Development of the cupyimg.skimage module in this repository has moved to a new open source RAPIDS project called cuCIM that was created by a collaboration between Quansight and NVIDIA. The cucim.skimage module there is an updated equivalent of cupyimg.skimage. The new repository has continuous integration testing on GPUs and generates binary packages via conda-forge. It also involves incorporation of additional image I/O functionality in a separate cucim.clara module.

Nearly all functions in cupyimg.numpy and cupyimg.scipy have been ported upstream to CuPy and are now available in CuPy >= 9.0.

Please migrate to using cuCIM and/or upstream CuPy in the coming weeks as this repository will eventually be archived.

cupyimg extends CuPy with additional functions for image/signal processing. This package implements a subset of functions from NumPy, SciPy and scikit-image with GPU support.

These implementations generally match the API and behavior of their corresponding CPU equivalents, although there are some limited exceptions. In some cases such as scipy.ndimage equivalents, complex-valued support is available on the GPU even though it is not present as part of the upstream library. See additional details under API Differences.

Ideally, NumPy/Scipy function implemented here will be submitted upstream to CuPy itself where they will benefit from a more comprehensive CI architecture on real GPU hardware and a broader set of maintainers. Currently, testing of this package on NVIDIA hardware has been done only on an NVIDIA 1080 Ti GPU using CUDA versions 9.2-10.2. However, it should work for all CUDA versions supported by the underlying CuPy library.

A more complete list of the implemented functions is available in the section below on Implemented Functions.

Functions tend to operate in the same manner as those from their upstream counterparts. If there are differences in dtype handling, etc. these should be noted within the corresponding function's docstring.

Aside from potential dtype differences, the primary difference with their CPU counterparts tends to be a requirement for cupy.ndarray inputs rather than allowing array-likes more generally. This behavior is consistent with CuPy itself where support for such array-likes is generally disallowed due to performance considerations. In cupyimg this cupy.ndarray rule is not yet consistently enforced everywhere, so some functions still accept numpy arrays as inputs and will transfer to the GPU automatically internally via cupy.asarray. Any such automatic coercion should not be relied upon and is subject to be removed in the future.

An simple example demonstrating applying of a uniform_filter to an array is:

import cupy as cp
from cupyimg.scipy.ndimage import uniform_filter

x = np.random.randn(128, 128, 128)
y = uniform_filter(x, size=5)

The RAPIDS project cuSignal provides an alternative implementation of functions from scipy.signal, including some not currently present here. Like cupyimg, it also depends on CuPy, but has an additional dependency on Numba. One other difference is at the time of writing, cuSignal does not support all of the new upfirdn and resample_poly boundary handling modes introduced in SciPy 1.4, while these are supported in cupyimg.

cupyimg supports Python 3.6, 3.7 and 3.8.

Requires:

• NumPy (>=1.14)
• CuPy (>=7.0)
• SciPy (>=1.2)
• scikit-image (>=0.16.2)
• fast_upfirdn (>=0.2.0)

To run the tests users will also need:

• pytest

Developers should see additional requirements for development in requirements-dev.txt.

Installation:

Packages for cupyimg are not yet on PyPI or conda-forge. Users should first configure a working CuPy environment. Then cupyimg can be installed from source.

An example installing cupyimg in a new conda environment is:

conda create -n cupyimg python=3.7
conda activate cupyimg
conda install numpy scipy scikit-image pytest cudatoolkit
pip install cupy-cuda101
pip install fast_upfirdn
pip install https://github.com/mritools/cupyimg/archive/master.zip

where cupy-cuda101 in the above will need to be changed to the appropriate version for the user's CUDA toolkit. See CuPy's documentation.

Example

import numpy as np
import cupy
import scipy.ndimage as ndi
from cupyimg.scipy.ndimage import uniform_filter
#from cupy.time import repeat

d = cupy.cuda.Device()
# separable 5x5x5 convolution kernel on the CPU
x = np.random.randn(256, 256, 256).astype(np.float32)
y = ndi.uniform_filter(x, size=5)
# %timeit y = ndi.uniform_filter(x, size=5)
#    -> 935 ms ± 69.2 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

# separable 5x5x5 convolution kernel on the GPU
xg = cupy.asarray(x)
yg = uniform_filter(xg, size=5)
# %timeit yg = uniform_filter(xg, size=5); d.synchronize()
#    -> 6.23 ms ± 24.8 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

As in CuPy itself, automatic conversion via cupy.asarray() is typically not performed on inputs in order to avoid unintended overheads from host/device transfers. Thus, many functions only accept CuPy arrays rather than more general array-likes as input. Enforcing this policy uniformly across cupyimg still needs some work, but in general one should not expected lists, numpy arrays or other iterables to be acceptable "image" inputs.

