as a fused HIP/CUDA kernel, to detrend light curves faster and more accurately than existing methods.

Performance

Benchmarked on AMD Radeon RX 9060 XT (16 GB VRAM) with real TESS sector 6 data (19,618 stars):

Method	Rate	Full Sector	Speedup
Celix wotan 12-worker	4.2 stars/sec	~78 min	baseline
TorchFlat v0.5.0 (hybrid)	59.3 stars/sec	~5.5 min	14.2x
TorchFlat + UMI kernel	154 stars/sec	~2.1 min	37x

Transit Depth Recovery Accuracy

Injection-recovery test on 1000 real TESS stars, median per-star error (lower = better):

Depth	wotan biweight	wotan welsch	TorchFlat UMI	Winner
0.05% (Earth-size)	23.5%	24.0%	22.3%	TorchFlat
0.1% (super-Earths)	20.5%	18.5%	15.8%	TorchFlat
0.3% (sub-Neptunes)	12.7%	5.9%	4.9%	TorchFlat
0.5% (Neptunes)	5.1%	1.8%	2.4%	welsch
1.0% (hot Jupiters)	0.8%	0.7%	1.2%	welsch
5.0% (deep transits)	0.1%	0.1%	0.3%	both perfect

TorchFlat is more accurate at the transit depths where most detectable planets live (0.05-0.3%). Per-star detrending is 69x faster (3.4ms vs 234ms).

Known Planet Recovery

Validated on 802 confirmed exoplanets (81 TESS + 721 Kepler). UMI recovers more planets (425) than biweight, Welsch, and Savitzky-Golay combined (377).

Speed

Validated on 8 TESS sectors, 4 Kepler quarters (Q2, Q5, Q9, Q17), K2, and 10,000-star parameter validation. All results with 95% bootstrap confidence intervals. Full data in results/.

The UMI Algorithm

UMI (Unified Median Iterative) is a three-phase robust location estimator:

1. Quickselect median -- exact median via O(n) selection algorithm, computed per-thread on GPU
2. Upper-RMS scale -- RMS of points above the median only. Transit dips never contaminate the scale estimate, giving a tighter and more accurate noise measurement than standard MAD
3. Asymmetric bisquare iterations -- weighted location refinement where downward deviations (transit dips) are penalized 2x more than upward ones

The asymmetric weight function exploits the fact that transits are always below the continuum. Standard biweight treats dips and spikes equally. UMI penalizes dips more aggressively, so the trend stays above the transit and transit depth is preserved.

All three phases run in a single fused GPU kernel call -- median, upper-RMS, and 5 iterations happen per-thread with zero global memory traffic between steps.

When the GPU kernel is not available (no ROCm/CUDA toolkit), UMI falls back to a pure-PyTorch path using torch.sort for median + upper-RMS scale.

Installation

pip install torchflat

Or from source:

git clone https://github.com/omarkhan2217/TorchFlat.git
cd TorchFlat
pip install -e .

Requirements: PyTorch >= 2.1.0, NumPy >= 1.24.0, SciPy >= 1.10.0

Works with both NVIDIA CUDA and AMD ROCm (via PyTorch's unified CUDA API). The UMI kernel compiles automatically on first use via JIT (requires ROCm SDK or CUDA toolkit).

Quick Start

Process a TESS sector

import numpy as np
import torchflat

star_data = [
    {
        "time": np.load("star_001_time.npy"),
        "pdcsap_flux": np.load("star_001_pdcsap.npy"),
        "sap_flux": np.load("star_001_sap.npy"),
        "quality": np.load("star_001_quality.npy"),
    }
    # ... for each star in the sector
]

results, skipped = torchflat.preprocess_sector(
    star_data,
    device="cuda",
)

for i, result in enumerate(results):
    if not result:
        continue
    windows = result["windows_2048"]
    trend = result["trend"]

Standalone UMI detrending

import torch
from torchflat import umi_detrend

# flux, time, valid_mask, segment_id are [B, L] tensors on GPU
detrended, trend = umi_detrend(
    flux, time, valid_mask, segment_id,
    window_length_days=0.5,
    asymmetry=2.0,       # 2.0=best accuracy, 1.0=variable stars, 1.5=mixed
)

Architecture

TorchFlat implements two processing tracks:

• Track A (Transit Search): Quality filter > gap interpolation > sigma clipping > UMI detrending > normalization > multi-scale window extraction
• Track B (Anomaly Detection): Quality filter > gap interpolation > conservative clipping > FFT highpass filter > MAD normalization > fixed-length padding

UMI kernel

The direct HIP/CUDA kernel (torchflat/csrc/umi_kernel.cu) runs one thread per (star, window position) pair. Each thread reads directly from the raw [B, L] flux array, no unfold or tensor copies needed:

1. Reads W values from raw flux, checks segment validity inline
2. Quickselect for exact median (O(n))
3. Upper-RMS scale from above-median points (no sort needed)
4. 5 asymmetric bisquare iterations
5. Writes the final location estimate

VRAM usage: 319 MB for a 50-star batch. The kernel compiles via JIT on first import and is cached for subsequent runs.

