# Tessellation¶

This step consists of defining time segments or space cells and assigning the individual molecule locations to one or more of these segments or cells/bins.

## Basic usage¶

The tramway command features the tessellate sub-command that handles this step. Let us consider again some of the example calls to this sub-command:

> tramway tessellate kmeans -i example.trxyt -l kmeans
> tramway tessellate gwr -i example.rwa -w -n 50 -N 50 -l gwr*
> tramway tessellate gwr -i example.rwa -d 0.1 -ss 10 -n 30 -l gwr*


The equivalent Python code is:

from tramway.helper import *

tessellate('example.trxyt', 'kmeans', label='kmeans')
tessellate('example.rwa', 'gwr', scaling=True, knn=(50, 50), label='gwr*')
tessellate('example.rwa', 'gwr', ref_distance=0.1, strict_min_location_count=10, knn=30, label='gwr*')


The above calls to the tessellate() helper read and write into the example.rwa file.

Alternatively, reads and writes to files can be limited by directly manipulating the analysis tree:

trajectories = load_xyt('example.trxyt')
analyses = tessellate(trajectories, 'kmeans', label='kmeans', return_analyses=True)
tessellate(analyses, 'gwr', ...
...
save_rwa('example.rwa', analyses)


The load_xyt() and save_rwa() functions respectivelly read the input file and write the output file.

The analysis tree is augmented by each further call to tessellate().

A spatial 2D partition can be displayed with the command-line:

> tramway draw cells -i example.rwa --label kmeans


or with the cell_plot() helper function:

cell_plot('example.rwa', label='kmeans')


## Standard methods¶

The available methods are:

• random: randomly located cell centers.
• grid: regular grid with equally sized square or cubic areas.
• hexagon: regular grid with equally shaped hexagons (2D only).
• kdtree: kd-tree tessellation with midpoint splits.
• kmeans: tessellation based on the k-means clustering algorithm.
• gwr (or gas): tessellation based on the Growing-When-Required self-organizing gas.

All the above methods can handle time as just another space dimension if the time_scale argument to tessellate() if defined. This argument scales the time axis so that it can be quantitatively related to the space axes. Note however that this feature has been very sparsely tested.

### The random method¶

random places cell centers at random within the bounding box of the (trans-)locations. Empty cells are discarded (unless allow_empty_cells=True is passed to tessellate()) and cell boundaries are defined as the Voronoi graph.

The corresponding tessellation class is RandomMesh.

From the command-line, a key argument is --cell-count. This is the desired number of cells.

The number of cells can alternatively be determined using the --location-count argument (or avg_location_count argument to the tessellate() Python helper). If defined, the number of cells is defined so that the average number of locations per cell (before empty cells are discarded) equals the specified value.

### The grid method¶

grid is a regular grid. Every cells are equal-size hypercubes.

The corresponding tessellation class is RegularMesh.

From the command-line, key arguments are --location-count and --distance.

Per default, --distance is set to the average translocation distance. If --location-count is not defined, neighbour cells are spaced by twice the value of --distance.

If --location-count is defined, the cells are sized so that the average number of locations per cell equals the specified value.

The cell size (or inter-cell distance) is bounded by $$0.8$$ times the value of --distance. This bound can be ignored with --min-location-count=0.

With the tessellate() Python helper only, one can adjust avg_distance (equivalent to commandline option --distance) without specifying an absolute value, with rel_avg_distance instead. The average translocation distance is multiplied by this value.

### The hexagon method¶

hexagon is a regular grid of hexagonal cells.

The corresponding tessellation class is HexagonalMesh.

From the command-line, key arguments are --location-count and --distance.

If --location-count is not defined, per default --distance is set to the average translocation distance and neighbour cells are spaced by twice this value.

If --location-count is defined, the cells are sized so that the average number of locations per cell equals the specified value.

Per default, the cell size (or inter-cell distance) is lower-bounded by $$0.8$$ times the average translocation distance. This bound can be ignored with --min-location-count=0.

With the tessellate() Python helper only, one can adjust avg_distance (equivalent to commandline option --distance) without specifying an absolute value, with rel_avg_distance instead. The average translocation distance is multiplied by this value.

### The kdtree method¶

kdtree is the quad-tree algorithm from InferenceMAP extended to any dimensionality greater than or equal to 2.

A main difference from the widely known k d-tree algorithm in that it recursively splits the cells in two equal parts along each dimension.

The corresponding tessellation class is KDTreeMesh.

Cells scale with the ref_distance input argument to tessellate() and equivalently the avg_distance attribute of the class.

The maximum cell size can be controlled with the max_level input argument that defines the maximum cell size as a multiple of the smallest cell size, in side length.

### The kmeans method¶

kmeans is a fast tessellation approach that usually displays different cell shapes and offers better resolutions along density borders.

The corresponding tessellation class is KMeansMesh.

The algorithm is initialized with a grid tessellation. As a consequence cells scale wrt the avg_probability argument (or command-line option --location-count) or ref_distance argument (or command-line option --distance).

### The gwr method¶

gwr stands for Grow(ing) When Required and is actually a largely modified version of the algorithm described by Marsland, Shapiro and Nehmzow in 2002.

The corresponding tessellation class is GasMesh.

The main arguments are min_probability (or command-line option -s/--min-location-count) and avg_distance (or command-line option -d/--distance).

gwr exhibits many more arguments. Some of them must be passed directly to the tessellate() method.

This method may be useful to build high resolution maps with the desired minimum number of locations per cell reasonably well approached in the low-density areas. The knn argument to cell_index may be very useful in combination to such high resolution tessellations.

gwr is more computer-intensive than the other methods. To prototype with this method, reasonnable under-trained solutions can be obtained passing a value less than 1 for argument pass_count or command-line option --pass-count.

