Skip to main content

Pure NumPy implementation of affine image transformations

Project description

affine_image.np

Affine transformations on (currently only 3D, maybe later 2D) arrays via NumPy, intended to be

🛠️ Install via: pip install affine-image

Usage 💡

How-to (Basic)

Given you can read affine matrices, the following conventions (taken from PyTorch) + an example should get you started

  • 1️⃣ batch and channel dim. prior to image dim. -> (batch, channel, x, y, z) for 3D images
  • 2️⃣ affine acts in inverse order on image dim. (first row acts on z, second row acts on y, third row acts on x)
  • 3️⃣ translation parameter (=value in last column of affine) of 1 moves the image by half its size in the respective dim.
import numpy as np
from affine_image import affine_transform_3d

b, c, x, y, z = (1, 3, 5, 5, 5)
im = np.random.rand(b, c, x, y, z)  # 1️⃣ shape: (batch, channels, x, y, z)
affine = np.array([[[1.5, 0, 0, 0],  # 2️⃣ acts on z-dim.: zoom of 150%
                    [0, 1, 0, 1.0],  # 2️⃣ acts on y-dim.: 3️⃣ translation by 2.5 pixels (=y/2)
                    [0, 0, 1, 0]]])  # 2️⃣ acts on x-dim.: zoom of 100% (i.e. no change)
shape = (x, y, z)
# Apply affine ✨ Since we are in the README, show all possible arguments (with default values)
im_out = affine_transform_3d(im, affine, shape, nearest=False, padding='zeros', align_corners=False, scipy_affine=False)

affine_transform_3d is the main function of this package and takes the arguments

  • im: Input image array with 5 dimensions (batch, channel, x, y, z)
  • affine: Affine transformation matrix with 3 dimensions (batch, 3, 4)
  • shape: Desired output shape
  • nearest: Use nearest-neighbor interpolation if True, otherwise use linear (=trilinear for 3D) interpolation
  • padding: Padding mode, either 'zeros', 'border', 'reflection' or int/float (=padding value). ('border' and 'reflection' are analogous to 'nearest' and 'mirror' in scipy)
  • align_corners: Align corners flag (see PyTorch's docs)
  • scipy_affine: Use SciPy affine convention if True

If scipy_affine is set to True, the conventions 2️⃣ and 3️⃣ are replaced with

  • 2️⃣* affine acts in normal order on image dim. (first row acts on x, second row acts on y, third row acts on y)
  • 3️⃣* translation parameter (=value in last column of affine) of 1 moves the image by one pixel in the respective dim.

Why? (Interlude for the Curious 🤓)

This subsection serves readers who are not familiar with PyTorch who probably ask: Why did affine_image (per default) follow the weird PyTorch conventions?

Let's start with a rewrite of the above example in torch (=PyTorch)

import torch
import torch.nn.functional as F

b, c, x, y, z = (1, 3, 5, 5, 5)
im = torch.rand(b, c, x, y, z) 
affine = torch.tensor([[[1.5, 0, 0, 0],
                        [0, 1, 0, 1.0],
                        [0, 0, 1, 0]]])
shape = (x, y, z)
# Apply affine in torch
grid = F.affine_grid(affine, size=[1, 3, *shape], align_corners=True)
im_out = F.grid_sample(im, grid, mode='bilinear', padding_mode='zeros', align_corners=True)

Note that torch requires two steps to apply an affine to an image

  1. Pass affine to F.affine_grid which returns a grid
  2. Apply the grid to the image using F.grid_sample

Let's look at the shape of the grid to understand it

print(grid.shape)  # Output: [1, 5, 5, 5, 3] = [1, *shape, 3]

The grid contains coordinates w.r.t the input image from which the output image is sampled, e.g.

print(grid[0, 0, 0, 0, :])  # Output: [-1.5000,  0.0000, -1.0000] (align_corners=True in code above made coordinates more understandable here)

are the z, y and x coordinate in the input image from which the first (a corner) pixel of the output image are sampled.

Ok, everything is set up to finally tackle the Why...?s:

  • Why two steps?: When applying an affine to multiple arrays/tensors, grid can be reused to avoid recalculation
  • Why 1️⃣?: Stacking images along the batch dim. enables parallel application of multiple affines
  • Why 2️⃣?: No idea 🤔 Probably something about speed in the underlying C++/CUDA code of torch
    • ...but why does affine-image follow it anyway?: To avoid the introduction of another set of conventions
  • Why 3️⃣?: Since grid coordinates of -1/1 indicate edges of the input images with 0 indicating the center...
    • ...but why?: Makes grid coordinates more general since they are independent of image shapes

How-to (Advanced)?

