fastremap

Renumber and relabel Numpy arrays at C++ speed and physically convert rectangular Numpy arrays between C and Fortran order using an in-place transposition.

import fastremap

uniq, cts = fastremap.unique(labels, return_counts=True) # may be much faster than np.unique

idxs = fastremap.indices(labels, 1231) # important for huge arrays

labels, remapping = fastremap.renumber(labels, in_place=True) # relabel values from 1 and refit data type
ptc = fastremap.point_cloud(labels) # dict of coordinates by label
ptc = fastremap.point_cloud(labels, shell=True) # get the surface contours only (3d images only)

labels = fastremap.refit(labels) # resize the data type of the array to fit extrema
labels = fastremap.refit(labels, value=-35) # resize the data type to fit the value provided

wider_dtype = fastremap.widen_dtype(np.uint32) # np.uint64
narrower_dtype = fastremap.narrow_dtype(np.uint32) # np.uint16

# remap all occurances of 1 -> 2
labels = fastremap.remap(labels, { 1: 2 }, preserve_missing_labels=True, in_place=True)

labels = fastremap.mask(labels, [1,5,13]) # set all occurances of 1,5,13 to 0
labels = fastremap.mask_except(labels, [1,5,13]) # set all labels except 1,5,13 to 0

mapping = fastremap.component_map([ 1, 2, 3, 4 ], [ 5, 5, 6, 7 ]) # { 1: 5, 2: 5, 3: 6, 4: 7 }
mapping = fastremap.inverse_component_map([ 1, 2, 1, 3 ], [ 4, 4, 5, 6 ]) # { 1: [ 4, 5 ], 2: [ 4 ], 3: [ 6 ] }

fastremap.transpose(labels) # physically transpose labels in-place
fastremap.ascontiguousarray(labels) # try to perform a physical in-place transposition to C order
fastremap.asfortranarray(labels) # try to perform a physical in-place transposition to F order

minval, maxval = fastremap.minmax(labels) # faster version of (np.min(labels), np.max(labels))

# computes number of matching adjacent pixel pairs in an image
num_pairs = fastremap.pixel_pairs(labels)  
n_foreground = fastremap.foreground(labels) # number of nonzero voxels
is_solid = fastremap.is_solid(labels, 1) # checks if all 1 w/ early exit, no copy
is_sorted = fastremap.is_sorted(labels, asc=True) # checks sorted in ascending order

# computes the cutout.tobytes(order) of each chunk and returns
# the binaries indexed by fortran order in the order specified (C or F)
# If the input image is F contiguous and F is requested, or C and C order,
# and the image is larger than a single chunk, this will be significantly
# faster than iterating and using tobytes.
binaries = fastremap.tobytes(labels, (64,64,64), order="F")

All Available Functions

• unique: Faster implementation of np.unique.
• renumber: Relabel array from 1 to N which can often use smaller datatypes.
• indices: Optimized search for matching values.
• remap: Custom relabeling of values in an array from a dictionary.
• refit: Resize the data type of an array to the smallest that can contain the most extreme values in it.
• narrow_dtype: Find the next sized up dtype. e.g. uint16 -> uint32
• widen_dtype: Find the next sized down dtype. e.g. uint16 -> uint8
• mask: Zero out labels in an array specified by a given list.
• mask_except: Zero out all labels except those specified in a given list.
• component_map: Extract an int-to-int dictionary mapping of labels from one image containing component labels to another parent labels.
• inverse_component_map: Extract an int-to-list-of-ints dictionary mapping from an image containing groups of components to an image containing the components.
• remap_from_array: Same as remap, but the map is an array where the key is the array index and the value is the value.
• remap_from_array_kv: Same as remap, but the map consists of two equal sized arrays, the first containing keys, the second containing values.
• transpose: Perform an in-place matrix transposition for rectangular arrays if memory is contiguous, apply the stock np.transpose function otherwise.
• asfortranarray: Perform an in-place matrix transposition for rectangular arrays if memory is contiguous, apply the stock np.asfortranarray function otherwise.
• ascontiguousarray: Perform an in-place matrix transposition for rectangular arrays if memory is contiguous, apply the stock np.ascontiguousarray function otherwise.
• minmax: Compute the min and max of an array in one pass.
• pixel_pairs: Computes the number of adjacent matching memory locations in an image. A quick heuristic for understanding if the image statistics are roughly similar to a connectomics segmentation.
• foreground: Count the number of non-zero voxels rapidly.
• is_solid: Check if the array is all one value w/ early exit
• is_sorted: Check if the array is sorted
• point_cloud: Get the X,Y,Z locations of each foreground voxel grouped by label.
• tobytes: Compute the tobytes of an image divided into a grid and return the resultant binaries indexed by their gridpoint in fortran order with the binary in the order requested (C or F).