For example, gwr is not suitable for controlling the average number of locations per cell. Argument avg_probability or command-line option -c/--location-count are taken as constraints that may not be satisfied. A trial-and-error approach may be necessary to generate a mesh with a suitable average location count per cell. Under-trained solutions tend to exhibit the same average location count as in more refined solutions.

### The window method¶

The window plugin implements a sliding time window.

The corresponding tessellation class is SlidingWindow.

The step (shift) and window width (duration) can be defined either as timestamps (default) or as frames (with frames=True or command-line option --frames).

Note that all the above spatial-only tessellation methods admit extra arguments related to time windowing, including time_window_duration and time_window_shift. This is equivalent to applying first the spatial method and then segmenting in time, all in one step.

An example is given in the Segmenting time subsection of the command-line primer section.

Arbitrary time segments can also be defined, although no helper is available for this. See also the Custom time segments subsection.

## Implementation details¶

The concept of tessellation or windowing in itself is independent of the proper partition of the molecule locations or translocations in subsets corresponding to cells or segments.

These data are combined together in a Partition.

Its tessellation attribute of type Tessellation is the “tessellation” or binning/sampling strategy grown from (trans-)location data and in principle suitable for partitioning/sampling any other similar set of (trans-)locations.

The partition itself, i.e. the proper assignment of locations to cells/bins or segments - whether these locations are those involved in growing the tessellation or others, is referred to as cell_index. This naming appears in the Tessellation class as a method, in the Partition class as an attribute (actually a property) and at other locations.

From a particular partition a series of derivative products are commonly extracted, such as the location count per cell, and some of these products are conveniently provided by the Partition.

Note that, because tessellating and partitioning are considered two different procedures, some input arguments to the tessellate() helper function may have multiple understandings. Some constraints may be taken as directions by the tessellation algorithm while the same constraints would typically be enforced by the partitioning.

As a consequence, tessellate() takes arguments with the strict_ prefix in their name. These arguments apply to the partition while arguments of similar names without this prefix apply to the tessellation.

### Tessellation nesting¶

Each cell of a tessellation can be tessellated again.

This is made possible by the NestedTessellations class.

The command-line also supports this extension. This can be useful for example to independently tessellate the space in each time segment:

> tramway tessellate window -i example.trxyt --shift 1 --duration 2 --output-label 2s_win
> tramway tessellate gwr -i example.rwa --input-label 2s_win --output-label windowed_gwr


### Custom cell centers¶

To define specific centroids and partition using knn, no explicit cells are needed.

The tessellation package exposes basic classes such as Delaunay and Voronoi.

Both can be used to implemented such a use case:

from tramway.core import *
from tramway.core.hdf5 import *
from tramway.tessellation import *
from numpy.random import rand
from pandas import DataFrame

n_centroids = 100
n_nearest_neighbours = 50

space_columns = ['x', 'y']

# find the bounding box
coordinates = translocations[space_columns]
xmin, xmax = coordinates.min(axis=0).values, coordinates.max(axis=0).values

# pick some centroids within the bounding box
centroids = rand(n_centroids, len(space_columns))
centroids *= xmax - xmin
centroids += xmin
centroids = DataFrame(data=centroids, columns=space_columns)

# grow the tessellation
tessellation = Delaunay()
tessellation.tessellate(centroids)

# find the nearest neighbours
cells = Partition(translocations, tessellation)
cells.cell_index = tessellation.cell_index(translocations,
knn=(n_nearest_neighbours, n_nearest_neighbours)) # knn is (min, max)

# assemble the analysis tree
analyses = Analyses(translocations)

# save it to a file
save_rwa('example.rwa', analyses)


The above example illustrates how to explicitly define cell centers. The specific case of random cell centers can be implemented using the ‘random’ plugin instead:

from tramway.helper import *

n_centroids = 100
n_nearest_neighbours = 50

tessellate('example. trxyt', 'random', cell_count=n_centroids,
knn=(n_nearest_neighbours, n_nearest_neighbours), label='random centroids')


### Custom time segments¶

The TimeLattice class is more flexible than the SlidingWindow class in that it admits arbitrary time segments.

The following example makes contiguous segments such that the total location count per segment is at least equal to a defined amount:

from tramway.helper import *
import numpy

min_location_count_per_segment = 10000

time_column = 't'

# count the rows (or locations) along time
timestamps = translocations[time_column].values
ts, counts = numpy.unique(timestamps, return_counts=True)

# pick the segment bounds
index = 0
bounds = [ts[index]]
while index < counts.size:
count = 0
while count < min_location_count_per_segment:
count += counts[index]
index += 1
if index == counts.size:
break
if index == counts.size: # or similarly if count < min_location_count_per_segment:
pass # let the loop end
else:
bounds.append(ts[index])

# grow a spatial tessellation
static_cells = tessellate(translocations, 'kmeans')

# associate the segments
segments = numpy.c_[bounds[:-1], bounds[1:]]
dynamic_cells = with_time_lattice(static_cells, segments)


The same example with different spatial tessellations for each segment can be implemented with the help of tessellation nesting.

If maps are inferred on such a tessellation, separate maps will be generated for each segment. These maps can be individualized as follows:

# infer the diffusivity
diffusivity_maps = infer(dynamic_cells, 'D')

# slice the maps (one map per segment) to plot each of them
for diffusivity_map in dynamic_cells.tessellation.split_frames(diffusivity_maps):
map_plot(diffusivity_map, cells=static_cells)

# assemble the analysis tree
analyses = Analyses(translocations)