Similar to PyTorch, affine_transform_3d behind the scenes uses a grid to resample the image.

Let's rewrite the first example to explicitly work with a grid via affine_image

import numpy as np
from affine_image import affine_grid_3d, sample_linear_3d, sample_nearest_3d

b, c, x, y, z = (1, 3, 5, 5, 5)
im = np.random.rand(b, c, x, y, z)
affine = np.array([[[1.5, 0, 0, 0],
                    [0, 1, 0, 1.0],
                    [0, 0, 1, 0]]])
shape = (x, y, z)
grid = affine_grid_3d(affine, shape, align_corners=False)
im_out = sample_linear_3d(im, grid, padding='zeros', align_corners=False)

To run nearest-neighbor interpolation, replace sample_linear_3d with sample_nearest_3d

im_out = sample_nearest_3d(im, grid, padding='zeros', align_corners=False)

If you have read the full Usage 💡 section, here, take a cookie 🍪

Speed 💨

Compared to torch and scipy, affine-image runs at ~25% the speed for trilinear interpolation and ~50% speed for nearest interpolation 🤓 Pretty OK for being a naive NumPy implementation!

Runtimes on AMD Ryzen 9 5950X CPU with 16 cores

Default runtime (in seconds)

Image size (Interpolation) torch scipy affine-image
64³ (nearest) 0.004 0.005 0.009
64³ (trilinear) 0.006 0.008 0.029
128³ (nearest) 0.029 0.043 0.096
128³ (trilinear) 0.048 0.064 0.262
256³ (nearest) 0.306 0.355 0.749
256³ (trilinear) 0.434 0.529 2.237

Single-thread runtime (in seconds)

Image size (Interpolation) torch scipy affine-image
64³ (nearest) 0.005 0.005 0.009
64³ (trilinear) 0.007 0.009 0.031
128³ (nearest) 0.042 0.043 0.093
128³ (trilinear) 0.062 0.064 0.251
256³ (nearest) 0.413 0.353 0.757
256³ (trilinear) 0.569 0.536 2.262

Compatibility 📏

affine_grid_3d is compatible with F.affine_grid (meaning their respective outputs match) 🎉

Besides that, the outputs of affine-image currently slightly differ from the outputs of torch and scipy. For nearest=True (especially with align_corners=True) affine-image almost matches torch:

issue1

The test script offers plots (like the one above) and colorful terminal output to chase the remaining mismatches. Contributions (see Issues) are much appreciated 🤗

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

affine_image-0.1.0.tar.gz (6.0 kB view details)

Uploaded Source

Built Distribution

affine_image-0.1.0-py3-none-any.whl (6.6 kB view details)

Uploaded Python 3

File details

Details for the file affine_image-0.1.0.tar.gz.

File metadata

  • Download URL: affine_image-0.1.0.tar.gz
  • Upload date:
  • Size: 6.0 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? No
  • Uploaded via: poetry/1.5.1 CPython/3.9.12 Linux/5.15.0-105-generic

File hashes

Hashes for affine_image-0.1.0.tar.gz
Algorithm Hash digest
SHA256 f382984bc5ef8128c6f0c437ab0b70ed7dc471102877bc0bf5ba7dc7e85faa65
MD5 1c1bf7a45a38fb10b0dd08ef3a25be62
BLAKE2b-256 78ecf947cf8cdb319b6be3b75c7de9fbd6041ffda9c82208e86168920784aea1

See more details on using hashes here.

File details

Details for the file affine_image-0.1.0-py3-none-any.whl.

File metadata

  • Download URL: affine_image-0.1.0-py3-none-any.whl
  • Upload date:
  • Size: 6.6 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? No
  • Uploaded via: poetry/1.5.1 CPython/3.9.12 Linux/5.15.0-105-generic

File hashes

Hashes for affine_image-0.1.0-py3-none-any.whl
Algorithm Hash digest
SHA256 09b2af92575bf2eac36ad49bfac3d0e10b780c91992e0faa828addbd67660814
MD5 4c86a877fce20497853486edd61e1be0
BLAKE2b-256 0f575c21a2d102691ba3ded827c2c450e86ee06605c8cf4ac573914ea47edaa6

See more details on using hashes here.

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page