pip Installation

pip install fastremap

If not, a C++ compiler is required.

pip install numpy
pip install fastremap --no-binary :all:

Manual Installation

A C++ compiler is required.

sudo apt-get install g++ python3-dev 
mkvirtualenv -p python3 fastremap
pip install numpy

# Choose one:
python setup.py develop  
python setup.py install 

The Problem of Remapping

Python loops are slow, so Numpy is often used to perform remapping on large arrays (hundreds of megabytes or gigabytes). In order to efficiently remap an array in Numpy you need a key-value array where the index is the key and the value is the contents of that index.

import numpy as np 

original = np.array([ 1, 3, 5, 5, 10 ])
remap = np.array([ 0, -5, 0, 6, 0, 0, 2, 0, 0, 0, -100 ])
# Keys:            0   1  2  3  4  5  6  7  8  9    10

remapped = remap[ original ]
>>> [ -5, 6, 2, 2, -100 ]

If there are 32 or 64 bit labels in the array, this becomes impractical as the size of the array can grow larger than RAM. Therefore, it would be helpful to be able to perform this mapping using a C speed loop. Numba can be used for this in some circumstances. However, this library provides an alternative.

import numpy as np
import fastremap 

mappings = {
  1: 100,
  2: 200,
  -3: 7,
}

arr = np.array([5, 1, 2, -5, -3, 10, 6])
# Custom remapping of -3, 5, and 6 leaving the rest alone
arr = fastremap.remap(arr, mappings, preserve_missing_labels=True) 
# result: [ 5, 100, 200, -5, 7, 10, 6 ]

The Problem of Renumbering

Sometimes a 64-bit array contains values that could be represented by an 8-bit array. However, similarly to the remapping problem, Python loops can be too slow to do this. Numpy doesn't provide a convenient way to do it either. Therefore this library provides an alternative solution.

import fastremap
import numpy as np

arr = np.array([ 283732875, 439238823, 283732875, 182812404, 0 ], dtype=np.int64) 

arr, remapping = fastremap.renumber(arr, preserve_zero=True) # Returns uint8 array
>>> arr = [ 1, 2, 1, 3, 0 ]
>>> remapping = { 0: 0, 283732875: 1, 439238823: 2, 182812404: 3 }

arr, remapping = fastremap.renumber(arr, preserve_zero=False) # Returns uint8 array
>>> arr = [ 1, 2, 1, 3, 4 ]
>>> remapping = { 0: 4, 283732875: 1, 439238823: 2, 182812404: 3 }

arr, remapping = fastremap.renumber(arr, preserve_zero=False, in_place=True) # Mutate arr to use less memory
>>> arr = [ 1, 2, 1, 3, 4 ]
>>> remapping = { 0: 4, 283732875: 1, 439238823: 2, 182812404: 3 }

The Problem of In-Place Transposition

When transitioning between different media, e.g. CPU to GPU, CPU to Network, CPU to disk, it's often necessary to physically transpose multi-dimensional arrays to reformat as C or Fortran order. Tranposing matrices is also a common action in linear algebra, but often you can get away with just changing the strides.

An out-of-place transposition is easy to write, and often faster, but it will spike peak memory consumption. This library grants the user the option of performing an in-place transposition which trades CPU time for peak memory usage. In the special case of square or cubic arrays, the in-place transpisition is both lower memory and faster.

• fastremap.asfortranarray: Same as np.asfortranarray but will perform the transposition in-place for 1, 2, 3, and 4D arrays. 2D and 3D square matrices are faster to process than with Numpy.
• fastremap.ascontiguousarray: Same as np.ascontiguousarray but will perform the transposition in-place for 1, 2, 3, and 4D arrays. 2D and 3D square matrices are faster to process than with Numpy.

import fastremap
import numpy as np 

arr = np.ones((512,512,512), dtype=np.float32)
arr = fastremap.asfortranarray(x)

arr = np.ones((512,512,512), dtype=np.float32, order='F')
arr = fastremap.ascontiguousarray(x)

C++ Usage

The in-place matrix transposition is implemented in ipt.hpp. If you're working in C++, you can also use it directly like so:

#include "ipt.hpp"

int main() {

  int sx = 128;
  int sy = 124;
  int sz = 103;
  int sw = 3;

  auto* arr = ....;

  // All primitive number types supported
  // The array will be modified in place, 
  // so these functions are void type.
  ipt::ipt<int>(arr, sx, sy);            // 2D
  ipt::ipt<float>(arr, sx, sy, sz);      // 3D
  ipt::ipt<double>(arr, sx, sy, sz, sw); // 4D

  return 0;
}

--
Made with <3