Many functions in cupyimg.scipy.ndimage support complex-valued floating point dtypes. These are not currently supported in the upstream scipy.ndimage module.

The functions in scipy.ndimage.filters have an additional dtype_mode keyword-only argument. When set to the default value of ndimage, convolutions are performed in double-precision as is done by scipy.ndimage. However, on the GPU single-precision operations are often substantially faster. The user can specify dtype_mode='float' to allow single-precision computations on single-precision inputs.

cupyimg.scipy.ndimage.convolve and cupyimg.scipy.ndimage.correlate have a couple of keyword-only arguments not present in scipy.ndimage. These are experimental and subject to change. For example, a crop kwarg allows performing "full" convolutions instead of one limited to the original image extent.

Some functions such as numpy.convolve support less common dtypes such as datetime64 or Decimal. These are not supported by upstream CuPy and are thus not available in cupyimg either.

cupyimg.numpy:

• apply_along_axis (upstream PR: 4008)
• convolve (upstream PR: 3371)
• correlate (upstream PR: 3525)
• gradient (upstream PR: 3963)
• histogram (upstream PR: 3124)
• histogram2d (upstream PR: 3947)
• histogramdd (upstream PR: 3947)
• ndim (upstream PR: 3060)
• quantile (upstream PR: 4370)
• ravel_multi_index (upstream PR: 3104)

cupyimg.scipy.interpolate:

• interpnd
• RegularGridInterpolator

cupyimg.scipy.ndimage.filters:

• convolve (upstream PR: PR 3184)
• convolve1d (upstream PR: PR 3184)
• correlate (upstream PR: PR 3184)
• correlate1d (upstream PR: PR 3184)
• gaussian_filter (upstream PR: PR 3505)
• gaussian_filter1d (upstream PR: PR 3505)
• gaussian_laplace (upstream PR: PR 3505)
• gaussian_gradient_magnitude (upstream PR: PR 3505)
• generic_laplace (upstream PR: PR 3505)
• generic_gradient_magnitude (upstream PR: PR 3505)
• laplace (upstream PR: PR 3505)
• prewitt (upstream PR: PR 3505)
• sobel (upstream PR: PR 3505)
• uniform_filter (upstream PR: PR 3505)
• uniform_filter1d (upstream PR: PR 3505)
• maximum_filter (upstream PR: PR 3239)
• maximum_filter1d (upstream PR: PR 3184, PR 3505)
• median_filter (upstream PR: PR 3500)
• minimum_filter (upstream PR: PR 3239)
• minimum_filter1d (upstream PR: PR 3184, PR 3505)
• percentile_filter (upstream PR: PR 3500)
• rank_filter (upstream PR: PR 3500)

cupyimg.scipy.ndimage.fourier:

• fourier_ellipsoid (upstream PR: PR 4361)
• fourier_gaussian (upstream PR: PR 3654)
• fourier_shift (upstream PR: PR 3654)
• fourier_uniform (upstream PR: PR 3654)

cupyimg.scipy.ndimage.interpolation:

• affine_transform (upstream PR: 3166)
• map_coordinates (upstream PR: 3166)
• rotate (upstream PR: 3166)
• shift (upstream PR: 3166)
• spline_filter (upstream PR: 4145)
• spline_filter1d (upstream PR: 4145)
• zoom (upstream PR: 3166)

other upstream enhancements for interpolation (SciPy 1.6 boundary modes and spline order 2-5): PR 4083 PR 4314 PR 4400 PR 4401 PR 4402

cupyimg.scipy.ndimage.measurements:

• center_of_mass (upstream PR: 4311)
• extrema (upstream PR: PR 3979)
• histogram (upstream PR: 4311)
• label (upstream PR: PR 3210)
• labeled_comprehension (upstream PR: 4311)
• maximum (upstream PR: PR 3979)
• maximum_position (upstream PR: PR 3979)
• mean (upstream PR: PR 3259, PR 3419)
• median (upstream PR: PR 3979)
• minimum (upstream PR: PR 3979)
• minimum_position (upstream PR: PR 3979)
• standard_deviation (upstream PR: PR 3259, PR 3419)
• sum (upstream PR: PR 3259, PR 3419, PR 3425)
• variance (upstream PR: PR 3259, PR 3419) related: (upstream PR: 4151)

cupyimg.scipy.ndimage.morphology:

• binary_erosion (upstream PR: 3907)
• binary_dilation (upstream PR: 3907)
• binary_opening (upstream PR: 3907)
• binary_closing (upstream PR: 3907)
• binary_hit_or_miss (upstream PR: 3907)
• binary_propagation (upstream PR: 3907)
• binary_fill_holes (upstream PR: 3907)
• black_tophat (upstream PR: 3946)
• generate_binary_structure (upstream PR: 3907)
• grey_closing (upstream PR: PR 3239)
• grey_dilation (upstream PR: PR 3216)
• grey_erosion (upstream PR: PR 3216)
• grey_opening (upstream PR: PR 3239)
• iterate_structure (upstream PR: 3907)
• morphological_gradient (upstream PR: 3946)
• morphological_laplace (upstream PR: 3946)
• white_tophat (upstream PR: 3946) related: 4058, 4059

cupyimg.scipy.signal:

