fuse-augmentations

Fuse consecutive geometric augmentation ops (rotation, scale, shear, flip) into a single grid_sample pass, eliminating redundant interpolation and improving image quality.

Drop-in replacement for Kornia's AugmentationSequential / Sequential - same call interface, fewer warps.

Installation

pip install fuse-augmentations

Requires: Python 3.10+, PyTorch 2.1+. Kornia is optional (needed only when passing Kornia transform objects).

Quick start

import torch
import kornia.augmentation as K
from fuse_aug import Compose

pipe = Compose(
    [
        K.RandomRotation(degrees=30, p=0.8),
        K.RandomHorizontalFlip(p=0.5),
        K.RandomAffine(degrees=0, scale=(0.8, 1.2), p=0.7),
    ]
)

image = torch.rand(4, 3, 256, 256)  # (B, C, H, W)
out = pipe(image)  # one grid_sample instead of three

Auxiliary targets - masks, boxes, keypoints

Pass data_keys to route auxiliary tensors through the same fused transform:

from fuse_aug import Compose
import kornia.augmentation as K

pipe = Compose(
    [
        K.RandomRotation(degrees=30, p=0.8),
        K.RandomHorizontalFlip(p=0.5),
    ],
    data_keys=["input", "mask"],
)

img_out, mask_out = pipe(image, mask)  # mask warped with mode='nearest'

Supported keys:

Bounding boxes and keypoints use the composed forward matrix; masks share the same sampling grid as the image.

Backend-free pipelines with from_params

Construct a fused pipeline from numeric parameter ranges - no Kornia import required:

from fuse_aug import Compose

pipe = Compose.from_params(rotation=(-30, 30), scale=(0.8, 1.2), hflip_p=0.5)
out = pipe(image)

from_params accepts: rotation, scale, scale_x, scale_y, shear_x, shear_y, translate_x, translate_y, hflip_p, vflip_p, interpolation, padding_mode, reorder, and data_keys.

How fusion works

Given a pipeline [Rotate, Scale, HFlip, GaussianBlur, Rotate]:

1. [Rotate, Scale, HFlip] are grouped into a FusedAffineSegment - their matrices are composed and one grid_sample call is made.
2. GaussianBlur is a spatial-kernel barrier - it passes through unchanged.
3. The trailing Rotate forms its own FusedAffineSegment.

print(pipe.fusion_plan)
# fused(RandomRotation, RandomResizedCrop, RandomHorizontalFlip) -> passthrough(RandomGaussianBlur) -> fused(RandomRotation)

print(pipe.n_warps_saved)
# 2  (saved 2 interpolation passes)

Reorder policy

from fuse_aug import Compose, ReorderPolicy
import kornia.augmentation as K

pipe = Compose(
    [
        K.RandomRotation(degrees=15, p=0.8),
        K.ColorJitter(brightness=0.3, p=0.5),  # POINTWISE
        K.RandomHorizontalFlip(p=0.5),
    ],
    reorder=ReorderPolicy.POINTWISE,
)

# ColorJitter is moved after HFlip, letting both geometric ops fuse:
# fused(RandomRotation, RandomHorizontalFlip) -> passthrough(ColorJitter)

Transform matrix access

out = pipe(image)
M = pipe.transform_matrix  # (B, 3, 3) composed forward matrix

Use M to transform stored coordinates (keypoints, boxes) that were not passed as data_keys.