paraug 0.1.1

Parity Augmentation — bit-exact CPU/GPU parity for image augmentation

paraug

Bit-exact CPU/GPU parity for image augmentation.

Languages: English | 繁體中文

paraug is a PyTorch-native augmentation library that guarantees the same seed produces the same output on CPU and CUDA. Per-primitive RNG is sampled on CPU regardless of tensor device, so a training run that randomly switches between CPU and GPU stages — or a unit test that swaps backends — stays deterministic.

Why parity matters

Most augmentation libraries (albumentations, kornia, torchvision) use device- local RNG. Same seed, different output across CPU/CUDA. This bites in three places:

1. Reproducibility: paper-to-code lineage breaks when a reviewer can't match published numbers.
2. Debugging: CPU-side unit tests don't catch GPU-only bugs and vice versa.
3. Distributed training: workers on heterogeneous hardware drift apart.

paraug fixes this by isolating RNG to CPU (torch.Generator(device="cpu")) and routing only the deterministic torch ops through device. Tolerance:

• Elementwise ops (gamma, noise, color jitter, …): atol 1e-6
• grid_sample-class ops (affine, perspective, tps, …): atol 2e-4 (bilinear ulp drift across ATen vs cuDNN)

Installation

pip install git+https://github.com/alieuidsh/paraug.git

PyPI release (pip install paraug) coming once the v0.1.x publish workflow is wired up — see issue tracker for status.

Quickstart

import torch
from paraug import AugPipeline

aug = AugPipeline({
    "geometric": {
        "affine": {"p": 1.0, "rot_deg": 15.0, "scale_range": (0.9, 1.1)},
        "tps":    {"p": 0.5, "max_disp": 12.0, "n_ctrl": 5},
    },
    "photometric": {
        "gamma":         {"p": 0.5},
        "color_jitter":  {"p": 0.5},
        "gaussian_blur": {"p": 0.3},
    },
})

img  = torch.rand(2, 3, 256, 256)         # (B, C, H, W)
mask = torch.ones(2, 1, 256, 256)         # optional

img_out, mask_out = aug(img, mask=mask, seed_base=42, epoch=0, step=0)

The same call on GPU is bit-exact within tolerance:

img_gpu  = img.cuda()
mask_gpu = mask.cuda()
img_cuda, mask_cuda = aug(img_gpu, mask=mask_gpu, seed_base=42, epoch=0, step=0)
assert (img_out - img_cuda.cpu()).abs().max() < 2e-4

Primitives

Geometric (7)

Name	Description
affine	Rotation + scale + translation via F.affine_grid
perspective	4-point homography from corner jitter
random_crop_pad	Scale-then-pad crop, area-preserving
elastic_transform	Bilinear-upsampled random displacement field
optical_distortion	Radial barrel / pincushion (k·r²)
random_shadow	Soft-blurred triangle multiplicative shadow
tps	Thin-plate-spline-like warp from low-res control grid

Photometric (24)

Intensity / color: gamma, color_jitter, hue_shift, random_grayscale, lighting, clahe, local_contrast, sharpness.

Noise: gaussian_noise, salt_pepper_noise, salt_patches.

Blur / artifacts: gaussian_blur, motion_blur, jpeg_approx.

Lighting: vignette, specular_highlight, specular_streaks.

Content overlays: cutout, paper_texture_overlay, watermark, random_text_overlay, background_compose, stains, creases.

Parity comparison

Library	Bit-exact CPU↔GPU	Per-item RNG	GPU native	Mask-aware	Batch-native	# Geometric¹	# Photometric¹	License
paraug	✓ (1e-6 / 2e-4)²	✓	✓ (torch)	✓	✓	7	24	Apache 2.0
albumentations	✗ (numpy-only)	✓	✗	✓	partial	~20	~50+	MIT
kornia	✗ (device-local RNG)	✓	✓ (torch)	✓	✓	~10	~45	Apache 2.0
torchvision.v2	✗ (device-local RNG)	✓	✓ (torch)	partial	✓	~18	~12	BSD-3
imgaug	✗ (numpy-only)	✓	✗	✓	partial	~20	~40	MIT
augly	✗ (PIL-only)	✓	✗	✗	✗	~5	~20	MIT

_{¹ External counts are approximate as of 2026-05 (sampled from each project's __init__.py / docs index). Versions move fast — consult each project's authoritative API reference for current numbers. paraug counts are code-exact (len(GEOMETRIC_PRIMITIVES) / len(PHOTOMETRIC_PRIMITIVES)).}

_{² Tolerance verified by tests/test_parity.py on NVIDIA 5060 Ti + 4080 at v0.1.0; exact bounds: 1e-6 for the 6 elementwise photometric ops listed in PHOTO_ELEMENTWISE (gamma / gaussian_noise / color_jitter / vignette / cutout / hue_shift), 2e-4 for grid_sample-class (geometric) and conv-class (blur) ops. GitHub free CI runners are CPU-only, so the 13 CUDA parity tests skip on CI — community verification on additional GPU SKUs is welcome (open a PR with the result, or run pytest tests/test_parity.py -k cpu_vs_cuda locally and post the output).}

When to use paraug

• Cross-device reproducibility (paper-grade ablation where CPU↔GPU drift breaks a baseline)
• Distributed training on heterogeneous hardware
• Unit-test-friendly augmentation pipelines (CPU-side RNG means a test on a free CI runner reproduces a developer's GPU result)

When NOT to use paraug

• You need 50+ primitive options out of the box → try albumentations or imgaug
• You need PIL-style per-image API → try augly
• You need built-in compositional ops like OneOf / SomeOf → try albumentations

Examples

See examples/:

• 01_quickstart.py — minimal load → augment → save
• 02_mask_aware.py — image + segmentation mask warped together
• 03_cpu_gpu_parity.py — same seed on CPU and CUDA, assert max_abs_diff < 2e-4

Citation

@software{paraug2026,
  author = {alieuidsh},
  title  = {paraug: Bit-exact CPU/GPU parity for image augmentation},
  year   = {2026},
  url    = {https://github.com/alieuidsh/paraug},
}

License

Apache 2.0 — see LICENSE.