• choose_conv_method (upstream PR: 3464)
• convolve (upstream PR: 3748)
• convolve2d (upstream PR: 3748)
• correlate (upstream PR: 3748)
• correlate2d (upstream PR: 3748)
• fftconvolve (upstream PR: 3828)
• hilbert
• hilbert2
• oaconvolve
• resample
• resample_poly
• upfirdn
• wiener (upstream PR: 3645)

cupyimg.scipy.special:

• entr (upstream PR: 2861)
• kl_div (upstream PR: 2861)
• rel_entr (upstream PR: 2861)
• huber (upstream PR: 2861)
• pseudo_huber (upstream PR: 2861)

cupyimg.scipy.stats:

• entropy (upstream PR: 4369)

skimage.color:

• All functions in this module are supported

skimage.exposure:

• adjust_gamma
• adjust_log
• adjust_sigmoid
• cumulative_distribution
• equalize_adapthist
• equalize_hist
• histogram
• is_low_contrast
• match_histograms
• rescale_intensity

skimage.feature:

• canny
• corner_harris
• corner_kitchen_rosenfeld
• corner_shi_tomasi
• corner_foerstner
• corner_peaks
• daisy
• hessian_matrix
• hessian_matrix_det
• hessian_matrix_eigvals
• match_template
• peak_local_max
• shape_index
• structure_tensor
• structure_tensor_eigvals

skimage.filters:

• apply_hysteresis_threshold
• difference_of_gaussians
• farid
• farid_h
• farid_v
• frangi
• gabor_kernel
• gabor
• gaussian
• hessian
• inverse
• laplace
• LPIFilter2D
• median (ndimage mode only)
• meijering
• prewitt
• prewitt_h
• prewitt_v
• rank_filter
• roberts
• roberts_pos_diag
• roberts_neg_diag
• sato
• scharr
• scharr_h
• scharr_v
• sobel
• sobel_h
• sobel_v
• threshold_isodata
• threshold_li
• threshold_local
• threshold_mean
• threshold_minimum
• threshold_multiotsu
• threshold_niblack
• threshold_otsu
• threshold_sauvola
• threshold_triangle
• threshold_yen
• try_all_threshold
• unsharp_mask
• wiener
• window

skimage.measure:

• approximate_polygon
• subdivide_polygon
• block_reduce
• compare_mse
• compare_nrmse
• compare_psnr
• compare_ssim
• inertia_tensor
• inertia_tensor_eigenvals
• label
• moments
• moments_central
• moments_coords
• moments_coords_central
• moments_hu
• moments_normalized
• perimeter
• profile_line
• regionprops
• regionprops_table
• shannon_entropy
• subdivide_polygon

skimage.metrics:

• mean_squared_error
• normalized_root_mse
• peak_signal_noise_ratio
• structural_similarity

skimage.morphology:

• ball
• binary_erosion
• binary_dilation
• binary_opening
• binary_closing
• black_tophat
• closing
• cube
• diamond
• dilation
• disk
• erosion
• octagon
• octahedron
• opening
• reconstruct
• rectangle
• remove_small_holes
• remove_small_objects
• square
• star
• white_tophat

skimage.registration:

• cross_correlate_masked
• optical_flow_tvl1
• phase_cross_correlation

skimage.restoration:

• calibrate_denoiser
• denoise_tv_chambolle
• richardson_lucy
• unsupervised_wiener
• wiener

skimage.segmentation:

• checkerboard_level_set
• circle_level_set
• clear_border
• disk_level_set
• find_boundaries
• inverse_gaussian_gradient
• mark_boundaries
• morphological_chan_vese
• morphological_geodesic_active_contour

skimage.transform:

• AffineTransform
• downscale_local_mean
• EssentialMatrixTransform
• estimate_transform
• EuclideanTransform
• FundamentalMatrixTransform
• integral_image
• integrate
• matrix_transform
• PolynomialTransform
• ProjectiveTransform
• pyramid_expand
• pyramid_gaussian
• pyramid_laplacian
• pyramid_reduce
• rescale
• resize
• rotate
• SimilarityTransform
• swirl
• warp
• warp_coords
• warp_polar

skimage.util:

• crop
• dtype_limits
• img_as_bool
• img_as_float
• img_as_float32
• img_as_float64
• img_as_int
• img_as_ubyte
• img_as_uint
• invert
• view_as_blocks
• view_as_windows