Validation

All validation results are saved as JSON in results/:

Validation	Result	File
Asymmetry train/test split	optimal=2.0, validated on 2000+10,000 stars	asymmetry_validation_2k.json
Known planet recovery	UMI wins 425/802 confirmed planets (53%), more than all others combined	known_planet_recovery_all.json
Multi-sector consistency	UMI wins 9/15 across sectors 6, 7, 12 (2000 stars each)	multisector_validation_2k.json
Multi-mission	Kepler: 10.5% vs wotan 36.6% at 0.1%. K2: 4.5% vs 46.6% at 0.5%	multi_mission.json
Kepler multi-quarter	Q2=3.7%, Q5=4.2%, Q9=4.5%, Q17=5.2% at 0.1% (consistent)	kepler_multi_quarter.json
Method comparison	UMI #1 at 0.1% vs 8 methods (biweight, welsch, lowess, etc.)	method_comparison.md

135/135 unit tests passing.

CLI

# Detrend a TESS sector
torchflat umi_detrend --input /path/to/fits/

# Detrend a single star
torchflat umi_detrend --input star.fits --output-format fits

# Kepler data
torchflat umi_detrend --input /path/to/kepler/ --mission kepler

# Plot a star
torchflat plot --fits star.fits --save output.png

# Speed benchmark
torchflat benchmark --input /path/to/fits/ --n-stars 500

Benchmarks

# Full sector speed benchmark
python benchmarks/bench_real_tess.py --data-dir /path/to/fits/sector_6 --n-stars 19618

# Asymmetry parameter validation
python benchmarks/validate_asymmetry.py

# Known planet recovery
python benchmarks/validate_known_planets.py

# Multi-sector validation
python benchmarks/validate_multisector.py

Note: Set $env:TORCHFLAT_NO_KERNEL = "0" (PowerShell) or export TORCHFLAT_NO_KERNEL=0 (bash) to enable the UMI kernel.

API Reference

Main Entry Points

• torchflat.preprocess_sector(star_data, ...) -- Full pipeline (Track A + Track B).
• torchflat.preprocess_track_a(times, fluxes, qualities, ...) -- Track A only.
• torchflat.preprocess_track_b(times, sap_fluxes, qualities, ...) -- Track B only.
• torchflat.umi_detrend(flux, time, valid_mask, segment_id, ...) -- Standalone UMI kernel.

Key Parameters

Parameter	Default	Description
device	"cuda"	Torch device
window_length_days	0.5	Sliding window width (days)
asymmetry	2.0	Dip penalty: 2.0 (TESS), 3.0 (Kepler), 1.0 (variable stars -- avoids -7240 ppm bias)
n_iter	5	Number of bisquare iterations
cval	5.0	Rejection threshold in MAD units
skip_track_b	False	Skip Track B (FFT highpass)
window_scales	4 scales	[(256,128), (512,256), (2048,512), (8192,2048)]
dtype	float32	Computation precision

Limitations

• Tested on AMD and NVIDIA. HIP kernel validated on AMD RX 9060 XT (ROCm 7.2). CUDA kernel validated on NVIDIA GTX 1650 Ti (CUDA 12.8). Requires CUDA toolkit 12.8+ and Visual Studio Build Tools on Windows.
• Fallback is slower. Without the compiled HIP/CUDA kernel, UMI uses torch.sort (20x slower, 44x more VRAM). Install the ROCm or CUDA toolkit to enable the fused kernel.
• Asymmetry bias. The default asymmetry=2.0 introduces a -451 ppm bias on flat stars. This is below TESS noise (~1000 ppm) but may matter for population-level radius studies. Use --bias-correct to remove it, or --asymmetry 1.0 for zero bias.
• Variable star bias. On stars with >1% variability, asymmetry=2.0 causes -7240 ppm bias. Use --asymmetry 1.0 for variable stars.
• Kepler long-cadence. Kepler 30-min cadence gives W=25 (vs TESS W=361). The min_segment_points parameter auto-scales to W//3 to avoid all-NaN output. Validated on Quarters 2, 5, 9, 17 (4000 stars).
• 8-hour transits. Both UMI and wotan fail on transits longer than ~5 hours with the default 0.5-day window. Use --window-length 1.5 for long-duration transits.

Development

git clone https://github.com/omarkhan2217/TorchFlat.git
cd TorchFlat
pip install -e ".[dev]"
pytest tests/ -v

Citation

If you use TorchFlat in your research, please cite:

@software{torchflat,
  author = {Khan, Omar},
  title = {TorchFlat: GPU-Accelerated Photometric Preprocessing with UMI Detrending},
  year = {2026},
  url = {https://github.com/omarkhan2217/TorchFlat}
}

License

MIT License. See LICENSE for details.